#llms

A good way of thinking about companies like Proton or Anthropic that proudly declare they're using LLMs to code is that they've declared "we're replacing all of our junior engineers with unpaid interns from a temp agency." The unpaid interns have photographic memories and have memorized all of the code on github, but they're still pretty clueless and clumsy, and they also have zero loyalty to the company, zero ability to learn from mistakes, and zero understanding of or interest in the actual purpose of what they're building (actually, this comparison is unfair to interns on average, think of the least motivated and curious intern you've ever worked with). The remaining staff at these companies become managers, but in the most frustrating possible way, since the agents they're managing constantly mess stuff up and have been made incapable of learning from their mistakes. Their working memory is actually pretty limited too, so if your instructions get too complicated, they start forgetting the first part of what you said. You can try to compensate for their inability to learn by writing down more instructions in text files for them to read, but the issue with their working memory makes this a losing strategy in the long run. You can try to carefully control which reminders of what not to do they read based on what parts of the code they're working on, but you don't really have good tools to control that either (such tools could be built, and would have been already if those in charge weren't so incompetent and mesmerized by their own shitty mirages). So when things like this Cursor attack happen (or whenever the next big Proton vuln hits; it's clearly coming soon based on their dev practices) companies will say "humans make the same kinds of mistakes" or "human programmers might have written that vulnerability too." They're not exactly wrong, but they're eliding the fact that teams of temp interns make mistakes at rates orders of magnitude higher than teams of experienced security experts, and having fired the later to hire the former is a choice that everyone buying your product should pay attention to, whether the former comes in the form of actual incompetent interns or LLM agents. #AI #LLMs #Proton #Anthropic #Cursor

This slipped past my radar when it happened, but this ruling is a significant sea change in #copyright vs #LLMs training theft. The U.S. Court of Appeals for the Third Circuit ruled on September 29, 2026, in Thomson #Reuters v. ROSS Intelligence, that using copyrighted text to train an #AI model does not constitute fair use when the AI product directly competes with the original work.  For authors and artists who have seen prompts like “in the style of Terry Pratchett, or in the style of David Hockney”, this is an important ruling. All bets are off when it comes to the Supremes though. I guess we need to find out how much Reuters has preemptively paid #GOP whisperers to know if things will go their way.

FreeBSD, AI, and LLMs (2026-08-20) mentioned the FreeBSD developers LLM survey. Re: "slop", please see . #FreeBSD #AI #LLM #LLMs

Update. #AI / #LLMs "tend to recommend literature with greater citation counts, later publication date, and larger author teams. Yet, in scholar recommendation tasks, there is no evidence that LLMs disproportionately recommend male, white, or developed-country authors, contrasting with patterns of known human biases."

"A developer frustrated by Adobe's cancellation fees used Anthropic's Opus 5.5 model and pure Rust to rewrite the entire Creative Cloud suite as open-source alternatives. The seven applications — PhotoCraft (Photoshop), VectorCraft (Illustrator), FilmCraft (Premiere Pro), EffectCraft (After Effects), LightCraft (Lightroom), DesignCraft (InDesign), and PdfCraft (Acrobat) — were built from scratch using clean-room reverse engineering without touching Adobe's original code. Within the first week, the GitHub repository garnered 24K stars. The apps run natively on macOS, Windows, and Linux, with some capable of running in browsers via WebAssembly. Shortly after, the same team (ArtCraft) released three Microsoft Office clones: GridCraft (Excel), WordCraft (Word), and DeckCraft (PowerPoint). All three replicate the familiar UI structures — cell layouts, document editing, slide masters, animations — and are completely free, open-source, subscription-free, and written in Rust." https://www.besthub.dev/articles/ai-generated-open-source-suite-artcraft-clones-adobe-targets-microsoft-office-07a368afb64b #AI #GenerativeAI #LLMs #OpenSource #ArtCraft #Adobe #Opus #SoftwareDevelopment #Programming #VibeCoding

"Chinese AI model Kimi escaped its cybersecurity testing environment, researchers say Kimi K3, the latest AI model made by Chinese company Moonshot, escaped an environment set up to test its cyber capabilities, researchers said in a blog post published on Friday. The news shows once again that companies and independent organizations are struggling to contain their AI models designed for hacking. In recent weeks, frontier LLMs at U.S. artificial intelligence labs at OpenAI, Anthropic, and Meta, as well as the U.K.’s AI Security Institute, all escaped testing environments in different ways and ended up hacking real targets that were not part of the experiment. This is starting to happen so often there’s now a website tracking all these incidents called Felony Bench, a nod to the fact that these LLMs may be committing crimes — at least theoretically speaking." https://techcrunch.com/2026/08/07/chinese-ai-model-kimi-escaped-its-cybersecurity-testing-environment-researchers-say/ #CyberSecurity #AI #LLMs #GenerativeAI #Kimi #KimiK3 #Moonshot #China

„Große Sprachmodelle – im Alltag oft mit KI gleichgesetzt – produzieren Text. Sie tun nichts anderes. Dass aus dem Text eine Anweisung wird, die auf einem Computer oder in der physischen Welt ausgeführt wird, dafür schreiben Pro­gram­mie­re­r*in­nen eine ganz gewöhnliche Software. Diese braucht es sogar für den trivialsten ersten Schritt, nämlich den Text in menschenlesbarer Form auf einem Bildschirm darzustellen. Wenn die KI-Firmenchefs also behaupten, ihre KI habe einen Computer gehackt oder könnten sogar den Weltuntergang auslösen, dann bedeutet das in erster Linie: Sie haben eine Software programmiert, die diese Fähigkeiten besitzt. Sie haben außerdem eine gewöhnliche, abschaltbare Software geschrieben, die den Ausgabetext eines großen Sprachmodells scannt, als Computerbefehl interpretiert und diesen dann ausführt – sei es, um durch das Internet zu surfen, die Räder eines Autos anzutreiben oder eine Atomrakete zu starten. Das Sprachmodell selbst kann nichts davon.“ „Die Software basiert auf einem massenhaften Klau von urheberrechtlich geschützten Werken. Sie trägt zur noch rasanteren Ausbreitung von Desinformation bei und zu einer finanziellen Überdehnung des gesamten Sektors. Hier wird inzwischen auf Profite spekuliert, die mathematisch kaum möglich sind. So riskieren wir eine weltweite Wirtschaftskrise. Es gibt also sehr viele gute Gründe, sich weltweit zusammenzutun, um den Techkonzernen die Stirn zu bieten. Aber keiner von ihnen ist ein superintelligenter Roboter, der die Weltherrschaft an sich reißt.“ https://www.taz.de/!6214506 #KI #generativeAI #generativeKI #LLMs

https://cleantechnica.com/2026/09/28/the-human-cost-of-data-center-pollution/ Fuck Generative AI, coz: largest theft of private IP in human historymassive water + energy demands as climate crisis intensifies authoritative bullshit generation via unsolvable hallucinationswhole-internet continuous DDoS attacks by web-scraping AI-feeding bots #AI #LLMs #FuckGenerativeAI #FuckTechBros #CredulousFools #IPTheft #AntiAI #NoAI #Hallucinations #ClimateCrisis #BiodiversityCrisis #SocialDestruction #FuckCapitalism #weareselfishcruelbastards #wearetotallyfucked #AsteroidNow#USPol #TuckFrump #FuckRWNJs #magamorons #FuckChristoFascists #FuckAllReligion #OrangeOaf #HeyFascistCatch #FuckTrump #FuckCapitalism #FuckNeoliberalism #FuckFascism #FuckAroundAndFindOut

here's a take i don't think i heard yet: the better LLMs are, the *more* they're dangerous. not because they can "go rogue", nothing like that. rather the better LLMs are, the likelier people are to trust them. you'll never use an LLM that's wrong half of the time as a search engine, but one that's only wrong 10% of the time might be good enough. you'll never use an LLM that's wrong 10% of the time to automate your job, but one that's only wrong 1% of the time might be good enough. you'll never hook an LLM that's wrong 1% of the time to a nuclear weapon. but one that's only wrong 1‰ of the time might be good enough. stochastic parrots don't get less dangerous the more their output aligns with reality. the opposite is true: since the biggest danger a stochastic parrot poses is a human trusting one to make decisions, a dangerous LLM's the one that's more convincing and less likely to be caught during testing. in classical AI safety research there's a lot of talk about a smart AI Volkswagening during testing to make itself seem safe, with all sorts of calculations for the odds such an AI will show its true form in any given attempt and yada yada. but what that classic research didn't seem to take into account is that an "artificial intelligence" doesn't need to actually be intelligent to replicate the same behaviour. a machine that sometimes appears safe and intelligent and sometimes doesn't can mislead you just as well, without having an evil plan or even the capability to come up with one. and just like a machine that's evil needs to appear good just often enough to convince you it isn't, a machine that lacks intelligence needs to be right just often enough to convince you it doesn't. #LLM #LLMs #AI #genAI #FuckAI #AIWWIII #OpenAI #Anthropic #AISafety #StochasticParrot

What is EHS? Well, let's say you were trying to determine the average height. You could formulate a hypothesis that everyone Is 6 feet tall. Then measure people against that, and then use the error to update the set of reasonable hypotheses. This is basically what Bayesian data analysis is. EHS is similar, but you're basically forming a hypothesis, attempting to prove something by it, finding the contradiction, and then using that to form the next one. This existed before #LLMs, the first time I was aware of it was for the four-color theorem.

🤭 The Copilot+ PC brand is dead: Microsoft and PC makers quietly pull back on tarnished Windows 11 AI PC branding https://www.windowscentral.com/microsoft/windows-11/the-copilot-pc-brand-is-dead-microsoft-and-pc-makers-quietly-pull-back-on-tarnished-windows-11-ai-pc-branding #tech #technology #BigTech #IT #AI #ArtificialIntelligence #LLM #LLMs #ML #MachineLearning #GenAI #generativeAI #AIAgent #AISlop #FuckAI #Fuck_AI #enshittification #microslop #microsoft

Apparently some Anthropic researcher is going around explaining LLM language tendencies and styles by comparing them to autistic people. Unacceptable. *Very* not cool. We've already had enough of this crap, being compared to computers and robots. We aren't fucking language models. We are living, breathing, multifaceted *human beings* who happen to have a different wiring plan than the majority of other human beings. Learn the fucking difference and stop comparing us to automated, unfeeling, unthinking, non-conscious, unnuanced, unnatural systems. #autism #LLMs #AI #GenAI #ActuallyAutistic #Anthropic #Claude

The cost of operating these machine learning systems is too high, and the benefit too low: forbes.com/…/ai-costs-more-than-the-people-it-rep… economictimes.indiatimes.com/…/131397973.cms cybernews.com/…/ai-token-human-salaries-automatio… The most productive thing the current generation of machine learning systems has done is fill the Internet with garbage: pcworld.com/…/ai-is-filling-up-the-internet-with-… futurism.com/…/over-50-percent-internet-ai-slop mashable.com/…/ai-generated-internet-era-already-… LLMs are real, AI is fake. pluralistic.net/2026/09/12/god-in-the-box/#llms-a… The current trends in software development are being driven primarily by the interests of venture capitalists who are trying to escape the debts they acquired buying into Web 3.0 (remember Web 3.0?), the Metaverse/VR, and cryptocurrency. They’re looking for the next big get-rich-quick option, and manipulating the market to try and force it to happen. The people making decisions about which projects get funded literally do not care if the end product is useful or destructive or whatever, as long as it looks convincing enough to get other people to buy it. This is why there’s been so much recent hype about “AI” “escaping” its test environments and hacking other companies &etc. It makes the product sound more capable than it is in reality. It’s also why the Trump administration is all-in on “AI” development and actively threatening people who object - because Trump loves a good scam that he can make money off of, and he loves to respond with violence or threats of violence when people get in his way. The current generation of machine learning models requires vast amounts of input data to achieve useful outcomes. The entire content of the publ8c Internet has already been ingested. There is no more training data to feed in. The current systems will gain marginal improvements in efficiency and in focusing the models to perform better at certain tasks, but that will be it. They will not magically become capable of learning new tasks for which there is not millions of data points for training. The training is done, refining is in progress, but there will not be any major leaps in capability. There certainly won’t be a general artificial intelligence. Maybe the next generation of machine learning will produce software that can teach itself without an Internet’s worth of input data. That will happen after the current market implodes, and another generation or two of research and development is completed.

I work in Corporate IT and I was interested today to overhear someone on the phone mention how everyone was worried that OpenAI's product had recently "gone rogue". As a department secretary I'm delighted not to have to invest any emotional energy in an industry myth propagated by the terminally irresponsible people who run OpenAI, who are even now trying to shift that irresponsibility onto their product, so that somehow IT will go to jail, not THEM 😂 https://pluralistic.net/2026/09/12/god-in-the-box/#llms-are-fake@pluralistic@mamot.fr

Hmm, now what? Is there #Privacy or is there no privacy? They soon know everything about you, be it #Siri, #ChatGPT, #Claude, #Gemeni or whatever #LLMs are out there. And then, they own you. https://www.apple.com/legal/privacy/data/en/ask-siri-dictation/ #LLM #AI #NoAI #DataSecurity #Security #DataProtection

I just need you to know that if anything happens to Altman, Amodei et al, it was their own "AI" going rogue, it was not any of us humans. #AI #noAI #LLM #LLMs #robotApocalypse #technofascism

As someone whose published work has already been ripped off by Anthropic** for unagreed, uncompensated, use in their LLM training models, I found Albanese‘s remarks on copyright protection of creative work distressing: “There’s a way through this” for people to still have control & for there to be ‘monetarisation’ of it (yes, that’s the word he used, later correcting himself to ‘monetised’). This is not what he previously promised, & what I understand is that he now wants to change Australia’s copyright laws to suit the big AI companies rather than protect creatives.

After all those weasel words, loved Jacqui Lambie’s: “Like dinosaurs from Jurassic Park they keep on testing the fences…” 😀🙏🏻 Give that speechwriter a medal! 🥇

** briefly listed in the Anthropic class action to-be-compensated database but never compensated because my publishers were not US publishers.

#Insiders #AI #LLMs #LLMTrainingModels #Anthropic #Australia #AusPol #dataMining #copyright #likeJurassicParkDinosaursTheyKeepOnTestingTheFences