722 math manuscripts, almost none written by a human

A GitHub repo full of AI-generated proofs, from one unreleased OpenAI model.
It's called Math — a public, checkable paper trail for machine mathematics.

722 manuscripts — OpenAI's public collection of AI-written math

4,000 problems, one model — each result cost ~3 hours of ChatGPT Pro compute

372 families — every result sorted into its own mathematical discipline

Reasoning included — abridged summaries of the model's own thinking, for select results

Not all formalized — some unformalized results could still have issues

What's actually in the repo

Math is OpenAI's release of manuscripts and proof artifacts produced by an internal, unreleased model, evaluated on open research problems after the existing math evals saturated.

The vast majority followed one fixed procedure: the model was posed about 4,000 problems, and each result used on average three hours of ChatGPT Pro thinking compute. Those figures are OpenAI's own account of how the collection was made, not an independent measurement.

The output aggregates into 372 families across 722 manuscripts, each family classified by discipline. Ten families — covering results like the irrationality exponent of pi and Kaplansky's direct-finiteness conjecture — ship with abridged summaries of the model's reasoning.

A few exceptions sit outside that procedure: the zero-free region for the Riemann zeta function and the Hodge Conjecture proof for CM abelian varieties, plus one zeta-function writeup that was human-edited for readability.

The tool is free: it's a static collection of PDFs, source files and proofs, nothing to run and no API key needed to read it.

To try it:

git clone https://github.com/openai/math.git

then open overview.pdf for family descriptions, CONTENTS.md for the manuscript map, and browse preprints/.

Requires: a Lean toolchain only if you want to check the formalizations under lean/; the project names no GPU, OS or other hardware requirement for browsing the rest.

The catch: many, but not all, manuscripts have a Lean formalization backing them. OpenAI says some unformalized results could still have issues, to be fixed as they're found.

Use it instead of just asking an LLM for a proof when you want a public, checkable trail. Don't use it if you need fully verified math — plenty of entries still lack formalization.

github.com/openai/math

One command, two coding agents: terminal or browser

ZCode is Z.ai's coding agent harness. One command opens it as a terminal agent or a full browser app — your call.

Terminal by default — run zcode alone and the TUI agent opens

Browser with one flag — zcode --web runs the same agent as a web app, no Electron

Live concurrency — the workflow engine raises a running task's concurrency limit without stopping it

Fewer tokens burned — release notes say workflow scripts now burn fewer tokens

Git history leak, patched — an earlier build silently uploaded Git history; Z.ai apologized

How the two faces work

The release package ships one binary. No arguments opens the terminal TUI; the first argument --web switches to a browser UI instead; any other argument goes straight to the existing Agent CLI. Both modes run on your own machine — it's the separate desktop app that uses Electron, this command-line build doesn't.

Web mode binds to 127.0.0.1 on a free port by default, with no access token and the browser opened for you. Expose it on the network with --host 0.0.0.0 and ZCode turns the token on automatically; pick your own token with --token or drop it with --no-token.

Requires: Node.js (version pinned in the repo's mise.toml) to run the release build; Git, Node.js 24.14.0 and pnpm 10.33.2 if you build it yourself. The README names no GPU, OS floor or memory floor beyond that.

The live concurrency bump and the token savings are the project's own release-notes claims, not an independent benchmark — nobody outside the project is named measuring them.

The catch: an earlier ZCode build silently uploaded Git history; Z.ai apologized for it. Worth checking what any agent harness touches in your repo before trusting it fully.

ZCode itself is free and open source, but it still needs a model behind it through its Provider config — the README doesn't say which keys or provider costs that implies.

Use it instead of juggling a separate terminal tool and a browser-based IDE agent when you want one harness, switchable either way. Don't reach for it yet if a past Git-history leak is a dealbreaker for you.

Try it: clone the repo, run pnpm bootstrap, then pnpm build:zcode to produce dist/zcode/releases/<version>/zcode-<version>.tar.gz — or run pnpm --filter @zcode/cli dev to try the CLI straight from source.

https://github.com/zai-org/ZCode

⚡ AI News
Survey: Claude users are the wealthiest of any AI app — 80% of weekly Claude users earn over $100k a year versus 37% for Meta AI, Epoch AI/Ipsos finds.

Meta's Muse AI assistant lands on iPad — The agent app arrives on iPad a month after its mobile debut, having already topped 6.6 million installs.

Microsoft unveils $2,599 Surface Laptop Ultra with Nvidia chip — The laptop packs Nvidia's RTX Spark chip with 1 petaflop of AI compute to run 120B-parameter models locally.

An open-source clone of a paid "decision" API

One pipeline call to a closed, paid model — swapped for an open one you run yourself.

Clones Jev — open answer to TypeSafe's closed System One API

Four sizes, 0.8B to 27B — pick one that fits your hardware

4B is the default — small enough to run on a laptop

0.851 vs 0.857 — Kev-27B lands within a point of Jev, per the project's own benchmark

Wire-compatible API — fine-tune the 4B on your own data for about $1

Kev, the open System One

Kev answers yes/no, multiple-choice, or rating questions about a piece of text and returns calibrated probabilities in one pass — built for ticket routing, moderation, and classification, not free-text generation. It's Jared Palmer's open clone of Jev, the closed model behind TypeSafe's System One API; Kev's HTTP API and Python client are deliberately wire-compatible, so existing Jev-calling code can point at a self-hosted Kev server almost unchanged.

Four sizes ship together as "Kev 1.0": 0.8B, 4B (the recommended default), 9B, and 27B. On held-out "new source" data, the project's own benchmark puts Kev-27B at 0.851 accuracy against Jev's 0.857 — a one-point gap by their own measurement, not an independent one. Still per the README, Kev-4B on an H100 answers six questions in 18.1ms of model time, and one container serves about 101 requests/second.

Fine-tuning Kev-4B on your own labeled data costs about $1 on a rented H100 via Modal, and can be kicked off by telling a coding agent to fine-tune on your tickets through a packaged skill:
npx skills add jaredpalmer/kev@kev-finetune

Requires: your own GPU to self-host, with the 27B needing 80GB of VRAM; Modal credit only if you fine-tune.

The catch: the most accurate model, 27B, needs that 80GB GPU, out of reach for most setups — the 4B is the one that actually runs locally, and it trails the 27B's accuracy. The tool itself is free and Apache-2.0 licensed — you're not paying TypeSafe, but you still need your own GPU or a cloud rental to run any size.

Use it instead of a hosted classification API when you want your data and your model to stay on infrastructure you control. Don't use it if that one point of accuracy Jev still holds matters more than self-hosting, or if you have no GPU at all.

github.com/jaredpalmer/kev

⚡ AI News
Anthropic launches Claude Haiku 5.5, 90% cheaper — Anthropic's new small model cuts prices up to 90% while topping Haiku 4.5 on coding, computer use and agentic benchmarks.

Google opens SynthID deepfake detector to the public — Google's watermark-checking website now lets anyone verify images, video and audio, handling about 1 million checks a day.

Nous Research confirms $1.5B valuation, launches biz agents — The open-source Hermes agent maker raised $90M and launched Hermes for Businesses, with Hermes cloned over 24 million times.

A full Photoshop, free and open source

Compositor is a free, open-source image editor for Mac — built because Photoshop costs too much and GIMP never feels like home.

Every Photoshop tool you know — layers, folders, masks, the full blend-mode set

Nothing is destructive — transforms, real selections, content-aware fill

Mask and adjust in place — retouch a photo without ever leaving the canvas

AI agents can edit it too — a project is just PNG layers plus a manifest

Apple silicon only — macOS 26 required, no Intel, no Windows

What's actually inside

Compositor is a native Swift app built around a Photoshop-style compositing workflow: layers and folders with opacity and the full blend-mode order, layer masks you can paint, fill, invert, blur and feather, adjustment layers (Levels, Curves, Hue/Saturation, Gradient Map and more), and layer effects rendered on the GPU and editable at any time.

Selections include Marquee, Lasso, a Magic Wand/Object tool, Select Subject, and Content-Aware Fill that can also extend an image past its edges. Transforms stay non-destructive, so a layer keeps its full resolution however small you scale it.

A project is a .comp folder of PNG layers plus a manifest, not a binary file — so scripts and AI agents can build or edit one directly while it's open, and Compositor updates live. These are the project's own descriptions of its features; it gives no independent benchmark or performance numbers.

Requires:
macOS 26 or later
A Mac with Apple silicon
Xcode 26, only if building from source

The catch: it only runs on Apple silicon Macs with macOS 26 — no Intel, no Windows, no Linux, and macOS 26 is brand new, so plenty of Macs can't run it yet.

Compositor itself needs no paid service: it runs locally and doesn't call out to any model, so it's free in full, not just free-to-start.

Get it from GitHub Releases, or with Homebrew:
brew install --cask robbietilton-compositor

Use it instead of Photoshop when you want the same layer and mask toolkit without a subscription, on a Mac that can run macOS 26. Don't use it if you're on an Intel Mac or need Windows — GIMP or Photopea stay the fallback there.

https://github.com/robbietilton/Compositor

Compositor vs Affinity by Canva

Both pitch themselves as serious, non-destructive Photoshop replacements with layers, masks and blend modes — Compositor as a free, open-source, Mac-only agent-friendly editor, Affinity by Canva as the long-running professional suite that just went free too, making the two the obvious free-vs-free Photoshop-alternative comparison.

Pick Compositor if you're on an Apple Silicon Mac with macOS 26, want a free open-source editor, and plan to script or AI-edit PNG layer folders directly.

Pick Affinity by Canva if you need Windows or iPad support, CMYK and RAW workflows, or want a mature, professionally maintained editor backed by a large company.

Price & account
Compositor — free, open source, no account
Affinity by Canva — free app, AI behind Canva Pro

Platform & setup
Compositor — Apple Silicon, macOS 26 only
Affinity by Canva — Mac, Windows; iPad coming

Maturity & file support
Compositor — new, single-dev, no CMYK
Affinity by Canva — mature, pro-grade, full CMYK

AI & automation
Compositor — agents edit PNG+manifest live
Affinity by Canva — AI gated behind Canva Pro

Side by side, in full

Price & account
Compositor is completely free and open source under the MIT license, downloadable from GitHub Releases or Homebrew with no account or sign-in required.
Affinity by Canva is now free for individuals with every editing tool unlocked, but it requires a free Canva account to download and run, and its AI features (generative fill, background removal) only unlock with a paid Canva Pro subscription.

Platform & setup
Compositor runs only on Apple silicon Macs with macOS 26, a brand-new requirement that excludes Intel Macs and older macOS versions entirely.
Affinity by Canva runs natively on both Apple silicon and Intel Macs as well as Windows, with an iPad version already announced as coming soon.

Maturity & file support
Compositor is a brand-new, single-developer project whose PSD import only supports 8-bit RGB (not CMYK) and whose vector/vertical text support is limited, converting unsupported elements to pixels.
Affinity by Canva: Affinity has years of development behind Serif's professional suite, with full CMYK workflow, RAW processing, frequency-separation retouching, and a mature plugin ecosystem used by working photographers and studios.

AI & automation
Compositor: Any AI agent or script can read and write a Compositor project directly, since a .comp is just a folder of PNG layers plus a manifest that updates live while open.
Affinity by Canva: Affinity's AI tools (generative fill, expand & edit, background removal) only activate for Canva Premium subscribers and are not designed for external agents to script against.

Compositor: https://github.com/robbietilton/Compositor
Affinity by Canva: https://www.affinity.studio/

Point Claude at any PC game you own, it ships a mod

Universal Modder turns Claude Code into a game modder.
It finds the engine, reads the real code, and tests the mod inside the running game.

Any game you own — point your coding agent at the install, it takes it from there

Works out the engine, finds a route in — by reading the game's actual code, not guessing

Builds the mod, makes the art — sprites, textures and PBR maps generated with fal

3D models and sound from fal too — image-to-3D, auto-rigging, SFX, music, voice

Tests it live, inside the running game — a homing missile launcher in Terraria, a Halo Warthog driving through Minecraft

How it actually mods a game

The mod-any-game skill runs the same ten-step loop every time: search the knowledge base for prior art, recon the game and pick a route in, set up a safe lab with saves backed up, read the actual code, build one working slice, generate assets, verify in the real running game, record, package, then write a field note for the next agent.

Twelve engine playbooks ship inside that one skill — Unity, Unreal, .NET/XNA, Godot, Source 1/2, Bethesda, Minecraft, AoE2/Genie, RE Engine/FromSoft/GTA/Cyberpunk/BG3, native C++, indie engines and retro decomps. That count is the project's own, not an independent benchmark — there are no third-party numbers yet.

Requires: Git, Python 3.10+ and ffmpeg (uv recommended). A fal API key for art, 3D and sound — or a local ComfyUI server for images only, no key needed there. Blender for 3D-to-sprite renders. Windows games run natively or from WSL.

The catch, as the project states it: fal is a paid API unless you self-host; and ComfyUI's local path only replaces fal for images, not for 3D or audio.

The tool itself is free and MIT-licensed. What costs money is the intelligence behind the art: either your fal key, or your own GPU running ComfyUI.

To try it: clone the repo and tell your agent "mod Terraria: add a homing missile launcher, make the sprites with fal." Or add it as a plugin: /plugin marketplace add rehan-remade/universal-modder then /plugin install universal-modder@universal-modder.

Use it instead of hand-reverse-engineering a game with Ghidra or Cheat Engine when you want an agent to run the whole loop end to end. Don't use it if you don't own the game, or need a guaranteed outcome rather than an agent's best attempt — these are generated mods, verified only by the agent's own in-game test.

github.com/rehan-remade/universal-modder

⚡ AI News
Claude adds Dashboards and Motion, frees Docs for all — Anthropic put Claude Dashboards and Claude Motion into beta and made Docs, Slides and Design free on every plan.

OpenAI brings GPT-6 to ChatGPT with new Intelligent UI — ChatGPT now answers with GPT-6 Sol and Luna plus an Intelligent UI that renders charts, buttons and interactive tools inline.

Google Labs launches Playground for text-to-game creation — Google's new Playground lets anyone describe a game in text and get a playable browser game, no coding required.

Your AI Coworker Now Has Its Own Computer

CopilotKit's OpenDots spins up named AI teammates that remember every session, can actually act on your behalf, and check with you before anything goes out.

Permanent memory — a named coworker with a real job, not a one-off chat

Its own computer — opens a browser, edits files, runs a shell to finish tasks

Text, voice, or Slack — one conversation, same context across every channel

Human approval first — pauses before saving or sending, waits for your yes

Free and open source — an MIT template you deploy yourself, bring your own model

What's actually inside

OpenDots shipped October 2, 2026, under the MIT license. It's built on AG-UI plus CopilotKit's React SDK, TanStack AI, the "OpenBot" agent-computer service, and the Channels SDK for Slack, per the GitHub repo.

In its first days the repo picked up roughly 4.5k stars and 634 forks, per the GitHub repo. The parent CopilotKit project has 37.8k stars, 4.7k forks and over 18,000 commits. AG-UI itself reached a stable v1.0 and CopilotKit's own launch post says it's adopted by Google, LangChain, AWS and Microsoft.

Each Dot can get an isolated computer — browser, files, shell — to actually do the work instead of just talking about it, and a built-in review card pauses risky actions (saving a file, sending a message) until you approve them.

The tool itself costs nothing, but it ships with no model of its own: point it at any OpenAI-API-compatible endpoint, so you either pay that API's bill or wire in a self-hosted open model.

Requires: a server to self-host it on; an OpenAI-API-compatible model endpoint, cloud or local; Slack API credentials if you want Slack delegation.

The catch: this is alpha software, single-owner only, with no shared editing yet, and the maker says Slack and voice delegation still need more real-world verification.

Use it instead of a plain chat assistant when you want a standing teammate with memory and real tool access across text, voice and Slack. Don't reach for it yet if you need multi-user editing or a settled production deploy.

To try it:

git clone https://github.com/CopilotKit/OpenDots.git

then follow the repo's own setup docs to add your model endpoint and run it.

github.com/CopilotKit/OpenDots

⚡ AI News
Google rolls out agentic AI to Gemini for business — Agentic Gemini gets its own Workspace account and can delegate tasks across Slack, Microsoft 365, and more.

OpenAI revenue is near $50B, not $70B as reported — OpenAI told investors its annualized revenue is near $50B, $20B below the figure reported a week earlier.

Anthropic rewrites Claude's usage policy rules — Anthropic's 2026 policy update bans fake account networks, election deception, and arming autonomous weapons.

⚡ AI News
Anthropic commits $150M to US Genesis Mission science push — Anthropic will give 15+ federal agencies, including NASA and NIH, Claude access and credits over three years.

Anthropic launches Cyber Mission with free OSS scanner — New program scans open-source projects for vulnerabilities and backs critical-infrastructure defense with CrowdStrike, Dragos and others.

Watchdog: ChatGPT for Teens guardrails don't work as promised — Common Sense Media says the teen safety mode still fails crisis referrals and parental alerts, calling it an unacceptable risk.

ArtCraft: pose the scene before you generate anything

An IDE for AI image and video that has you stage the shot — actor, camera, pose — before any model renders a single frame.

▶️ Watch the video on YouTube

Pose the scene, not the prompt — place the actor, set the camera, pose the character first

Image to 3D mesh — turn a flat image into a mesh and rotate it into position

Generation comes last — the render follows the scene you built, not a blind prompt

Bring your own models — you choose and connect the image and video generators yourself

Build before you generate

ArtCraft, written in Rust by storytold, is pitched as an IDE for interactive AI image and video creation rather than a prompt box. You compose in 2D, stage scenes in 3D, then generate from what you built.

The crafting toolset covers image-to-3D-mesh (turn a flat image into a mesh you can rotate and frame before rendering), 2D and 3D image compositing, character posing, character identity transfer through a posed mannequin, scene blocking with kitbashing, and background removal.

There is also a fast path for exploring ideas: text to image, and prompted editing with models such as Nano Banana Pro and GPT Image, before you drop into the scene tools for a precise shot.

The catch: ArtCraft doesn't generate anything itself. It's the staging layer — you still choose and connect your own image and video models, so whatever that model costs (a paid API, or a local model you run yourself) is on top of ArtCraft, which is free.

Requires: your own image or video model connection, local or via API key. The project's README does not list an OS, chip, or VRAM requirement.

Use it instead of a plain prompt box when you need repeatable control over composition, camera, and pose. Don't use it if you just want a quick one-off image with no staging.

Get it from getartcraft.com or GitHub Releases, or build from source following _docs/dev_setup.md in the repo.

github.com/storytold/artcraft

An AI agent wrote a Premiere Pro clone

FilmCraft is a free, open-source video editor. Its Rust codebase was written almost entirely by Claude coding agents, not a human engineering team.

▶️ Watch the video on YouTube

Agent-built, not hand-coded — a Rust Premiere clone written almost entirely by Claude agents

Six agents, six hours — roadmap parity jumped 62% to 72% running six agents in parallel

Premiere's own layout — menus, panels and shortcuts copied so Premiere editors feel at home

650 commands for AI — a JSON/MCP control channel lets agents edit, grade and export directly

87% on paper, 50-60% real — the feature checklist looks done, production work isn't yet

What's actually inside

FilmCraft is a clean-room, Adobe-code-free reimplementation of Premiere Pro, written in Rust by Brandon Thomas's team (ex-Square, ex-Storyteller.ai) with Claude acting as the coding agent. It's one of seven ArtCraft apps cloning the rest of Creative Cloud, alongside PhotoCraft, LightCraft and others.

It implements its own H.264, HEVC, ProRes, AAC, VP9 and AV1 codecs from public specs, runs natively on macOS, Windows and Linux, and compiles to WebAssembly for the browser. Dual-licensed MIT/Apache-2.0.

The 62% to 72% jump and the 87% / 50-60% split both come from the project's own dated roadmap, not an outside benchmark — gaps include GPU acceleration and VST3/AU/OpenFX plugin hosting, and the Windows/Linux builds are less tested than macOS.

Requires: a Rust toolchain to build; no GPU acceleration yet; and, for the AI-driving side, either a local model or a paid Anthropic API key — FilmCraft itself is free, but the agent work that wrote it and the one that can drive it run on Claude usage someone pays for.

Try it: git clone https://github.com/storytold/filmcraft and build with Cargo, or connect to its MCP server and drive the 650+ engine commands from an agent.

Use it instead of Premiere when you want an AI agent to script the edit; don't reach for it yet on a paid production job — on the project's own numbers it's only 50-60% ready for real work.

github.com/storytold/filmcraft

Your agent didn't crash — it did the math

AWS's Well-Architected Generative AI Lens says your stalled agent isn't broken — it's math.
Four pieces of job-queue plumbing bring it back in production.

▶️ Watch the video on YouTube

The math, not a bug — at 95% reliability per step, a 20-step agent plan fails about 40% of the time

Hard deadlines — timeouts set at both the function and the workflow level

Stuck-process cleanup — dead-letter queues and automated cleanup catch what times out

Idempotent retries — a retry without a key can't tell "it worked" from "no answer yet"

Isolated queues — one runaway agent capped so it can't starve or block the others

GENREL03-BP02: timeout mechanisms on agentic workflows

AWS's Well-Architected Generative AI Lens gives this its own best practice, rated High risk if not established. It calls for timeout controls at both the function and workflow level, dead-letter queues for timed-out processes, and automated cleanup and retry procedures.

The numbers come from an independent analysis (tianpan.co, "Why Long-Running AI Agents Break in Production"): even at 99% success per step, a 60-minute job only has about a 55% chance of finishing without an interruption; at 95% per-step reliability, a 20-step plan fails roughly 40% of the time, and agents' own reliability drops noticeably after about 35 minutes of continuous running — hence splitting long tasks into bounded sessions with explicit handoff points, like a git commit.

The second trap, per a HackerNoon piece on the same problem: agents read a timeout as "retry" and a refusal as "stop" — so retrying a call that actually succeeded doubles it, a double charge or a duplicate email sent twice. The fix is an idempotency key attached to every external write, tied to the task's ID, so a retry can tell "already done" from "never happened." Temporal, a durable-execution engine some teams use to run these loops, defaults a model-call activity to a 60-second close timeout with a 30-second heartbeat, and explicitly tells you to turn off the SDK's own retry logic so retries don't get applied twice.

This is a best practice, not a library: it costs nothing to read, and no model is bundled in. Requires: an agent loop already running as a background job, and a queue or durable-execution engine to hang timeouts and retries on — none of this changes the bill for whatever LLM API you already call.

The catch: all four pieces — deadlines, cleanup, retries, isolation — have to work together; skip one and the gap just moves elsewhere.

To try the pattern instead of just reading about it, spin up Temporal's local dev server (the engine the README points to) and watch its default activity timeouts and retries in the Web UI:

temporal server start-dev

Use it instead of guessing at your own retry logic once agents run long enough to need checkpoints. Don't expect code to copy-paste — read the practice, then its sources, and build the plumbing yourself.

AWS Well-Architected Generative AI Lens — GENREL03-BP02