Trust Me, I'm Lying (re: AI Bullshit)

If you're unfamiliar with the OpenAI Navier-Stokes thing, you can read The Guardian's ludicrously generous summary of events here and the mathematician's statement that they mysteriously refuse to link here. The tl;dr version is that OpenAI claimed to have solved a very complex math problem with 10,000 agents in 88 hours but they almost certainly misappropriated human work to do so and are lying out their asses about the level of effort involved on the human side within OpenAI.

Scientists like Timnit Gebru were arguing with random Bluesky people about the nature of LLMs because of OpenAI's rumored claim; someone I know was depression-posting about how the world would soon be over because of the machines; Bluesky's CTO was posting about how the machines are obviously thinking and that upsets people's sense of self; anti-AI people were posting about how this was all because of IPOs. Conveniently, after a couple days of this, an Anthropic researcher named Jacob Coxon quit his job and posted on X: The Nazi App about how he quit because the AI companies were 10 years away from ending all humanity by unsafely developing superintelligence. One of the other cultists at Anthropic, Evan Hubinger, then quoted his thread to agree and say he personally thinks there's an >10% chance of ending humanity with LLMs in the next decade. Credulous journalists circulated these claims with little to no pushback. Bernie Sanders released statements promoting his AI safety legislation and is apparently convening Senators to discuss it.

Neither Coxon nor Hubinger substantiated their claims. OpenAI, in claiming to have both solved Navier-Stokes in the manner that they did and in claiming not to have used any user data to do it, didn't substantiate their claims. Bernie also didn't substantiate his claims. None of this has stopped the hysteria from continuing to grow, of course.

I was shocked to find that a number of people I follow on social media are more open to the idea of the machine being a thinking creature than I would have guessed, and I encountered the "what would convince you otherwise" type of question more than once over the weekend and into this week. I'm not particularly interested in litigating whether or not AI can think, mainly because it obviously can't, but I am interested in discussing barriers to my mind ever being changed on this question. The answer, for me, is very simple; it's the paragraph directly proceeding this one. The machines can't think, because machines historically have not been capable of thinking. "Machines can't think" is currently a paradigm all rational people agree was true as recently as 2022. If the machines can now think, then I require evidence demonstrating as much. Science relies on falsifiability and repeatability, neither of which Anthropic or OpenAI are willing to commit to. If you ask them to do so, you'll be informed it's impossible, which of course simply isn't true. They could provide a lot more information about their inputs and outputs than they do, it just wouldn't benefit them as privately held companies seeking IPOs, so they don't.

You can make a number of pseudo-sociological inferences from this, like "capitalism inhibits good science". So true. But, again, this doesn't really interest me. I'm not a sociologist and I already knew corporations were evil.

If you push back on the assertion that machines can think by pointing out that no one has provided scientific evidence otherwise, you will often encounter people willing to challenge you on the basis of your own expertise. What about this paper or that one? Well, what is thinking, anyway? Are you a cognitive scientist or a psychologist? Who are you to say the machine isn't thinking? And this, finally, is the phenomenon I am most interested in.

We live in a renewed age of charlatans (see above), and in an unprecedented age of information availability. Anyone in the entire world can use the internet to find peer-reviewed papers and non-peer-reviewed commentary on any topic scientists have ever talked about. (Anyone in the entire world can conflate the two, too.) Given this relative ease of access, it's unsurprising that online argumentation often includes requests for proof. Is there a study showing that American literacy has declined? Where are the articles talking about how food addiction is or isn't a real thing? What authoritative body decided writing by hand is cognitively different from typing on a computer? You get the gist. It's always good to substantiate your claims with appropriate references when arguing in an online forum, but what I've noticed about the past several years is that people have slowly stopped accepting argumentation bolstered with supporting evidence and started treating argumentation as a duel of references, "evidence" free of independent thought. This culminates most ridiculously in people retrieving arguments from ChatGPT or another LLM and using it in place of their own reasoning.

Unfortunately, the phenomenon also isn't limited to online arguments. I've encountered it while having political discussions, talking to people who will claim allegiance to one analytical framework over others but won't articulate time-and-place specific theories of change for our specific problems. I've encountered it while reading the news, too, where people try to answer questions like "is Donald Trump insane?" by referring to a panel of experts who defer based on wording or lack of examination; the question of what we can all see and hear is treated as secondary. I've even had people try to tell me I'm wrong about a movie being bad because of its Rotten Tomatoes score and then get huffy when I say my opinion is that the movie was bad, not that most people agree with me that the movie was bad.

These anecdotes range from petty complaints to signs of societal apocalypse, granted. But still. I would like to see more people approach critical conversations first and foremost through a perspective of valuing their own human brains. Human beings ship with the ability to reason enabled out of the box. I am capable, through some modest subject matter expertise, of arguing that the Hugging Face attack isn't evidence of sentience because of what a CTF exercise actually is and how a sandbox works; but you, random nontechnical person, are capable of arguing that it's not evidence because OpenAI simply hasn't substantiated any of their claims. Neither of us are obligated to defer to AI researchers who insist upon the deference due to legitimate scientists while refusing to meet the evidentiary standards that distinguish scientific research from fiction or fraud.

Expertise matters, including the kind of expertise you get after a long stint doing a PhD or working on bleeding edge research for 20 years, but if someone asks you to turn your brain off and use theirs instead they should be required to justify it. What's happening at AI companies right now isn't anything like climate science or the development of the atom bomb. These people have a history of making claims that are later revealed to be wildly exaggerated or straight-up fabrication. In my day to day work, I'm a computer programmer; when I do performance tuning, I am always on the lookout for improvements that seem a little too good. Tweaking an algorithm and getting 200% faster code is great, but if it's 800% faster then I start double checking myself, because "too good to be true" is as real in computers as any other industry. This is just a normal, rational thing that most people who are decent at their jobs do—and yet, AI executives regularly announce that they've either achieved AGI or are extremely close to it, over and over, always surprised and disappointed when it turns out they missed something, yet seemingly never capable of learning from their past mistakes.

Here is my leap of deductive reasoning, a feat I am capable of by virtue of my big beautiful human brain: they aren't actually surprised. They don't actually think they've invented AGI. They're lying; some of them are lying while also freaking themselves out, campfire story style, but they are all, to a man, lying. They are lying and making shit up. You do not owe them deference and you shouldn't grant it to them; they're trying to bully people into it by misappropriating scientific terms and aesthetics, but that too is part of the con. Anthropic's "research papers" aren't peer reviewed papers, they're blog posts with a table of contents. OpenAI's swarms of agents aren't "civilizations", they are programs running on servers. Roko never had a fucking basilisk. "AI safety", start to finish, is made up.

You can trust me, though. I'd never lie to you.