Huggingface cofounder pushes against LLM hype, really softly. Not especially worth reading except to wonder if high profile skepticism pieces indicate a vibe shift that can’t come soon enough. On the plus side it’s kind of short.

The gist is that you can’t go from a text synthesizer to superintelligence, framed as how a straight-A student that’s really good at learning the curriculum at the teacher’s direction can’t really be extrapolated to an Einstein type think-outside-the-box genius.

The world ‘hallucination’ never appears once in the text.

New ultimate grift dropped, Ilya Sutskever gets $2B in VC funding, promises his company won’t release anything until ASI is achieved internally.

Before focusing on AI he was going off about what he called the rot economy, which also had legs and seemed to be in line with Doctorow’s enshitification concept. Applying the same purity standard to that would mean we should be suspicious if he ever worked with a listed company at all.

Still I get how his writing may feel inauthentic to some, personally I get preacher vibes from him and he often does a cyclical repetition of his points as the article progresses which to me sometimes came off as arguing via browbeating, and also I’ve had just about enough of reading performatively angry internet writers.

Still, he must be getting better or at least coming up with more interesting material, since lately I’ve been managing to read them all the way through.

What else though, is he being secretly funded by the cabal to make convolutional neural networks great again?

That he found his niche and is trying to make the most of it seems by far the most parsimonious explanation, and the heaps of manure he unloads on the LLM both business and practices weekly surely can’t be helping DoNotPay’s bottom line.

i think yud at some point claimed this (preventing the robot devil from developing alignment countermeasures) as a reason his EA bankrolled think tanks don’t really publish any papers, but my brain is too spongy to currently verify, as it was probably just some tweet.

I don’t think him having previously done undefined PR work for companies that include alleged AI startups is the smoking gun that mastopost is presenting it as.

Going through a Zitron long form article and leaving with the impression that he’s playing favorites between AI companies seems like a major failure of reading comprehension.

It’s adorable how they let the alignment people still think they matter.

Should be noted that it’s mutual, Hanania has gone to great lengths to suck up to siskind, going back to at least the designer mouth bacteria thing.

And GPT-4.5 is terrible for coding, relatively speaking, with an October 2023 knowledge cutoff that may leave out knowledge about updates to development frameworks.

This is in no way specific to GPT4.5 but remains a weirdly undermentioned albatross about the neck of the entire LLM code-guessing field, probably because the less you know about what you told it to generate the likelier you are to think it’s doing a good job, and the enthusiastically satisfied customer reviews in social media that I’ve interacted with certainly seemed to skew toward less-you-know types.

Even when the up-to-date version release happened before the cut-off point you are probably out of luck, since the newer version is likely way underrepresented in the training data compared to the previous versions that people may have been using for years by that point.

Nothing in my experience with LLMs or my reading of the literature has ever led me to believe that prompting one to numerically rate something and treating the result as meaningful would be a productive use of someone’s time.

Still occasionally think about that bit in the o1 white paper where the openai researchers innocuously pose the question of what if our benchmarks for detecting hallucinations are shit actually, wouldn’t that be something.

Implicitly assuming that the technology to terraform Mars is just around the corner is the we’ll become profitable once we hit AGI of space exploration.

In todays ACX comment spotlight, Elon-anons urge each other to trust the plan:

Could also be don’t worry about deepseek type messaging that addresses concerns without naming names, to tell us that a drastic reduction in infrastructure costs was foretold by the writing of St Moore and was thus always inevitable on the way to immanentizing the AGI, ἀλληλούϊα.

It’s like you founded a combination of an employment office and a cult temple, where the job seekers aren’t expected or required to join the cult, but the rites are still performed in the waiting room in public view.

chef’s kiss

The surface claim seems to be the opposite, he says that because of Moore’s law AI rates will soon be at least 10x cheaper and because of Mercury in retrograde this will cause usage to increase muchly. I read that as meaning we should expect to see chatbots pushed in even more places they shouldn’t be even though their capabilities have already stagnated as per observation one.

2. The cost to use a given level of AI falls about 10x every 12 months, and lower prices lead to much more use. You can see this in the token cost from GPT-4 in early 2023 to GPT-4o in mid-2024, where the price per token dropped about 150x in that time period. Moore’s law changed the world at 2x every 18 months; this is unbelievably stronger.

Saltman has a new blogpost out he calls ‘Three Observations’ that I feel too tired to sneer properly but I’m sure will be featured in pivot-to-ai pretty soon.

Of note that he seems to admit chatbot abilities have plateaued for the current technological paradigm, by way of offering the “observation” that model intelligence is logarithmically dependent on the resources used to train and run it (i = log( r )) so it’s officially diminishing returns from now on.

Second observation is that when a thing gets cheaper it’s used more, i.e. they’ll be pushing even harded to shove it into everything.

Third observation is that

The socioeconomic value of linearly increasing intelligence is super-exponential in nature. A consequence of this is that we see no reason for exponentially increasing investment to stop in the near future.

which is hilarious.

The rest of the blogpost appears to mostly be fanfiction about the efficiency of their agents that I didn’t read too closely.

Penny Arcade weighs in on deepseek distilling chatgpt (or whatever actually the deal is):

You misunderstand, they escalate to the max to keep themselves (including selves in parallel dimensions or far future simulations) from being blackmailed by future super intelligent beings, not to survive shootouts with border patrol agents.

I am fairly certain Yud had said something very close to that effect in reference to preventing blackmail from the basilisk, even though he tries to no-true-scotchman zizians wrt his functional decision ‘theory’ these days.

Distilling is supposed to be a shortcut to creating a quality training dataset by using the output of an established model as labels, i.e. desired answers.

The end result of the new model ending up with biases inherited from the reference model should hold, but using as a base model the same model you are distilling from would seem to be completely pointless.