404 is iffy, but this isn't something that should be necessary if the network is recovering the algorithm(s) that generated the data in the first place. Apply this to text data, etc etc.

Good that it's happening, bad that these standards are being developed after agents are being made available on unclassified networks. www.diu.mil/work-with-us...

Becoming a parallel reality where the coding heavy people I follow on here are constantly talking about the void they're stepping into while almost nobody I talk to off this platform has LLMs on their mind.

You, a cynic (hates technology): I am have concerns about trying to fully automate the economy. Me, an optimist (bravely): I think we *should* fully automate the economy so we can do whatever we want 😌💪

Interesting counterpoint to the arguments that LLMs capture linguistic complexity: "..model behavior is mainly driven by the richness of linguistic resources rather than by sensitivity to grammatical complexity, (thus) resembles human (language) only superficially." doi.org/10.1371/jour...

Posting a screenshot of an LLM response saying someone is wrong about something, but without the prompt included, is just poasting with extra steps.

Probably a bearish signal for OpenAI that Noam Brown said this in his post about DoD work a week ago. Senior researcher, extremely talented, getting involved in "policy." x.com/polynoamial/...

Had a meeting at work recently where it was said that, because there's so much happening in the field right now, it's better to have a few public-facing pieces that say something interesting, than to have something new every day.

I know the quotes on this are dunks on agents, but those aside, surely stuff like this makes some engineers less worried about the scale of possible displacement?

Always worth a reminder that some People In The Know - pundits etc - were *very* confident that Trump had a specific rationale for what he did in Venezuela that would not carry over to Iran.

A brief intermission from poasting: I wrote an addendum. Have you ever heard someone say: "Your experience of free will is an illusion because all human behavior is pre-determined?" This one is about that, and why it's bad. vincentcarchidi.substack.com/p/addendum-e...

Yesterday, I was a software engineer. Today, I am an agriculture expert. Tomorrow, I will be whatever the AI Discourse needs me to be.

I might need to do a short, 500-word addendum to this on the idea that, if you're trying to pin down relevant distinctions between human and machine behavior, it is impermissible to make observations about one's internal states.

Consciousness is just a skill issue. I haven't had subjective experience since I started using Claude Code

"We collated emerging evidence to support our position that intermediate tokens are not guaranteed to have any end user semantics, and that their interpretability and solution accuracy are often at loggerheads..." arxiv.org/abs/2504.09762

The AI consciousness discourse just ain't it.

I just never really cared for the GTA games. Any of them

Aha! I see you fell for my AI-generated text. Now you have to accede to all my premises.

I don't want to press the points in these screenshots too firmly out of context, but the increasing reliability of LLMs (in specific domains) - now self-evident for many, and long-sought - comes to the detriment of arguments that LLMs are not part of long lineage of what constitutes a "machine."

LLMs help people improve their writing, I say, as the construction "X isn't just Y; it's Z" shows up for the fifth time in the article I'm reading.

Yeah, this is never happening. Well, not anytime soon at least. There's too much other shit flying at the will at the same time to even pretend to know where AI's/LLMs' practical impact is going to begin and end. And more things just keep coming.

Back to more important topics. Here's why I think Claude has the potential to revolutionize the ancient art of warf-

Trying to avoid this, but if Trump actually attempted a draft (not sure anybody in his admin. actually said this btw), the plan would be rescinded faster than some of the other half-baked shit he's tried.

Studies show my homunculus knows when it's being evaluated

I did a quick search, didn't see anything relevant. But I wonder if the DoD firing and/or sidelining of a bunch of JAGs in 2025 is making a noticeable effect on US Iran strikes? Not an area I'm familiar with, but kind of surprised not to see any talk of it. Maybe it's not as relevant as I think.

If you, like me, are tired of seeing the same three thoughts about LLMs - or, also like me, tired of seeing lazy dismissals that use the words "dualism" and "consciousness" and "how do I know I'm not an LLM?" - this piece might be for you.

Well, it happened again. I started writing a short piece and became a long piece. People have questioned why I insist that human linguistic behavior is not part of a causal structure, whereas LLM behavior is. This piece provides justification. 🧵 vincentcarchidi.substack.com/p/the-obscur...

Denying that I'm conscious to own the skeptics

Step 1: Hypothesize that LLMs can do scary things Step 2: Ask an LLM how it feels about life Step 3: Notice that the chain of thought is a little fucky Step 4: Conclude that LLMs are liars and can do scary things This👏 is👏science

Just found this absolute gem of a paper. If I'm reading this right, the author is saying that if computers look like they're thinking, then it means *we* can *stop* thinking 🤯 academic.oup.com/mind/article...

At last, my report on AI safety is live. In a follow-up to a report published last year, I argue that the high-bar for AI policy is to make complex, SOTA models suitable for safety-critical domains - where failures during operation would cause direct harm. 🧵 newlinesinstitute.org/tech-econ-so...

I’ve for a long time been trying to pin down and articulate a feeling I have that we are completely wrong in seeing producing words as taking actions. Speech acts are a lie. Haven’t succeeded yet. But anyway, this is funny:

The higher effort version of this is that much of the research that purports to find human capabilities in LLMs neglects basic principles of scientific explanation. A 🧵out of personal interest, not so much about specific uses.

we want functional tools, not to have to offer prayers to the machine spirit to get excel to work

I *do not know* if large language models are conscious, but if you criticize Claude one more time, I *will* defend his honor.

FWIW, there's no substantive connection based on public info between the use of Claude and the elementary school strike in Iran, unless I missed some news (possible).

Surprised we're still talking about this guy. Still feel this way about him.

I actually 100% believe my writing across different venues improved significantly as a result of LLMs exposing how much repetition there is in a given field. I became much more conscious of it.

This is on my mind a lot, and I don't think the flip-side of what I'm saying can be dismissed: there really is a way to glean some kind of productivity gains or otherwise personal gains by using LLMs in this post-hoc fashion (basically figuring it out as one goes).

So Yudkowsky must be loving the Iranian strikes on data centers right

Wonder what Ilya's seeing these days

In light of Claude's use in the Iran strikes, does this make Amodei One Of The Good Ones or Barely Better Than Altman? Please register your hot takes below. They will be evaluated objectively. www.ft.com/content/97bd...

At last, my report on AI safety is live. In a follow-up to a report published last year, I argue that the high-bar for AI policy is to make complex, SOTA models suitable for safety-critical domains - where failures during operation would cause direct harm. 🧵 newlinesinstitute.org/tech-econ-so...

Jesus christ man archive.is/Bwy4p

Thinking about this old Geoffrey Pullum quote criticizing generative linguistics. And while I think it was sort of wrong in the direction it was pointing, I really like the framing. There's definitely something to it. www.lel.ed.ac.uk/~gpullum/EAC...

“The Bitter Lesson” is a fine little essay. Anyone in software/ML/AI should read it. But I see a lot of people run with a maximalist overstatement of it, essentially: “structure and domain knowledge are always a mistake”. This is historically dubious and theoretically ungrounded. A thread:

Swamp weather in Philly. Feel like a bag of bones