curious_cat_163

I think condensing intelligence into smaller models is the way to go.

Also, “smaller” can mean many different things. The cost is not in storing the weights on disk.

It is in the power required to do the inference with the “active parameters”.

There is increasingly more evidence that those two can be decoupled and more power to those who are pushing on that lever!

Where is this thing? Would like to visit!

We are headed towards the “trough of disillusion” of this particular cycle.

Hey! Love the idea. I think a lot of skepticism here would be addressed if you discussed your plans to monetize. People just want to know how you will (eventually) make money in a way that is aligned with how they expect this to evolve.

That makes me sad. I will miss his voice. I loved how he interrupted his guests and kept them honest and on point. I loved the casual offer for tea/coffee at the end. I would love how it had this encore bit at the end, sometimes!

This podcast chose its listeners and kept it real. Thanks to everyone who makes it possible. Hope they get a fitting replacement for Melvyn and keep it going!

Give it time, it should [1] go down to where it belongs. :-)

[1] https://news.ycombinator.com/item?id=1781013

We don't really have very many open source models. We have "open weights" models. Ai2 is one of the very few labs that actually make their entire training/inference code AND datasets AND training run details public. So, that this investment is happening is a welcome step.

Congratulations to the team at Ai2!

So what is final state here for us? Return to menial not-yet-automated work? And when this would be eventually automated, what's left? Plug our brains to personalized autogenerated worlds that are tailored to trigger related neuronal circuitry for producing ever increasing dopamine levels and finally burn our brains out (which is arguably already happening with tiktok-style leasure)? And how you are supposed to pay for that, if all work is automated? How economics of that is supposed to work?

Wow. What a picture! Here's an optimistic take, fwiw: Whenever we have had a paradigm shift in our ability to process information, we have grappled with it by shifting to higher-level tasks.

We tend to "invent" new work as we grapple with the technology. The job of a UX designer did not exist in 1970s (at least not as a separate category employing 1000s of people; now I want to be careful this is HN, so there might be someone on here who was doing that in the 70s!).

And there is capitalism -- if everyone has access to the best-in-class model, then no one has true edge in a competition. That is not a state that capitalism likes. The economics _will_ ultimately kick in. We just need this recent S-curve to settle for a bit.

That's an interesting way to look at it.

Aren't layoffs a version of that? Are we seeing any evidence that folks who have been let go from Intel have resulted in spin-offs and startups?

I know at least one person who went to work at Nvidia from Intel but that is neither of those things.

Oh, I don't know. Maybe build chips that do things 10x more efficiently and sell them a lower cost to compete?

It _is_ a hype bubble but it is also an S-curve. Intel has missed the AI boat so far, if they are trying to catch up, I would encourage them to try. Building marginally better x86 chips might not cut it anymore.

And, so what?

It is fairly rare to see an ex-employee put a positive spin on their work experience.

I liked my jobs and bosses!

Maybe you'll want to try Techne: https://techne.app.

Rather than fine-tuning models on a small number of environments, we expect the field will shift toward massive-scale training across thousands of diverse environments.

This is a great hypothesis for you to prove one way or the other.

Doing this effectively will produce RL models with strong few-shot, task-agnostic abilities capable of quickly adapting to entirely new tasks.

I am not sure if I buy that, frankly. Even if you were to develop radically efficient means to create "effective and comprehensive" test suites that power replication training, it is not at all a given that it will translate to entirely new tasks. Yes, there is the bitter lesson and all that but we don't know if this is _the_ right hill to climb. Again, at best, this is a hypothesis.

But achieving this will require training environments at a scale and diversity that dwarf anything currently available.

Yes. You should try it. Let us know if it works. All the best!

100% agreed. It is all about removing friction for me. Case in point: I would not have touched React in my previous career without the assist that LLMs now provide. The barrier to entry just _felt_ to be too large and one always has the instinct to stick with what one knows.

However, it is _fun_ to go over the barrier if it is chatting with a model to get a quick tutorial and produce working code for a prototype (for your specific needs) where the understanding that you just developed is applied. The alternative (without LLMs) is to first do the ground work of learning via tutorials in text/video form and then do the cognitive mapping of applying the learning to one's prototype. I would make a lot of mistakes that expert/intermediate React developers don't make on this path.

One could argue that it shortcuts some learning and perhaps the old way results in better retention. But, our field changes so fast... and when it remains static for too long, projects die. I think of all this as accelerant for progress in adoption of new ways of thinking about software and diffusing that more quickly across the developer population globally. Code is always fungible, anyway. The job is about all the other things that one needs to do besides coding.

That’s very conservative.

Rituals define a school of thought (or a religion). These are rituals of folks who want to prevent catastrophe through conservation. To each their own.

Ultimately, individual habits do add up. But with climate, one would be hard pressed to find evidence that conservation is the path forward. It does not work, unfortunately.

Nathan Lambert provides a counterpoint to the recent "The Illusion of Thinking" paper by Apple [1]:

"On one of these toy problems, the Tower of Hanoi, the models structurally cannot output enough tokens to solve the problem — the authors still took this as a claim that “these models cannot reason” or “they cannot generalize.” This is a small scientific error."

"it appears that a majority of critiques of AI reasoning are based in a fear of no longer being special rather than a fact-based analysis of behaviors."

[1] https://www.arxiv.org/pdf/2506.06941

I am not sure why this ought to require "pump another $100 Billion". Could you elaborate?

Yes, the more recent generation of GPUs optimize for attention math. But they are still fairly "general-purpose" accelerators as well. So when I see papers like this (interesting idea, btw!), my mental model for costs suggests that the CapEx to buy up the GPUs and build out the data centers would get re-used for this and 100s of other ideas and experiments.

And then the hope is that the best ideas will occupy more of the available capacity...

It is hard to compare models with humans so not sure how to answer it for both. :)

But, for models, this is an interesting finding because a lot of LRMs are LLMs with a _bunch_ of post-training done on top. We know this about DeepSeek R1 (one of the models evaluated in the Apple paper) for sure. They write extensively about how they took DeepSeek-V3-Base and made R1 with it. [1]

If the post-training is resulting in lower performance on simpler tasks then it ought to inspire more research on how to make it so that it doesn't -- i.e., with more training (of any kind), we should be gaining more capabilities. This has been a problem with DNNs historically, btw. We had these issues when fine-tuning text/image classifiers as well. Some weight changes can be destructive. So, it has to be done with a _lot_ of care. And, I am sure folks are working on it, to be honest. Maybe some of them will say something here. :-)

[1] https://github.com/deepseek-ai/DeepSeek-R1

Rather than standard benchmarks (e.g., math problems), we adopt controllable puzzle environments that let us vary complexity systematically

Very clever, I must say. Kudos to folks who made this particular choice.

we identify three performance regimes: (1) low complexity tasks where standard models surprisingly outperform LRMs, (2) medium-complexity tasks where additional thinking in LRMs demonstrates advantage, and (3) high-complexity tasks where both models experience complete collapse.

This is fascinating! We need more "mapping" of regimes like this!

What I would love to see (not sure if someone on here has seen anything to this effect) is how these complexity regimes might map to economic value of the task.

For that, the eval needs to go beyond puzzles but the complexity of the tasks still need to be controllable.

That’s a good example of a worst case scenario. This is why we would still need humans loitering about.

The question is do they still need 10? Or 2 would suffice? How about 5?

This does not need to be a debate about the absolutes.

To me, it reads like a survey paper intended for (and maybe written by) a researcher about to start a new project. I am not a researcher in this space but I have dabbled elsewhere, so it is somewhat accessible. The degree to which one leverages existing jargon in their writing is a choice, of course.

I am curious -- what would have made it more effective at conveying information to you? Different people learn differently but I wonder how people get beyond the hurdles of jargon.

Why we don’t have an LLM based search tool for our pc / smartphones?

I'll offer my take as an outside observer. If someone has better insights, feel free to share as well.

In market terms, I think it is because Google, Microsoft and Apple are all still trying with varied success. It has to be them because that's where a big bulk of the users are. They are all also public companies with impatient investors wanting the stock to go up into the right. So, they are both cautious about what ship to billions of devices (brand protection) and cautious about "opening up" their OS beyond that they have already done (fear of disruption).

In technical terms, it is taking a while because if the tool is going to use LLMs, then they need to solve for 99.999% of the reliability problems (brand protection) that come with that tech. They need to solve for power consumption (either on edge or in the data centers) due to their sheer scale.

So, their choices are ship fast (which Google has been trying to do more) and iterate in public; or partner with other product companies by investing in them (which Microsoft has been doing with Open AI and Google is doing with Anthropic, etc.).

Apple is taking some middle path but they just fired the person who was heading up the initiative [1] so let's see how that goes.

My two cents.

[1] https://www.reuters.com/technology/artificial-intelligence/a...

If I ask Claude to do a basic operation on all files in my codebase it won't do it.

Not sure exactly how you used Claude for this, but maybe try doing this in Cursor (which also uses Claude by default)?

I have had pretty good luck with it "reasoning" about the entire codebase of a small-ish webapp.

This assumes that inference is needed 24/7.

That may or may not be true for use-cases that require asynchronous, bulk inference _and_ require some task-specific post-training.

FWIW, my approach towards tasks like the above is to

1. start with using an off-the-shelf LM API until

2. one figures out (using evals that capture product intent) what the failure modes are (there always are some) and then

3. post-train against those (using the evals)

Working on it. :-)

I find the AT protocol interesting.

But, now they have network effects; > 30 million users -- that's a big market to build for.

Yes -- LLMs can write a lot of code and after some reviewing it can also go to prod -- but I have not nearly enough applications of LLMs on the post-prod phase; like dealing with evolution in requirements, ensuring security as zero days get discovered, etc.

Would love to hear folks' experience around "managing" all this new code.

Oh, absolutely. That's how we are going to deal with the current crop of agents here -- some combination of updates to the weights, prompt-tuning and sandboxing so bad things cannot happen. So, I am not one of those people who is against doing those things to mitigate risks.

However, shouldn't we ask for more? Even writing the paragraph above feels exhausting. We asked for AGI -- and we got a bunch of ugly hacks to make things kinda, sorta work? Where is the elegance in all that?

And the thing is, when we try to solve narrow problems with neural networks -- we do have the elegance. AlphaFold, AlphaGo, Text Embeddings, etc. All that stuff just works.

But, somehow, with agents (which are LLM calls using tools in a loop), we have given up on any hope of them being more elegantly designed to do the right thing. Why is that?