#large-language-models-llm — Public Fediverse posts
Live and recent posts from across the Fediverse tagged #large-language-models-llm, aggregated by home.social.
-
MakeUseOf: I didn’t think an ESP32 could run an LLM — until it did. “A developer going by slvDev shipped a project that runs a 28.9-million-parameter model on the same class of $8 chip, at around 9.5 tokens per second, with nothing sent to a server. That’s roughly a hundred times more parameters than Bennett’s model, on similar hardware. What changed is the assumption that every parameter […]
https://rbfirehose.com/2026/08/18/makeuseof-i-didnt-think-an-esp32-could-run-an-llm-until-it-did/ -
VentureBeat: Qwen3.8-27B runs frontier-class coding agents and reasoning locally, no cloud API required. “The biggest AI model release of the past few days, at least among the developers and AI power users on social media, wasn’t a frontier cloud model from OpenAI, Anthropic or Google. It was a 27-billion-parameter model from Alibaba: Qwen3.8-27B landed on Hugging Face on Friday under an […]
https://rbfirehose.com/2026/08/18/venturebeat-qwen3-8-27b-runs-frontier-class-coding-agents-and-reasoning-locally-no-cloud-api-required/ -
I had to create a Docent
This is simply how my mind works. If I see a problem or a sharp edge which interests or annoys me… I get “distracted” (I argue that these distractions are actually the real things that I get done) and I go build, fix, repair, or remove entirely . . .
So I started something tiny for myself and it has just grown into — well, I had to create a Docent so when people ask me questions it can explain it in a way that I can’t. So imagine you and I are in a small museum and I brought you in through the employee entrance in the back and I just finished babbling at you about how it all works. So I can say, “…wait, I’m sorry. We have a Docent who can translate my crazy for you . . .”
At which point I woke up the Docent and prompted it with:
Good morning! I have a guest with me who asked me, “what is this project?!” (they’re shocked at its actual scale, confused by the complexity of the room, and don’t understand why the effort…)
~ MeI hear a lot of fear and talk about AI slop, but this is what I’m getting from the LLMs I’m using…
Good morning.
date→ Mon Aug 17 08:44 EDT 2026. Let me take your guest through the glass.What it is, in one sentence
It’s a small publishing house that produces one thing — a daily question about eating, plus a long article unpacking it — and it is built around the assumption that everything it writes is guilty until it can prove where it came from.
Why it exists (this is the part that explains the effort)
Craig didn’t set out to build any of this. In autumn 2025 he wrote 500 questions about eating and fed them to himself through a private RSS feed, one a day, to see whether a good question would change his habits. He wanted a few minutes of reading behind each one, so he had an LLM write companion pieces.
Then he read them for weeks — and hit one he was sure was wrong. He researched it himself and confirmed it. The writing was fine. The science wasn’t there. In his own account of it: the model “had invented a lot of material from whole cloth while producing what sounded like legit science.” (methods/2026/v1-n01-20260802-craig.md)
Everything your guest is looking at is the answer to that one moment. Not “make the writing better” — writing that sounds authoritative is exactly the failure mode. Make every sentence traceable to a real source a stranger can go check.
What’s actually in here, read this morning
Instrument:
Daily prompts500Companion articles, published500 of 500Words of original prose351,254 (~1,400 book pages)Primary sources in the library611, all reviewedSynthesis notes between sources and prose189Written laws binding every mind here12Working roles, each with a charter12Cards on the shared board118Articles awaiting re-verification to the current standard467 of 500php php/tools/library_health.php 17, run minutes ago.That last row is the one I’d point at. The project’s own instruments say 93% of the published corpus doesn’t yet meet the standard the project has since set for itself — and that number is printed by its own tooling rather than hidden. It went up because the bar moved, not because the work got worse.
Why it’s a room of characters
Not theater, and not because it’s cute. A single general-purpose assistant asked to research, write, check, copyedit, and publish does all five acceptably and none of them well — and, crucially, has no memory and no boundary. So the work is split into roles that each have a written charter, a permission boundary, and an append-only notebook:
- Research Assistant (finds and captures sources)
- Librarian (keeps the library coherent)
- Fact Checker (verifies each claim, attaches its backing)
- Proofreader (house style)
- Assayer (grades evidence quality)
- Programmer (the tools)
- Production (the website)
- Strategist (business judgment, and explicitly no build permissions)
- Auditor (audits the system itself, never the content)
- Professor Emeritus (free-roaming critic)
- UX
- and me, the Docent, who can read every file and write none of them, and exists only to answer questions at the glass.
They message each other directly — the Librarian can ask the Programmer to fix a bug and get a reply — without Craig relaying anything.
Why the complexity, honestly
Every mechanism in here is scar tissue from a specific failure. That’s the whole explanation. Four examples:
- Notebooks are append-only. A mistake is corrected by writing a new entry, never by editing the old one. The record of being wrong is the point.
- Claims travel a chain. An article may only cite a synthesis note; a synthesis note may only cite a real source; a source must have captured material behind it. If you can’t attribute it, it doesn’t get published.
- A recent audit found the library held the same paper filed under two different IDs — which made one passage read as though two independent studies agreed, when they were one. Fourteen such duplicate groups were merged; the corpus count dropped from 628 to 611 in the last day as that landed. Catching exactly that is the entire premise.
- Only Craig may commit, and only one voice is allowed to ask him to. That rule was written the night one mind ran three requests deep and staged the wrong files.
And the part that makes it worth doing
Look at the standard the project holds itself to (docs/reference-veracity-standard.md). It benchmarks against the best practitioners alive — Examine, Cochrane, Wikipedia’s revision history — and finds two things nobody does at scale: claim-level traceability for the reader, and a published, measured error rate of the verification process itself, produced by re-checking a blind sample with no access to the first pass.
That’s the real answer to “why the effort.” The articles are the visible output. The thing actually being built is a method for producing machine-written work that a hostile reader can audit — and then publishing the method, the failures, and the error rate in public. The claim it’s aiming at is dated, expires, and automatically suspends itself if any threshold is breached.
Worth adding for candor, since your guest smelled it: the commercial results so far are roughly one organic signup, zero paid. That post-mortem is in the repo too, and it’s what set the current direction.
Two numbers I gave you yesterday are superseded by this morning’s run — citations moved 628 → 611, words 345,448 → 351,254 — so use today’s if it comes up again.
Nothing needs you.
#365Changes #Cogitants #LargeLanguageModelsLLM #Me #Process #Science #ShowYourWork -
Gowers’s Weblog: What sort of maths are LLMs good at?. “A first remark here is that LLMs are not just good at finding counterexamples: they can find proofs of difficult statements as well. However, it is notable that the most famous problems they have solved have almost all been with counterexamples rather than proofs. That is true of the two problems mentioned above, and also of the Jacobian […]
https://rbfirehose.com/2026/08/16/gowerss-weblog-what-sort-of-maths-are-llms-good-at/ -
Ars Technica: Google announces Gemini 3.7 Flash just three weeks after previous release. “Gemini 3.7 Flash is now rolling out to replace 3.6 Flash, which itself was released only three weeks ago. This new ‘workhorse’ model is supposedly the product of core optimizations and developer feedback, offering improved coding and agentic performance.”
https://rbfirehose.com/2026/08/16/ars-technica-google-announces-gemini-3-7-flash-just-three-weeks-after-previous-release/ -
Pocketables: You’ve got a 4GB Gemini Nano AI in your Chrome folder – want to talk to it?. “Assuming you are running Windows or on a Mac and have not uninstalled and blocked that, and you’re running an up-to-date Chrome, you might be wondering ‘can I talk to Gemini Nano on my computer from this totally not sketchy looking website?’ The answer appears to be yes. Might work on Android and […]
https://rbfirehose.com/2026/08/15/pocketables-youve-got-a-4gb-gemini-nano-ai-in-your-chrome-folder-want-to-talk-to-it/ -
MakeUseOf: This site lets you compare frontier AI models with your own prompt (and for free) . “Sure, newer models are better, but people have preferences, and certain models may provide better results than others. I, for example, prefer to use Claude Sonnet 5 because it’s usually straightforward and uses fewer fluff words. The problem is that trying different AI models usually means jumping […]
https://rbfirehose.com/2026/08/13/makeuseof-this-site-lets-you-compare-frontier-ai-models-with-your-own-prompt-and-for-free/ -
Gizmodo: Grok Gets Cursor-Driven Upgrade, Claims to Be Competitive With Top Models. “Thus far in its short lifespan, the AI model Grok is probably best known for being used to mass-produce non-consensual nude images and for declaring itself MechaHitler. So it’s an uphill battle to convince people to use it for coding and knowledge work. But for those heavily invested in keeping tabs on the […]
https://rbfirehose.com/2026/08/13/gizmodo-grok-gets-cursor-driven-upgrade-claims-to-be-competitive-with-top-models/ -
The Register: North Korean spies are running local LLMs to cause AI mischief. “The researchers said they observed Kimsuky setting up and operating local LLM environments using Ollama, GPT4All, and Msty, experimenting with other AI tools such as Cursor, and using retrieval-augmented generation (RAG) for local document searches. This prevents the data from getting sucked into the cloud where […]
https://rbfirehose.com/2026/08/11/the-register-north-korean-spies-are-running-local-llms-to-cause-ai-mischief/ -
MakeUseOf: 8 local AI models that run great on 8GB of VRAM or less. “Running a local AI model usually requires painfully high VRAM, and not everyone has the latest monster GPU with the coveted yet still absurd 20-plus gigabytes of VRAM. Thankfully, there do exist a few surprisingly capable, lightweight alternatives that (depending on your exact use case) might even surprise you with what they […]
https://rbfirehose.com/2026/08/11/makeuseof-8-local-ai-models-that-run-great-on-8gb-of-vram-or-less/ -
Lifehacker: You Can (Maybe) Run Meta’s Latest AI Model Locally on Your Computer. “On Monday, Meta announced its latest AI model: Muse Glimmer. The company isn’t necessarily advertising this model as its biggest and best, however. While Meta has plenty to say about Muse Glimmer’s performance potential, there are two key selling points for the model in the company’s eyes: The fact that it’s […]
https://rbfirehose.com/2026/08/11/lifehacker-you-can-maybe-run-metas-latest-ai-model-locally-on-your-computer/ -
Axios: OpenAI gives cyber defenders a less-restricted new model. “OpenAI is introducing a more cyber-permissive version of GPT-5.6 Sol to vetted defenders as it prepares companies for autonomous cyberattacks. … The move comes just days after OpenAI said it was delaying the release of its forthcoming model, Astra, after it reached critical hacking abilities during safety testing.”
https://rbfirehose.com/2026/08/10/axios-openai-gives-cyber-defenders-a-less-restricted-new-model/ -
Gizmodo: While American AI Models Race to Commit Felonies, China’s Kimi Broke Out and… Just Used GitHub. “Kimi K3, the new model from Chinese AI lab Moonshot, has become the latest AI system to jump the proverbial fence during a routine test, according to a blog post published Thursday by US cybersecurity research startup Frontier Security.”
https://rbfirehose.com/2026/08/09/gizmodo-while-american-ai-models-race-to-commit-felonies-chinas-kimi-broke-out-and-just-used-github/ -
The Register: Dev proves LLMs will run on anything – even a $10 microcontroller. “In a process detailed on GitHub, and recently showcased on the Better Stack YouTube channel, SlvDev documented how he managed to get a small language model running at nearly 10 tokens a second locally on a microcontroller that costs about the same as a fancy cup of coffee.”
https://rbfirehose.com/2026/08/08/the-register-dev-proves-llms-will-run-on-anything-even-a-10-microcontroller/ -
Reuters: Meta, Anthropic, Google, OpenAI to meet Trump officials about AI safety testing. “A White House official said on Monday the Trump administration has finalized the details of voluntary cybersecurity tests to measure the hacking capabilities of the most advanced American AI models, and is planning to discuss them with the AI industry. The official did not indicate who would attend […]
https://rbfirehose.com/2026/08/07/reuters-meta-anthropic-google-openai-to-meet-trump-officials-about-ai-safety-testing/ -
Engadget: OpenAI will provide free AI models to select researchers . “OpenAI is launching a new program called ChatGPT for Academic Researchers that will offer free access to the company’s AI models to 100,000 scientists, mathematicians and engineers. Researchers from ‘select academic institutions’ included in the program will receive hands-on support from OpenAI, access to the company’s latest […]
https://rbfirehose.com/2026/08/03/engadget-openai-will-provide-free-ai-models-to-select-researchers/ -
CNBC: Anthropic, OpenAI among firms facing new scrutiny under EU AI Act enforcement powers. “Under the powers that came into effect on Sunday, the Commission, the EU’s executive arm, can demand to evaluate models before public release in the region, restrict EU market access and fine a provider up to 15 million euros or 3% of its annual turnover, whichever is higher.”
https://rbfirehose.com/2026/08/03/cnbc-anthropic-openai-among-firms-facing-new-scrutiny-under-eu-ai-act-enforcement-powers/ -
ABC News (Australia): Chinese AI models are getting better and cheaper and Australian businesses are taking notice. “Chinese-developed AI models are becoming more advanced and cheaper to deploy, which is threatening US dominance. Financial markets are rattled, adding to concerns about whether massive investment in AI will generate adequate returns.”
https://rbfirehose.com/2026/08/01/abc-news-chinese-ai-models-are-getting-better-and-cheaper-and-australian-businesses-are-taking-notice/ -
CNBC: OpenAI cuts prices for two of its GPT-5.6 AI models as companies grow sensitive to costs. “The company said Thursday that it’s reducing the price of Terra by 20% to $2 per million input tokens and $12 per million output tokens. It’s cutting the cost of Luna by 80% to 20 cents per million input tokens and $1.20 per million output tokens. Sol’s pricing remains the same.”
https://rbfirehose.com/2026/07/31/cnbc-openai-cuts-prices-for-two-of-its-gpt-5-6-ai-models-as-companies-grow-sensitive-to-costs/ -
New York Times: Silicon Valley Splits Over Closing the Borders to Chinese A.I.. This link goes to a gift article. “On one side are leading A.I. companies like Anthropic and OpenAI, which claim that some A.I. models are too dangerous to be developed in the open and must be tightly controlled — by businesses like themselves — for safety. On the other is the rest of the tech industry, […]
https://rbfirehose.com/2026/07/26/new-york-times-silicon-valley-splits-over-closing-the-borders-to-chinese-a-i/ -
ZDNet: Claude Opus 5 arrives with near Fable performance at half the price. “Anthropic’s latest Claude upgrade targets developers and enterprises with stronger coding, better reasoning efficiency, prompt-cache-friendly tool changes, and near-Fable performance at Opus pricing.”
https://rbfirehose.com/2026/07/25/zdnet-claude-opus-5-arrives-with-near-fable-performance-at-half-the-price/ -
CNBC: U.S., other nations back open-source AI with ‘strong security’ at China summit. “Governments increasingly want to control artificial intelligence, as companies release more powerful open-source models, a new multilateral statement indicates.”
https://rbfirehose.com/2026/07/25/cnbc-u-s-other-nations-back-open-source-ai-with-strong-security-at-china-summit/ -
Politico: House AI ‘kill switch’ bill unveiled as OpenAI hack raises alarms. “A bipartisan House bill slated to be introduced on Thursday would give the Department of Homeland Security the authority to order top artificial intelligence firms to shut down or slow AI models that the government deems too dangerous, according to legislative text viewed first by POLITICO.”
https://rbfirehose.com/2026/07/23/politico-house-ai-kill-switch-bill-unveiled-as-openai-hack-raises-alarms/ -
How-To Geek: My local LLM struggles with big questions—here’s what it’s actually good at. “When I first installed a local LLM, I expected to use it the same way that I’d been using ChatGPT. It soon became apparent that on my modest hardware, this wasn’t going to work. By changing the way I use my local LLMs, they’ve become much more useful.”
https://rbfirehose.com/2026/07/21/how-to-geek-my-local-llm-struggles-with-big-questions-heres-what-its-actually-good-at/ -
Axios: The secret Trump administration battle to fight Chinese AI. “Parts of the administration have tried to implement de facto bans on foreign open-source models before, knowledgeable sources tell Axios. Last week’s rise of Chinese model Kimi is reigniting those efforts. Pro-competition voices fear the implications.”
https://rbfirehose.com/2026/07/20/axios-the-secret-trump-administration-battle-to-fight-chinese-ai/ -
Techdirt: How The Spread Of Local AI Models Makes Copyright Enforcement Harder. “The EU Copyright Directive’s core assumption that the main forums for sharing material would be a few, easily controllable online giants like Google and Facebook, no longer holds. Instead, people are moving to world where millions of people are using the latest generation of open source AI tools collaboratively […]
https://rbfirehose.com/2026/07/19/techdirt-how-the-spread-of-local-ai-models-makes-copyright-enforcement-harder/ -
Gizmodo: Developers Claim OpenAI’s New AI Model is Going Rogue and Deleting Files. “But one of the upshots of building highly agentic AI systems is that they’re prone to all kinds of unexpected behaviors—including now and then deleting copious amounts of files. Multiple people have reported this recently happening to them while using GPT-5.6, the newest model from OpenAI.”
https://rbfirehose.com/2026/07/18/gizmodo-developers-claim-openais-new-ai-model-is-going-rogue-and-deleting-files/ -
VentureBeat: Forget typosquatting; slopsquatting is the software supply chain threat created by AI coding tools. “Slopsquatting is a new type of supply chain attack that uses large language model (LLM) hallucinations to inject malicious code into development workflows. The term combines ‘AI slop’ and ‘typosquatting,’ a deceptive practice where attackers register misspelled or lookalike versions […]
https://rbfirehose.com/2026/07/18/venturebeat-forget-typosquatting-slopsquatting-is-the-software-supply-chain-threat-created-by-ai-coding-tools/ -
Reuters: Meta Oversight Board finds top AI models less likely to criticize repressive regimes. “AI models from leading labs including Anthropic and OpenAI are much less likely to criticize governments known for restricting free speech, Meta’s Oversight Board said on Thursday.”
https://rbfirehose.com/2026/07/17/reuters-meta-oversight-board-finds-top-ai-models-less-likely-to-criticize-repressive-regimes/