📰 News
-
It will call the shop — Gemini’s Call for Me. US Pixel 11 subscribers, Phone app beta. Stock, a reservation, a moved appointment. It can say what you approved, you watch the transcript, and you can take the line. It dials from your number. → TechCrunch
-
Claude found an enzyme. CRISPR people shrugged — 950 agents, 21 hours, 200,000 enzymes that turn RNA into DNA. They flagged ART. It looks a bit like CRISPR, and nobody knows what it does. Harrington, from the CRISPR side, says this is genome mining people have done since 2008. → The Decoder
-
Schmidhuber, to Sakana — The person behind 1990 world models (a model that rolls the world forward) and early 1991 deep learning is now chief scientific advisor. Tokyo’s RSI lab wants a loop where machines keep improving their own research. He says the next step is physical AI, not more chat. → The Decoder
-
A 7B model that moves a robot — Black Forest Labs’ FLUX 3 Action. Several cameras, then the next action and the next frame. Record success on RoboLab-120, under half the last best open model, up to 3.95x faster. Weights on Hugging Face. → The Decoder
-
The experts’ calendar was late — Forecasting Research Institute. July 2025 math-olympiad gold landed five years before the median expert, ten before the median superforecaster. They capped the biggest AI company’s 2026 revenue near $25 billion. FRI puts Anthropic around $100 billion this September. A virology bench parked in 2030 looks like April 2025. They aimed high on people finishing bio-lab tasks with a chatbot. → The Decoder
-
Last year’s score got cheap. Today’s best did not — Epoch. Hold the score still and the price falls about 47% a quarter, roughly 13x a year. o3 at 75% on GPQA, a PhD-level science set, was 30 cents a question. Eighteen months later GPT-5.6 hits that for 0.04 cents. MIT says strip cheaper chips and price wars, and the algorithms are only about 3x a year. The current best can cost more per question, because it spends extra compute thinking. → The Decoder
🔥 GitHub Trending
-
anthropics/financial-services — Claude drafts for finance. Pitches, meeting briefs, sector notes, ledger checks. Not advice. A person still signs.
-
obra/superpowers — For coding agents. A short spec, a plan that starts with tests, then subagents follow it.
-
BuilderIO/agent-native — The button and the agent call the same action. What you do on screen is visible to the agent, and the other way around.
-
dream-num/univer — An office runtime for agents. Sheets, docs, and slides in one place, people and agents editing the same files.
-
google/ax — Declare an agent task. It sandboxes the job, preloads the repo and tools, and only allows the hosts you listed.
▶️ YouTube
-
Claude Opus 5.5 AI: An Incredible Leap Forward — Two Minute Papers. In the walking-creatures test Astra used a simpler model, and more generations did not fix it.
-
8 Jev Use Cases That Feel Like Cheating — Matthew Berman. Jev in the browser: catch AI slop, strip ads, rank mail, search inside a page.
💬 Community
-
ChatGPT, via the ad pixel — A write-up on ChatGPT learning what you do on other sites through an ad collector.
-
Math on ciphertext — Apple research on homomorphic encryption, computing without decrypting, paired with on-device machine learning.
-
Continual learning on 8GB — mini-AGI. Trained from scratch on a laptop, one example at a time, 8GB of VRAM.