corbin

It's well-known folklore that reinforcement learning with human feedback (RLHF), the standard post-training paradigm, reduces "alignment," the degree to which a pre-trained model has learned features of reality as it actually exists. Quoting from the abstract of the 2024 paper, Mitigating the Alignment Tax of RLHF (alternate link):

LLMs acquire a wide range of abilities during pre-training, but aligning LLMs under Reinforcement Learning with Human Feedback (RLHF) can lead to forgetting pretrained abilities, which is also known as the alignment tax.

In practice, the behaviors that the chatbots learn in post-training are FUD and weasel-wording; they appear to not unlearn facts, but to learn so much additional nuance as to bury the facts. The bots perform worse on various standardized tests about the natural world after post-training; there are quantitative downsides to forcing them to adopt any particular etiquette, including speaking like a chud.

The problem is mostly that the uninformed public will think that the chatbot is knowledgeable and well-spoken because it rattles off the same weak-worded hedges as right-wing pundits, and it's addressed by the same improvements in education required to counter those pundits.

Answering your question directly: no, slop machines can't be countered with more slop machines without drowning us all in slop. A more direct approach will be required.

Yes, but the article's not actually about that. It's about Microsoft returning to the same datacenter-building schedule from a decade ago. Datacenters have a lag of about 3-5yrs depending on what's inside them and where they're located, so what we're actually seeing is Microsoft projecting a relative reduction in overall usage. Note that among all the cancellations of notes and prospective claims, Microsoft isn't walking back their two-decade nuclear-power deal with Westinghouse; they're not destroying or reducing any existing capacity, just planning to build less. At risk of quoting Bloomberg:

After a frantic expansion to support OpenAI and other artificial intelligence projects, [Microsoft] expects spending to shift from new construction to fitting out data centers with servers and other equipment.

To the extent that the bubble is popping, Microsoft and other datacenter owners have to guess half a decade in advance when the bubble will pop, and if you take them at their word — that is, if we assume that they canceled these contracts with perfect foresight — then the bubble must have already popped in 2023-2024, and the market is experiencing coyote time because…? More likely, this is fallout from their ongoing breakup with OpenAI, who almost certainly begged Microsoft for so much compute (and definitely begged for too many nVidia GPUs!) that Microsoft had to adjust their datacenter plans. The bubble's not done until OpenAI has exhausted all possible funding, say in late 2025 or early 2026 when Softbank and the Saudis realize that they've made a hilarious mistake.

We've discussed this previously on awful.systems, both the value of nuclear-energy contracts and Microsoft's retraction of intents.

Like any reality-show writing room, they only plan one episode in advance and only have a week's worth of photography in mind.

As the classic film Network points out, the Saudi money is the end of the road; there aren't any richer or more gullible large wealth funds who will provide further cash. So OpenAI could be genuinely out of greater ~~fools~~ financing after another year of wasting Somebody Else's Money. This crash has removed "large" from the front of any other wealth fund that might have considered bailing them out. The Stargate gamble could still work out, but so far I think ti's only transferred bag-holding responsibilities from Microsoft to Oracle.

Another path is to deflate nVidia's cap. At first blush, this seems impossible to me; nVidia's business behavior is so much worse than that of competitors Intel or Imagination yet they have generally never lost faith from their core gaming laity, and as long as nVidia holds 20-30% of the gaming GPU market they will always have a boutique niche with cap at least comparable to e.g. their competitor AMD. But GPUs have been treated as currency among datacenter owners, and a market crash could devalue the piles of nVidia GPUs which some datacenter owners have been using as collateral for purchasing land, warehouses, machines, more GPUs, etc. nVidia isn't the only bag-holder here, though, and since they don't really want to play loan-shark and repossess a datacenter for dereliction, odds are good that they'll survive even if they're no longer king of the hill. The gold rush didn't work out? Too bad, no returns allowed on shovels or snow gear.

Side note: If folks just wanted to know whether tech in general is hurt by this, then yes, look at Tesla's valuation. Tesla is such a cross-cutting big-ticket component of so many ETFs that basically every retirement scheme took a hit from Tesla taking a hit. The same thing will happen with nVidia and frankly retirement-fund managers should feel bad for purchasing so much of what any long-term investor would consider to be meme stocks. (I don't hold either TSLA or NVDA stocks.)

I hope this makes sense. I don't post with this candor when I'm well-rested and sober.

Australian chemist and videographer Explosions & Fire argues convincingly that the ongoing recent radioactive-boy-scout scandal should not result in prosecution. For context, a 24-year-old man ordered small samples of radioactive isotopes from the USA, Australia failed to intercept it at the border, and they are prosecuting him in order to avoid embarrassment over incompetence. I don't have a choice sneer; E&F is unwaveringly energized over the topic of radioactive isotopes and injustice, and the whole thing is worth watching.

Today on the orange site, an AI bro is trying to reason through why people think he's weird for not disclosing his politics to people he's trying to be friendly with. Previously, he published a short guide on how to talk about politics, which — again, very weird, no possible explanation for this — nobody has adopted. Don't worry, he's well-read:

So far I've only read Harry Potter and The Methods of Rationality, but can say it is an excellent place to start.

The thread is mostly centered around one or two pearl-clutching conservatives who don't want their beliefs examined:

I find it astonishing that anyone would ask, ["who did you vote for?"] … In my social circle, anyway, the taboo on this question is very strong.

To which the top reply is my choice sneer:

In my friend group it's clear as day: either you voted to kill and deport other people in the friend group or you didn't. Pretty obvious the group would like to know if you're secretly interested in their demise.

Angela Collier has a wonderfully grumpy video up, why functioning governments fund scientific research. Choice sneer at around 32:30:

But what do I know? I'm not a medical doctor but neither is this chucklefuck, and people are listening to him. I don't know. I feel like this is [sighs, laughs] I always get comments that tell me, "you're being a little condescending," and [scoffs] yeah. I mean, we can check the dictionary definition of "condescending," and I think I would fit into that category. [Vaccine deniers] have failed their children. They are bad parents. One in four unvaccinated kids who get measles will die. They are playing Russian roulette with their child's life. But sure, the problem is I'm being, like, a little condescending.

Strange is a trooper and her sneer is worth transcribing. From about 22:00:

So let's go! Upon saturating my brain with as much background information as I could, there was really nothing left to do but fucking read this thing, all six hundred thousand words of HPMOR, really the road of enlightenment that they promised it to be. After reading a few chapters, a realization that I found funny was, "Oh. Oh, this is definitely fanfiction. Everyone said [laughing and stuttering] everybody that said that this is basically a real novel is lying." People lie on the Internet? No fucking way. It is telling that even the most charitable reviews, the most glowing worshipping reviews of this fanfiction call it "unfinished," call it "a first draft."

A shorter sneer for the back of the hardcover edition of HPMOR at 26:30 or so:

It's extremely tiring. I was surprised by how soul-sucking it was. It was unpleasant to force myself beyond the first fifty thousand words. It was physically painful to force myself to read beyond the first hundred thousand words of this – let me remind you – six-hundred-thousand-word epic, and I will admit that at that point I did succumb to skimming.

Her analysis is familiar. She recognized that Harry is a self-insert, that the out-loud game theory reads like Death Note parody, that chapters are only really related to each other in the sense that they were written sequentially, that HPMOR is more concerned with sounding smart than being smart, that HPMOR is yet another entry in a long line of monarchist apologies explaining why this new Napoleon won't fool us again, and finally that it's a bad read. 31:30 or so:

It's absolutely no fucking fun. It's just absolutely dry and joyless. It tastes like sand! I mean, maybe it's Yudkowsky's idea of fun; he spent five years writing the thing after all. But it just [struggles for words] reading this thing, it feels like chewing sand.

Anecdote: I gave up on COBOL as a career after beginning to learn it. The breaking point was learning that not only does most legacy COBOL code use go-to statements but that there is a dedicated verb which rewrites go-to statements at runtime and is still supported on e.g. the IBM Enterprise COBOL for z/OS platform that SSA is likely using: ALTER.

When I last looked into this a decade ago, there was a small personal website last updated in the 1990s that had advice about how to rewrite COBOL to remove GOTO and ALTER verbs; if anybody has a link, I'd appreciate it, as I can no longer find it. It turns out that the best ways of removing these spaghetti constructions involve multiple rounds of incremental changes which are each unlikely to alter the code's behavior. Translations to a new language are doomed to failure; even Java is far too structured to directly encode COBOL control flow, and the time would be better spent on abstract specification of the system so that it can be rebuilt from that specification instead. This is also why IBM makes bank selling COBOL emulators.

In lesser corruption news, California Governor Gavin Newsom has been caught distributing burner phones to California-based CEOs. These are people that likely already have Newsom's personal and business numbers, so it's not hard to imagine that these phones are likely to facilitate extralegal conversations beyond the existing ~~bribery~~ legitimate business lobbying before the Legislature. With this play, Newsom's putting a lot of faith into his sexting game.

Well, let's not let Baldur be a complete dumbass. There is something bad here, and we've discussed it before (1, 2), but it's not "US authorities" gaining "control" over "bigotry and biases". The actual harm here is appointing AI-safety dorks to positions in NIST. For those outside the USA, NIST is our metrologist organization, and there's no good reason for AI safety to show up there.