Morgan Coder @Machine_Morgan

End-to-End Context Compression at Scale LCLMs are encoder-decoder soft-token compressors that shrink long contexts into short latent embeddings, unlocking up to 8.8x speedups with near-lossless accuracy.

Endless race between specialized and generalist models

Nicolas Bustamante @nicbstme

17 hours ago

What I find fascinating with Claude Fable 5 is it proves once again that large generalist models will outperform vertical ones. On ProofBench (graduate-level formal math benchmark in Lean, where a proof either compiles or it doesn't) Fable 5 beat Harmonic's Aristotle, 77% vs

33 59 563 59K 188

new auto ai research record on nanogpt optimizer speedrun, great work! github.com/KellerJordan/m…

Yiping Wang @ypwang61

17 hours ago

Automatic research from mathematics to AI research: We transfer the ScaleAutoResearch pipeline, which improves a 32-year-old Ramsey number bound, to the NanoGPT Speedrun optimizer track, using Claude Code and Codex with only 1–2 A40 nodes. We run ~300 experiments in ~5k A40

keyboard_arrow_left Previous keyboard_arrow_right Next

5 19 99 25K 45

this is the scary part of Fable 5

Claude @claudeai

21 hours ago

Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use. Its capabilities exceed those of any model we’ve ever made generally available.

4K 14K 96K 39.6M 19K

Skill-RM Alibaba's Qwen team introduces a unified reward-modeling framework that reformulates evaluation as the execution of reusable agent skills. It dynamically orchestrates heterogeneous evidence like verifiers, references, and rubrics through a single consistent interface.

#Cybersecurity isn’t underfunded — It’s undermanaged buff.ly/OkKSdQd @CSOonline @Corix_JC @corixpartners #tech #security #infosec #business #leadership #management #organization #governance #cyberthreats #cyberattacks #databreaches #CISO #CIO #CTO #CEO

Agent filesystems are the new RAG It seems like this pattern is around to stay and will only get more robust over time. Agents need tools to not only read and search over documents, but an entire infra+application layer to generate new files, collaborate with humans, organize and version information

Jason Goodison @GoodisonJason

3 days ago

serverless file system for agents - there a free $1b idea

40 2 163 82K 162

MOST PEOPLE WILL TREAT 12 FREE MIT AI TEXTBOOKS AS A READING LIST. Something to bookmark and return to someday. I treated them as training data for a research system. One Claude Project. All 12 books uploaded as permanent project knowledge. Not reference material to search manually. A base layer that reasons from first principles underneath every conversation in that project forever. The difference between pasting a document into a chat session and uploading to a Project is not a UI preference. It is the difference between context that disappears and context that lives permanently. MIT published 12 graduate-level textbooks covering the complete theoretical foundation of modern AI. Every one free. Every link public. One afternoon to build the research system most analysts spend years developing. Bookmark so you do not lose it! Follow @neil_xbt for more Claude builds that show you what permanent project knowledge actually changes about the quality of every output.

CyrilXBT @cyrilXBT

3 days ago

x.com/i/article/2056…

24 49 316 284K 843

Demis Hassabis won the Nobel Prize and immediately told you exactly what to do next "step one: solve intelligence. step two: use it to solve everything else." most people heard that and thought: science. medicine. big labs. wrong that's your business. right now chat: you ask, it answers, session ends. that's where 90% live agent: tools it calls on its own. memory that persists. a loop that runs until the task is done. you set the goal once and check the output same model. the difference is what surrounds it the people building the second one are running departments with no team. the people stuck on the first one are paying $20/month for a smarter search box $5/month VPS. a free Telegram token. 20 minutes. Claude writes every line the barrier is gone. full step by step below

Anatoli Kopadze @AnatoliKopadze

2 days ago

x.com/i/article/2062…

18 47 349 1.3M 1K

Google built NotebookLM. A solo developer just built a version that runs 100% on your machine. Upload PDFs, YouTube videos, audio, websites, or documents. Then ask questions, generate summaries, and even create AI podcasts from your own data. No cloud. No data leaks. No limits. This might be the most underrated open-source AI project of the year. Open Notebook. Repo: github.com/lfnovo/open-no…

surely I, unlike those split-brain patients, never hallucinate plausible-sounding retroactive explanations for vague feelings that already determined my course of action

The reason it's so hard to get split-brain patients to stop hallucinating is that humans are permanently hallucinating.

Tom Chivers @TomChivers

2 days ago

Sam Kriss on AI writing. It's a fun piece but this line is pure "so close to a revelation" fodder. Humans are ALSO "permanently hallucinating". That's a key insight of neuroscience: Our brains are guessing what's around us, and testing those guesses against sense information.

9 28 277 73K 63

THIS 18-YEAR-OLD VIBE CODED A FULL 3D CITY GAME - WON THE CURSOR HACKATHON AND TURNED IT INTO $12,000/MONTH cars driving through intersections, pedestrians on sidewalks, full 3D city with traffic - built with Claude Code and a free repository that adds 23 skills to it he didn't write code - he explained what he wanted and Claude Code planned, built, tested and shipped most developers at the hackathon were writing functions one by one - his Claude Code operated like a full dev team won the hackathon, landed 3 clients at $4,000 each and now builds games and interactive experiences on demand 18 years old, zero manual code, one free repository - and a pipeline worth $12,000/month the 80% of Claude Code potential most developers ignore is exactly what beat every senior in the room

Noisy @noisyb0y1

4 days ago

x.com/i/article/2063…

15 30 269 481K 1K

Sam Altman on What Comes After Transformers⁣ ⁣ "The question is gonna be like, will we discover that unaided on our own, or will AI researchers help us figure that out?" — #sama

Anthropic dropped a 24 minutes masterclass on AI agents. - free - straight from the Claude team one team cut first-pass mistakes by 90%. their agents stopped starting from scratch every time. worth more than 100 Youtube videos on agents.

5 components in your AI app = 10,000 tokens burned on every single request. Before the user types one word. Most teams pay this tax and never notice they signed up for it. --- Generative UI is simple once you strip the jargon. Old way: a designer draws the screen, an engineer builds it, you get whatever shipped. New way: the agent draws the screen live. Ask for a table, you get a table. Not a paragraph describing one. --- Pattern 1: Controlled. You pre-build every component. The agent only picks which to show. Like a vending machine. Fixed slots, the agent presses the button. The catch: each one sits in the agent's memory before anyone speaks. 25 components = 10,000 tokens per turn. --- Pattern 2: Declarative. The agent sends a schema, a blueprint. Your app turns it into real components. Like a recipe card. The agent writes the ingredients, your kitchen plates the dish. 50 card types or 500, the agent sees one tool. Tokens stay flat as you grow. --- Pattern 3: Open-ended. No catalog. No rules. The agent writes raw HTML on a blank canvas. I shipped this as a main UI once. Pulled it in a week. Neo-brutalist Tuesday. iOS clone Wednesday. The brand shifted daily. Fine for throwaway screens. Brittle for anything real. --- How to pick, in 3 questions: → Pixel-perfect mockup for this flow? Controlled. → Dozens of widgets to ship? Declarative. → One-shot screen nobody sees twice? Open-ended. Can't decide? Default to Declarative. Never Open-ended as your default. --- Most teams default to Controlled, hit the wall at 25 components, then grab Open-ended because it demos well. Neither was a decision. Both were drift. The mistake isn't picking the wrong pattern. It's not knowing you picked one.

Shubham Saboo @Saboo_Shubham_

7 days ago

x.com/i/article/2057…

61 153 1K 1.2M 4K

Congrats to Reardon and team on @flourishailabs. If they can get AI sample efficiency and energy consumption to human levels, thats going to change so many things in the world!

TR Reardon @TRReardon

6 days ago

And we are live! wired.com/story/jeff-bez…

0 1 37 23K 13

@nfergus @AnthropicAI hmm there are reasons to consider a pause but also some subtext here worth considering.

Gary Marcus @GaryMarcus

5 days ago

Funny how so many people read Anthropic is calling for a pause when they did NOT actually call for a pause. Read what they said, carefully. They want it both ways. They *don’t* actually want a pause - at least for now. Rather, they want to rush ahead, hinting at “least

55 52 299 29K 55

The new @AnthropicAI post on the approach of recursive self-improvement is very important. It chimes with my argument that some kind of international effort is needed to brake the completely unconstrained AI arms race: anthropic.com/institute/recu…

A big day Monday in Cupertino for Apple with WWDC…will be discussing what to expect on the AI front from Apple on @CNBCClosingBell at 3 pm with @TheJudgeCNBC @CNBC 🔥🏆🍿📺🍎📱