Good morning. Google DeepMind's Genie 4 builds worlds you can walk around in, Artificial Analysis rebuilt its leaderboard mostly on private tests, Liquid packed six models into about 2B parameters each, and the BEAM now fits in your browser in under a megabyte. Poteto, meanwhile, shared tibot, the bot that reads her mentions so she doesn't have to.
Google DeepMind calls it its most realistic and interactive world model yet. You explore scenes in real time, and they react, make sound, and the people in them move and speak. Demis Hassabis says it brings together real-time physics, reasoning and audio. X's news summary says it starts as a research preview for trusted testers.
Bigger worlds, smaller models.
Google DeepMind calls it its most realistic and interactive world model yet. You explore scenes in real time, and they react, make sound, and the people in them move and speak. Demis Hassabis says it brings together real-time physics, reasoning and audio. X's news summary says it starts as a research preview for trusted testers.
60% of the new Index uses private test sets or held-out answers, and FrontierCode 2.0 and Terminal-Bench-Science join it. Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra at 57 and GPT-6.1 Sol at 55. The open-weights gap grew from 3 points to 12, and GPT-6.1 Terra is the cheapest top-10 model at $0.96 per task.
The family covers text (1.7B, 256K context), vision with video input, audio in 18 languages, an omni model, and two decision models with typed outputs. It has open weights, you can fine-tune it on one GPU, and Liquid says LFM3-VL-2B keeps up with Gemma 4 12B, a model 6x its size.
It says its scientists worked with six organizations that develop medicines and started drug discovery work of its own, and it shared early results. Separately, a new OneGov deal brings Claude to US federal, state, local, tribal and territorial governments at 40% off.
Ant Ling's open-weights Ling 3.1 Flash 1019 scores 44 on AA's v4.3 index, up from 41. Prince Canuma's Neywa Labs came out of stealth with Nativ, which runs open models offline on Apple Silicon. And X's news summary says Swiss startup Giotto.ai launched “sovereign” models, though its CEO says some of the weights come from Kimi.
Draw on it, reset it, merge it (maybe).
Tibo: “Just kidding, today is a reset.” He quoted ChatGPT's new iCloud Mail and Calendar plugins, then signed off with “And it's done. See you tomorrow for some good news.” So something is coming Thursday.
It's new in design mode: edit elements, draw on the page and check designs at different screen sizes as visual prompts. Michael Truell also says Cursor is “beginning to hire Bots onto our team.”
Cognition's new benchmark gives deliberately vague prompts so models have to use judgment, and grades whether a maintainer would merge the change. The top model scores just 45%, and AA added it to Index v5 the same night.
Felix Rieseberg says it “no longer provides signal”, so he took it out of his hiring loops this month, and he has an optimistic take for junior engineers.
Thariq is building an AI-native macOS dock with modes that his agent can rewrite, and he wants a way for open source to take agent edits and still merge upstream automatically. Theo: “If I have to go to a dashboard to use your product, I'm not going to use your product.”
Grok Bot 0.77.0 adds permission settings for auto-review rules, lets you pick which payment method bots use first in the vault, and puts several questions on one card. Zed's Delta makes Claude Code multiplayer with `/delta share`, and Replit Cloud Agents turn plain-language requests into tested PRs.
It's the bot she uses to run her X account. Tibot finds every post about you or your keywords, sends one short sorted digest a day (or every hour), skips spam, translates, can auto-block trolls, and never posts for you. “i still write all my posts and articles by hand.” Later she posted a profile picture to use for it (~1.9k).
Tibo posted “poteto.com has landed” (~6.1k), and she answered “💀💀💀 i cant afford” (~1.6k) before showing off a domain of her own.
She says engineers, designers and PMs each contribute hundreds of PRs a day, because the code is built to be agent-friendly and the easy thing is the right thing. She calls Grok Bot the industry's best example of where all those tokens go: “quality and velocity, without compromise.”
That was her whole caption on the Grok Bot 0.77.0 release.
The BEAM, now in your browser.
Software Mansion's Popcorn now runs the actual BEAM, which fits in under 1 MB gzipped, with the latest Elixir and better stability and speed. Within hours, someone had Elixir text-to-CAD running entirely in the browser, which means Jido runs there too.
Zach Daniel's source-code marketplace made its first sale, now lists four codebases, and its blog spotlights PG Orbit by @acrogenesis, the first one Zach didn't add himself.
Elixir Tip #51 says to drop dbg/2 at the end of a pipeline instead of IO.inspect. Douglas Correa built a Grok Bot CLI in Elixir, and a forum post walks through a reconciler GenServer for OTP processes a deploy killed.
80,000 packages, a few seconds.
The evaluation-time WebAssembly docs now include use cases and how to ship Wasm modules in flakes, for when “you've ever run up against the limits of the Nix language.”
Determinate's CI builds about 80,000 packages across three architectures, from the compiler to the kernel to Chrome. It only rebuilds what changed, so sometimes it takes a few seconds. He's also hiring.
It finishes Kitty graphics with animation, compresses scrollback by ~90% on average, adds a new OpenGL renderer on Linux and ghostty +ssh, and fixes 10 security advisories, so update. Mitchell also reports a big six months for libghostty.
@mbarneyjr is now a nixpkgs maintainer and says the CloudFormation language server is coming to nixpkgs. SatsBridge open-sourced its Hermes Agent NixOS distribution.
Components and charts.
It's one library with styled LiveView components, headless ones you style with your own classes, and now native iOS and Android components for Mob.
It was just announced on the Elixir Forum. Also, src_rip is making a video on underrated recent Phoenix and LiveView features: colocated hooks, scopes, and the new layout pattern.
One thread, still counts.
Our Ash query found exactly one post in 24h, an Elixir Forum chat about fulltext search across Ash resources. Ash creator Zach Daniel spent the day on Moatless (see Elixir).
Current = Tue Oct 27 08:00 CET → Wed Oct 28 08:00 CET; previous = Mon Oct 26 08:00 CET → Tue Oct 27 08:00 CET (24h each, back to back, from X post counts). These are the same queries saved on Oct 27, so the numbers compare directly. * = low volume.
AA v5 now runs 60% on private tests, and FrontierCode 2.0 asks whether you should merge a PR, not just whether it passes the spec.
LFM3 fits six models around 2B each, Popcorn runs the BEAM in your browser in under 1 MB, and Nativ runs open models offline on a Mac.
Cursor is “hiring Bots”, Poteto shared tibot, and an Anthropic team dropped the coding interview.