stevenhuang

Anthropic need not care about this space. It's possible it's more worth it for them to dedicate their resources into making better base models.

Doesn't sound like you've been following the space. They exist, it's called LoRas. The Unsloth project is huge and whole purpose is to make it easy to train to train LLM LoRas. The reason you don't see LoRas on their own is because people who fine tune models merge them back into the base and ship the full model weights instead

Also see LoRAX, it's an inference engine that exactly does this, it can serve many LoRas off a single base model

Sounds like a skill issue

Some are better at reading comprehension than others.

It's more likely that token costs will keep falling over time.

So drive then. You can also move. The point is there are options. Many cities have strong infrastructure and have embraced walk ability. If not driving is such a big deal to you, why not do something about it?

No one forces you to drive. You can still walk.

Anyone who has used local LLMs will come to this conclusion.

Too much context confuses the LLM. Reduce unnecessary context and performance improves. Not to mention, too much context and you just run out of vram.

Such limitations may be mitigated in future architectures, but it certainly isn't something you can ignore currently.

Why should it be asinine? The universe is under no obligation to be understandable to humans.

> just to clear space for the things you need to comprehend.

That's really it. I've been finding the same limitations with the default PR code review tools. We currently make do by having a section in our PRs "recommended review order" with a list of files/dirs that should be looked at first, which not only is much smaller Loc than everything else but is the main relevance of the PR.

Within a given project there are inevitably certain files/concepts that are more important, and a better review tool will help emphasize the structure of a large PR to make things easier to review.

I suppose you created a web based front end to GitHub PRs for your team to use? Or is it a local tool? Wondering how yours works.

We use gitlab and there's probably some way to hook up a custom external review tool easily (a bot that posts a link to view the current PR in the tool, for example)

Thanks for the idea!

I wouldn't take stock in what LeCun says. He's already been wrong many times about what LLMs can be capable of.

We're in new grounds, no one knows which way it'll go.

I don't think anyone would particularly care if only one of them is wrong, if most are correct.

If they are all wrong, that's when it would backfire.

They're all here https://github.com/openai/math/tree/main/preprints

At this point it takes some weird epistemics to think they'd need to outright fabricate the outcome of an investigation when countless other mathematical breakthroughs have been solved the same.

They are certainly biased, but I don't think such bias is strong enough to cause them to lie about facts. We may agree to disagree here.

You don't believe OpenAI given their assessment, so it seems you're not operating on evidence in the first place

It's not that murky anymore. It's been determined none of the chat contents were used. If you don't believe OpenAI, that's your choice.

See https://openai.com/index/navier-stokes-solution/

> Following an investigation, we have confirmed that Buckmaster’s Codex prompts over the two months preceding this announcement and paper on September 8, 2026, could not have influenced the system in any way, including through training. The OpenAI internal model used for this result was developed through large-scale reinforcement learning on top of a previously pretrained model. Our proofs also differ significantly. In the Euler case, Alpöge and Buckmaster proved a result with external forcing, while OpenAI’s system proved a result without external forcing.

Also called surprisal minimization in Karl Fristons free energy principle.

If by now you still think it's all just hype, it's safe to say you've succumbed to a mind virus that renders you unable to think critically about AI. Otherwise you'd have some level of awareness of just how far this technology has developed, and you should find these developments more than plausible.

This comment makes no sense. You realize how much lsp is used in other editors like vim/neovim? The existence of lsp is exactly what lets custom editors to flourish. Any language/lsp client can add their own extensions too. Any new editor can come up with it's own way of doing things and implement it on top of lsp, as long as the lsp client supports those extra features then there's no problem. I don't know how you can get it so wrong.

> I suspect that the average user doesn’t make a distinction between Uber and driver, and possibly assume that drivers are trained to act in a safe manner.

Are you for real?

So it moves from both "my implementation and specification is incorrect", to just "my specification is incorrect".

I don't understand this type of thinking. Proving what you can is still better. Don't let perfect be the enemy of good.

You have poor epistemics if you are still assigning exactly 0% probability given developments of late.

I work on firmware and deep embedded systems. Trust me when I say LLMs are getting just as capable in this domain. Eg LLMs can do RF filter design. Tell the LLM your target metrics and it will give very reasonable design constraints, especially when you come back with simulation results and relevant feedback.

No proponents of RSI state they will be operating outside of reality. Said another way, they will operate within the confines of what's possible and still be RSI. I'm quite surprised this is something that needs to be clarified.

You are constructing a straw man of your own making.

I guess self contained RSI can only possible if the information contained in all of recorded human knowledge to date is "reality-complete", ie sufficiently captures enough about reality that a "perfectly optimum learning algorithm" is theoretically able to reconstruct everything there is to know about our physical reality.

If the algorithms are insufficiently optimum or the recorded knowledge is of insufficient fidelity, then we'd find ourselves at a local optimum and would need to interface with reality.

A huge part of learning is to probe reality and observe effects, so I think even for current RSI to increase chances of success we would structure it so it can interact with an external environment of some sort, and receive inputs. It would be needlessly limiting otherwise.

> Recursive self improvement is ultimately limited by everything else that cannot move at the speed of electricity.

I'm not sure what your point is. No one thought RSI would break the laws of physics.

It takes quite a lack of foresight to think RSI is completely speculative when it's already been demonstrated how capable agents are at long horizon tasks given suitable harness and unambiguous success criteria. It's hardly a leap to give LLM the goal of improving itself on benchmarks and let it conduct it's own experiments and spin up training runs completely unsupervised.

It's strange you believe this can't happen when a weaker form of it is already happening. And to be so certain RSI can't happen when there really is no technical basis why it can't.

LLMs are advancing the frontiers of knowledge. Is it really that hard for you to believe that people are simply genuinely fascinated about these developments?

People are burnt twice shy and have become comically cynical. Your discernment has been completely shot.

I wouldn't rule out pilot testimony of UFO sightings, nor the possibility we're indeed developing a machine God.

There's ample evidence to support both by now.

Both