Comment by mattlutze
12 hours ago
A lot of people here are responding to the message but not to the meaning.
It would be a good idea for young people to deeply know how these programs work. Not so that they can spend their career building them, but so that they can approach the next class of problems we'll all start trying to solve, with intuition all the way down to the weights and underlying mathematics. And also, to develop a healthy intuition of when "Just LLM it" will not be the right choice.
"Build an OS" wasn't a common university project because we were all expected to go out and work on Windows, but because understanding the bare-metal firmware for a computer helps you deeply understand how to intuit building for a whole class of problems.
The funny thing about LLM is, you can build one from scratch and yet you still won't understand how it works. You would understand what kind of matrix multiplications the neural network performs (in fact that's not that hard. An OS is orders of magnitudes more complex), but you would still have no idea why it does what it does.
This is a great point. The complexity sources are entirely different. An OS is complex because it’s solving a difficult problem and we’ve seen decades of evolution both in knowledge and-how and, at the frontier, in performance standards.
The abilities of LLMs are emergent. You can experiment with LLMs and know as much about their observable behavior as the experts. But there’s no way to “crack open” an LLM and see precisely where each skill or tendency lives; as far as we currently understand, it’s all mashed together.
Learning how to build emergent systems is also a skill kids should learn these days. The closest I got was coding up game of life for CSE 142 (intro programming).
If stochastic gradient descent isn't taught in whatever CS Theory 101 is now, it really should be these days.
> there’s no way to “crack open” an LLM and see precisely where each skill or tendency lives
Mechanistic Interpretability has entered the chat.
For a classic example, see https://www.anthropic.com/research/tracing-thoughts-language...
The spirit of your point stands, though. This kind of research is interesting to read about, but it's very hard, more like neuroscience or biology than computer science ("LLMs are grown, not made"). You're dealing with a lot of extremely _messy_ complexity, for which organic life is really the only good point of comparison. Most of us here are't really equipped for that kind of work; it's not at all like, say, reverse-engineering a piece of software written by humans. And of course the only people who can do it on frontier models from Anthropic and OpenAI are people within the labs themselves. (But I'm optimistic we'll see more of this work on open weights models...)
Yah, "building" it is not sufficient. But a lot of times when I build I want to know the why. "Why does gradient Descent have some clever tricks that easily translate to matrix math"?
Lot's of neat stuff to learn.
I'm not sure it's possible to have intuition about systems that work in thousands of orthogonal dimensions. In fact I'm pretty sure most of the research is people trying fairly arbitrary things and testing them and then post rationalising implied understanding of what is really happening on top of good outcomes.
I think it’s reasonable to have a shallow understanding of most parts and a deep understanding of a small number of parts. That’s how most engineers are.
Most software engineers do not have a deep understanding of CPU architectures. In fact they probably don’t even have a shallow understanding and get around just fine. How many of them are looking up the instruction set for the CPUs they deploy their CRUD app to in EC2?
But in the case of CPU architecture there are SOME people who understand how things work 100%, and they've built and vetted abstractions/mental models that enable other engineers and scientists to have that kind of mixed shallow/deep understanding in a way that works. On the side of LLMs we're still lacking an expertise which could flawlessly explain how these things operate; the abstractions that we're using are instead derived inductively and are totally unvetted.
2 replies →
In fact, one of the jobs of an engineer is to make sure that other engineers who don't work in his or her area do not need to understand that area deeply, yet build something reliable with it. They need just the summary that he or she writes up into the datasheet for the part. Ensure these conditions are met for safe/reliable operation, give it these inputs, expect these outputs, these timings, this energy consumption, this heat generation, frequency response, tensile strength, whatever.
I disagree. It's absolutely possible to develop an intuition about extremely complex mathematical ideas, including llms or high dimensional systems. Learning to build an llm is a great way to start building that intuition.
Developing an intuition about high dimensional systems is pretty different from understanding the character of some specific point on a 1e9+ dimensional manifold of parameters, in my professional opinion (setting aside all the degrees of freedom that come from the structure of the thing). Sure one can understand generic principles like the curse of dimensionality, but truly groking how an LLM works is basically an open problem as far as I'm aware. I'm not saying there's no benefit for amateurs to study how LLMs work, but let's be realistic about how far mere intuition can truly take anyone in this space.
1 reply →
What provable conclusions have you intuited around how LLMs work? Give me some examples to prove your point? I'm absolutely happy to change my view with enough data.
4 replies →
You may be right. Nonetheless, it's possible that the act of trying to get your brain to understand a system that works in more than a few orthogonal dimensions will help develop useful intuitions about how LLMs, and the many systems/products/robots that will be built on top of them, work.
It is sometimes the opposite - a large number of things makes the system easier to predict and reason about (statistics, behavior of gases etc).
I'd have to push back, though not on the part you'd expect. Your description of human researchers is roughly right: a lot of the field is try-things-and-narrativize-after.
But the load-bearing assumption is that intuition has to be human-shaped intuition. Humans can't intuit thousands of orthogonal directions because we project everything down into a 3D metaphor and hope it holds. That's a fact about our hardware, not about the systems.
And the reason why is the most interesting part: nothing requires the compression step. A model or an agent can operate over the actual objects, holding thousands of runs and ablations in context and noticing regularities in the native dimensionality, without translating them into a picture of a ball rolling down a hill. No bottleneck at "can you visualize it."
So the narrower claim: it's not that intuition here is impossible full-stop, it's that human intuition is unreliable. Your post-hoc rationalization point is evidence for that, not against it. The story exists because a person needs something to hold in their head. Drop that requirement and the failure mode goes with it.
You apparently have no intuition on what the intuition is
This is no longer funny.
What corpus would one be advised to use for training? Should one limit oneself to legally downloadable documents or should one follow the examples of Altman, Amodei, and Zuckerberg and use LibGen and Anna's Archive?
> A lot of people here are responding to the message but not to the meaning.
Well it is framed as quite specific advice.
(I'm done with mining PG tweets for meaning)
[dead]
Agreed. During comp sci we got to re-implement various algos of networks, OS, database, firmware.. and it all gave complimentary intuitions that were useful when tackling practical implementations and bottlenecks.
> understanding the bare-metal firmware for a computer
IMO this is still relevant, everything surrounding the LLMs needs such a vast infrastructure that I don't know if I would find it more useful to learn the maths behind ML than CS
As a 17 year old, I agree with this. Ofc I'm against all the hate directed at PG, I believe that all knowledge has value regardless of its economic utility, but I understand where the hate is coming from. Personally, I find LLMs boring for now, and I'm more focused on CS and electrical engineering.
The next class of problems to solve for future generations is how to survive on a hot planet and how to recover from the damage done by reckless lunatic profiteering assholes.
Maybe people want to take it as cynicism but I think you are factually correct that personal cooling, infectious disease, atypical weather and fire breaks are all more realistic topics for people too young to already have patents, connections or their own oil well.
Couldn't have said it better
Depends, the assumption things are predictable always negatively affects both Market Bears and Bulls alike.
Indeed, if credulous folks look to the world expecting people to bestow success upon them... than the disillusionment with reality will hit their savings harder.
The Shrek movie market correction correlations are undeniably funny, and a new film is due July 2027. OpenAI may be going public in the next few months while still losing $2.25 for every $1 of customer revenue, and with 6 other firms sharing over $4Tn in debt disclosed to investors in a footnote.
There is only one direction things can go at the Peak of inflated expectations. Popcorn ready. =3
https://en.wikipedia.org/wiki/Gartner_hype_cycle
> "Build an OS" wasn't a common university project because we were all expected to go […] helps you deeply understand how to intuit building for a whole class of problems.
Sure, but it was for a specific degree with a syllabus that taught you the foundational knowledge. It was not expected from the law students to learn how to build one.
> It would be a good idea for young people to deeply know how these programs work.
It would be a good idea for _everyone in the industry_ to deeply know how LLM training, inference and "agents" work, not least because it removes the ability of shysters to bamboozle with bullshit.
But, as much as a good idea it is for the young to understand this, it's the elderly who will be really taken advantage of if they do not keep up - just look at Facebook for good examples of why.
[dead]