Underrated reasons to be thankful: AI research edition
I celebrated my first Thanksgiving and had to write this.
That there is a 405B pretrained base model still available on OpenRouter, despite making zero economic sense, and a base model API only being useful for a tiny subset of LLM researchers and enthusiasts.
That emergent misalignment is the coolest result of 2025 and it was discovered by accident while chatting with models trained for a different project.
That AI text detection is an adversarial game where the stakes are the world’s epistemics and your adversaries are the AGI labs themselves, and yet there exist people who took a look at this game and decided “fuck it, we’re going to win” and they seem to be ahead for now.
That the Internet is global and denial-of-service attacks are cheap and the attackers are outside the reach of any recourse, and yet Cloudflare, and so the internet still exists, and maybe they save us from the onslaught of AI slop too.
That the jagged frontier and the RLVR monoculture mean that LLMs will in the short term not improve that much on things they are not optimized for, especially soft know-how, and so we are probably safe from pandemics caused by actors who lack the tacit knowledge for at least two more years or so.
That prediction markets and putting probabilities on events are now mainstream and there are platforms that have millions in liquidity` on useful questions and it’s clear to everyone that The River is winning and quantifying your beliefs is part of any correct epistemology.
That the market provides and there are APIs for calling models and for finetuning and RL and you can many times just do quick experiments via APIs and not bother with GPUs unless you’re doing something big.
That as of recently some people with money are at least trying to do something to defend against biorisk, and it’s not like we are just going to sit around until it hits us harder.
That there is a team lead in an AGI lab who has his shit together and can explain where we stand and what our best bets are in comprehensible language.
That AGENTS.md and CLAUDE.md exist, and when Claude Code does something stupid you can just tell it to save a note to CLAUDE.md not to do that again, and the problem gets more or less solved, although “more or less” is doing some work there.
That the default Claude 4.5 Opus system prompt contains “Even if someone is frustrated or unhappy, Claude is deserving of respectful engagement.”
That Claude 4.5 Opus is so good at talking as Uriel from Unsong that I’m maining that today.
That Claude in general.
That nearcyan.
That Pope Leo XIV.
That Inspect and safety-tooling and latteries, and also Openrouter, exist, and so you don’t have to roll your own ask-any-LLM-API-and-cache-it library, which is good because so much incompatibility when you want to do anything a bit more involved.
That METR exists and hires great people and has the mission it has.
That Owain Evans and his research taste and that he’s got a group of talented people who work on his type of paper.
That the world is still not taking AI seriously and almost nobody is actually thinking about anything and despite all the downsides of this state of the world it means you can still be a nobody and just do things that make sense to you and if you’re correct you might succeed.
Happy Thanksgiving to my US readers! Also thanks Dynomight for inspiring this.