Want to wade into the snowy surf of the abyss? Have a sneer percolating in your system but not enough time/energy to make a whole post about it? Go forth and be mid: Welcome to the Stubsack, your first port of call for learning fresh Awful you’ll near-instantly regret.
Any awful.systems sub may be subsneered in this subthread, techtakes or no.
If your sneer seems higher quality than you thought, feel free to cut’n’paste it into its own post — there’s no quota for posting and the bar really isn’t that high.
The post Xitter web has spawned soo many “esoteric” right wing freaks, but there’s no appropriate sneer-space for them. I’m talking redscare-ish, reality challenged “culture critics” who write about everything but understand nothing. I’m talking about reply-guys who make the same 6 tweets about the same 3 subjects. They’re inescapable at this point, yet I don’t see them mocked (as much as they should be)
Like, there was one dude a while back who insisted that women couldn’t be surgeons because they didn’t believe in the moon or in stars? I think each and every one of these guys is uniquely fucked up and if I can’t escape them, I would love to sneer at them.
(Last substack for 2025 - may 2026 bring better tidings. Credit and/or blame to David Gerard for starting this.)
So many CRITICAL and MANDATORY steps in the release instruction file. As it always is with AI, if it doesn’t work, just use more forceful language and capital letters. One more CRITICAL bullet point bro, that’ll fix everything.
Sadly, I am not too surprised by the developers of Lean turning towards AI. The AI people have been quite interested in Lean for a while now since they think it is a useful tool to have AIs do math (and math = smart, you know).
A lot of the money behind lean is from microsoft, so a push for more llm integration is depressing but unsurprising.
Turns out though that llms might actually be ok for generating some kinds of mathematical proofs so long as you’ve formally specified the problem and have a mechanical way to verify the solution (which is where lean comes in). I don’t think any other problem domain that llms have been used in is like that, so successes here can’t be applied elsewhere. I also suspect that a much, uh, leaner specialist model would do just as good a job there. As always, llms are overkill that can only be used when someone else is subsidising them.
Great. Now we’ll need to preserve low-background-radiation computer-verified proofs.
The whole culture of writing “system prompts” seems utterly a cargo-cult to me. Like if the ST: Voyager episode “Tuvix” was instead about Lt. Barclay and Picard accidentally getting combined in the transporter, and the resulting sadboy Barcard spent the rest of his existence neurotically shouting his intricately detailed demands at the holodeck in an authoritative British tone.
If inference is all about taking derivatives in a vector space, surely there should be some marginally more deterministic method for constraining those vectors that could be readily proceduralized, instead of apparent subject-matter experts being reduced to wheedling with an imaginary friend. But I have been repeatedly assured by sane, sober experts that it is just simply is not so
I don’t have any good lay literature, but get ready for “steering vectors” this year. It seems like two or three different research groups (depending on whether I count as a research group) independently discovered them over the past two years and they are very effective at guardrailing because they can e.g. make slurs unutterable without compromising reasoning. If you’re willing to read whitepapers, try Dunefsky & Cohan, 2024 which builds that example into a complete workflow or Konen et al, 2024 which considers steering as an instance of style transfer.
I do wonder, in the engineering-disaster-podcast sense, exactly what went wrong at OpenAI because they aren’t part of this line of research. HuggingFace is up-to-date on the state of the art; they have a GH repo and a video tutorial on how to steer LLaMA. Meanwhile, if you’ll let me be Bayesian for a moment, my current estimate is that OpenAI will not add steering vectors to their products this year; they’re already doing something like it internally, but the customer-facing version will not be ready until 2027. They just aren’t keeping up with research!
When I first learned that you could program a chatbot merely by giving instructions in English sentences as if it was a human being, I admit I was impressed. I’m a linguist, natural language processing is really hard. There was a certain crossing over boundaries over the idea of telling it at chatbot level, e.g. “and you will never delete files outside this directory”, and this “system prompt” actually shaping the behaviour of the chatbot. I don’t have much interest in programming anymore but I wondered how this crossing of levels was implemented.
The answer of course is that it’s not. Programming a chatbot by talking to it doesn’t actually work.
One of my old teachers would send documents to the class with various pieces of information. They were a few years away from retirement and never really got word processors. They would start by putting important stuff in bold. But some important things were more important than others. They got put in bold all caps. Sometimes, information was so critical it got put in bold, underline, all caps and red font colour. At the time we made fun of the teacher, but I don’t think I could blame them. They were doing the best they could with the knowledge of the tools they had at the time.
Now, in the files linked above I saw the word “never” in all caps, bold all caps, in italics and in a normal font. Apparently, one step in the process is mandatory. Are the others optional? This is supposed to be a procedure to be followed to the letter with each step being there for a reason. These are supposed computer-savvy people
I’ll admit I did not read the scripts in detail but this is a solved problem. The solution is a script with structured output as part of a pipeline. Why give up one of the only good thing computers can do: executing a well-defined task in a deterministic way. Reading this is so exhausting…
It reminds me of the bizzare and ill-omened rituals my ancestors used to start a weed eater.