Comment by kozikow

The incentives seem to be lined up to push users to the frontier of "how much can the harness around the model do and claim without most people pushing back on the claims."

The bigger the pile of stuff you have the tool create unsupervised, the more incomplete/inaccurate the overall picture of it is gonna be. That doesn't matter for most hobbyist/helper-tool/game-port/etc type of tasks.

I get a ton of value out of the tools and yet part of that value has come from being hyper-aware of the shape of the limitations, which changes much less and more slowly than the specific of the limitations.

"How'd you notice that one!?" - well, it helps to be the only person on the team who took the time to interrogate the e2e interaction of all the code...

The median piece of OSS software was already pretty rough once you had to depend on it in production at scale to make money. The median piece of the new AI-generated wave of software, also pretty rough.

The bet that OpenAI/Anthropic is making is that they can automate the interactions generally/globally enough to work around this and eliminate the need for humans in the loop. But it's a tricky problem because it's precisely where the permutations and special cases get nasty and much harder to deal with than more broadly-applicable cases of codegen/test-gen/mash-it-until-the-tests-pass-and-the-code-review-angent-is-happy.

This has all started to feel very Emperor's New Clothes-ish. "Well, if it's not working for you, then it's your fault. You're just not using it properly. You're not worthy of its wonders".

Not to say there's no value to be had. However, it's also been oversold and its error-proneness is a real problem.

But, it seems there's a steady stream of commenters and creators who are quick to remind us that it's never a problem with the tech. It's always a user problem.

Meanwhile, the errors are renamed "hallucinations", but let's not focus on those. In fact, if you just look right over here at this .md file...

using skills with opus 5.5 and agents for each individual task that have shared knowledge of all of your records does not lead to a significant increase in trustworthiness

> When people use words like “chatbot” to talk about their AI use and can’t really name basic things like the model they were using, it reveals a wide chasm between the people who are deep into understanding how to use these things and those who poke at easy options every few months and then look for ways to confirm their priors instead of trying to understand how other people are getting value.

While the latest and newest should be needed for greater capability, the average should be enough to give a glimpse pf what is possible. Like a basic netbook is enough to enjoy computing, no need to purchase the latest supercomputer.

If you need to switch your setup every quarter or so, it’s not a tool, it’s a rat race.

me too, but I recently blew the transformer on my boiler by following its directions to wire up a C wire to power a smart thermostat. oops. $500 mistake.

How often do you check the citations? I’ve found with the Google search AI summary it will provide answers and link to their citations, but a good chunk of the time the answer isn’t on the linked page.

It just makes stuff up (or gets it from somewhere, who knows) and then links to a tangentially related page as if it was the source. Either that or it identifies the page as being a source it wants to use, and then summarizes it incorrectly. Not sure the order of operations it’s doing internally, but regardless the citations are bullshit.

That sounds terrible. I like to live like a human being

Every time I use Google search AI for anything more complicated than a search query I feel that way. A paid model from pretty much anyone is much better.

It’s sort of weird how

1) open weights and cheap models are only about 6 months behind the frontier in evals, but…

2) it’s always the premium model that only the biggest fans pay for, that’s just now finally achieved enlightenment.

Oh well, I guess we’ll see in another 6 months.

opus 5.5 uses agents and skills in chat sessions as well and maintains internal context that it tries to distill from what you provide

I tried to have it write a synopsis website for the different types of dance patterns in Dance Dance Revolution, with guides on their execution, explanations, and provided song sources. I provided all of the source articles and search websites needed for it to be able to do that, and gave it a sample set I wanted it to utilize and build off of.

It has bits and pieces of information wrong in 90% of the sections it created. Poor song choices, missing song choices, incorrect explanations, incorrect step-patterns, the list goes on.

If you're not an expert in a given field for which you're using it, it's going to look like a genius. If you are, it's untrustworthy and borderline useless.

Maybe it wouldn't be if there were ample training data on my topic, but there isn't.

It utilized the research from millions of papers many very close to the destination answer to make those proofs. What would it do on something that isn't supported like that? Be confidently wrong while sounding right. That's what we've seen repeatedly.