Introducing the «AI Influence Level (AIL) v1.0» by Daniel Meissler
> A transparency framework for labeling AI involvement in content creation
https://danielmiessler.com/blog/ai-influence-level-ail
What do you think? Are you using it?
llm
Re: https://lipn.info/@mevenlennonbertrand/116997917683191056
To riff on this a bit from a very different field at a very different scale, as a pretext to share an anecdote:
When I was a graduate student I spent a fair amount of time (an embarrassing amount of time) trying to get a computer to learn how to play Tic-Tac-Toe (naughts and crosses) by self play, given minimal help. Why? Because it turns out to be significantly harder than you might think based on the simplicity of the game, and I was curious why. It's not that hard in the grand scheme of things, but it's weirdly harder than you'd (well, I'd) expect (1). So, I went about coding up a Tic-Tac-Toe implementation that could play as quickly as possible, since gameplay would be the inner loop of a search algorithm, tested it some, and set a self-play-based search I was working on loose. Before too long my search was logging data that suggested it had found at least one optimal player. I was excited because I had opinions about how to use self-play to guide strategy search and it seemed like I might have some data supporting those opinions (calling them "hypotheses" would be giving my past self too much credit).
After quite a bit of testing, I discovered that my Tic-Tac-Toe implementation had a bug: in a small set of configurations of the board, my code would flag two in a row as a win instead of three in a row. My "optimal players" turned out to exploit this fact, choosing moves to drive the game into that set of configurations, then "winning" with two in a row. Oops. (2)
I noticed the problem because some of my discovered players were beating a minimax player during post hoc testing, which I knew to be impossible. I was able to sort out why because Tic-Tac-Toe is small enough to exhaustively test: you can enumerate the < 6,000 board configurations (3), run a function or strategy on all of them, and analyze the output in a reasonable amount of time and without too much trouble. I was able to spot by eye what was wrong.
One might ask: why didn't you have unit tests to do this sort of thing? Well, after that experience I did do more routine testing. However, speaking more generally now, there's an infinite regress when it comes to checking whether or not an implementation of a set of rules is faithful to the intention of that set of rules. Barring some kind of theoretical guarantee you'd need some other implementation of the rules to perform that sort of test. How can you be certain that implementation is comprehensive and also doesn't have a bug? You sometimes can't, practically speaking, and sometimes you can't theoretically speaking either. When the domain becomes large and complicated enough that both the above features about Tic-Tac-Toe I took advantage of---a sure indicator grounded in theory, and the ability to comprehensively and satisfactorily test---fail to hold, you're a bit adrift. (4)
I'm not very familiar with Lean, but I do know a bit about type theory, and modern type checkers tend to be Turing complete. Formal verification of a Turing complete programming language is not easy. I understand that practical type systems are designed to make type judgments of interest decidable, but that does not mean they're easy nor that the rules are easy to implement faithfully or test. It seems to me to be a tough problem, which is one of many reasons why I've taken the claims about LLMs doing this or that math thing with a big grain of salt. (5) I've expected to see the analog of my two-in-a-row "optimal" Tic-Tac-Toe players emerge to much fanfare for awhile now, and I expect to see many more.
#AI #LLM #FormalizedMath #math #Lean #ComputerScience #LearningBySelfPlay #SelfPlay #TicTacToe
(1) Donald Michie reported his MENACE system, which was a rudimentary form of reinforcement learning, could learn a decent game of Tic-Tac-Toe within a few hundred games. Christopher Rosin reported his coevolution-based Tic-Tac-Toe learner, a variant of learning by self play, required tens of billions of games to achieve less competence. Arthur Samuel's system learned master-level checkers by self play faster than this.
(2) Debugging a system by wrapping a search or optimization algorithm around it is an underappreciated technique I've used many times. It's a bit like fuzzing with an incentive to cause trouble.
(3) When I said "given minimal help", I meant it: I didn't give the players knowledge of the board symmetries.
(4) Personally I think it's incumbent on computer science professionals to flag when this is the case, since that's something we're educated to know (at least in principle).
(5) "Under the assumption that the implementation of system Y is correct, a human-written formalization of problem X in system Y produced output that a team of humans verified could be adapted into a proof of theorem Z" would grab fewer headlines than "AI proves Z", though.
To riff on this a bit from a very different field at a very different scale, as a pretext to share an anecdote:
When I was a graduate student I spent a fair amount of time (an embarrassing amount of time) trying to get a computer to learn how to play Tic-Tac-Toe (naughts and crosses) by self play, given minimal help. Why? Because it turns out to be significantly harder than you might think based on the simplicity of the game, and I was curious why. It's not that hard in the grand scheme of things, but it's weirdly harder than you'd (well, I'd) expect (1). So, I went about coding up a Tic-Tac-Toe implementation that could play as quickly as possible, since gameplay would be the inner loop of a search algorithm, tested it some, and set a self-play-based search I was working on loose. Before too long my search was logging data that suggested it had found at least one optimal player. I was excited because I had opinions about how to use self-play to guide strategy search and it seemed like I might have some data supporting those opinions (calling them "hypotheses" would be giving my past self too much credit).
After quite a bit of testing, I discovered that my Tic-Tac-Toe implementation had a bug: in a small set of configurations of the board, my code would flag two in a row as a win instead of three in a row. My "optimal players" turned out to exploit this fact, choosing moves to drive the game into that set of configurations, then "winning" with two in a row. Oops. (2)
I noticed the problem because some of my discovered players were beating a minimax player during post hoc testing, which I knew to be impossible. I was able to sort out why because Tic-Tac-Toe is small enough to exhaustively test: you can enumerate the < 6,000 board configurations (3), run a function or strategy on all of them, and analyze the output in a reasonable amount of time and without too much trouble. I was able to spot by eye what was wrong.
One might ask: why didn't you have unit tests to do this sort of thing? Well, after that experience I did do more routine testing. However, speaking more generally now, there's an infinite regress when it comes to checking whether or not an implementation of a set of rules is faithful to the intention of that set of rules. Barring some kind of theoretical guarantee you'd need some other implementation of the rules to perform that sort of test. How can you be certain that implementation is comprehensive and also doesn't have a bug? You sometimes can't, practically speaking, and sometimes you can't theoretically speaking either. When the domain becomes large and complicated enough that both the above features about Tic-Tac-Toe I took advantage of---a sure indicator grounded in theory, and the ability to comprehensively and satisfactorily test---fail to hold, you're a bit adrift. (4)
I'm not very familiar with Lean, but I do know a bit about type theory, and modern type checkers tend to be Turing complete. Formal verification of a Turing complete programming language is not easy. I understand that practical type systems are designed to make type judgments of interest decidable, but that does not mean they're easy nor that the rules are easy to implement faithfully or test. It seems to me to be a tough problem, which is one of many reasons why I've taken the claims about LLMs doing this or that math thing with a big grain of salt. (5) I've expected to see the analog of my two-in-a-row "optimal" Tic-Tac-Toe players emerge to much fanfare for awhile now, and I expect to see many more.
#AI #LLM #FormalizedMath #math #Lean #ComputerScience #LearningBySelfPlay #SelfPlay #TicTacToe
(1) Donald Michie reported his MENACE system, which was a rudimentary form of reinforcement learning, could learn a decent game of Tic-Tac-Toe within a few hundred games. Christopher Rosin reported his coevolution-based Tic-Tac-Toe learner, a variant of learning by self play, required tens of billions of games to achieve less competence. Arthur Samuel's system learned master-level checkers by self play faster than this.
(2) Debugging a system by wrapping a search or optimization algorithm around it is an underappreciated technique I've used many times. It's a bit like fuzzing with an incentive to cause trouble.
(3) When I said "given minimal help", I meant it: I didn't give the players knowledge of the board symmetries.
(4) Personally I think it's incumbent on computer science professionals to flag when this is the case, since that's something we're educated to know (at least in principle).
(5) "Under the assumption that the implementation of system Y is correct, a human-written formalization of problem X in system Y produced output that a team of humans verified could be adapted into a proof of theorem Z" would grab fewer headlines than "AI proves Z", though.
Edited 51d ago
I'm in a #github internal group for high-profile FOSS projects (due to @leaflet having a few kilo-stars), and the second most-wanted feature is "plz allow us to disable copilot reviews", with the most-wanted feature being "plz allow us to block issues/PRs made with copilot".
Meanwhile, there's a grand total of zero requests for "plz put copilot in more stuff".
This should be significative of the attitude of veteran coders towards #LLM creep.
Speaking of #CES, you may have noticed my coverage is very thin this year. There's a reason for this: I'm doing my level best to *not* give the oxygen of publicity to large language models and related "AI" tech this year.
An #LLM is not #AI. It will never be AI, no matter how big. Its output is statistical mediocrity at best, confident falsities at worst. The only ones worth using are trained on stolen data. Their environmental damage is staggering and growing, as is their mental impact.
Edited 251d ago
I am bigoted against LLMs.
I know it and it is how I deal with the hype.
Meaning I can't constantly be reasonable and
figure out what is meant and the context of LLMs
and what people think about them.
It drains my brain power from other things.
And I have continually found nothing compelling.
Worse, I have typically found very frustrating
examples of people using very strong but implied
assumptions and using logic that depends utterly
on using blinders and ignoring reason.
Until the hype dies, I am not interested in them.
I am still interested in the old AI stuff like
for example path finding, NNs, and markov chains.
#butlerianjihad #LLM #hype
I know it and it is how I deal with the hype.
Meaning I can't constantly be reasonable and
figure out what is meant and the context of LLMs
and what people think about them.
It drains my brain power from other things.
And I have continually found nothing compelling.
Worse, I have typically found very frustrating
examples of people using very strong but implied
assumptions and using logic that depends utterly
on using blinders and ignoring reason.
Until the hype dies, I am not interested in them.
I am still interested in the old AI stuff like
for example path finding, NNs, and markov chains.
#butlerianjihad #LLM #hype
Edited 246d ago
A mere week into 2026, OpenAI launched “ChatGPT Health” in the United States, asking users to upload their personal medical data and link up their health apps in exchange for the chatbot’s advice about diet, sleep, work-outs, and even personal insurance decisions.
This is the probably inevitable endgame of FitBit and other "measured life" technologies. It isn't about health; it's about mass managing bodies. It's a short hop from there to mass managing minds, which this "psychologized" technology is already being deployed to do (AI therapists and whatnot). Fully corporatized human resource management for the leisure class (you and I are not the intended beneficiaries, to be clear; we're the mass).
Neural implants would finish the job, I guess. It's interesting how the tech sector pushes its tech closer and closer to the physical head and face. Eventually the push to penetrate the head (e.g. Neuralink) should intensify. Always with some attached promise of convenience, privilege, wealth, freedom of course.
#AI #GenAI #GenerativeAI #LLM #OpenAI #ChatGPT #health #HealthTech
Edited 244d ago
This is quite a piece
https://www.planetearthandbeyond.co/p/reality-is-breaking-the-ai-revolution
I appreciate videos like this one from Nature that collect expert viewpoints, but sometimes the experts should be challenged.
Jared Kaplan of Anthropic had some very misleading claims.
LLMs do not democratize access to expertise. It feels like that because they sounds like an expert, but only when you ask them questions in domains you don't know. Really, they're just making shit up, and you don't notice in areas you're not an expert in.
LLMs will not solve open problems in STEM. Researchers may use machine learning tools to do that, but ML is for finding patterns in data. It can't "solve" or make "insights." It only applies when we already have vast amounts of the right kind of data.
And if we want to talk about LLMs as a cybersecurity threat, we should talk about how vulnerable they are to attackers. Imagining a genius AI hacker is nothing more than a distraction!
started listening The Tucker Carlson Show - Sam Altman on God, Elon Musk and the Mysterious Death of His Former Employee
absolutely epic episode
#LLM #AI
iPadを何とかできるのは、うちの艦長くらいですよ
Apple's Cheap MacBook: What to Expect in 2026 https://www.macrumors.com/2025/11/07/low-cost-macbook-rumors-2026/
Since I'm job and work hunting I tend to see the absurd new job titles that are bouncing around in the tech sector. The latest, which I've seen twice today, is "artificial general intelligence engineer" or some permutation thereof. I do my best to spend the minimum possible time on these and have no guess about whether they're legitimate.
#AI #GenAI #GenerativeAI #LLM #AGI
#AI #GenAI #GenerativeAI #LLM #AGI
I put the text below on LinkedIn in response to a post there and figured I'd share it here too because it's a bit of a step from what I've been posting previously on this topic and might be of some use to someone.
In retrospect I might have written non-sense in place of nonsense.
If you're in tech the Han reference might be a bit out of your comfort zone, but Andrews is accessible and measured.
In retrospect I might have written non-sense in place of nonsense.
If you're in tech the Han reference might be a bit out of your comfort zone, but Andrews is accessible and measured.
It's nonsense to say that coding will be replaced with "good judgment". There's a presupposition behind that, a worldview, that can't possibly fly. It's sometimes called the theory-free ideal: given enough data, we don't need theory to understand the world. It surfaces in AI/LLM/programming rhetoric in the form that we don't need to code anymore because LLM's can do most of it. Programming is a form of theory-building (and understanding), while LLMs are vast fuzzy data store and retrieval systems, so the theory-free ideal dictates the latter can/should replace the former. But it only takes a moment's reflection to see that nothing, let alone programming, can be theory-free; it's a kind of "view from nowhere" way of thinking, an attempt to resurrect Laplace's demon that ignores everything we've learned in the >200 years since Laplace forwarded that idea. In that respect it's a (neo)reactionary viewpoint, and it's maybe not a coincidence that people with neoreactionary politics tend to hold it. Anyone who needs a more formal argument can read Mel Andrews's The Immortal Science of ML: Machine Learning & the Theory-Free Ideal, or Byung-Chul Han's Psychopolitics (which argues, among other things, that this is a nihilistic).
We're going to put together a resource for musicians on how they can have album art without using large language models. Do you have any resources you'd like to see in that? Free image libraries? Friends who can help do design? Basic how-tos?
#mutualaid #solidarity #helpingeachother #mirlo #resources #ai #llm
#lispyGopherClimate #lisp #technology #podcast #archive, #climate #haiku by @kentpitman
https://communitymedia.video/w/c3GdAXe7BQTbK3VrcXCm7E
& @ramin_hal9001
On the #climate I would like to talk about the company that found #curl and #openssl's #deeplearning many (10ish) 0-day vulns "using #ai ". (#llm s were involved).
This obviously relates to my #lisp #symbolic #DL https://screwlisp.small-web.org/conditions/symbolic-d-l/ (ffnn equiv). Thanks to everyone involved with that so far.
I implemented that using #commonLisp #condition handling viz KMP.
Edited 225d ago
Compare and contrast
This:
and this:
This:
In the year of the city 2274, the remnants of human civilization live in a sealed city beneath a cluster of geodesic domes, a utopia run by computer. The citizens live a hedonistic lifestyle, but when they turn 30 must enter the "Carrousel", a public ritual that destroys their bodies, under the pretense they would be "Renewed" or reborn.
and this:
In the year of the city 2274, the colony of human beings on Mars live in a sealed city beneath a cluster of geodesic domes, a utopia run by generative AI. The citizens live a hedonistic lifestyle, but when they turn 30 must enter the "Cloud", a public ritual that destroys their bodies, under the pretense their consciousness would be uploaded to a computer and live forever.
#lispyGopherClimate #live #technology #podcast hosted by @ramin_hal9001 https://tilde.town/~ramin_hal9001/articles/ai-cannot-live-up-to-the-hype.html this week!
https://communitymedia.video/w/sBNPeWFJ7NuDVCjkAv7fkX
My guess as to the topic from the main show toot which is
https://fe.disroot.org/objects/88f34eba-ccaf-407e-8505-8a2a3bd1613b <- visit
@kentpitman 's poem
Repudiating the #ai #llm hype!
And #environmental damage #climate
#commonLisp #lisp #typeTheory #coalton common lisp static typing DSL https://coalton-lang.github.io/ ,
#ELS 2025 talk by Robert Smith. Robert emailed a note.
Edited 218d ago
Good AI research should tell us something about life, or it should help people. I hate seeing research about automating what people do. It's not a good goal for science or society! I was recently reminded of this by a paper applying LLMs to math.
This domain has many good questions: what do we mean when we say a person "solves math problems"? What are they actually doing? How is this like or not like what an LLM does? How might mathematicians benefit from this?
Instead, we get papers that pit an LLM against a human on a math problems dataset. This is great for claiming "AI has superhuman math abilities now!", but it's debatable whether good answers in a test-taking environment have anything to do with logic, reasoning, or creative problem solving. Instead of exploring to what extent LLMs are "really intelligent" vs. "stochastic parrots" (and perhaps the same question for humans), it reduces everything down to a number, one that hides the deeper problem and seems far more definitive than it is.
#ai #llm
Edited 214d ago
I hate OpenAI but I had to use Whisper to help someone make accessible content. I hate that I had to use Whisper to do it because it comes from OpenAI.
But I don't know of any other way to get a text transcription from a media file that is free/open. (Besides doing it manually.)
I tell myself because it's for education and accessibility it's okay, but I still don't like it.
#TIL that HDD prices also increased (because of "AI") like the SSD/memory prices before. I was slowly building my home NAS, thinking that if I use HDD, not SSD, then I should race only with the possibility of Internet shutdown in my country and hope that I'll build the machine before the Internet will become unusable or completely turned off, and I will be able to preserve at least something from human knowledge and creativity. But … looks like for now I should race not only with censorship, but also with fucking "AI" corporations 
The worst timeline ever, I never saw how the price of something was decreased in my life. When I moved to my city at 2008, the bus fare was near 18 roubles. For now it is 88 roubles — the 389% rise. Fuck this shit.
I think, one time the novel "Walkaway" from Cory Doctorow will become not a novel but a real-life. Because there are no mass-adopted solution for now — you could choose only in between "you own nothing and be happy" and "you own nothing and welcome to GULAG" 
. So the good enough solution — not to participate in that circus at all and walk away.
OpenClaw founder Steinberger joins OpenAI, open-source bot becomes foundation
From https://www.reuters.com/business/openclaw-founder-steinberger-joins-openai-open-source-bot-becomes-foundation-2026-02-15/
Everything I've read about OpenClaw suggests it's the NFT of AI. These folks need the fiction that AI is approaching "consciousness", or at least "agency", to continue.
#AI #GenAI #GenerativeAI #LLM #AgenticAI #VibeCoding #OpenAI #OpenClaw
From https://www.reuters.com/business/openclaw-founder-steinberger-joins-openai-open-source-bot-becomes-foundation-2026-02-15/
Everything I've read about OpenClaw suggests it's the NFT of AI. These folks need the fiction that AI is approaching "consciousness", or at least "agency", to continue.
#AI #GenAI #GenerativeAI #LLM #AgenticAI #VibeCoding #OpenAI #OpenClaw
ICYMI: some dork used #Claude #LLM to 'fix' an issue in #rsync and broke it (v3.4.3).
#GenAI is shit and shld not be used for code that needs to work, even if you have the morals of a billionaire skunk.
Lots of critical tasks, such as backup 🤦♂️ will be broken as people update their #Debian #Linux thanks to this stupid obsession and psychosis.
How much will that cost? On top of all the costs inflicted on people who never touch the stuff and don't want to run anything built with it, like me.
看到这句😂
我用象棋软件或者围棋软件可以下赢任何一个人类的世界冠军，但我不会满大街的说我是下棋高手。#LLM
https://t.me/hyi0618/10750
Gente de #Mastodon:
Me encanta el término que usan @alex y @emilymbender sobre los #LLM > "Text extruder"
Se os ocurre una traducción al español?
@javisamo @ana_valdi @efialto
@CorioPsicologia
Extruidor de textos me suena raro
Edited 108d ago