upvote
I think future is probably more similar to speculative execution (inference/decoding). A small LLM is used to speculate and a large LLM is used to confirm if needed. If the small LLM is accurate enough on N tokens it’s cheap for the large LLM to say looks good and keep moving along.
reply
[dead]
reply
General purpose models are always more robust and generally better than smaller narrower models. My bet is that compute will catch up and any “small” model will still be generally capable, just smaller than sota, rather than intentionally narrow. The exception would be for very well defined tasks where the data distribution never varies, but these are rare and don’t really need “AI” anyway when they do exist.
reply
> General purpose models are always more robust and generally better than smaller narrower models

I feel like this is just the marketing conflation of AI=LLM, versus regular old ML? We're never going to need to deploy a full reasoning model on a low-power device just to do some fancy image recognition in the field. Specialised ML models are just intrinsically able to be a lot more efficient than their generalist equivalents

reply
> General purpose models are always more robust and generally better than smaller narrower models.

What do you mean with more robust?

reply
Less weird unexpected failures, more innate ability to handle edge cases gracefully. Quite important when you're running high on automation and low on oversight.
reply
This may be speculative, but couldn't robustness emerge by having a number of specialized models, that are interconnected and feed into each other? Are there any arguments from ML that would speak against this?
reply
That does work. Even if you drop the "specialized" part. Ensembles of the same architecture at the same scale trained on the same data do outperform a singular model of the same line - especially on corner cases. Successes of an ensemble correlate stronger than failures do.

The usual argument against is that if you have "a number of specialized models" that perform well in ensemble, you can take that ensemble, and distill it into a single larger model (dense or integrated sparse, like MoE), and get the same improvement in performance with an efficiency win.

This works because having those "specialized models" duplicates a lot of the highly conserved "low level" wiring that's required for a model to function at all. As such, you end up running a small scale version of the same "backbone" computational processes many times. "Merging" those models into a larger, denser model allows for a singular strong "backbone" to be used for everything.

reply
deleted
reply
You're getting downvoted, but you're completely right. There are very few cases in which narrowing a model down is buying you anything worthwhile.

It seems like for LLMs, "general intelligence" is expensive, but "one more domain" is fairly cheap.

reply
The demand for smaller models and single purpose AI agents will only get bigger in the future. It’s no different than going for mainframe computers back in the day to personal computers. It’s inevitable…..
reply
Has nothing to do with demand, it’s a question of the most efficient way to train a good model. Narrower models are in general worse than general purpose ones, so you’re better using the largest general purpose model that fits in your compute budget vs trying to somehow remove capability or knowledge and assume the model stays as capable for the task you want.
reply
> It's actually how organic brains work - specialized tasks are offloaded to local cortical columns.

How are small isolated language models more similar to that than MoE in LLMs?

reply
MoEs don't route like most people imagine. They aren't learning topic based experts despite the name

The original Mixtral paper [0] (in the "Routing analysis" section) found:

"surprisingly, we do not observe obvious patterns in the assignment of experts based on the topic"

A quick skim of more recent analysis on MoE shows that this hasn't changed. MoE models do appear to work, but don't appear to do what the name implies, if anything they're routing based on the structure of the text and not the semantic content (and we're still not entirely sure what they're doing).

0. https://arxiv.org/pdf/2401.04088

reply
Right MoE is a tradeoff between efficiency and intelligence.
reply
I think the harness and local context should supply that missing piece between general model and bespoke application. Each application has its own context and action quirks that don't generalize well. Maybe it's just 5% but that is genuinely specific. So its rightful place is in context engineering.

I have a long-ass post about how this could be implemented. https://old.reddit.com/r/VisargaPersonal/comments/1um9uyv/st...

reply
What about recent models providing correct proofs to open math problems?
reply
I haven't tried it, but I saw Leanstral, an LLM specialized in writing Lean proofs, posted on HN recently and it claims to outperform some larger general purpose models. It didn't beat Claude Opus, but it seems to do decently at one tenth the cost. It's plausible that further research could yield other models that are smaller and more effective at limited tasks, reversing the trend of ever growing models.
reply
What about it?
reply
No this will never work. Domain specific models will never be a thing because intelligence carries over and compounds.

Why didn’t OpenAI release a math specific model? Why not a literature specific one? Why do they instead have generic models of different sizes? And how did all labs converge on this?

Why does Fable just not train on non cybersec and non biology data but instead have clearly costly and annoying classifiers?

reply
Your examples (math, literature) involve natural language. It stands to reason that a general language model will be more competitive in those domains. If you want examples of successful domain-specific models, look at AlphaZero and AlphaFold. LLMs aren't anywhere close to achieving that level of competence at abstract strategy games or protein folding.

"This will never work" is a pretty confident assertion for a field that's so young and rapidly evolving.

reply
> I don't foresee AGI arising out training bigger LLMs (Though investors won't realise that for a while yet).

This is what the parent said. AGI won't rise out of AlphaZero and AlphaFold in the same way AGI won't rise out of Houdini chess engine. This is the industry consensus.

reply
>AGI won't rise out of AlphaZero and AlphaFold in the same way AGI won't rise out of Houdini chess engine.

That's a straw man. Nobody thinks AGI will rise out of domain-specific systems. The question is whether domain-specific systems are necessary for AGI.

Of course, the problem is that AGI isn't a well-defined concept. But if we define it as achieving superhuman performance across several hundred domains where there are objective measures of success, it doesn't seem far-fetched to predict that it will involve some general reasoning system paired with a bunch of specialized modules.

reply
The parent said

> I don't foresee AGI arising out training bigger LLMs

I agree that AGI will involve tool usage but not only involving domain specific AI models.

But lets try to find the discriminating point in the discussion - do you believe AGI will necessarily involve training bigger LLM's or not?

I believe they are necessary. WBU?

reply
You're still intentionally misreading the OP's statement. If you read it again in context, they're clearly saying that they think training bigger LLMs is not sufficient. I think I agree with that statement, but my confidence is pretty low.

No, I don't think LLMs are necessary for AGI at all. I think there are multiple paths to AGI, some of which involve LLMs and some which don't.

reply
AGI is a macguffin (or a shaggy dog) for a story told to investors. It has never been plausible on the timescales suggested and it almost certainly will not emerge from LLMs.
reply
> No this will never work.

This bet is too early.

> Why didn’t OpenAI release a math specific model? Why not a literature specific one? Why do they instead have generic models of different sizes? And how did all labs converge on this?

Because they have a very early product and they could train it, brute force, with access to an extraordinarily large pool of money. So did all the other labs. Because it was thus easier to scrape everything rather than spend enormous effort (with tools that did not really exist) to partition the training set. Any number of other "because"s.

It's just what they are doing now and it showed the earliest results.

LLMs are still less intelligent than rats, which have tiny brains.

reply
Rats are a weird choice for that comparison. They're some of the smartest animals on the planet, some say smarter than dogs.
reply
Which is exactly what makes them a good comparison, IMO. They are arguably smarter than dogs with about two fifths as many neurons. A shorter lifespan, faster breeding and many more threats has made them better.

The point I am getting to elliptically is that larger models aren't necessarily the solution. They are one solution pathway.

It is fully possible (I think actually likely) that the ultimately successful path for LLMs will emerge from the pressure of keeping them small, not making them large. Very small, domain specific models could well outperform large models in their domains, and they might even show that domain specialisation is not necessarily much of a limitation, just a useful impetus to stay small. Like how rats can drive those little cars.

(I think the frontier models are potentially already too big. Can't prove it or even close to it, but it feels like this is going to be a story.)

reply
DeepMind did release a math specific model. And OpenAI has released a coding specific model.

The answer to your question is “because the market isn’t big enough”, not because it doesn’t work. Why would knowing about 2019 internet memes help you in any way at coding?

reply
> Why would knowing about 2019 internet memes help you in any way at coding?

99.99% of the knowledge an LLM has is useless for a given scenario, the hard part is knowing what the .01% that’s needed is. Knowing as much as it can means the model can handle edge cases, turns of phrase, etc.

Put another way, it avoids overfitting. That’s basically the insight that’s given way to the current AI boom.

reply
> Why would knowing about 2019 internet memes help you in any way at coding?

https://github.com/Brainrotlang/brainrot

"Brainrot is a meme-inspired programming language that translates common programming keywords into internet slang and meme references."

reply
> And OpenAI has released a coding specific model

They did and retracted it because they found that GPT 5.5 beat codex pareto optimally. This keeps happening.

> because the market isn’t big enough

Huuh? market isn't big enough for AGI? The parent suggested that AGI would emerge from this process.

reply
[flagged]
reply