Top
Best
New

Posted by jaykru 15 hours ago

Why I'm still bearish on LLMs after Navier-Stokes(dank.systems)
209 points | 221 commentspage 2
yunwal 8 hours ago|
> those who need done a small set of narrowly defined tasks with existing clear guardrails: repetitive physical labor in a controlled environment, call center and customer service chat work, etc.

I have no idea how people can so confidently say that call center work is a “controlled environment” or “repetitive”. It’s almost by definition not repetitive or controlled. Customer support is what I go to when the controlled environment has failed

vachina 7 hours ago||
Depends on what customer support means.

Typically it means knowledge retrieval from a KB or manipulating a control surface not visible to you.

fhe 5 hours ago||
came here to say exactly this. in fact, this is probably why we are not seeing a lot of AI application on customer service use case, and when we see one, it's almost always frustrating.
jumploops 8 hours ago||
LLMs are basically multi-dimensional magic mirrors.

Depending on where you point them, they can be incredibly useful.

They can even be useful when you point them at each other (though increasingly difficult to get good results).

I'm excited for the promise of RSI and a future where models have inherently "live" weights, but it's not clear to me that the transformer is more than a useful tool to help us get there.

againstapples 10 hours ago||
> the models generalize well only on tasks within a small neighborhood of the specific tasks they've been trained on, and even then with severe caveats. the frontier labs have developed a general recipe to teach models almost any specific task enjoying clearly defined levels of task performance; many tasks are covered in the training data

Is this really any different to how humans learn, it takes a lot of training on one specific task to make a human expert as well?

bravoetch 9 hours ago||
I was a young child when I learned chess by reading a short book, then practicing with a friend. That is not how LLMs learn. I'm no expert on LLMs, but if you showed a human all chess games and books in all history and then said 'play chess' and they still kept making illegal moves, they would have to have a brain injury.
harimau777 9 hours ago|||
It seems like an LLM potentially could learn that way if each practice game it participated in was added to its training data.
willmarch 8 hours ago||
Yes, this is essentially how AlphaGo and AlphaZero algorithms work to train superhuman Go/chess/shogi agents. It’s an elegant algorithm that is analogous to how humans learn games.
zug_zug 7 hours ago||
Well except AlphaZero played 44 million chess games in that time (and actually played with a 44 core computer). So I'd like to point out that the human is still just a few orders of magnitude more efficient.
willmarch 6 hours ago||
Yes, we all know that biological systems are more efficient than machines through billions of years of evolution and natural selection but the overall process is largely the same (interacting with an environment, learning from results, improving underlying architecture, etc); efficiencies will come with more time and improvements.
fhe 5 hours ago|||
but maybe... the said human has also read every other piece of text ever written, including ones about other (similar?) board games, which in aggregate vastly dwarfs whatever he has read on chess, that non-chess reading could have corrupted whatever he's learned about chess?
danielmarkbruce 9 hours ago|||
Humans wear a lot of hats when the do work. They don't even realize how many. My experience with building real systems using LLMs is that you have to be very explicit about such hats and you don't realize how many are worn until you see edge case after edge case after edge case. Check this. Check that. Check this. Check that. Check check check.
JohnMakin 10 hours ago|||
> Is this really any different to how humans learn

yes.

knuppar 9 hours ago||
being a bit more specific: the sample efficiency of humans is orders of magnitude larger for more abstract concepts. the same doesn't hold for memory-intensive tasks though (like any kind of trivia), but that only takes you so far.
danpalmer 7 hours ago||
We've had technology beating humans on memory for millennia, and we've had technology beating humans on computation for many decades now.

The tricky thing with LLMs is describing what they actually do. They are too clearly beating humans on some things, but what exactly? Memory – already done, they're bad at basic computation (all LLMs just write code for actual computation/calculation). And as you say, they do badly at more abstract concepts.

bananzamba 9 hours ago||
Also doesn't the very good ARC AGI 2 score of GPT-6 Astra kinda contradict this, since each problem is its own game with very different rules
someguynamedq 8 hours ago||
> current frontier models need laborious oversight and guardrails on even the simplest task

As models advance, we shift the goalpost for what "simplest task" means. Before, "simplest task " meant "write a coherent English sentence." Now, "simplest task" means autonomously fix, review, and merge a bugfix.

abeppu 8 hours ago|
Eliza wrote coherent English sentences.
Gregkion 4 minutes ago||
And you know compare Eliza with what an LLM can do today?

Or do i miss the point you are trying to do?

vatsachak 9 hours ago||
I agree with the caveat that it's more like a cracked junior engineer who can manage swarms of interns.

Frontier Labs will probably survive off hype valuations but will serve the important purpose of discovering architectures/techniques that will probably spread through rumors/transfers to the rest of the world.

melvinroest 2 hours ago||
Yea I get the bearishness from my own personal experience.

Personally, I use LLMs for a lot of things. Oftentimes, I'm a think out loud type of person so even having something that feels like a rubber duck, but more competent, is already amazing for me. And LLMs are a lot more competent than a rubber duck.

But especially sometimes I've noticed that LLMs can be unbelievably stupid. It recently happened a few times with Fable 5.1 as well. Ultimately, I think it comes down to that LLMs can't think broadly. In software development one can usually see this too. For example, a whole app might be built by an LLM and it didn't spend a single token thinking about security because the prompter is at the level of "build a dating app for dogs, make no mistakes". Now you have a dating app for dogs that is insecure.

Since I prompt for almost everything in my life to have an LLM as a sounding board, I'm usually not an expert either. I've noticed LLMs are amazing at "bulk search engine information aggregation" (or whatever you want to call it). So if I need something from the Dutch government, I can find it way more quickly. But oftentimes I've noticed that going for a walk and thinking about a particular thing I'm facing is a more effective way of finding a good solution.

Other times times they are not incredibly stupid, but can't form a strong opinion. This usually happens when I'm tackling a wicked problem [1]. When that's the case, prepare for LLMs to sway with you for every small change in your opinion that you ever will experience.

So I agree: drop in replacement for knowledge workers? No. Rigorous specification is usually needed yes. Though, the small win here is that it doesn't always need to be as rigorous as programming is and it can happen in natural language. It depends on the topic/problem being tackled.

I really like them as UX tools though. Amazing for interactive prototyping and requirements elicitation. And that also corresponds with what the author is saying. Though I find it a bit of a disservice saying "just 3". You know how hard requirements elicitation is? It became a whole lot easier thanks to LLMs (I might change this opinion in a year, haha, but this is the opinion I hold now).

[1] https://en.wikipedia.org/wiki/Wicked_problem

slibhb 7 hours ago||
> the frontier labs are priced according to the narrative that they have produced or will in the very near future produce a fully automated drop-in replacement for most knowledge workers

That's a reason to be bearish about AI companies, not LLMs. But is it even true? OpenAI and Anthropic have each reported ~50 billion in revenue with ~900 billion valuations. That's a high ratio but I'm not sure if follows that the only way it pans out is if we get "fully automated drop-in replacement for most knowledge workers".

It wouldn't shock me to see those revenue numbers scaling up to where they need to be over the next decade ( to, say, ~400 billion) without ever achieving drop-in worker replacements.

zug_zug 7 hours ago|
I looked at the math and I think it's true. Remember revenue is just sales, not profit. These labs are shooting for > $1T valuations, which traditionally means your PROFIT is at least 1/20th or 1/30th of that (so let's say minimum 30B$/year PROFIT).

These companies however are LOSING money (anthropic tries to make it sound like it's profit by deviating from accepted accounting principles) and subsidizing these models. When accounting for all the engineering salaries, training, GPUs, etc, what's their best-case realistic margin three years out, 10%?

So to we'd need a scenario where companies are spending a collective 300B annually on AI (believable) but ALSO that these companies jack up their margins WITHOUT companies switching to the cheaper open-source models (even when there's a $300B incentive to do so).

alain94040 8 hours ago||
Not convinced by those points.

In particular, I found this very misleading or irrelevant:

a typical CPU project anecdotally has about three times as many specification and validation engineers as design engineers and a 5:1 ratio is not unheard of

The reason silicon design has such verification to design ratio is because the cost of one bug is many, many orders of magnitude higher than software. Both in dollar cost and in schedule cost (it takes months to fab a chip, and if you messed up and need to spin a fix, it costs tens of millions of dollars, not counting any design engineering cost).

I don't think you can extrapolate these very industry-specific facts to judging LLMs.

danpalmer 7 hours ago|
> The reason ... is because the cost of one bug is many, many orders of magnitude higher than software. Both in dollar cost and in schedule cost (it takes months ... and if you messed up and need to spin a fix, it costs tens of millions of dollars, not counting any design engineering cost).

Aren't you just describing waterfall? That's still very prevalent in software engineering, and pretty much any other type of engineering – civil, chemical, building, architecture, drug discovery.

It's typically true that software can fail faster and cheaper, but it's also true that the costs are still vastly higher to fix later in the process.

alain94040 7 hours ago||
No. Silicon is on another level. Which is why the EDA verification is an industry on its own.

Sure, there are some software that have similar "can't have bugs" requirements. I imagine the computers on Moon missions also had that kind of high bar. I wouldn't use NASA requirements as a proof for how LLMs should be used.

pfdietz 10 hours ago||
Specifically: bearish on LLMs generally, not bearish on LLMs for pure math.
jaykru 10 hours ago|
yes, huge for pure math and activities that look like it.
danielmarkbruce 9 hours ago||
Doesn't really even need to look like it. If you can verify rewards, RLVR will optimize really really well. If you can't... it's a struggle. There are probably fewer fields where you can verify rewards than one might hope.
skydhash 8 hours ago||
> There are probably fewer fields where you can verify rewards than one might hope.

2 tasks I've done today that I believe robots are nowhere near being able to do: Cleaning my wardrobe and draining bad fuel out of my generator. As in generic use cases.

camd32 8 hours ago|
> current frontier models need laborious oversight and guardrails on even the simplest tasks.

This is only true if you are concerned about the intermediate steps of the model as opposed to the outcome. The huggingface hack was a perfect example of the model doing whatever it takes to accomplish the goal of maximizing its score.

willy_k 5 hours ago|
So, if you are concerned about what the model does? Yeah.
More comments...