You've never routed like this before. @OpenRouter is bringing Jev to all of your LLM calls, so your agentic workflows never have to waste a token again. As always, faster, cheaper, more intelligent. Go build the future.

this guy literally explained the truth about the "ADHD superpower"

New ARC Prize 2026 - ARC-AGI-2 High Score 83.06% by @tufalabs The Grand Prize is unlocked at 85%

DOOM now runs in Bend 2! Copy the command to run it in your terminal from the README on GitHub: github.com/eliesgalvira/bend… Or run it directly with nix: nix run github:eliesgalvira/bendoom Available for Linux x86-64, Linux ARM64 and Apple Silicon macOS. Linux requires X11 or Xwayland. Controls: - WASD to move - Left and right arrows to turn - SHIFT to run - CTRL to shoot - SPACE to use doors, lifts, switches and to restart after dying You play Hangar, the first level of the original shareware Doom, start to finish on Hurt Me Plenty. No menus, saves or other levels yet. Some interesting tidbits: - Maps, textures and palettes load at startup from a WAD file, Doom's asset format. WAD stands for "Where's All the Data?" - The rendering targets a pixel-perfect match with Chocolate Doom - The renderer is a pure function that updates its pixel buffer in place, thanks to affine types - Bend 2’s dependent types let us prove that no sequence of inputs can move the player into a position that overlaps a solid wall.

How good is Jev at designing a design system? You give it one sentence, and it picks from over 1 quadrillion combinations. Nothing in the video is sped up, each restyle is 4 parallel calls: about 2 seconds, $0.0007 29 flags control corner shape, borders, icon family, chart paths, page texture, density and motion, all on @shadcn. Colour, radius, spacing and type scale are continuous on top of that. Interesting little Sunday experiment

Jev is an incredible model router. Ask Jev which models can solve a user’s task, and route it to the cheapest one. We have a new frontier model overnight.

> wakes up > there are 5 open issues "wow someone is actually using Bend 🥳" > spends the Saturday fixing them > done! now I can play some League > there are 20 open issues 😐

Today we rolled out Astra to every engineer at Databricks (N=~3500). Some notes that may be helpful to others: 1. Astra unambiguously out performs our previous highest-end models (Opus 5, Sol 5.6) on highly complex tasks, especially those related to high level system design or long range horizontal tasks. 2. Engineers given Astra increased overall coding spend by around 60% compared to baseline. 3. It is not clear Astra meaningfully improves on medium/low complexity coding tasks compared to earlier models. We suspect those tasks are mostly saturated (i.e. perfectly executed) by existing models. 4. We learned above by piloting Astra with around 200 users to gain signal on both quality and cost. We use Unity Gateway to do cohort-based experiments for all new models. 5. We give engineers a sub-budget specific to Astra to encourage them to use Astra selectively on complex tasks while preferring lower cost models for everyday tasks. Our engineers are able to mix-and-match tools and models within their overall budget envelope (we also allow for increased budgets through various mechanisms). These budgets are defined in Unity Gateway and regularly revisited. Note: We do not have robust comparisons of Astra-vs-Fable because we have net yet rolled out Fable widely due to data retention policies.

multiple camera angles from a single video using Krea Agent. see examples and prompt below 👇

ReScript 13.0.0-alpha.6 is out! I contributed parser improvements that cut parsing time by ~5% on real code (~10% in some synthetic benchmarks). Long-awaited source map support is here too. Give this pre-release a try! github.com/rescript-lang/res…

Formalized possible world semantics (allocation) with step-indexed logical relations (recursive types and functions stored in objects) to model sound rules for TypeScript. Because why not.

On the 25 ARC-AGI-3 public games, GPT-6 Astra scored 100% using the new provider adapter harness. But that score alone doesn't show just how efficiently it solved them - or how close it came to the fewest-action solutions anyone has found, including runs produced through extensive software-assisted search. The chart below compares three data points for each Public game: 1. The human baseline, calculated by summing the upper-median human action count for each level 2. The number of actions Astra used at max reasoning with the new provider adapter harness, which preserves opaque reasoning across requests and enables auto-compaction 3. The best known valid fewest-action run on each game - aka the speedrun I asked Peter Findley, an engineer who holds or ties the speedrun record on all 25 games, how he produced his runs. He said he used Codex CLI and a custom harness combining object-oriented memory, layered metacognition, skills, and hooks. He used extensive breadth-first search on a 256 GB RAM machine, but says Monte Carlo tree search with negative perturbations found what he described as the global minima. In short, he put a lot of work into programmatically achieving speedruns. So how did Astra do? It crushed the human baseline, using 6,485 actions compared with 17,135. That efficiency, combined with completing every level, is how it achieved 100%. Almost as remarkable is how close it came to several speedrun records. It tied three games (ar25, ft09, and sb26) and came within 5% on three more (s5i5, tu93, and dc22). Summed across all 25 games, the speedruns used 5,502 actions. Astra used 6,485 - 62% fewer than the human baseline and just 18% more than the speedruns. The fact that it came this close in a single recorded run per game, without executing code or repeatedly replaying the games, is incredible. It speaks to how capable Astra is, at least on ARC-AGI-3-like tasks requiring world modeling, spatial reasoning, rule induction, state tracking, planning, exploration, adaptation, and efficient problem-solving.

I've tried this on my own sudoku model and seeing some beautiful viz but no fractals yet! Though theoretically it should be much more easily analyzed through your method? github.com/chenglou/sotaku

Doing research on semantics is so satisfying these days. Especially when the agents spontaneusly bring up old friends — @PeterOHearn12, @hyang144, Cardelli, Appel, Ahmed, Pfenning, Reus. Or even suggest to "use symbolic execution in the style of Berdine/Calcagno/@PeterOHearn12".

first ray tracing demo on Bend2! looks unimpressive til you realize there's no shader nor opengl buffers on EVERY frame, it: - allocates a "quadtree JSON" with 1 million objects (??) - ray-traces over it, using a Python-ish marcher - renders every pixel to the screen - collects the entire structure and it still renders at ~80 FPS on 1024x1024 with no leaks and no stop-the-world garbage collection. it just keeps going again this is not optimized in any sensible way. this is rendering 3d graphics by allocating a massive JSON on every frame and then ray tracing and consuming it whole inside the Apple M4 GPU. and somehow it works I'm soo happy I finally have time to play with my creation 🥺

Last week we published a factorization of RSA-260. Today, we’re sharing the methodology of how Devin and a Cognition researcher built the world’s fastest GPU optimized lattice siever, to make factoring numbers 10x cheaper than the previous state of the art: cognition.com/blog/factoring…

A Millenium Prize isn't cool. You know what's cool? A Billenium Prize.

Hey folks. Sotaku 2 is out! Timely, given the speculation about GPT Astra’s recurrence Tldr: Sotaku is a tiny laptop-grade neural net that learns to solve sudoku without being given its rules. The broad goal is to see if we can solve bigger symbolic problems purely neurally, and sidestep “neurosymbolic” architectures - New SOTA: solves 99.12% of the hardest sudokus in the world (benchmark: sudoku-extreme) - Still the same 800K params looped transformer, trained from scratch in <90 minutes on a single H200 - No inference-time hacks (no verification, no beam search, no damping). Iterates stably til >4096 iterations! - Using evolutionary strategies for fine-tuning helped, but it ended up not needed anyway The training change is tiny: for a subset of batches, let the model run some iterations without tracking gradients, then train regularly (with gradients) through the next few. That lets it practice later iterations without backpropagating through the whole sequence. Otherwise, same architecture & loss as v1 Some folks see weights tying through a loop as premature optimization (aka anti-bitter lesson). My current hunch is that it isn’t. The properties of a loop vs its unrolled counterparts are not the same, for both traditional programs and neural nets. Sure, an unrolled loop can happen to learn the same set of weights, but I’d argue that tying weights in a loop is a good forcing function in the optimization process. It kinda changes the learning goals Fully reproducible recipes, experiments & checkpoints: github.com/chenglou/sotaku

turns out arc-agi-3 might actually be a skill issue. we saturated the benchmark at 100% rhae with a general purpose coding agent - claude code + opus 5, and no additional harness, but just one skill. the key idea was simple: force a falsifiable prediction before every action. that turns every move into an experiment, every miss into a precise correction, and lets the agent learn the game by being wrong on the record. arc-skill.vercel.app/