corbin

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.

Hacker News is truly a study in masculinity. This brave poster is willing to stand up and ask whether Bluey harms men by making them feel emotions. Credit to the top replies for simply linking him to WP's article on catharsis.

Here's some food for thought; ha ha, only serious. What if none of this is new?

If this is a dealbreaker today, then it should have been a dealbreaker over a decade ago, when Google first rolled out Knowledge panels, which were also often inaccurate and unhelpful.

If this isn't acceptable from Google, then it shouldn't be acceptable from DuckDuckGo, which has the same page-one results including an AI summary and panels, nor any other search engines. If summaries are unacceptable from Gemini, which has handily topped the leaderboards for weeks, then it's not acceptable using models from any other vendor, including Alibaba, High-Flyer, Meta, Microsoft, or Twitter.

If fake, hallucinated, confabulated, or synthetic search results are ruining the Web today, then they were ruining the Web over two decades ago and have not lessened since. The economic incentives and actors have shifted slightly, but the overall goal of fraudulent clicks still underlies the presentation.

If machine learning isn't acceptable in collating search results today, then search engines would not exist. The issue is sheer data; ever since about 1991, before the Web existed, there has been too much data available on the Internet to search exhaustively and quickly. The problem is recursive: when a user queries a popular search engine, their results are populated by multiple different searchers using different techniques to learn what is relevant, because no one search strategy works at scale for most users asking most things.

I'm not saying this to defend Google but to steer y'all away from uncanny-valley reactionism. The search-engine business model was always odious, but we were willing to tolerate it because it was very inaccurate and easy to game, like a silly automaton which obeys simple rules. Now we are approaching the ability to conduct automated reference interviews and suddenly we have an "oops, all AI!" moment as if it weren't always generative AI from the beginning.

For posterity: English Wikipedia is deletionist, so your burden of proof is entirely backwards. I know this because I quit English WP over it; the sibling replies are from current editors who have fully internalized it. English WP's notability bar is very high and not moved by quantity of sources; it also has suffered from many cranks over the years, and we should not legitimize cranks merely because they publish on ArXiv.

We can read between the lines for ourselves. From OpenAI's announcement of Stargate in January, the only equity-holder who has built datacenters is Oracle, and the only other technology partner who has built datacenters is Microsoft. They claim that OpenAI will be operationally responsible, but OpenAI doesn't have a team dedicated to building out and staffing datacenters. In related reporting, Microsoft relaxed its exclusive rights to OpenAI's infrastructure specifically for Oracle and Stargate. As for the motives, I'll highlight Ed's reporting:

The Oracle/Stargate situation was a direct result — according to reporting from The Information — of OpenAI becoming frustrated with Microsoft for not providing it with servers fast enough, including an allotment of 300,000 of NVIDIA's GB200 chips by the end of 2025.

Why is Microsoft canceling a Gigawatt of data center capacity while telling everybody that it didn’t have enough data centers to handle demand for its AI products? I suppose there’s one way of looking at it: that Microsoft may currently have a capacity issue, but soon won’t, meaning that further expansion is unnecessary.

This is precisely it. Internally, Microsoft's SREs perform multiple levels of capacity planning, so that a product might individually be growing and requiring more resources over the next few months, but a department might be overall shrinking and using less capacity over the next few years. A datacenter requires at least 4yrs of construction before its capacity is available (usually more like 5yrs) which is too long of a horizon for any individual product...unless, of course, your product is ChatGPT and it requires a datacenter's worth of resources. Even if OpenAI were siloed from Microsoft or Azure, they would still know that OpenAI is among their neediest customers and include them in planning.

Source: Scuttlebutt from other SREs, mostly. An analogous situation happened with Google's App Engine product: App Engine's biggest users impacted App Engine's internal capacity planning at the product level, which impacted datacenter planning because App Engine was mostly built from one big footprint in one little Oklahoma datacenter.

Conclusion: Microsoft's going to drop OpenAI as a customer. Oracle's going to pick up the responsibility. Microsoft knows that there's no money to be made here, and is eager to see how expensive that lesson will be for Oracle; Oracle is fairly new to the business of running a public cloud and likely thinks they can offer a better platform than Azure, especially when fueled by delicious Arabian oil-fund money. Folks may want to close OpenAI accounts if they don't want Oracle billing them someday.

Reading through the docket, he is entitled to a hearing for relief and has a modicum of standing due to the threat of deportation from the USA to China; it's not unreasonable to go to federal court. The judge was fairly courteous in referring him to the Pro Se Project a week ago. I'm a little jealous of how detached he is from reality; from 36(a) of the Amended Complaint:

The Plaintiff asserts that completing a Ph.D. in Health Services Research significantly increases earning potential. The average salary for individuals with such a Ph.D. is $120,000 annually, compared to $30,000 annually in China, where Plaintiff’s visa cancellation forces him to seek employment. Over an estimated 30-year working career, this represents a lifetime income loss of $2,700,000.

He really went up to the judge and said, "your honor, my future career is dependent on how well I prompt ChatGPT, but statistically I should be paid more if I have a second doctorate," and the judge patted him on his head and gave him a lollipop for being so precocious.