Haven’t read the source paper yet (apparently it came out two weeks ago, maybe it already got sneered?) but this might be fun: OpenAI admits AI hallucinations are mathematically inevitable, not just engineering flaws.

Full of little gems like

Beyond proving hallucinations were inevitable, the OpenAI research revealed that industry evaluation methods actively encouraged the problem. Analysis of popular benchmarks, including GPQA, MMLU-Pro, and SWE-bench, found nine out of 10 major evaluations used binary grading that penalized “I don’t know” responses while rewarding incorrect but confident answers.

I had assumed that the problem was solely technical, that the fundamental design of LLMs meant that they’d always generate bullshit, but it hadn’t occurred to me that the developers actively selected for bullshit generation.

It seems kinda obvious in retrospect… slick bullshit extrusion is very much what is selling “AI” to upper management.

Woke up to some hashtag spam this morning

AI’s Biggest Security Threat May Be Quantum Decryption

which appears to be over of those evolutionary “transitional forms” between grifts.

The sad thing is the underlying point is almost sound (hoarding data puts you at risk of data breaches, and leaking sensitive data might be Very Bad Indeed) but it is wrapped up in so much overhyped nonsense it is barely visible. Naturally, the best and most obvious fix — don’t hoard all that shit in the first place — wasn’t suggested.

(it also appears to be a month-old story, but I guess there’s no reason for mastodon hashtag spammers to be current 🫤)

One to watch from a safe distance: dafdef, an “ai browser” aimed at founders and “UCG creators”, named using the traditional amazon-keysmash naming technique and and following the ai-companies-must-have-a-logo-suggestive-of-an-anus style guide.

Dafdef learns your browsing patterns and suggests what you’d do next After watching you fill out similar forms a few times, Dafdef starts autocompleting them. Apply with your startup to YC, HF0 and A16z without wasting your time.

So… spicy autocomplete.

But that’s not all! Tired of your chatbot being unable to control everything on your iphone, due to irksome security features implemented by those control freaks at apple? There’s a way around that!

Introducing the “ai key”!

A tiny USB-C key that turns your phone into a trusted AI assistant. It sees your screen, acts on your behalf, and remembers — all while staying under your control.

I’m sure you can absolutely trust an ai browser connected to a tool that has nearly full control over your phone to not do anything bad, because prompt injection isn’t a thing, right?

(I say nearly full, because I think Apple Pay requires physical interaction with a phone button or face id, but if dafdef can automate the boring and repetitive parts of using your banking app then having full control of the phone might not matter)

h/t to ian coldwater

It is related, inasmuch as it’s all generated from the same prompt and the “answer” will be statistically likely to follow from the “reasoning” text. But it is only likely to follow, which is why you can sometimes see a lot of unrelated or incorrect guff in “reasoning” steps that’s misinterpreted as deliberate lying by ai doomers.

I will confess that I don’t know what shapes the multiple “let me just check” or correction steps you sometimes see. It might just be a response stream that is shaped like self-checking. It is also possible that the response stream is fed through a separate llm session when then pushes its own responses into the context window before the response is finished and sent back to the questioner, but that would boil down to “neural networks pattern matching on each other’s outputs and generating plausible response token streams” rather than any sort of meaningful introspection.

I would expect the actual systems used by the likes of openai to be far more full of hacks and bodges and work-arounds and let’s-pretend prompts that either you or I could imagine.

It’s just more llm output, in the style of “imagine you can reason about the question you’ve just been asked. Explain how you might have come about your answer.” It has no resemblance to how a neural network functions, nor to the output filters the service providers use.

It’s how the ai doomers get themselves into a flap over “deceptive” models… “omg it lied about its train of thought!” because if course it didn’t lie, it just edited a stream of tokens that were statistically similar to something classified as reasoning during training.

Um, actually it’s called a typestate.