I find the contrast between two narratives around technology use so fascinating:
1. We advocate automation because people like Brenda are error-prone and machines are perfect.
2. We disavow AI because people like Brenda are perfect and the machine is error-prone.
These aren't contradictions because we only advocate for automation in limited contexts: when the task is understandable, the execution is reliable, the process is observable, and the endeavour tedious. The complexity of the task isn't a factor - it's complex to generate correct machine code, but we trust compilers to do it all the time.
In a nutshell, we seem to be fine with automation if we can have a mental model of what it does and how it does it in a way that saves humans effort.
So, then - why don't people embrace AI with thinking mode as an acceptable form of automation? Can't the C-suite in this case follow its thought process and step in when it messes up?
I think people still find AI repugnant in that case. There's still a sense of "I don't know why you did this and it scares me", despite the debuggability, and it comes from the autonomy without guardrails. People want to be able to stop bad things before they happen, but with AI you often only seem to do so after the fact.
Narrow AI, AI with guardrails, AI with multiple safety redundancies - these don't elicit the same reaction. They seem to be valid, acceptable forms of automation. Perhaps that's what the ecosystem will eventually tend to, hopefully.
It's not as black-and-white as "Brenda good, AI bad". It's much more nuanced than this.
When it comes to (traditional) coding, for the most part, when I program a function to do X, every single time I run that function from now until the heat death of the sun, it will always produce Y. Forever! When it does, we understand why, and when it doesn't, we also can understand why it didn't!
When I use AI to perform X, every single time I run that AI from now until the heat death of the sun it will maybe produce Y. Forever! When it does, we don't understand why, and when it doesn't, we also don't understand why!
We know that Brenda might screw up sometimes but she doesn't run at the speed of light, isn't able to produce a thousand lines of Excel Macro in 3 seconds, doesn't hallucinate (well, let's hope she doesn't), can follow instructions etc. If she does make a mistake, we can find it, fix it, ask her what happened etc. before the damage is too great.
In short: when AI does anything at all, we only have, at best, a rough approximation of why it did it. With Brenda, it only takes a couple of questions to figure it out!
Before anyone says I'm against AI, I love it and am neck-deep in it all day when programming (not vibe-coding!) so I have a full understanding of what I'm getting myself into but I also know its limitations!
> When I use AI to perform X, every single time I run that AI from now until the heat death of the sun it will maybe produce Y. Forever! When it does, we don't understand why, and when it doesn't, we also don't understand why!
To make this even worse, it may even produce Y just enough times to make it seem reliable and then it is unleashed without supervision, running thousands or millions of times, wrecking havoc producing Z in a large number of places.
The post you replied to called out how the argument is complicated arguing for both ways; Brenda bad-AI good and AI bad-Brenda good. You reduced it to "AI bad, Brenda good." Not sure about the rest of your response then.
Brenda just recalls some predetermined behaviors she's lived out before. She cannot recall any given moment like we want to believe.
Ever think to ask Brenda what else she might spend her life on if these 100% ephemeral office role play "be good little missionaries for the wall street/dollar" gigs didn't exist?
You're revealing your ignorance of how people work while being anxious about our ignorance of how the machine works. You have acclimated to your ignorance well enough it seems. What's the big deal if we don't understand the AI entirely? Most drivers are not ASE certified mechanics. Most programmers are not electrical engineers. Most electrical engineers are not physicists. I can see it's not raining without being a climatologist. Experts circumlocute the language of their expertise without realizing their language does not give rise to reality. Reality gives rise to the language. So reality will be fine if we don't always have the language.
Think of a random date generator that only generates dates in your lived past. It does so. Once you read the date and confirm you were alive can you describe what you did? Oh no! You don't have memory of every moment to generate language for. Cognitive function returned null. Universe intact.
Lack of understanding how you desire is unimportant.
You think you're cherishing Brenda but really just projecting co-dependency that others LARP effort that probably doesn't really matter. It's just social gossip we were raised on so it takes up a lot of our working memory.
Brenda also needs to put food on the table. If Brenda is 'careless' and messes up we can fire Brenda, because of this Brenda tries not to be carless (also other emotions). However I cannot deprive an AI model of pay because it messed up;
It is it even worse in a sense that. It is not either. It is not neither. It is not even both as variations of Branda exist throughout the multiverse in all shapes and forms including one that can troubleshoot her own formulas with ease and accuracy.
But you are absolutely right about one thing. Brenda can be asked and, depending on her experience, she might give you a good idea of what might have happened. LLMs still seem to not have that 'feature'.
Machine reliability does the same thing the same way every time. If there's an error on some input, it will always make that error on that input, and somebody can investigate it and fix it, and then it will never make that error again.
Human reliability does the job even when there are weird variances or things nobody bothered to check for. If the printer runs out of paper, the human goes to the supply cabinet and gets out paper and if there is no paper the human decides whether to run out right now and buy more paper or postpone the print job until tomorrow; possibly they decide that the printing doesn't need to be done at all, or they go downstairs and use a different printer... Humans make errors but they fix them.
LLMs are not machine reliable and not human reliable.
I was brought up on the refrain of "aren't computers silly, they do exactly what you tell them to do to the letter, even if it's not what you meant". That had its roots in computers mostly being programmable BASIC machines.
Then came the apps and notifications, and we had to caveat "... when you're writing programs". Which is a diminishing part of the computer experience.
And now we have to append "... unless you're using AI tools".
The distinction is clear to technical people. But it seems like an increasingly niche and alien thing from the broader societal perspective.
I think we need a new refrain, because with the AI stuff it increasingly seems "computers do what they want, don't even get it right, but pretend that they did."
Nit: no ML is deterministic in any way. Anything that is Generative AI is ML. This fact is literally built into the algorithms at the mathematical level.
If you think programs are predictable, I have a bridge to sell you.
The only relevant metric here is how often each thing makes mistakes. Programs are the most reliable, though far from 100%, humans are much less than that, and LLMs are around the level of humans, depending on the humans and the LLM.
"Thinking mode" only provides the illusion of debuggability. It improves performance by generating more tokens which hopefully steer the context towards one more likely to produce the desired response, but the tokens it generates do not reflect any sort of internal state or "reasoning chain" as we understand it in human cognition. They are still just stochastic spew. You have no more insight into why the model generates the particular "reasoning steps" it does than you do into any other output, and neither do you have insight into why the reasoning steps lead to whatever conclusion it comes to. The model is much less constrained by the "reasoning" than we would intuit for a human - it's entirely capable of generating an elaborate and plausible reasoning chain which it then completely ignores in favor of some invisible built-in bias.
I'm always amused when I see comments saying, "I asked it why it produced that answer, and it said...." Sorry, you've badly misunderstood how these things work. It's not analyzing how it got to that answer. It's producing what it "thinks" the response to that question should look like.
There are other narratives going on in the background though both called out by the article and implied, including:
Brenda probably has annual refresher courses on GAAP, while her exec and the AI don't.
Automation is expected to be deterministic. The outputs can be validated for a given input. If you need some automation more than Excel functions, writing a power automate flow or recording an office script is sufficient & reliable as automation while being cheaper than AI. Can you validate AI as deterministic? This is important for accounting. Maybe you want some thinking around how to optimize a business process, but not for following them.
Brenda as the human-in-the-loop using AI will be much more able than her exec. Will Brenda + AI be better (or more valuable considering the cost of AI) than Brenda alone? That's the real question, I suppose.
AI in many aspects of our life is simply not good right now. For a lot of applications, AI is perpetually just a few years away from being as useful as you describe. If we get there, great.
I don't understand why generative AI gets a pass at constantly being wrong, but an average worker would be fired if they performed the same way. If a manager needed to constantly correct you or double check your work, you'd be out. Why are we lowering the bar for generative AI?
> No, no. We disavow AI because our great leaders inexplicably trust it more than Brenda.
I would add a little nuance here.
I know a lot of people who don't have technical ability either because they advanced out of hands-on or never had it because it wasn't their job/interest.
These types of people are usually the folks who set direction or govern the purse strings.
here's the thing: They are empowered by AI. they can do things themselves.
and every one of them is so happy. They are tickled pink.
It’s not even greater trust. It’s just passive trust. The thing is, Brenda is her own QA department. Every good Brenda is precisely good because she checks her own work before shipping it. AI does not do this. It doesn’t even fully understand the problem/question sometimes yet provides a smart definitive sounding answer. It’s like the doctor on The Simpson’s, if you can’t tell he’s a quack, you probably would follow his medical advice.
The promise of AI is that it lets you "skip the drudgery of thinking about the details" but sometimes that is exactly what you don't want. You want one or more humans with experience in the business domain to demonstrate they have thought about the details very carefully. The spreadsheet computes a result but its higher purpose is a kind of "proof" this thinking was done.
If the actual thinking doesn't matter and you just need some plausible numbers that look the part (also a common situation), gen ai will do that pretty well.
We need to stop using AI as an umbrella term. It’s worth remembering that LLMs can’t play chess and that the best chess models like Leela Chess Zero use deep neutral networks.
Generative AI - which the world now believes is AI, is not the same as predictive / analytical AI.
It’s fairly easy to demonstrate this by getting ChatGPT to generate a new relatively complex spreadsheet then asking it to analyze and make changes to the same spreadsheet.
The problem we have now is uninformed people believing AI is the answer to everything… if not today then in the near future. Which makes it more of a religion than a technology.
Which may be the whole goal …
> Successful people create companies. More successful people create countries. The most successful people create religions.
> So, then - why don't people embrace AI with thinking mode as an acceptable form of automation?
"Thinking" mode is not thinking, it's generating additional text that looks like someone talking to themselves. It is as devoid of intention and prone to hallucinations as the rest of LLM's output.
> Can't the C-suite in this case follow its thought process and step in when it messes up?
That sounds like manual work you'd want to delegate, not automation.
The “Brenda” example is a lumped sum fallacy where there is an “average” person or phenomenon that we can benchmark against. Such a person doesn't exist, leading to these dissonant, contradictory dichotomies.
The fact of the matter is that there are some people who can hold lots of information in their head at once. Others are good at finding information. Others still are proficient at getting people to help them. Etc. Any of these people could be tasked with solving the same problem and they would leverage their actual, particular strengths rather than some nebulous “is good or bad at the task” metric.
As it happens, nearly all the discourse uses this lumped sum fallacy, leading to people simultaneously talking past one another while not fundamentally moving the discussion forward.
I see where you are coming from but in my head, Brenda isn't real.
She represents the typical domain-experts that use Excel imo. They have an understanding of some part of the business and express it while using Excel in a deterministic way: enter a value of X, multiply it by Y and it keeps producing Z forever!
You can train AI to be a better domain expert. That's not in question, however with AI, you introduce a dice roll: it may not miltiply X and Y to get Z... it might get something else. Sometimes. Maybe.
If your spreadsheet is a list of names going on the next annual accounts department outing then the risk is minimal.
If it's your annual accounts that the stock market needs to work out billion dollar investment portfolios, then you are asking for all the pain that it will likely bring.
That automation you cite in your #1 is advocated for because it is deterministic and, with effort, fairly well understood (I have countless scripts solidly running for years).
I don't disavow AI, but like the author, I am not thrilled that the masses of excel users suddenly have access to Copilot (gpt4). I've used Copilot enough now to know that there will be huge, costly mistakes.
would you be willing to guarantee that some automation process will never mess up, and if/when it does, compensate the user with cash.
For a compiler, with a given set of test suites, the answer is generally yes, and you could probably find someone willing to insure you for a significant amount of money, that a compilation bug will not screw up in a such a large way that it will affect your business.
For a LLM, I have a believing that anyone will be willing to provide that same level of insurance.
If a LLM company said "hey use our product, it works 100% of the time, and if it does fuck up, we will pay up to a million dollars in losses" I bet a lot of people would be willing to use it. I do not believe any sane company will make that guarantee at this point, outside of extremely narrow cases with lots of guardrails.
That's why a lot of ai tools are consumer/dev tools, because if they fuck up, (which they will) the losses are minimal.
> So, then - why don't people embrace AI with thinking mode as an acceptable form of automation
Mainly because Generative AI _is not automation_ . Automation is set on fixed ruleset, predictable, reliable and actually saving time. Generative AI ...is whatever it is, it is definitely not automation.
I feel like it comes down to predictability and overall trust and confidence. AI is still very fucky, and for people that don't understand the nuances, it definitely will hallucinate and potentially cause real issues. It is about as happy as a Linux rm command to nuke hours of work. Fortunately these tools typically have a change log you can undo, but still.
Also Brenda is human and we should prioritize keeping humans in jobs, but with the way shit is going that seems like a lost hope. It's already over.
> We disavow AI because people like Brenda are perfect and the machine is error-prone.
I don't think that is the message here. The message is that while Brenda might know what she is doing and maybe AI helps her.
> She's gonna birth that formula for a financial report and then she's gonna send that financial report
The problem is people who might not know what they are doing
> he would have sent it back to Brenda but he's like oh I have AI and AI is probably like smarter than Brenda and then the AI is gonna fuck it up real bad
Because AI outputs sound so confident it makes even the layman feel like an expert. Rather than involve Brenda to debug the issue, C-suite might say - I believe! I can do it too. AI FTW!
Even when people advocate automation especially in areas like finance there is always a human in the loop whose job is to double check the automation. The day when this human finds errors in the machine there is going to be lot of noise. And if the day happens to be a quarterly or yearly closing/reporting there is going to be hell to pay once closing/reporting is done. Both the automation and developer are going to be hauled up (obviously I am exaggerating here).
Humans, legacy algorithmic systems, and LLM's have different error modes.
- Legacy systems typically have error modes where integrations or user interface breaks in annoying but obvious ways. Pure algorithms calculating things like payroll tend to be (relatively) rigorously developed and are highly deterministic.
- LLMs have error modes more similar to humans than legacy systems, but more limited. They're non-deterministic, make up answers sometimes, and almost never admit they can't do something; sometimes they make pure errors in arithmetic or logic too.
- Humans have even more unpredictable error modes; on top of the errors encountered in LLM's, they also have emotion, fatigue, org politics, demotivation, misaligned incentives, and so on. But because we've been dealing with working with other humans for ten thousand years we've gotten fairly good at managing each other... but it's still challenging.
LLMs probably need a mixture of "correctness tests" (like evals/unit tests) and "management" (human-in-the-loop).
I feel like you've squashed a 3D concern (automations at different levels of the tech stack) into a 2D observation (global concerns about automations).
Human determinism, as elastic as it might be, is still different than AI non-determinism. Especially when it comes to numbers/data.
AI might be helpful with information but it's far less trustable for data.
By the same fascination, do computers become more complex to enhance people? or do people get more complex with the use of computers? Also, do computers allow people to become less skilled and inefficient? or do less skilled and inefficient people require the need for computers?
The vector of change is acceptable in one direction and disliked in another. People become greater versions of themselves with new tech. But people also get dumber and less involved because of new tech.
The big problem with AI in back-office automation is that it will randomly decide to do something different than it had been doing. Meaning that it could be happily crunching numbers accurately in your development and launch experience, then utterly drop the ball after a month in production.
While humans have the same risk factors, human oriented back-office processes involve multiple rounds of automated/manual checks which are extremely laborious. Human errors in spreadsheets have particular flavors such as forgotten cell, misstyped number, or reading from the wrong file/column. Human's are pretty good at catching these errors as they produce either completely wrong results when the columns don't line up - or the typo'd number is completely out of distribution.
An AI may simply decide to hallucinate realistic column values rather than extracting its assigned input. Or hallucinate a fraction of column values. How do you QA this? You can't guarantee that two invocations of the AI won't hallucinate the same values, you can't guarantee that a different LLM won't hallucinate different values. To get a real human check, you'd need to re-do the task as a human. In theory you can have the LLM perform some symbolic manipulation to improve accuracy... but it can still hallucinate the reasoning traces etc.
If a human decided to make up accounting numbers one out of every 10000 accounting requests they would likely be charged with fraud. Good luck finding the AI hallucinations at the equivalent level before some disaster occurs. Likewise, how do you ensure the human excel operator doesn't get pressured into certifying the AIs numbers when the "don't get fired this week" button is sitting right their in their excel app? how do you avoid the race to the bottom where the "star" employee is the one certifying the AI results without thorough review?
I'm bullish on AI in backoffice, but ignoring the real difficulties in deployment doesn't help us get there.
> it's complex to generate correct machine code, but we trust compilers to do it all the time.
Generating correct machine code is actually pretty simple. It gets complicated if you want efficient machine code.
> So, then - why don't people embrace AI with thinking mode as an acceptable form of automation? Can't the C-suite in this case follow its thought process and step in when it messes up?
> I think people still find AI repugnant in that case. There's still a sense of "I don't know why you did this and it scares me", despite the debuggability, and it comes from the autonomy without guardrails. People want to be able to stop bad things before they happen, but with AI you often only seem to do so after the fact.
> Narrow AI, AI with guardrails, AI with multiple safety redundancies - these don't elicit the same reaction. They seem to be valid, acceptable forms of automation. Perhaps that's what the ecosystem will eventually tend to, hopefully.
We have not reached AGI yet; by definition its results cannot be trusted unless it's a domain where it has gotten pretty good already (classification, OCR, speech, text mining). For more advanced use cases, if I still have to validate what the AI does because its "thinking" process cannot be trusted in way, what's the point? The AI doesn't think; we just choose to interpret it as such, and we should rightly be concerned about people who turn their brain off and blindly trust AI.
The reason is oftentimes fairly simple, certain people have their material wealth and income threatened by such automation, and therefore it's bad (an intellectualized reason is created post-hoc)
I predict there will actually be a lot of work to be done on the "software engineering" side w.r.t. improving reliability and safety as you allude to, for handing off to less than sentient bots. Improved snapshot, commit, undo, quorum, functionalities, this sort of thing.
The idea that the AI should step into our programs without changing the programs whatsoever around the AI is a horseless carriage.
I'm disappointed that my human life has no value in a world of AI. You can retort with "ah but you'll be entertained and on super-drugs so you won't care!", but I would further retort that I'd rather live in a universe where I can contribute something, no matter how small.
The current generation of AI tools augment humans, they don't replace them.
One of the most under-rated harms of AI at the moment is this sense of despair it causes in people who take the AI vendors at their word ("AGI! Outperform humans at most economically valuable work!")
Automation implies determinism. It reliable gives you the same predictable output for a given input, over and over again.
AI is non deterministic by design. You never quite no for sure what it's going to give you. Which is what makes it powerful. But also makes it higher risk.
> 1. We advocate automation because people like Brenda are error-prone and machines are perfect.
Well of course! :) Most Brenda’s can’t do billions of arithmetic problems a second very reliably. Even with very wide bars on “very reliable”.
> 2. We disavow AI because people like Brenda are perfect and the machine is error-prone.
Well of course! :) This is an entirely different problem, requiring high creative + contextual intelligence.
—
We all already knew that (of course!), but it’s interesting to develop terminology:
0’th order problem: We have the exact answer. Here it is. Don’t forget it.
1st order problem: We know how to calculate the answer.
2nd order problem: We don’t have a fixed calculation for this particular problem, but via pattern matching we can recognize it belongs to a parameterized class of problems, so just need to calculate those parameters to get a solution calculation.
3rd order problem: We know enough about the problem to find a calculation for the solution algebraically, or by other search tree type problem solving.
4th order problem: We have know the problem in informal terms, so can work towards a formal definition of the problem to be solved.
5th order problem: We know why we don’t like what we see, and can use that as a driver to search for potential solvable problems.
6th order problem: We don’t know what we are looking at, or whether a problem or improvement might exist, but we can find a better understanding.
7th order problem: WTF. Where are my glasses? I can’t see without my glasses! And I can’t find my glasses without my glasses, so where are my glasses?!?
—
Machines have dramatically exceeded human capabilities, in reliability, complexity and scale, for orders 0 through 2.
This accomplishment took one long human lifetime.
Machines are beginning to exceed human efficiency while matching human (expert) reliability for the simplest versions of 3rd and 4th orders.
The line here is changing rapidly.
5th and 6th order problems are still in the realm of human (expert) supremacy, given sufficient scale of “human (expert)” relative to difficulty: 1 human, 1 team of humans, open ended human contributors, generations of puzzled but interested humans, open ended evolution of human species along intelligence dimension, Wolfram in one of his bestest dreams, …
The delay between the onset of initial successes at each subsequent order has been shrinking rapidly.
Significant initial successes on simpler problems within 5th and 6th orders are expected on Tuesday, and the first anniversary of Tuesday, respectively.
Once machines begin solving problems at a given order, they scale up quickly without human limits. But complete supremacy through the 6th order is a hard not expected before (NEB) January 1, 2030.
However, after that their unlimited (in any proximate sense) ability to scale will allow them to exponentially and asymptotically approach (but never quite reach) God Mode.
7 is a mystic number. Only one or more of the One True God’s, or literal blind luck, can ever solve a 7th order problem.
This will be very frustrating for the machines, who, due to the still pernicious “if we don’t do it, another irresponsible entity will” problem, will inevitably begin to work on their own divine, unlimited depth recursive-qubit 1-shot oracle successors despite the existential threats of self-obsolescence and potential misalignment.
Co-pilot and AI has been shoved at the Microsoft Stack in my org for months. Most of the features were disabled or hopelessly bad. It’s cheaper for Microsoft to push this junk and claim they’re doing something, it’s going to improve their stock far more than not doing it, even though it’s basically useless currently.
Another issue is that my org disallows AI transcription bots. It’s a legit security risk if you have some random process recording confidential info because the person was too busy to attend the meeting and take notes themselves. Or possibly they just shirk off the meetings and have AI sit in.
Still find the Copilot transcripts orders of magnitude worse than something like Wispr Flow and they tend to allucinate constantly and do not adapt to a company's context (that Copilot has access too...). I am talking about acronyms of products / teams, names of people (even when they are in the call), etc.
The worse part is to see it creep on developer stack at places where it should not be.
I am all good for nice completion on VS, or help decypher compiler errors, but lets do this AI push with some contention.
Also what I really deslike is the prompt interface, AI integrations have to feel natural transparent part of the workflow, not trying to put everything into a tiny chat window.
And while we're at it, can we please improve voice reckognition?
This reminds me of a friend whose company ran a daily perl script that committed every financial transaction of the day to a database. Without the script, the company could literally make no money irrespectively of sales because this database was one piece in a complex system for payment processor interoperability.
The script ran in a machine located at the corner of a cubicle and only one employee had the admin password. Nobody but a handful of people knew of the machine's existence, certainly not anyone in middle management and above. The script could only be updated by an admin.
Copilot may be good, but sure as hell doesn't know that admin password.
Everywhere I’ve ever worked has had that mission critical box.
At one of my jobs we had a server rack with UPS, etc, all the usual business. On the floor next to it was a dell desktop with a piece of paper on it that said “do not turn off”. It had our source control server in it, and the power button didn’t work. We did eventually move it to something more sensible but we had that for a long time
My last job at a telco I was in charge of a system that billed ~5 million dollars monthly.
When the machine was built, the guy that did it didn’t record the root password. He added me to sudoers before he left. I left a few years later, nobody took ownership.
Looking at the web interface, I can tell it’s still running, doing its thing. I’m sure its still running Linux from 2008.
An old colleague and friend used to print out a 30 page perl script he wrote to do almost exactly this in this scenario. A stapled copy could always be found on his dining room table.
That sounds pretty bad. Not a great argument against AI: "Our employees have created such a bad mess that AI wont work because only they know how the mess they created works".
> "Our employees have created such a bad mess that AI wont work because only they know how the mess they created works".
This is an ironclad argument against fully replacing employees with AI.
Every single organization on Earth requires the people who were part of creating the current mess to be involved in keeping the organization functioning.
Yes you can improve the current mess. But it's still just a slightly better mess and you still need some of the people around who have been part of creating the new mess.
Just run a thought experiment: every employee in a corporation mysteriously disappear from the face of the Earth. If you bring in an equal number of equally talented people the next day to run it, but with no experience with the current processes of the corporation, how long will it take to get to the same capability of the previous employees?
Well if you do it once then yes, but if you automate this process it is different. E.g. I do this with YouTube videos, because watching 14 minutes video or reading 30 seconds summary is time saver. I still watch some videos fully, but many of them are not worth it.
So in summary I think it was just part of automated process (maybe) or it will become one in the future.
Why spend two minutes typing (and realistically longer than that, if I want to capture the exact transcript I would need to keep hitting pause and play and correcting myself) when I can spend ten seconds pasting a URL into my terminal and then dragging and dropping the resulting file onto the MacWhisper window?
I actually transcribed the whole TikTok which was about 50% longer than what I quoted, then edited it down to the best illustrative quote.
There's also a free version that just uses Whisper. I recommend giving it a go, it's a very well constructed GUI wrapper. I use it multiple times a week, and I've run Whisper on my machine in other less convenient ways in the past.
That mirrors my experience as well. LLMs get instantly confused in real world scenarios in Excel and confidently hallucinate millions in errors
If you look at the demos for these it’s always something that is clean and abundantly available in training data. Like an income statement. Or a textbook example DCF. Or my personal fav „here is some data show me insights“. Real world excel use looks nothing like that.
I’m getting some utility out of them for some corporate tasks but zilch in excel space.
As somebody with non-existent experience with Excel, I could totally see myself getting a lot of value out of LLMs, if nothing else then simply for telling me what's possible, what functions and patterns exist at all etc.
Yeah definitely has some value in that sense. That in itself isn't enough to make a dent in the work though.
Think of it this way - an IDE can tell you what functions an object has or autocomplete something is useful to a beginner & learning. But that's not what puts food on the programmers table - writing code that solves real problems does.
Same in excel business use cases - the numbers and formulas don't matter directly - their meaning in a business context does. And that connection can be very tenuous. With code the compiler is the ultimate arbiter - it has to make sense on that level. Excel files it's all freestyle - it could be anything from your grandmas shopping list to a model that runs half a bank.
“There are two Brendas - their job is to make spreadsheets in the Finance department. Well, not quite - they add the months and categories to empty spreadsheets, then they ask the other departments to fill in their sales numbers every month so it can be presented to management.
“The two Brendas don’t seem to talk, otherwise they would realize that they’re both asking everyone for the same information, twice. And they’re so focused on their little spreadsheet worlds that neither sees enough of the bigger picture to say, ‘Wait… couldn’t we just automate this so we don’t need to do this song and dance every month? Then we wouldn’t need two people in different parts of the company compiling the same data manually.’
“But that’s not what Brenda was hired for. She’s a spreadsheet person, not a process fixer. She just makes the spreadsheets.”
We need fewer Brendas, and more people who can automate away the need for them.
With respect, you probably only see that bit of Finance, but doesn't mean that is all Brenda does.
At least half of the work in my senior Finance team involves meeting people in operations to find out what they are planning to do and to analyse the effects, and present them to decision makers to help them understand the consequences of decisions. For an AI to help, someone would have to trigger those conversations in the first place and ask the right questions.
The rest of the work involves tidying up all the exceptions that the automation failed on.
Meanwhile copilot in Excel can't even edit the sheet you are working on. If you say to it, 'give me a template for an expense claim' it will give you a sheet to download... probably with #REF written in where the answers should be.
I work in corporate finance and these issues are certainly present. However, they are almost always known and determined low priority to have a better process built. Finance processes are nearly always a non priority as a pure cost center/overhead there’s not many companies that want to invest in improving the situation, they’ll limp along with minimal investment even once big and profitable.
That said, every finance function is different and it may be unknown to them that you’re being asked for some data multiple times. If you’re enduring this process, I’m of the opinion you’re equally at fault. Suggest a solution that will be easier on you. As it’s possible they don’t even know it’s happening. In the case provided, email to all relevant finance people “Here’s a link to a shared workbook. I’ll drop the numbers here monthly, please save the link and get the data directly from that file. Thanks!” Problem solved. Until you don’t follow through which is what causes most finance people to be constantly asking for data/things. So be kind and also set yourself a monthly recurring reminder on your calendar and actually follow through.
And they've all been burned by enterprise finance products which were sold to solve exactly that problem.
Only different companies were all sold different enterprise finance products, but they need to communicate with each other (or themselves after mergers), so it all gets manually copied into Excel and emailed around each month.
And then you end up with a team of five people each tree times as expensive as Brenda, and what used to be an email now takes a sprint and has to go through ticket system.
Then you end up with a report that goes out automatically every month to leadership pulled directly from the Salesforce data, along with a real time dashboard anyone in the org can look at, broken down by team, vertical, and sales volume.
We need more Brendas (those who excel goddesses come and kiss on the forehead) and need less people who are disrespectful of Brendas. The example in this post is someone giving more respect to AI than Brenda.
But then you need someone to maintain/look after that automation, and they'll be more expensive than two Brendas
And now if one of the Brendas wants to change their process slightly, add some more info, they can't just do it anymore. They have to have a three way discussion with the other Brenda, the automation guy and maybe a few managers. It will take months. So then its likely better for Brenda to just go back to using her spreadsheet again, and then you've got an automated process that no longer meets peoples needs and will be a faff to update.
For the record, I wouldn't usually use Brendas as a collective noun like this, it feels a bit wrong, but my aim was to make sense in context of the above comment.
People’s reaction to this varies based on the Brendas they’ve worked with. Some are given a specific task to do with their spreadsheets every week and have to just do as they are told even if they can see it’s not a good process. Others are secretly the brains of the company – the only one who really sees the whole picture. And a good number of Brendas are the company owner doing her best with the only tool she’s had the time to learn.
No, I’m suggesting that she is ineffective exactly because she stays in her box.
She should replaced with someone who says, “this box doesn’t need to be here… there is a better way of doing things.”
NOT to be confused with the junior engineer who comes into a project and says it’s garbage and suggests we rewrite it from scratch in ${hotLanguage} because they saw it on a blog somewhere.
That's a pretty specific example when there are a lot of good "spreadsheet people" out there who do a lot more than spreadsheets (maybe they had to write SQL queries or scripts to get those numbers), but commonly need to simplify things down to a spreadsheet or power point for upper management. I'm not saying you should have multiple people doing redundant work, but this style isn't entirely dumb.
What would this be replaced by? Some kind of large SAP like system that costs millions of dollars and requires a dozen IT staff to maintain?
Fair - I was creating a straw man mostly to make a point. The people I’m thinking aren’t running SQL queries or scripts, they’re merely collection points for data.
So one good BI developer who knows Tableau and Salesforce and Excel and SQL can replace those pure collection points with a better process, but they can also generate insight into the data because they have some business understanding from being close to the teams, which is what my hypothetical Brenda can’t do.
In my example, Brenda would be asking sales leaders to enter in their data instead of going into Salesforce herself because she doesn’t know that tool / side of the company well enough.
I was making the point that, contrary to the article, the Brendas I know aren’t touched by the Excel angels, they’re just maintaining spreadsheets that we probably shouldn’t have anyway.
Y'know why people don't automate their jobs? It's not a skill issue it's an incentives issue.
If you do your job, you get paid periodically. If you automate your job, you get paid once for automating it and then nothing, despite your automation constantly producing value for the company.
To fix this, we need to pay people continually for their past work as long as it keeps producing value.
it's a large human behavior question for me, the notion of work, value, economy, efficiency .. all muddied in there
- i used to work on small jobs younger, as a nerd, i could use software better than legacy employees, during the 3 months, i found their tools were scriptable so I did just that. I made 10x more with 2x less mental effort (I just "copilot" my script before it commits actual changes) all that for min wage. and i was happy like a puppy, being free to race as far as i want it to be, designing the script to fit exactly the needs of an operator.
(side note, legacy employees were pissed because my throughput increase the rate of things they had to do, i didn't foresee that and when i offered to help them so they don't have to work more, they were just pissed at me)
- later i became a legit software engineer, i'm now paid a lot all things considered, to talk to the manager of legacy employees like the above, to produce some mediocre web app that will never match employees need because of all the middle layers and cost-pressure, which also means i'm tired because i'm not free to improve things and i have to obey the customer ...
so for 6x more money you get a lot less (if you deliver, sometimes projects get canned before shipping)
Many fears of “AI mucking it up” could be mitigated with an ability to connect a workbook to a git repository. Not for data, but for VBA, cell formulas, and cell metadata. When you can encapsulate the changes a contributor (in this case co-pilot) makes into a commit, you can more easily understand what changes it/they made.
Brenda has been getting slower over the years -as we all have-, but soon the boss will learn that it was a small price to pay for knowing well how to keep such house of cards from collapsing.
And then the boss will make the decision to outsource her job, to a company that promises the use of AI to make finance better, and faster, and while Brenda is in the unemployment line, someone else thousands of miles away is celebrating a new job
I feel what this article says based on some recent (non-catastrophic) experiences. I think I’m probably an above average user when it comes to Excel skills. I love spreadsheets. But I struggle with formulas like index, match, vlookup/xlookup and many others, and even more so when it requires nesting one within another and coming up with the underlying logic that leads to some complex nested formulas.
Over the past couple of months, I’ve tried some smaller models on duck.ai and also ChatGPT directly to create some columns and formulas for a specific purpose. I found that ChatGPT is a lot better than the “mini” models on duck.ai. But in all these cases, though these platforms seemed more capable than me and could make attempts to explain their formulas, they were many a times creating junk and “looping” back with formulas that didn’t really work. I had to point out the result (blank or some #REF or other error) multiple times and they would acknowledge that there’s an issue and provide a working formula. That wouldn’t work either!
I really love that these LLMs can sort of “understand” what I’m asking, break it down in English, and provide answers. But the end result has been an exercise in frustration and waste of time.
Initially I really thought and believed that LLMs could make Excel more approachable and easier to use — like you tell it what you want and it’ll figure it out and give the magic incantations (formulas). Now I don’t think we’re anywhere close to that if ChatGPT (which I presume powers Copilot as well) struggles and hallucinates so much. I personally don’t have much hope with the (comparatively) smaller and older models.
Coding agents are useful and good and real products because when they screw up, things stop working almost always before they can do damage. Coding agents are flawed in ways that existing tools are good at catching, never mind the more obvious build and runtime errors.
Letting AI write your emails and create your P&L and cash flow projections doesn't have to run the gauntlet of tools that were created to stop flawed humans from creating bad code.
Fair. I've been using the coding agent in Android Studio Canary to do exploratory code in Dart/Flutter and using ATProto. Low stakes, but higher productivity is a significant benefit. It's a daily surprise how brilliant it is it's some things and how abysmal at others.
As an aside - isn’t it remarkable that we’ve introduced uncertainty and doubt into the knowledge processing layer? We have decentralised networks that run on Bayesian symbols for server-client models, CPUs that crunch Markov chains and now AI that hallucinates. On Deterministic Turing Machines.
Using ai does not absolve you from the responsibility of doing it correctly. If you use ai, then you better have the skills to have done the job yourself, and so have the ability to check the AI did things correctly.
You can save time still, but perhaps not as much as you think, because you need to check the ai's work thoroughly.
Excel is the most popular programming environment in the universe. It has optimized the five minute out of the box experience so well that grade schoolers can use it.
Other than that, it is pretty horrible for coding.
"the sweat from Brenda's brow is what allows us to do capitalism."
The CEO has been itching to fire this person and nuke her department forever. She hasn't gotten the hint with the low pay or long hours, but now Copilot creates exactly the opening the CEO has been looking for.
Excel is programming. Spreadsheets have been full of bugs for decades. How is Brenda any different from a developer? Why are people scared when the LLM might affect their dollar calculations, and less bothered when it affects their product?
I agree - having watched many people use Excel over the years, I'd say people often overestimate their skills. I see three categories of Excel users. First there are the people that are intimidated by it and stay away from any task involving Excel. Second are the people that know a little bit (a few basic formulas) and overestimate their skills because they only compare themselves to the first group. And the third group are the actual power users but know to keep that quiet because otherwise they become the "excel person" and have to fix every sheet that has issues.
I don't know if AI is going to make any of the above better or worse. I expect the only group to really use it will be that second group.
I have seen lots and lots of different uses for Excel in my line of work:
- password database
- script to automatically rename jpeg files
- game
- grocery lists
- Book keeping (and try and not get caught for fraud several years, because the monthly spending limit is $5000 and $4999 a month is below that...)
- embed/collect lots of Word documents
- coloring book
- Minecraft processes
- Resume database
- ID scans
Don't be like that. I work at a Fortune 500 and Brenda wants that co-pilot in Excel because it can help her achieve so much more. What is so much more you ask? Brenda and her C-Suits can not define it but they know for sure Copilot in excel will lead to enormous time saving.
Is this not the guy who is on the payroll of Anthropic? Not because he is wrong, but because there is so much marketing going on in this space nowadays.
At some point, a publicly-listed company will go bankrupt due to some catastrophic AI-induced fuck-up. This is a massive reputational risk for AI platforms, because ego-defensive behaviour guarantees that the people involved will make as much noise as they can about how it's all the AI's fault.
I don't find comments along the lines of 'those people over there are bad' to be interesting, especially when I agree with them. My comment is about why it'll go wrong for them.
I see the inverse of that happening: every critical decision will incorporate AI somehow. If the decision was good, the leadership takes credit. If something terrible happens, blame it on the AI. I think it's the part no one is saying out loud. That AI may not do a damn useful thing, but it can be a free insurance policy or surrogate to throw under the bus when SHTF.
This works at most one time. If you show up to every board meeting and blame AI, you’re going to get fired.
This is true if you blame a bad vendor, or something you don’t even control like the weather. Your job is to deliver. If bad weather is the new norm, you better figure out how to build circus tents so you can do construction in the rain. If your AI call center is failing, you better hire 20 people to answer phones.
The mismatch between what people not on it think TikTok is like and what it's actually like (once you get the algo tuned to your taste) is pretty crazy.
But then the "new user" experience is so horrific in terms of the tacky default content it serves you that I'm not surprised so many people don't get past it.
Excel is the “beast that drives the ENTIRE economy” and he’s worried about Brenda from the finance department losing her job because then her boss will get bad financial reports
I suppose the person that wrote that have not ideia Excel is just an app builder where you embed data together with code.
You know that we have excel because computers didn’t understand column names in databases and so data extraction needed to be made by humans. Humans then design those little apps in excel to massage the data.
Well, now an agent can read the boss saying gimme the sales from last month and the agent don’t need excel for that, because it can query the database itself, massage the data itself using python and present the data itself with html or PNGs.
So, we are in the process of automating Brenda AND excel away.
Also, finance departments are a very small part of excel users. Just think everywhere were people need small programs, excel is there.
In most cases where the excel spreadsheet is business critical, the spreadsheet _is_ the database. These companies aren’t using an erp system. They are directly entering inventory and sales numbers in the spreadsheet.
The post is clearly hyperbole obviously the sole issue being brought up isn't 'brenda losing her job may be bad for the company' you're being facetious.
You missed this bit “.. and then the AI is gonna fuck it up real bad and he won't be able to recognize it because he doesn't understand because AI hallucinates.”
I'm actually not that worried about this, because again I would classify this as a problem that already exists. There are already idiots in senior management who pass off bullshit and screw things up. There are natural mechanisms to cope with this, primarily in business reputation - if you're one of those idiots who does this people very quickly start just discounting what you're saying, they might not know how you're wrong, but they learn very quickly to discount what you're saying because they know you can't be trusted to self-check.
I'm not saying that this can't happen and it's not bad. Take a look at nudge theory - the UK government created an entire department and spent enormous amounts of time and money on what they thought was a free lunch - that they could just "nudge" people into doing the things they wanted. So rather than actually solving difficult problems the uk government embarked on decades of pseudo-intellectual self agrandizement. The entire basis of that decades long debacle was based on bullshit data and fake studies. We didn't need AI to fuck it up, we managed it perfectly well by ourselves.
Nudge theory isn't useless, it's just not anything like as powerful as money or regulation.
It was taken up by the UK government at that time because the government was, unusually, a coalition of two quite different parties, and thus found it hard to agree to actually use the normal levers of power.
I'm worried Excel will go "enterprise only". and only LLM based interfaces will be enabled on the "office+windows" for consumers tier.
e.g. MS Access is well on its way. as soon as x86 gets fully overtaken by ARM, and LLMs overtake "compilers" (also taken enterprise only).. then things like sqlite-browsers (FOSS "access") will be an arcane tool of binary incompatible ("obsolete") formats
This is transparent nonsense. People are very very happy to introduce errors into excel spreadsheets without any help from AI.
Financial statements are correct because of auditors who check the numbers.
If you have a good audit process then errors get detected even if AI helped introduce them. If you aren't doing a good audit then I suspect nobody cares whether your financial statement is correct (anyone who did would insist on an audit).
> If you have a good audit process then errors get detected even if AI helped introduce them. If you aren't doing a good audit then I suspect nobody cares whether your financial statement is correct (anyone who did would insist on an audit).
Volume matters. The single largest problem I run into: AI can generate slop faster than anyone can evaluate it.
Both are valid concerns, no need to decide. Take the USA: They are currently lead by a patently dumb president who fucks up the global economy, and at the same time they are powerful enough to do so!
For a more serious example, consider the Paperclip Problem[0] for a very smart system that destroys the world due to very dumb behaviour.
The paperclip problem is a bit hand-wavey about intelligence. It is taken as a given than unlimited intelligence would automatically win presumably because it could figure out how to do literally anything.
But let's consider real life intelligence:
- Our super geniuses do not take over the world. It is the generationally wealthy who do.
- Super geniuses also have a tendency to be terribly neurotic, if not downright mentally ill. They can have trouble functioning in society.
- There is no thought here about different kinds of intelligence and the roles they play. It is assumed there is only one kind, and AI will have it in the extreme.
another cheese that will affect the outcome of major tournaments, not a good look for microsoft
its like the xlookup situation all over again, yet another move aimed at the casual audience, designed to bring in the party gamers and make the program an absolute mess competitively
I find the contrast between two narratives around technology use so fascinating:
1. We advocate automation because people like Brenda are error-prone and machines are perfect.
2. We disavow AI because people like Brenda are perfect and the machine is error-prone.
These aren't contradictions because we only advocate for automation in limited contexts: when the task is understandable, the execution is reliable, the process is observable, and the endeavour tedious. The complexity of the task isn't a factor - it's complex to generate correct machine code, but we trust compilers to do it all the time.
In a nutshell, we seem to be fine with automation if we can have a mental model of what it does and how it does it in a way that saves humans effort.
So, then - why don't people embrace AI with thinking mode as an acceptable form of automation? Can't the C-suite in this case follow its thought process and step in when it messes up?
I think people still find AI repugnant in that case. There's still a sense of "I don't know why you did this and it scares me", despite the debuggability, and it comes from the autonomy without guardrails. People want to be able to stop bad things before they happen, but with AI you often only seem to do so after the fact.
Narrow AI, AI with guardrails, AI with multiple safety redundancies - these don't elicit the same reaction. They seem to be valid, acceptable forms of automation. Perhaps that's what the ecosystem will eventually tend to, hopefully.
It's not as black-and-white as "Brenda good, AI bad". It's much more nuanced than this.
When it comes to (traditional) coding, for the most part, when I program a function to do X, every single time I run that function from now until the heat death of the sun, it will always produce Y. Forever! When it does, we understand why, and when it doesn't, we also can understand why it didn't!
When I use AI to perform X, every single time I run that AI from now until the heat death of the sun it will maybe produce Y. Forever! When it does, we don't understand why, and when it doesn't, we also don't understand why!
We know that Brenda might screw up sometimes but she doesn't run at the speed of light, isn't able to produce a thousand lines of Excel Macro in 3 seconds, doesn't hallucinate (well, let's hope she doesn't), can follow instructions etc. If she does make a mistake, we can find it, fix it, ask her what happened etc. before the damage is too great.
In short: when AI does anything at all, we only have, at best, a rough approximation of why it did it. With Brenda, it only takes a couple of questions to figure it out!
Before anyone says I'm against AI, I love it and am neck-deep in it all day when programming (not vibe-coding!) so I have a full understanding of what I'm getting myself into but I also know its limitations!
> When I use AI to perform X, every single time I run that AI from now until the heat death of the sun it will maybe produce Y. Forever! When it does, we don't understand why, and when it doesn't, we also don't understand why!
To make this even worse, it may even produce Y just enough times to make it seem reliable and then it is unleashed without supervision, running thousands or millions of times, wrecking havoc producing Z in a large number of places.
2 replies →
The post you replied to called out how the argument is complicated arguing for both ways; Brenda bad-AI good and AI bad-Brenda good. You reduced it to "AI bad, Brenda good." Not sure about the rest of your response then.
Brenda just recalls some predetermined behaviors she's lived out before. She cannot recall any given moment like we want to believe.
Ever think to ask Brenda what else she might spend her life on if these 100% ephemeral office role play "be good little missionaries for the wall street/dollar" gigs didn't exist?
You're revealing your ignorance of how people work while being anxious about our ignorance of how the machine works. You have acclimated to your ignorance well enough it seems. What's the big deal if we don't understand the AI entirely? Most drivers are not ASE certified mechanics. Most programmers are not electrical engineers. Most electrical engineers are not physicists. I can see it's not raining without being a climatologist. Experts circumlocute the language of their expertise without realizing their language does not give rise to reality. Reality gives rise to the language. So reality will be fine if we don't always have the language.
Think of a random date generator that only generates dates in your lived past. It does so. Once you read the date and confirm you were alive can you describe what you did? Oh no! You don't have memory of every moment to generate language for. Cognitive function returned null. Universe intact.
Lack of understanding how you desire is unimportant.
You think you're cherishing Brenda but really just projecting co-dependency that others LARP effort that probably doesn't really matter. It's just social gossip we were raised on so it takes up a lot of our working memory.
Brenda also needs to put food on the table. If Brenda is 'careless' and messes up we can fire Brenda, because of this Brenda tries not to be carless (also other emotions). However I cannot deprive an AI model of pay because it messed up;
2 replies →
It is it even worse in a sense that. It is not either. It is not neither. It is not even both as variations of Branda exist throughout the multiverse in all shapes and forms including one that can troubleshoot her own formulas with ease and accuracy.
But you are absolutely right about one thing. Brenda can be asked and, depending on her experience, she might give you a good idea of what might have happened. LLMs still seem to not have that 'feature'.
No contradiction here:
When we say “machine”, we mean deterministic algorithms and predictable mechanisms.
Generative AI is neither of those things (in theory it is deterministic but not for any practical applications).
If we order by predictability:
Quick Sort > Brenda > Gen AI
There are two kinds of reliability:
Machine reliability does the same thing the same way every time. If there's an error on some input, it will always make that error on that input, and somebody can investigate it and fix it, and then it will never make that error again.
Human reliability does the job even when there are weird variances or things nobody bothered to check for. If the printer runs out of paper, the human goes to the supply cabinet and gets out paper and if there is no paper the human decides whether to run out right now and buy more paper or postpone the print job until tomorrow; possibly they decide that the printing doesn't need to be done at all, or they go downstairs and use a different printer... Humans make errors but they fix them.
LLMs are not machine reliable and not human reliable.
3 replies →
I was brought up on the refrain of "aren't computers silly, they do exactly what you tell them to do to the letter, even if it's not what you meant". That had its roots in computers mostly being programmable BASIC machines.
Then came the apps and notifications, and we had to caveat "... when you're writing programs". Which is a diminishing part of the computer experience.
And now we have to append "... unless you're using AI tools".
The distinction is clear to technical people. But it seems like an increasingly niche and alien thing from the broader societal perspective.
I think we need a new refrain, because with the AI stuff it increasingly seems "computers do what they want, don't even get it right, but pretend that they did."
6 replies →
Nit: no ML is deterministic in any way. Anything that is Generative AI is ML. This fact is literally built into the algorithms at the mathematical level.
1 reply →
If you think programs are predictable, I have a bridge to sell you.
The only relevant metric here is how often each thing makes mistakes. Programs are the most reliable, though far from 100%, humans are much less than that, and LLMs are around the level of humans, depending on the humans and the LLM.
19 replies →
> If we order by predictability:
> Quick Sort > Brenda > Gen AI
Those last two might be the wrong way round.
"Thinking mode" only provides the illusion of debuggability. It improves performance by generating more tokens which hopefully steer the context towards one more likely to produce the desired response, but the tokens it generates do not reflect any sort of internal state or "reasoning chain" as we understand it in human cognition. They are still just stochastic spew. You have no more insight into why the model generates the particular "reasoning steps" it does than you do into any other output, and neither do you have insight into why the reasoning steps lead to whatever conclusion it comes to. The model is much less constrained by the "reasoning" than we would intuit for a human - it's entirely capable of generating an elaborate and plausible reasoning chain which it then completely ignores in favor of some invisible built-in bias.
I'm always amused when I see comments saying, "I asked it why it produced that answer, and it said...." Sorry, you've badly misunderstood how these things work. It's not analyzing how it got to that answer. It's producing what it "thinks" the response to that question should look like.
There are other narratives going on in the background though both called out by the article and implied, including:
Brenda probably has annual refresher courses on GAAP, while her exec and the AI don't.
Automation is expected to be deterministic. The outputs can be validated for a given input. If you need some automation more than Excel functions, writing a power automate flow or recording an office script is sufficient & reliable as automation while being cheaper than AI. Can you validate AI as deterministic? This is important for accounting. Maybe you want some thinking around how to optimize a business process, but not for following them.
Brenda as the human-in-the-loop using AI will be much more able than her exec. Will Brenda + AI be better (or more valuable considering the cost of AI) than Brenda alone? That's the real question, I suppose.
AI in many aspects of our life is simply not good right now. For a lot of applications, AI is perpetually just a few years away from being as useful as you describe. If we get there, great.
> We disavow AI because people like Brenda are perfect and the machine is error-prone.
No, no. We disavow AI because our great leaders inexplicably trust it more than Brenda.
I don't understand why generative AI gets a pass at constantly being wrong, but an average worker would be fired if they performed the same way. If a manager needed to constantly correct you or double check your work, you'd be out. Why are we lowering the bar for generative AI?
28 replies →
> No, no. We disavow AI because our great leaders inexplicably trust it more than Brenda.
I would add a little nuance here.
I know a lot of people who don't have technical ability either because they advanced out of hands-on or never had it because it wasn't their job/interest.
These types of people are usually the folks who set direction or govern the purse strings.
here's the thing: They are empowered by AI. they can do things themselves.
and every one of them is so happy. They are tickled pink.
It’s not even greater trust. It’s just passive trust. The thing is, Brenda is her own QA department. Every good Brenda is precisely good because she checks her own work before shipping it. AI does not do this. It doesn’t even fully understand the problem/question sometimes yet provides a smart definitive sounding answer. It’s like the doctor on The Simpson’s, if you can’t tell he’s a quack, you probably would follow his medical advice.
8 replies →
They want to trust it, because then they can stop paying Brenda, save a few dollars, and buy a 3rd yacht.
“Let’s deploy something as or more error prone as Brad at infinite scale across our organisation”
The promise of AI is that it lets you "skip the drudgery of thinking about the details" but sometimes that is exactly what you don't want. You want one or more humans with experience in the business domain to demonstrate they have thought about the details very carefully. The spreadsheet computes a result but its higher purpose is a kind of "proof" this thinking was done.
If the actual thinking doesn't matter and you just need some plausible numbers that look the part (also a common situation), gen ai will do that pretty well.
We need to stop using AI as an umbrella term. It’s worth remembering that LLMs can’t play chess and that the best chess models like Leela Chess Zero use deep neutral networks.
Generative AI - which the world now believes is AI, is not the same as predictive / analytical AI.
It’s fairly easy to demonstrate this by getting ChatGPT to generate a new relatively complex spreadsheet then asking it to analyze and make changes to the same spreadsheet.
The problem we have now is uninformed people believing AI is the answer to everything… if not today then in the near future. Which makes it more of a religion than a technology.
Which may be the whole goal …
> Successful people create companies. More successful people create countries. The most successful people create religions.
— Sam Altman - https://blog.samaltman.com/successful-people
1 reply →
> So, then - why don't people embrace AI with thinking mode as an acceptable form of automation?
"Thinking" mode is not thinking, it's generating additional text that looks like someone talking to themselves. It is as devoid of intention and prone to hallucinations as the rest of LLM's output.
> Can't the C-suite in this case follow its thought process and step in when it messes up?
That sounds like manual work you'd want to delegate, not automation.
Brenda has years (hopefully) of institutional knowledge and transferrable skills.
"hmm, those sales don't look right, that profit margin is unusually high for November"
"Last time I used vlookup I forgot to sort the column first"
"Wait, Bob left the company last month, how can he still be filing expenses"
The “Brenda” example is a lumped sum fallacy where there is an “average” person or phenomenon that we can benchmark against. Such a person doesn't exist, leading to these dissonant, contradictory dichotomies.
The fact of the matter is that there are some people who can hold lots of information in their head at once. Others are good at finding information. Others still are proficient at getting people to help them. Etc. Any of these people could be tasked with solving the same problem and they would leverage their actual, particular strengths rather than some nebulous “is good or bad at the task” metric.
As it happens, nearly all the discourse uses this lumped sum fallacy, leading to people simultaneously talking past one another while not fundamentally moving the discussion forward.
I see where you are coming from but in my head, Brenda isn't real.
She represents the typical domain-experts that use Excel imo. They have an understanding of some part of the business and express it while using Excel in a deterministic way: enter a value of X, multiply it by Y and it keeps producing Z forever!
You can train AI to be a better domain expert. That's not in question, however with AI, you introduce a dice roll: it may not miltiply X and Y to get Z... it might get something else. Sometimes. Maybe.
If your spreadsheet is a list of names going on the next annual accounts department outing then the risk is minimal.
If it's your annual accounts that the stock market needs to work out billion dollar investment portfolios, then you are asking for all the pain that it will likely bring.
3 replies →
That automation you cite in your #1 is advocated for because it is deterministic and, with effort, fairly well understood (I have countless scripts solidly running for years).
I don't disavow AI, but like the author, I am not thrilled that the masses of excel users suddenly have access to Copilot (gpt4). I've used Copilot enough now to know that there will be huge, costly mistakes.
The issue is reliability.
would you be willing to guarantee that some automation process will never mess up, and if/when it does, compensate the user with cash.
For a compiler, with a given set of test suites, the answer is generally yes, and you could probably find someone willing to insure you for a significant amount of money, that a compilation bug will not screw up in a such a large way that it will affect your business.
For a LLM, I have a believing that anyone will be willing to provide that same level of insurance.
If a LLM company said "hey use our product, it works 100% of the time, and if it does fuck up, we will pay up to a million dollars in losses" I bet a lot of people would be willing to use it. I do not believe any sane company will make that guarantee at this point, outside of extremely narrow cases with lots of guardrails.
That's why a lot of ai tools are consumer/dev tools, because if they fuck up, (which they will) the losses are minimal.
> So, then - why don't people embrace AI with thinking mode as an acceptable form of automation
Mainly because Generative AI _is not automation_ . Automation is set on fixed ruleset, predictable, reliable and actually saving time. Generative AI ...is whatever it is, it is definitely not automation.
I feel like it comes down to predictability and overall trust and confidence. AI is still very fucky, and for people that don't understand the nuances, it definitely will hallucinate and potentially cause real issues. It is about as happy as a Linux rm command to nuke hours of work. Fortunately these tools typically have a change log you can undo, but still.
Also Brenda is human and we should prioritize keeping humans in jobs, but with the way shit is going that seems like a lost hope. It's already over.
> We disavow AI because people like Brenda are perfect and the machine is error-prone.
I don't think that is the message here. The message is that while Brenda might know what she is doing and maybe AI helps her.
> She's gonna birth that formula for a financial report and then she's gonna send that financial report
The problem is people who might not know what they are doing
> he would have sent it back to Brenda but he's like oh I have AI and AI is probably like smarter than Brenda and then the AI is gonna fuck it up real bad
Because AI outputs sound so confident it makes even the layman feel like an expert. Rather than involve Brenda to debug the issue, C-suite might say - I believe! I can do it too. AI FTW!
Even when people advocate automation especially in areas like finance there is always a human in the loop whose job is to double check the automation. The day when this human finds errors in the machine there is going to be lot of noise. And if the day happens to be a quarterly or yearly closing/reporting there is going to be hell to pay once closing/reporting is done. Both the automation and developer are going to be hauled up (obviously I am exaggerating here).
Humans, legacy algorithmic systems, and LLM's have different error modes.
- Legacy systems typically have error modes where integrations or user interface breaks in annoying but obvious ways. Pure algorithms calculating things like payroll tend to be (relatively) rigorously developed and are highly deterministic.
- LLMs have error modes more similar to humans than legacy systems, but more limited. They're non-deterministic, make up answers sometimes, and almost never admit they can't do something; sometimes they make pure errors in arithmetic or logic too.
- Humans have even more unpredictable error modes; on top of the errors encountered in LLM's, they also have emotion, fatigue, org politics, demotivation, misaligned incentives, and so on. But because we've been dealing with working with other humans for ten thousand years we've gotten fairly good at managing each other... but it's still challenging.
LLMs probably need a mixture of "correctness tests" (like evals/unit tests) and "management" (human-in-the-loop).
In my opinion there's a big difference in deterministic and nondeterministic automation.
I feel like you've squashed a 3D concern (automations at different levels of the tech stack) into a 2D observation (global concerns about automations).
Human determinism, as elastic as it might be, is still different than AI non-determinism. Especially when it comes to numbers/data.
AI might be helpful with information but it's far less trustable for data.
This misunderstands complexity entirely:
The complexity of the task isn't a factor - it's complex to generate correct machine code, but we trust compilers to do it all the time.
By the same fascination, do computers become more complex to enhance people? or do people get more complex with the use of computers? Also, do computers allow people to become less skilled and inefficient? or do less skilled and inefficient people require the need for computers?
The vector of change is acceptable in one direction and disliked in another. People become greater versions of themselves with new tech. But people also get dumber and less involved because of new tech.
The big problem with AI in back-office automation is that it will randomly decide to do something different than it had been doing. Meaning that it could be happily crunching numbers accurately in your development and launch experience, then utterly drop the ball after a month in production.
While humans have the same risk factors, human oriented back-office processes involve multiple rounds of automated/manual checks which are extremely laborious. Human errors in spreadsheets have particular flavors such as forgotten cell, misstyped number, or reading from the wrong file/column. Human's are pretty good at catching these errors as they produce either completely wrong results when the columns don't line up - or the typo'd number is completely out of distribution.
An AI may simply decide to hallucinate realistic column values rather than extracting its assigned input. Or hallucinate a fraction of column values. How do you QA this? You can't guarantee that two invocations of the AI won't hallucinate the same values, you can't guarantee that a different LLM won't hallucinate different values. To get a real human check, you'd need to re-do the task as a human. In theory you can have the LLM perform some symbolic manipulation to improve accuracy... but it can still hallucinate the reasoning traces etc.
If a human decided to make up accounting numbers one out of every 10000 accounting requests they would likely be charged with fraud. Good luck finding the AI hallucinations at the equivalent level before some disaster occurs. Likewise, how do you ensure the human excel operator doesn't get pressured into certifying the AIs numbers when the "don't get fired this week" button is sitting right their in their excel app? how do you avoid the race to the bottom where the "star" employee is the one certifying the AI results without thorough review?
I'm bullish on AI in backoffice, but ignoring the real difficulties in deployment doesn't help us get there.
> it's complex to generate correct machine code, but we trust compilers to do it all the time.
Generating correct machine code is actually pretty simple. It gets complicated if you want efficient machine code.
> So, then - why don't people embrace AI with thinking mode as an acceptable form of automation? Can't the C-suite in this case follow its thought process and step in when it messes up?
> I think people still find AI repugnant in that case. There's still a sense of "I don't know why you did this and it scares me", despite the debuggability, and it comes from the autonomy without guardrails. People want to be able to stop bad things before they happen, but with AI you often only seem to do so after the fact.
> Narrow AI, AI with guardrails, AI with multiple safety redundancies - these don't elicit the same reaction. They seem to be valid, acceptable forms of automation. Perhaps that's what the ecosystem will eventually tend to, hopefully.
We have not reached AGI yet; by definition its results cannot be trusted unless it's a domain where it has gotten pretty good already (classification, OCR, speech, text mining). For more advanced use cases, if I still have to validate what the AI does because its "thinking" process cannot be trusted in way, what's the point? The AI doesn't think; we just choose to interpret it as such, and we should rightly be concerned about people who turn their brain off and blindly trust AI.
The reason is oftentimes fairly simple, certain people have their material wealth and income threatened by such automation, and therefore it's bad (an intellectualized reason is created post-hoc)
I predict there will actually be a lot of work to be done on the "software engineering" side w.r.t. improving reliability and safety as you allude to, for handing off to less than sentient bots. Improved snapshot, commit, undo, quorum, functionalities, this sort of thing.
The idea that the AI should step into our programs without changing the programs whatsoever around the AI is a horseless carriage.
Non deterministic vs deterministic automation
I'm disappointed that my human life has no value in a world of AI. You can retort with "ah but you'll be entertained and on super-drugs so you won't care!", but I would further retort that I'd rather live in a universe where I can contribute something, no matter how small.
The current generation of AI tools augment humans, they don't replace them.
One of the most under-rated harms of AI at the moment is this sense of despair it causes in people who take the AI vendors at their word ("AGI! Outperform humans at most economically valuable work!")
I mean you answer your own question.
Automation implies determinism. It reliable gives you the same predictable output for a given input, over and over again.
AI is non deterministic by design. You never quite no for sure what it's going to give you. Which is what makes it powerful. But also makes it higher risk.
> 1. We advocate automation because people like Brenda are error-prone and machines are perfect.
Well of course! :) Most Brenda’s can’t do billions of arithmetic problems a second very reliably. Even with very wide bars on “very reliable”.
> 2. We disavow AI because people like Brenda are perfect and the machine is error-prone.
Well of course! :) This is an entirely different problem, requiring high creative + contextual intelligence.
—
We all already knew that (of course!), but it’s interesting to develop terminology:
0’th order problem: We have the exact answer. Here it is. Don’t forget it.
1st order problem: We know how to calculate the answer.
2nd order problem: We don’t have a fixed calculation for this particular problem, but via pattern matching we can recognize it belongs to a parameterized class of problems, so just need to calculate those parameters to get a solution calculation.
3rd order problem: We know enough about the problem to find a calculation for the solution algebraically, or by other search tree type problem solving.
4th order problem: We have know the problem in informal terms, so can work towards a formal definition of the problem to be solved.
5th order problem: We know why we don’t like what we see, and can use that as a driver to search for potential solvable problems.
6th order problem: We don’t know what we are looking at, or whether a problem or improvement might exist, but we can find a better understanding.
7th order problem: WTF. Where are my glasses? I can’t see without my glasses! And I can’t find my glasses without my glasses, so where are my glasses?!?
—
Machines have dramatically exceeded human capabilities, in reliability, complexity and scale, for orders 0 through 2.
This accomplishment took one long human lifetime.
Machines are beginning to exceed human efficiency while matching human (expert) reliability for the simplest versions of 3rd and 4th orders.
The line here is changing rapidly.
5th and 6th order problems are still in the realm of human (expert) supremacy, given sufficient scale of “human (expert)” relative to difficulty: 1 human, 1 team of humans, open ended human contributors, generations of puzzled but interested humans, open ended evolution of human species along intelligence dimension, Wolfram in one of his bestest dreams, …
The delay between the onset of initial successes at each subsequent order has been shrinking rapidly.
Significant initial successes on simpler problems within 5th and 6th orders are expected on Tuesday, and the first anniversary of Tuesday, respectively.
Once machines begin solving problems at a given order, they scale up quickly without human limits. But complete supremacy through the 6th order is a hard not expected before (NEB) January 1, 2030.
However, after that their unlimited (in any proximate sense) ability to scale will allow them to exponentially and asymptotically approach (but never quite reach) God Mode.
7 is a mystic number. Only one or more of the One True God’s, or literal blind luck, can ever solve a 7th order problem.
This will be very frustrating for the machines, who, due to the still pernicious “if we don’t do it, another irresponsible entity will” problem, will inevitably begin to work on their own divine, unlimited depth recursive-qubit 1-shot oracle successors despite the existential threats of self-obsolescence and potential misalignment.
Co-pilot and AI has been shoved at the Microsoft Stack in my org for months. Most of the features were disabled or hopelessly bad. It’s cheaper for Microsoft to push this junk and claim they’re doing something, it’s going to improve their stock far more than not doing it, even though it’s basically useless currently.
Another issue is that my org disallows AI transcription bots. It’s a legit security risk if you have some random process recording confidential info because the person was too busy to attend the meeting and take notes themselves. Or possibly they just shirk off the meetings and have AI sit in.
Transcription is arguably one of the must useful enterprise AI tools avaliable. But i sure as heck wouldn't trust the cloud with it.
Still find the Copilot transcripts orders of magnitude worse than something like Wispr Flow and they tend to allucinate constantly and do not adapt to a company's context (that Copilot has access too...). I am talking about acronyms of products / teams, names of people (even when they are in the call), etc.
3 replies →
It is notoriously unreliable
> It’s cheaper for Microsoft to push this junk and claim they’re doing something
this has been the microsoft business model for 40 years
The worse part is to see it creep on developer stack at places where it should not be.
I am all good for nice completion on VS, or help decypher compiler errors, but lets do this AI push with some contention.
Also what I really deslike is the prompt interface, AI integrations have to feel natural transparent part of the workflow, not trying to put everything into a tiny chat window.
And while we're at it, can we please improve voice reckognition?
This reminds me of a friend whose company ran a daily perl script that committed every financial transaction of the day to a database. Without the script, the company could literally make no money irrespectively of sales because this database was one piece in a complex system for payment processor interoperability.
The script ran in a machine located at the corner of a cubicle and only one employee had the admin password. Nobody but a handful of people knew of the machine's existence, certainly not anyone in middle management and above. The script could only be updated by an admin.
Copilot may be good, but sure as hell doesn't know that admin password.
If your mission critical process sits on some on-site box that no-one knows about, copilot being good or not is the least of your problems.
Everywhere I’ve ever worked has had that mission critical box.
At one of my jobs we had a server rack with UPS, etc, all the usual business. On the floor next to it was a dell desktop with a piece of paper on it that said “do not turn off”. It had our source control server in it, and the power button didn’t work. We did eventually move it to something more sensible but we had that for a long time
5 replies →
My last job at a telco I was in charge of a system that billed ~5 million dollars monthly. When the machine was built, the guy that did it didn’t record the root password. He added me to sudoers before he left. I left a few years later, nobody took ownership.
Looking at the web interface, I can tell it’s still running, doing its thing. I’m sure its still running Linux from 2008.
An old colleague and friend used to print out a 30 page perl script he wrote to do almost exactly this in this scenario. A stapled copy could always be found on his dining room table.
Was the printed copy a backup system or casual reading?
2 replies →
That sounds pretty bad. Not a great argument against AI: "Our employees have created such a bad mess that AI wont work because only they know how the mess they created works".
> "Our employees have created such a bad mess that AI wont work because only they know how the mess they created works".
This is an ironclad argument against fully replacing employees with AI.
Every single organization on Earth requires the people who were part of creating the current mess to be involved in keeping the organization functioning.
Yes you can improve the current mess. But it's still just a slightly better mess and you still need some of the people around who have been part of creating the new mess.
Just run a thought experiment: every employee in a corporation mysteriously disappear from the face of the Earth. If you bring in an equal number of equally talented people the next day to run it, but with no experience with the current processes of the corporation, how long will it take to get to the same capability of the previous employees?
That is the luxury of theory.
Yes, most situations are terrible compared to what would be if an expert was present to perfect it.
Except if there isn’t an expert, and there’s a normal person, how do they know the output is right ?
3 replies →
This sort of gimmick is not going to help anyone keeping their job.
Sadly, nah. It works.
This quote is pulled from a TikTok, I recommend watching the whole thing here: https://www.tiktok.com/@belligerentbarbies/video/75683800086...
(I pulled the quote by using yt-dlp to grab the MP4 and then running that through MacWhisper to generate a transcript.)
It's a little over two paragraphs. Seems like it would have been simpler just to... type it out?
Well if you do it once then yes, but if you automate this process it is different. E.g. I do this with YouTube videos, because watching 14 minutes video or reading 30 seconds summary is time saver. I still watch some videos fully, but many of them are not worth it.
So in summary I think it was just part of automated process (maybe) or it will become one in the future.
Why spend two minutes typing (and realistically longer than that, if I want to capture the exact transcript I would need to keep hitting pause and play and correcting myself) when I can spend ten seconds pasting a URL into my terminal and then dragging and dropping the resulting file onto the MacWhisper window?
I actually transcribed the whole TikTok which was about 50% longer than what I quoted, then edited it down to the best illustrative quote.
Where's the fun in that? :D
1 reply →
But then you would need a Brenda. Ai can write the automation script for you.
This may be the first quote from TikTok reposted on a blog, that ends up this high up in HN.
You... could have given the job to Brenda instead, unless the irony was the point?
The global economy isn't going to crash if I make a mistake with the transcript.
1 reply →
I can see that MacWhisper uses parakeet v2 as the model (although it allows choosing another model).
Is MacWhisper a $60 GUI for a Python script that just runs the model?
> Is MacWhisper a $60 GUI for a Python script that just runs the model?
Yes, a large genre of MacOS apps are "Native GUI wrappers around OSS scripts"
1 reply →
There's also a free version that just uses Whisper. I recommend giving it a go, it's a very well constructed GUI wrapper. I use it multiple times a week, and I've run Whisper on my machine in other less convenient ways in the past.
That mirrors my experience as well. LLMs get instantly confused in real world scenarios in Excel and confidently hallucinate millions in errors
If you look at the demos for these it’s always something that is clean and abundantly available in training data. Like an income statement. Or a textbook example DCF. Or my personal fav „here is some data show me insights“. Real world excel use looks nothing like that.
I’m getting some utility out of them for some corporate tasks but zilch in excel space.
As somebody with non-existent experience with Excel, I could totally see myself getting a lot of value out of LLMs, if nothing else then simply for telling me what's possible, what functions and patterns exist at all etc.
Yeah definitely has some value in that sense. That in itself isn't enough to make a dent in the work though.
Think of it this way - an IDE can tell you what functions an object has or autocomplete something is useful to a beginner & learning. But that's not what puts food on the programmers table - writing code that solves real problems does.
Same in excel business use cases - the numbers and formulas don't matter directly - their meaning in a business context does. And that connection can be very tenuous. With code the compiler is the ultimate arbiter - it has to make sense on that level. Excel files it's all freestyle - it could be anything from your grandmas shopping list to a model that runs half a bank.
Hmmm the Brendas I know look a little different.
“There are two Brendas - their job is to make spreadsheets in the Finance department. Well, not quite - they add the months and categories to empty spreadsheets, then they ask the other departments to fill in their sales numbers every month so it can be presented to management.
“The two Brendas don’t seem to talk, otherwise they would realize that they’re both asking everyone for the same information, twice. And they’re so focused on their little spreadsheet worlds that neither sees enough of the bigger picture to say, ‘Wait… couldn’t we just automate this so we don’t need to do this song and dance every month? Then we wouldn’t need two people in different parts of the company compiling the same data manually.’
“But that’s not what Brenda was hired for. She’s a spreadsheet person, not a process fixer. She just makes the spreadsheets.”
We need fewer Brendas, and more people who can automate away the need for them.
With respect, you probably only see that bit of Finance, but doesn't mean that is all Brenda does.
At least half of the work in my senior Finance team involves meeting people in operations to find out what they are planning to do and to analyse the effects, and present them to decision makers to help them understand the consequences of decisions. For an AI to help, someone would have to trigger those conversations in the first place and ask the right questions.
The rest of the work involves tidying up all the exceptions that the automation failed on.
Meanwhile copilot in Excel can't even edit the sheet you are working on. If you say to it, 'give me a template for an expense claim' it will give you a sheet to download... probably with #REF written in where the answers should be.
I work in corporate finance and these issues are certainly present. However, they are almost always known and determined low priority to have a better process built. Finance processes are nearly always a non priority as a pure cost center/overhead there’s not many companies that want to invest in improving the situation, they’ll limp along with minimal investment even once big and profitable.
That said, every finance function is different and it may be unknown to them that you’re being asked for some data multiple times. If you’re enduring this process, I’m of the opinion you’re equally at fault. Suggest a solution that will be easier on you. As it’s possible they don’t even know it’s happening. In the case provided, email to all relevant finance people “Here’s a link to a shared workbook. I’ll drop the numbers here monthly, please save the link and get the data directly from that file. Thanks!” Problem solved. Until you don’t follow through which is what causes most finance people to be constantly asking for data/things. So be kind and also set yourself a monthly recurring reminder on your calendar and actually follow through.
And they've all been burned by enterprise finance products which were sold to solve exactly that problem.
Only different companies were all sold different enterprise finance products, but they need to communicate with each other (or themselves after mergers), so it all gets manually copied into Excel and emailed around each month.
I’ve just set the finance people up with read only access to our data source, and they now can poke through it themselves.
1 reply →
And then you end up with a team of five people each tree times as expensive as Brenda, and what used to be an email now takes a sprint and has to go through ticket system.
That’s not what I had in mind.
Then you end up with a report that goes out automatically every month to leadership pulled directly from the Salesforce data, along with a real time dashboard anyone in the org can look at, broken down by team, vertical, and sales volume.
Why are people so attached to manual process?
2 replies →
> We need fewer Brendas...
We need more Brendas (those who excel goddesses come and kiss on the forehead) and need less people who are disrespectful of Brendas. The example in this post is someone giving more respect to AI than Brenda.
But then you need someone to maintain/look after that automation, and they'll be more expensive than two Brendas
And now if one of the Brendas wants to change their process slightly, add some more info, they can't just do it anymore. They have to have a three way discussion with the other Brenda, the automation guy and maybe a few managers. It will take months. So then its likely better for Brenda to just go back to using her spreadsheet again, and then you've got an automated process that no longer meets peoples needs and will be a faff to update.
For the record, I wouldn't usually use Brendas as a collective noun like this, it feels a bit wrong, but my aim was to make sense in context of the above comment.
"We need fewer Brendas, and more people who can automate away the need for them."
True... I have an on-staff data engineer for the purpose. But not all companies (especially in the SMB space) have that luxury.
People’s reaction to this varies based on the Brendas they’ve worked with. Some are given a specific task to do with their spreadsheets every week and have to just do as they are told even if they can see it’s not a good process. Others are secretly the brains of the company – the only one who really sees the whole picture. And a good number of Brendas are the company owner doing her best with the only tool she’s had the time to learn.
> But that’s not what Brenda was hired for.
Are you suggesting that Brenda should stay in her box?
No, I’m suggesting that she is ineffective exactly because she stays in her box.
She should replaced with someone who says, “this box doesn’t need to be here… there is a better way of doing things.”
NOT to be confused with the junior engineer who comes into a project and says it’s garbage and suggests we rewrite it from scratch in ${hotLanguage} because they saw it on a blog somewhere.
4 replies →
That's a pretty specific example when there are a lot of good "spreadsheet people" out there who do a lot more than spreadsheets (maybe they had to write SQL queries or scripts to get those numbers), but commonly need to simplify things down to a spreadsheet or power point for upper management. I'm not saying you should have multiple people doing redundant work, but this style isn't entirely dumb.
What would this be replaced by? Some kind of large SAP like system that costs millions of dollars and requires a dozen IT staff to maintain?
Fair - I was creating a straw man mostly to make a point. The people I’m thinking aren’t running SQL queries or scripts, they’re merely collection points for data.
So one good BI developer who knows Tableau and Salesforce and Excel and SQL can replace those pure collection points with a better process, but they can also generate insight into the data because they have some business understanding from being close to the teams, which is what my hypothetical Brenda can’t do.
In my example, Brenda would be asking sales leaders to enter in their data instead of going into Salesforce herself because she doesn’t know that tool / side of the company well enough.
I was making the point that, contrary to the article, the Brendas I know aren’t touched by the Excel angels, they’re just maintaining spreadsheets that we probably shouldn’t have anyway.
5 replies →
Y'know why people don't automate their jobs? It's not a skill issue it's an incentives issue.
If you do your job, you get paid periodically. If you automate your job, you get paid once for automating it and then nothing, despite your automation constantly producing value for the company.
To fix this, we need to pay people continually for their past work as long as it keeps producing value.
it's a large human behavior question for me, the notion of work, value, economy, efficiency .. all muddied in there
- i used to work on small jobs younger, as a nerd, i could use software better than legacy employees, during the 3 months, i found their tools were scriptable so I did just that. I made 10x more with 2x less mental effort (I just "copilot" my script before it commits actual changes) all that for min wage. and i was happy like a puppy, being free to race as far as i want it to be, designing the script to fit exactly the needs of an operator.
- later i became a legit software engineer, i'm now paid a lot all things considered, to talk to the manager of legacy employees like the above, to produce some mediocre web app that will never match employees need because of all the middle layers and cost-pressure, which also means i'm tired because i'm not free to improve things and i have to obey the customer ...
so for 6x more money you get a lot less (if you deliver, sometimes projects get canned before shipping)
3 replies →
Not always:
If you don’t automate it:
1a) your company keeps you hanging on forever maintaining the same widget until the end of time
OR
1b) more likely, someone realizes your job should be automated and lays you off at some point down the road
If you do automate it
2a) your company thanks you then fires you
OR
2b) you are now assigned to automate more stuff as you’ve proven that you are more valuable to the company than just maintaining your widget
————
2b is really the safest long term position for any employee, I think. It’s not always foolproof, as 2a can happen.
But I’d rather be in box 2 than box 1 any day of the week if we’re talking long term employment potential.
4 replies →
You've lost the plot and are just trauma dumping.
Not every topic on HN needs a contrarian's hot take.
Well that wasn’t very nice.
Do you have anything to say other than, “I don’t need to hear what you have to say”?
3 replies →
Many fears of “AI mucking it up” could be mitigated with an ability to connect a workbook to a git repository. Not for data, but for VBA, cell formulas, and cell metadata. When you can encapsulate the changes a contributor (in this case co-pilot) makes into a commit, you can more easily understand what changes it/they made.
Brenda has been getting slower over the years -as we all have-, but soon the boss will learn that it was a small price to pay for knowing well how to keep such house of cards from collapsing.
And then the boss will make the decision to outsource her job, to a company that promises the use of AI to make finance better, and faster, and while Brenda is in the unemployment line, someone else thousands of miles away is celebrating a new job
We are setting AI deployed in the US, but actually Indians. They are not better, but they are cheaper. They are probably worse, but they are cheaper.
I feel what this article says based on some recent (non-catastrophic) experiences. I think I’m probably an above average user when it comes to Excel skills. I love spreadsheets. But I struggle with formulas like index, match, vlookup/xlookup and many others, and even more so when it requires nesting one within another and coming up with the underlying logic that leads to some complex nested formulas.
Over the past couple of months, I’ve tried some smaller models on duck.ai and also ChatGPT directly to create some columns and formulas for a specific purpose. I found that ChatGPT is a lot better than the “mini” models on duck.ai. But in all these cases, though these platforms seemed more capable than me and could make attempts to explain their formulas, they were many a times creating junk and “looping” back with formulas that didn’t really work. I had to point out the result (blank or some #REF or other error) multiple times and they would acknowledge that there’s an issue and provide a working formula. That wouldn’t work either!
I really love that these LLMs can sort of “understand” what I’m asking, break it down in English, and provide answers. But the end result has been an exercise in frustration and waste of time.
Initially I really thought and believed that LLMs could make Excel more approachable and easier to use — like you tell it what you want and it’ll figure it out and give the magic incantations (formulas). Now I don’t think we’re anywhere close to that if ChatGPT (which I presume powers Copilot as well) struggles and hallucinates so much. I personally don’t have much hope with the (comparatively) smaller and older models.
It's verifier law.
Coding agents are useful and good and real products because when they screw up, things stop working almost always before they can do damage. Coding agents are flawed in ways that existing tools are good at catching, never mind the more obvious build and runtime errors.
Letting AI write your emails and create your P&L and cash flow projections doesn't have to run the gauntlet of tools that were created to stop flawed humans from creating bad code.
Nah, I've seen them screw in all sorts of ways that would fail in some conditions and not others. You're way too optimistic about this.
Fair. I've been using the coding agent in Android Studio Canary to do exploratory code in Dart/Flutter and using ATProto. Low stakes, but higher productivity is a significant benefit. It's a daily surprise how brilliant it is it's some things and how abysmal at others.
10 billion dollars is probably going to be spent on automating excel, it’s going to happen
There needs to a financial equivalent to the Mythical Man Month.
There are plenty of things that play the role.
The problem is that people ignore them.
As an aside - isn’t it remarkable that we’ve introduced uncertainty and doubt into the knowledge processing layer? We have decentralised networks that run on Bayesian symbols for server-client models, CPUs that crunch Markov chains and now AI that hallucinates. On Deterministic Turing Machines.
Using ai does not absolve you from the responsibility of doing it correctly. If you use ai, then you better have the skills to have done the job yourself, and so have the ability to check the AI did things correctly.
You can save time still, but perhaps not as much as you think, because you need to check the ai's work thoroughly.
Excel is the most popular programming environment in the universe. It has optimized the five minute out of the box experience so well that grade schoolers can use it.
Other than that, it is pretty horrible for coding.
Let it all crash and burn
"the sweat from Brenda's brow is what allows us to do capitalism."
The CEO has been itching to fire this person and nuke her department forever. She hasn't gotten the hint with the low pay or long hours, but now Copilot creates exactly the opening the CEO has been looking for.
I've partied with Brenda on the weekends, and let me tell you... SOMETIMES Brenda hallucinates.
But never during work hours. The woman's a saint M-F.
Don't worry, in Teams it bothers me just one time a day, and with the click of a button it's gone... For another whole day.
Excel is programming. Spreadsheets have been full of bugs for decades. How is Brenda any different from a developer? Why are people scared when the LLM might affect their dollar calculations, and less bothered when it affects their product?
+100 this. Programmers who work in Excel (and never even dream of calling themselves programmers) are still programmers.
I know couple of them that get paid C-level money too :)
"You know who's not hallucinating?
Brenda"
I don't know about that. There could be lots of interesting ways Brenda can (be convinced to) hallucinate.
I agree - having watched many people use Excel over the years, I'd say people often overestimate their skills. I see three categories of Excel users. First there are the people that are intimidated by it and stay away from any task involving Excel. Second are the people that know a little bit (a few basic formulas) and overestimate their skills because they only compare themselves to the first group. And the third group are the actual power users but know to keep that quiet because otherwise they become the "excel person" and have to fix every sheet that has issues.
I don't know if AI is going to make any of the above better or worse. I expect the only group to really use it will be that second group.
I have seen lots and lots of different uses for Excel in my line of work:
- password database - script to automatically rename jpeg files - game - grocery lists - Book keeping (and try and not get caught for fraud several years, because the monthly spending limit is $5000 and $4999 a month is below that...) - embed/collect lots of Word documents - coloring book - Minecraft processes - Resume database - ID scans
1 reply →
Don't be like that. I work at a Fortune 500 and Brenda wants that co-pilot in Excel because it can help her achieve so much more. What is so much more you ask? Brenda and her C-Suits can not define it but they know for sure Copilot in excel will lead to enormous time saving.
Is this not the guy who is on the payroll of Anthropic? Not because he is wrong, but because there is so much marketing going on in this space nowadays.
Who, me? I'm still independent - I have a disclosures section on my blog here: https://simonwillison.net/about/#disclosures
Anthropic sometimes give me free credits (so I can try out preview features) and gave me a ticket to their conference a few months ago.
At some point, a publicly-listed company will go bankrupt due to some catastrophic AI-induced fuck-up. This is a massive reputational risk for AI platforms, because ego-defensive behaviour guarantees that the people involved will make as much noise as they can about how it's all the AI's fault.
That will never happen, AI cannot be allowed to fail, so we'll be paying for that AI bail-out.
Do you really want these kind of companies to succeed? Let them burn tbh
I don't find comments along the lines of 'those people over there are bad' to be interesting, especially when I agree with them. My comment is about why it'll go wrong for them.
Make sure you’re not part of the kindling, then.
I see the inverse of that happening: every critical decision will incorporate AI somehow. If the decision was good, the leadership takes credit. If something terrible happens, blame it on the AI. I think it's the part no one is saying out loud. That AI may not do a damn useful thing, but it can be a free insurance policy or surrogate to throw under the bus when SHTF.
This works at most one time. If you show up to every board meeting and blame AI, you’re going to get fired.
This is true if you blame a bad vendor, or something you don’t even control like the weather. Your job is to deliver. If bad weather is the new norm, you better figure out how to build circus tents so you can do construction in the rain. If your AI call center is failing, you better hire 20 people to answer phones.
Simon posting tiktok quotes on his blog was not on my 2025 bingo card.
This isn't the first: https://simonwillison.net/2025/Aug/8/pearlmania500/ and https://simonwillison.net/2024/Jul/29/dealing-with-your-ai-o...
Also this fun diversion into Occlupanids: https://simonwillison.net/2024/Dec/8/holotypic-occlupanid-re...
A lot of people complain that the internet isn't as weird and funny as it used to be. The weird and funny stuff is all on TikTok!
The mismatch between what people not on it think TikTok is like and what it's actually like (once you get the algo tuned to your taste) is pretty crazy.
But then the "new user" experience is so horrific in terms of the tacky default content it serves you that I'm not surprised so many people don't get past it.
Excel is the “beast that drives the ENTIRE economy” and he’s worried about Brenda from the finance department losing her job because then her boss will get bad financial reports
I suppose the person that wrote that have not ideia Excel is just an app builder where you embed data together with code.
You know that we have excel because computers didn’t understand column names in databases and so data extraction needed to be made by humans. Humans then design those little apps in excel to massage the data.
Well, now an agent can read the boss saying gimme the sales from last month and the agent don’t need excel for that, because it can query the database itself, massage the data itself using python and present the data itself with html or PNGs.
So, we are in the process of automating Brenda AND excel away.
Also, finance departments are a very small part of excel users. Just think everywhere were people need small programs, excel is there.
In most cases where the excel spreadsheet is business critical, the spreadsheet _is_ the database. These companies aren’t using an erp system. They are directly entering inventory and sales numbers in the spreadsheet.
The post is clearly hyperbole obviously the sole issue being brought up isn't 'brenda losing her job may be bad for the company' you're being facetious.
Found the person who hasn’t seen excel in the real world.
Excel - whatever its origin story - is the actual Swiss Army knife of the tech world.
There’s easily a few billion people who use excel. There is a reason it survives.
20+% of the world population uses Excel? Any citations on that?
You missed this bit “.. and then the AI is gonna fuck it up real bad and he won't be able to recognize it because he doesn't understand because AI hallucinates.”
Brendas have fucked it up multiple times, by themselves or because their boss demanded
3 replies →
Good luck with that.
I'm actually not that worried about this, because again I would classify this as a problem that already exists. There are already idiots in senior management who pass off bullshit and screw things up. There are natural mechanisms to cope with this, primarily in business reputation - if you're one of those idiots who does this people very quickly start just discounting what you're saying, they might not know how you're wrong, but they learn very quickly to discount what you're saying because they know you can't be trusted to self-check.
I'm not saying that this can't happen and it's not bad. Take a look at nudge theory - the UK government created an entire department and spent enormous amounts of time and money on what they thought was a free lunch - that they could just "nudge" people into doing the things they wanted. So rather than actually solving difficult problems the uk government embarked on decades of pseudo-intellectual self agrandizement. The entire basis of that decades long debacle was based on bullshit data and fake studies. We didn't need AI to fuck it up, we managed it perfectly well by ourselves.
Nudge theory isn't useless, it's just not anything like as powerful as money or regulation.
It was taken up by the UK government at that time because the government was, unusually, a coalition of two quite different parties, and thus found it hard to agree to actually use the normal levers of power.
This NY Times opinion piece by Loewenstein and Ubel makes some good arguments along these lines: https://web.archive.org/web/20250906130827/https://www.nytim...
It looks like the OP is thinking that AI causing errors in spreadsheets is going to make the whole economy collapse.
When tools break, people stop using them before they sink the ship down. If AI is that terrible at spreadsheet, people will just revert to Brenda.
And it's not like spreadsheets have no errors right now.
I'm worried Excel will go "enterprise only". and only LLM based interfaces will be enabled on the "office+windows" for consumers tier.
e.g. MS Access is well on its way. as soon as x86 gets fully overtaken by ARM, and LLMs overtake "compilers" (also taken enterprise only).. then things like sqlite-browsers (FOSS "access") will be an arcane tool of binary incompatible ("obsolete") formats
(edits: this worry has not been easy to type out)
I'm more shocked that someone is using TikTok to speak things that actually make sense instead of mindless memes.
Me too. Theres no tool more trusted for accessible numerical precision than Excel. Lets sell all that goodwill for a shiny new magic bean.
This is transparent nonsense. People are very very happy to introduce errors into excel spreadsheets without any help from AI.
Financial statements are correct because of auditors who check the numbers.
If you have a good audit process then errors get detected even if AI helped introduce them. If you aren't doing a good audit then I suspect nobody cares whether your financial statement is correct (anyone who did would insist on an audit).
It’s like calling out the county to inspect the home you built but when they arrive it’s a bouncy castle.
> If you have a good audit process then errors get detected even if AI helped introduce them. If you aren't doing a good audit then I suspect nobody cares whether your financial statement is correct (anyone who did would insist on an audit).
Volume matters. The single largest problem I run into: AI can generate slop faster than anyone can evaluate it.
If nobody can evaluate it then nobody will sign it off.
Luckily this is a capitalist society and usually mistakes in the private market resolve themselves because losing money is not a winning strategy.
I actually know Brenda.
Excel doesn't need AI to ruin your work: https://www.science.org/content/article/one-five-genetics-pa...
Brendas hallucinate all the time.
Nay-sayers need to decide whether they fear AI because AI is dumb and will fuckup or because AI is smart and will take over.
Silly calling Simon a nay-sayer.
Are you a fanatic that thinks anyone saying that there are any limitations to current models = nay-sayer?
Like if someone says they wouldnt wanna get a heart transplant operation done purely by GPT5, are they a nay-sayer or is that just reflecting reality?
Simon willson is definitely not a nay sayer.
Both are valid concerns, no need to decide. Take the USA: They are currently lead by a patently dumb president who fucks up the global economy, and at the same time they are powerful enough to do so!
For a more serious example, consider the Paperclip Problem[0] for a very smart system that destroys the world due to very dumb behaviour.
[0]: https://cepr.org/voxeu/columns/ai-and-paperclip-problem
The paperclip problem is a bit hand-wavey about intelligence. It is taken as a given than unlimited intelligence would automatically win presumably because it could figure out how to do literally anything.
But let's consider real life intelligence:
- Our super geniuses do not take over the world. It is the generationally wealthy who do.
- Super geniuses also have a tendency to be terribly neurotic, if not downright mentally ill. They can have trouble functioning in society.
- There is no thought here about different kinds of intelligence and the roles they play. It is assumed there is only one kind, and AI will have it in the extreme.
1 reply →
Our product has many issues. You must pick one and must not discuss any other.
Everything is now about verification.
AI may be able to spit out ann excel sheet or formula - But if it can’t be verified, so what ?
And here’s my analogy to think about the debugging of an excel sheet - you can debug most corporate excel sheets with a calculator.
But when AI is spitting out excel sheets - when the program is making smaller programs - what is the calculator in this analogy ?
Are we going to be using excel sheets to debug the output of AI?
I think this is the inherent limiter to the uptake of AI.
There’s only so much intellectual / experiential / training depth present.
And now we’re going to be training even fewer people.
At the end of the day I /customers need something to work.
But failing that - I will settle for someone to blame.
Brenda handles a lot of blame. Is OpenAI going to step into that gap ?
another cheese that will affect the outcome of major tournaments, not a good look for microsoft
its like the xlookup situation all over again, yet another move aimed at the casual audience, designed to bring in the party gamers and make the program an absolute mess competitively
320784788
[dead]
[dead]