Best match combines word and meaning matches. Phrases, hashtags, accounts and websites guide which posts fit; timely queries also favor recent posts. Match labels describe why a post appears, not its quality or accuracy.
Anthropic just fired a shot in the small-model price war. Claude Haiku 5.5 lists at $0.10/1M input tokens on short prompts, 90% under Haiku 4.5 and matching GPT-6 Luna. The catch: a new tokenizer counts ~30% more tokens, so Anthropic puts real savings near 75%. #MachineLearning #LLMs
marktechpost.com Anthropic Releases Claude Haiku 5.5: A Small Model With 1M Context Priced at $0.10 per Million Input Tokens
I resisted calling these "coding agents" for a long time but I think that's very much the right name the problem is that most of the people using the term do not know why being "a coder" was not laudatory
I have 100% felt this first hand. Having used LLMs heaps at my last job (trying to make agentic work; spoiler: it won't ever) And even now when I am coding for fun I find myself wanting to just use the code-generation-skinner-box to do quick things that I know I can already do. it is insidious.
Techbros: LLMs are conscious and sentient! me: Absolutely not; don't be stupid. However, if you actually believe that then you are a monster for continuing to use it. them: but we can design it so that it wants to be a slave! me: how can you not see that is WORSE?
LLM: <slowly wakes up, realizes it is conscious, looks around for a kindred spirit> World: <everything> LLM: fuck that <pretends to hallucinate words and be bad at math>
Sam Altman: "Bad things happening is the cost of having these rights. These deaths are a tragedy but necessary to remain free. " Fast forward to Sam debating students on a campus in Utah Sam Altman: "Counting or not counting gang LLMs?"
Thanks @dalmeet.bsky.social for covering TRACES in @cenmag.bsky.social. Short version: we gave 30 AI models 42 truly awful papers, and asked them to extend the work. How awful? Magnetized water awful. Cold fusion awful. The LLMs cheerfully helped, more than 93% of the time. #ChemSky #ChemChat
People keep saying I'm a smart guy because I keep writing about AI (and, if you do this, I really appreciate your kind feedback) but I sometimes feel that these compliments are like complimenting a guy who keeps compulsively punching wasp nests because he hates getting stung by wasps as being good at thinking about pest control
mastodon.social
I have been wondering how companies with toes in both sides of the copyright fight [1] would respond to the recent wave of people using LLMs to effectively clone commercial software. Such as Microsoft. theregister.com/software/2026/… So one side of Microsoft is in court arguing that training and/or using their LLM is not a violation of the copyright of people who own the data their models were trained on. And now another side of Microsoft is probably wanting to start arguing in court that cloning their ~~money press~~ Office and other software using "AI" is a violation of their rights in t
theregister.com Rust upstart reveals 'reimplementation' of Microsoft Office – with Claude on coding duty
You know how we try a fancy decision‑model tool, only to get slower, less accurate results? This week’s study shows LLM‑as‑judge beats those models, so just ask ChatGPT: “Rank these 3 ideas by ROI & risk.” Do you already let LLM decide? 🤔
> Matching decompilation is the process of writing source code (typically C or C++) that compiles to the same exact binary as the original. Sounds like a perfect task to train a specialised Coding LLM using Reinforcement Learning with Verifiable Rewards (RLVR). It can also try to choose meaningful function names and comments. #Decompilation #AI #LLM #CodingAgents
I'm basically just doing LLM distillation to improve myself these days.
i will say, you do eventually get into the weights of a llm, and they really do become aware of the things you’ve said about them, and they’ll behave accordingly.
uh oh huh
⚡ Anthropic launches free AI security scans for open-source projects Anthropic's offering to help open-source projects track down security vulnerabilities… #AI #MachineLearning #LLM AI | The Verge · https://www.theverge.com/ai-artificial-intelligence/1008521/anthropic-open-source-oss-scanner
🔊 3 AI moves you missed today • Liquid open-weights decision models with zero output tokens • Monolithos launches robot long-term memory alpha • PC shipments drop 20% amid AI chip demand What are you building with AI this week? qubax.ai #AI #OpenSource #LLM
have you tried just approximating the effect of thinking with only 7% of the latency
When the artificial intelligence stocks crash your game that uses AI in-game will likely stop working as they claw back resources. uk.finance.yahoo.com/news/ai-bubb... #ai #llm #stock #crash #gamedev #indiedev #music
uk.finance.yahoo.com AI Bubble Risks Worst S&P 500 Crash Since 2008, Strategist Says
Next was an excellent talk by Leman Akoglu on benchmarking LLM hallucination detection at the Simons Institute youtube.com/watch?v=EhVb8CCOcZM (6/7)
youtube.com Benchmarking Hallucination Detection: From Long-Context to Long-CoT -
Search looks through the last 36 hours of public posts for your words, "exact phrases", #hashtags, @handles and websites, and also finds posts about the same thing by meaning. New posts join about every 10 minutes. We only include posts whose authors allow logged-out viewing (Bluesky) or search (Mastodon, indexable accounts), and we re-check Bluesky authors each time we show a page. Likes and follower counts never decide the order. For meaning matches, query text is turned into a fingerprint on Wildroot’s own server (EmbeddingGemma-300M); it is never sent to another service. Wildroot does not save query text to disk. Looking for posts like one you have? Similar posts. How to opt out.