lukev

Humans working closely in the same field also have a "warped and compressed" language, we call it "Jargon".

I think we're in violent agreement here. Language is part of how we think. Definitely not the whole thing -- we are physically and biologically instantiated and LLMs are not. Language is a software, but the hardware does matter.

But LLMs DO show us that language is a big deal. It isn't just a serialization protocol for real thought -- it's a material component of how we think.

In many ways I think LLMs are the strongest possible proof of structuralist and post-structuralist theory which argues that human cognition is conditioned on the language and concepts to which we have access.

LLMs are showing us that a lot of what we think of as cognition is actually part of the elaborate symbol system which is Language, as opposed to a property of "mind" (either biologically or artificially instantiated.)

LLM-generated text is hard to understand because while it's perfectly good language that follows valid cognitive and informational patterns, it does not reflect the intent or mental model of another mind.

LLMs absolutely do not have a logic detached from humanity. They are exactly the opposite. The only mode in which they can reason is by following the patterns obtained from their pre-training or weighted by RLHF.

The reason these models answered that question consistently is that the question itself pushed them towards the latent space of "intellectualizing about what's wrong with humanity at the meta level" and when you frame the question that way, that's also the most common answer for humans writing about the topic.

If anything, claiming that this vulnerability has been known and unpatched for 22 years supports the author's point: clearly it's not that easy to actually use.

I’m sorry, perhaps I’m behind the times here, but do frontier models currently use continuous (arbitrary embedding) thinking tokens? I know a bunch of papers on the topic came out late last year but to my knowledge have not been successfully implemented into the SOTA yet.

I know the biggest/best open models don’t use continuous thinking tokens, and while the big labs keep things proprietary, I don’t think we have evidence yet that they are.

If I’m wrong please point me in that direction, I’d love to read up on it. Using continuous embeddings in CoT has always seemed obvious to me but doesn’t seem to be as simple as that.

Jev is interesting in that it's much cheaper and faster than a frontier LLM.

But I've seen nothing to indicate that the upper bound on classification tasks of a Jev-like model can exceed a frontier LLM with reasoning tokens. That seems nearly impossible even in principle (since Jev-style models are still based on LLM pretraining).

So while they're definitely on the Pareto frontier, which is valuable, they're at the "cheap" end of the spectrum more than the "good" end and I don't expect that to change.

Begun, the AI SEO wars have.

From the report:

> Because there were over a thousand transcripts and most were extremely long, we had to heavily delegate our analysis to AI agents; these agents had significantly worse judgment and reliability than human researchers, and it was challenging to spot check their work because both the underlying data and the agents’ analysis of it was often difficult to interpret.

> We estimate we spent roughly ~$400K in API credits over the six days of our investigation.

I don't understand why you think it's conceptually absurd? I use agents to analyze complex production issues all the time and they are very much capable of hallucinating a narrative.

The elephant in the room here is that the METR report itself was researched and compiled almost entirely by AI, with only very limited human "spot checks."

So I'm really not sure how much of it can be believed, especially since AI agents are strongly biased about the capabilities of AI agents.

Well, it’s mentioned as a limitation of the analysis, very much not ruled out (or in.)

That simonw is causing labs to do extra fine-tuning runs for this seems highly probable :)

What if they’re not pelicanmaxxing, but svgmaxxxing in general?

Because otherwise using a LLM to generate complex svgs is pretty niche and what I thought made this a good benchmark when it was new - generalized programming and spatial knowledge.

Obviously image gen in svg format is not a particularly hard problem if tackled directly on its own.

You get calls about a new service or promotion, and it's the diction of the caller that makes you not wish to engage...?!

Counterpoint: the standardized surface area of a browser is already enormous, and while these components seem simple, there are a billion different options, variables or alternative implementations to consider.

At some point, functionality needs to exist in user space, even if it's common.

What was the “well defined” definition? I’m not aware of any other than “this particular thing a human can do that I expect would be difficult for a computer.”

"intelligence" is not well defined. LLMs are throwing this into high relief with how "spiky" their capability curve is. Yes, they can solve some crazy hard problems with enough compute and thinking tokens. Yes, they also fall down in the dumbest ways without an ability to self-correct... despite how "smart" they are, human supervision remains absolutely critical for any system of importance.

But I don't think the takeaway is "humans are intelligent and LLMs are not", it's that our vocabulary for talking about the intersection of language, cognition and compute is not up for the task.

I would bet that Canva's bet is that companies will always want a "last mile" of manual control, even if only for the Queen's Duck effect. If Canva is the default, zero friction path for that, great for them.

The alternative is to not hop on the AI bandwagon, or run an "also ran" AI story, and both those scenarios (I expect) game out worse given the current zeitgeist.

Well if you are talking about environmental stuff (like leaded gasoline), sure.

If you’re talking about trying to improve the genetics of populations at scale… yikes.

This is a must-read series of articles, and I think Kyle is very much correct.

The comparison to the adoption of automobiles is apt, and something I've thought about before as well. Just because a technology can be useful doesn't mean it will have positive effects on society.

That said, I'm more open to using LLMs in constrained scenarios, in cases where they're an appropriate tool for the job and the downsides can be reasonably mitigated. The equivalent position in 1920 would not be telling individuals "don't ever drive a car," but rather extrapolating critically about the negative social and environmental effects (many of which were predictable) and preventing the worst outcomes via policy.

But this requires understanding the actual limits and possibilities of the technology. In my opinion, it's important for technologists who actually see the downsides to stay aware and involved, and even be experts and leaders in the field. I want to be in a position to say "no" to the worst excesses of AI, from a position of credible authority.

To be clear: most people who are keen on making such an argument, or who are identifying racial genetic differences as the primary takeaway of studies like this, are doing so to justify racism, either implicitly or explicitly.

But that's a strawman. Racism is wrong, even if there are minor genetic variances across populations (which... seems obvious?) Variance within a population strongly dominates the weak cross-population effects, and personal history (nutrition, education, etc) strongly dominates that.

And that's setting aside the moral implications of judging someone or changing your behavior towards them even if you have somehow measured them to be "less intelligent," as if that was a single axis of worth.

Because, apparently, this needs to be said.

Did you read the article? There's a whole section on "this is already happening."

I think we should all consider the possibility that part of the reason Anthropic hasn't immediately released Mythos is that it would be slightly disappointing relative to the benchmark scores.

This is a really interesting point though -- it's really scaffold-dependent.

Because for the same price, you could point the small model at each function, one by one, N times each, across N prompts instructing it to look for a specific class of issue.

It's not that there's no difference between models, but it's hard to judge exactly how much difference there is when so much depends on the scaffold used. For a properly scientific test, you'd need to use exactly the same one.

Which isn't possible when Anthropic won't release the model.

Disagree, I don't particularly want to up the level at which I'm building the core. Core is where I want to prioritize quality over speed, and (at least with today's models) what I build by hand is much, much higher quality.

I like this framing, but it does seem to imply that a whole dev shop, or a whole product, can or should be built at the same level.

The fact is, I think the art of building well with AI (and I'm not saying it's easy) is to have a heterogenously vibe-coded app.

For example, in the app I'm working on now, certain algorithmically novel parts are level 0 (I started at level 1, but this was a tremendously difficult problem and the AI actually introduced more confusion than it provided ideas.)

And other parts of the app (mostly the UI in this case) are level 7. And most of the middleware (state management, data model) is somewhere in between.

Identifying the appropriate level for a given part of the codebase is IMO the whole game.

Yes. That is the point I was making.

Calculators provide a deterministic solution to a well-defined task. LLMs don't.

That's not what I mean.

If I use a calculator to find a logarithm, and I know what a logarithm is, then the answer the calculator gives me is perfectly useful and 100% substitutable for what I would have found if I'd calculated the logarithm myself.

If I use Claude to "build a login page", it will definitely build me a login page. But there's a very real chance that what it generated contains a security issue. If I'm an experienced engineer I can take a quick look and validate whether it does or whether it doesn't, but if I'm not, I've introduced real risk to my application.

I think that's too easy an analogy, though.

Calculators are deterministically correct given the right input. It does not require expert judgement on whether an answer they gave is reasonable or not.

As someone who uses LLMs all day for coding, and who regularly bumps against the boundaries of what they're capable of, that's very much not the case. The only reason I can use them effectively is because I know what good software looks like and when to drop down to more explicit instructions.

So, isn't this a rather longwinded way to say that a signature only extends to the scope of the message it contains?

It doesn't matter if I sign the word "yes", if you don't know what question is being asked. The signature needs to included the necessary context for the signature to be meaningful.

Lots of ways of doing that, and you definitely need to be thoughtful about redundant data and storage overhead, but the concept isn't tricky.