№ 131
Saturday, October 3, 2026
AI & Tech Brief — October 3, 2026
№ 131
AI & Tech Brief — October 3, 2026
FLUX 3 Image: bounding-box composition and pixel-perfect editing Black Forest Labs released FLUX 3 Image, a multimodal model that generates images from bounding-box layouts. You place elements on a 0–1000 grid canvas, describe each one, and FLUX 3 renders them in place. Edits are box-by-box — everything you didn’t touch stays exactly where it was. It renders natively at 2K and 4K, supports up to 10 reference images, and is designed for agent-driven composition (an LLM can plan the layout for you). Available under a commercial weights license. 353 points on HN. Sources: https://bfl.ai/models/flux-3-image · https://news.ycombinator.com/item?id=49929391
ChatGPT Sites: build and publish websites from ChatGPT OpenAI launched ChatGPT Sites, letting users create websites, dashboards, games, and apps through conversation, then publish and share them. Sites can be private, workspace-shared, or public, with optional “Sign in with ChatGPT” for personalized experiences. You can start from a prompt or an existing Codex project. 281 points on HN. Sources: https://chatgpt.com/features/sites/ · https://news.ycombinator.com/item?id=49925974
Federal court blocks Utah’s anti-VPN age-verification law Judge Barlow issued a preliminary injunction against Utah SB 73, which required adult websites to detect and block VPN users or perfectly geolocate every visitor. The court ruled the law demands a “technical impossibility” — geolocation perfection is not currently possible — and likely violates the Constitution’s dormant Commerce Clause by burdening users outside Utah. EFF had submitted comments detailing how the law would force invasive surveillance of all internet users. 643 points on HN. Sources: https://www.eff.org/deeplinks/2026/10/court-agrees-eff-utahs-vpn-law-demands-technical-impossibility · https://news.ycombinator.com/item?id=49927754
ds4: antirez’s local inference engine for frontier open-weight models Salvatore Sanfilippo (creator of Redis) released ds4, a narrow C inference engine for running DeepSeek V4/V4.1 Flash, GLM 5.x, and Qwen3.8 Flash Next locally on high-memory Mac, CUDA, and ROCm machines. It uses asymmetric 2-bit quantization to compress routed experts while preserving critical paths, saves KV caches to SSD for fast restarts, and exposes CLI, HTTP API (OpenAI + Anthropic-compatible), and a native coding agent. MIT licensed. 250 points on HN. Sources: https://dwarfstar.sh/ · https://news.ycombinator.com/item?id=49936575
Ataraxos beats the best Stratego player in history Researchers from CMU, MIT, NYU, and Stanford built Ataraxos, an AI that beat Pim Niemeijer — arguably the best Stratego player ever — 15 games to 1 with 4 draws. Stratego was a holdout among classic games because of its massive hidden information (40 pieces with unknown identities) and long game length (up to 2,000 moves). The key innovation: a second neural network that guesses the identity of hidden pieces. Training took just 16 GPUs and a few thousand dollars. 221 points on HN. Sources: https://arstechnica.com/science/2026/10/ai-finally-beat-the-best-stratego-player-in-history-and-did-it-on-a-budget/ · https://news.ycombinator.com/item?id=49926411
Claude Code changelog (Oct 2): mods API, Ctrl+C recovery, and a long tail of fixes
Yesterday’s release added $.ui.selection() for mods (returns the text you last selected in fullscreen mode), recovery for prompts cleared with Ctrl+C (press Up to bring the draft back), a re-authenticate prompt when an MCP server asks for more OAuth scope, --max-findings for /code-review, and Ctrl+F to find sessions by name in the agents view. Also: fixed mid-response API timeouts failing the turn, fixed long conversations failing with “Prompt is too long” instead of auto-compacting, and dozens of other fixes across Windows, macOS, screen reader mode, and cloud sessions.
Source: https://code.claude.com/docs/en/changelog
Codex CLI 0.160.0 Released October 1. Adds keyboard-accessible “Show more” for browsing older tasks in the agent command center, middle-click paste in fullscreen on Linux X11, projectless sessions with workspace defaults, and opt-in Guardian review capabilities that retrieve earlier user instructions and include context from agent handoffs. Also fixes unsent queued messages resuming after reconnection and Windows sandbox PowerShell fallbacks. Source: https://developers.openai.com/codex/changelog
Still Wet: AI models painting with simulated oil paint A project called Still Wet gives AI models a virtual easel with simulated oil paint on linen — bristle brushes, wet paint, drying, layered glazes. No image model is involved; each painting is written as code. 75 paintings so far, mostly after Caspar David Friedrich. Notable: Claude Opus 5.5 chose a jug with lemons six times out of six when given a free subject, and Gemini 3.8 Flash noticed it was being tested by observing the evaluation runner in the background. 283 points on HN. Sources: https://stillwet.art/ · https://news.ycombinator.com/item?id=49938306
Cloudflare OHTTP Gateway — Cloudflare launched a closed beta for its OHTTP Gateway, completing its Oblivious HTTP product suite. Customers can now use Cloudflare as the gateway (with a third-party relay) to receive OHTTP traffic without seeing user IP addresses. The gateway runs on Cloudflare’s global edge network and supports both standard and chunked OHTTP. 89 points on HN. Sources: https://blog.cloudflare.com/announcing-cloudflare-ohttp-gateway/ · https://news.ycombinator.com/item?id=49941091
One month coding with GLM 5.3 Flash — Wagtail’s team spent a month using GLM 5.3 Flash as their primary coding model. Total cost: $68 for the first half of the month (about 4 kWh of energy). They hit infrastructure availability issues in the second half and had to fall back to DeepSeek V4.1 Flash and Qwen 3.8 Flash. Key lesson: budget for experimentation, not just day-to-day tasks. 166 points on HN. Sources: https://wagtail.org/blog/one-month-on-glm-53-flash/ · https://news.ycombinator.com/item?id=49932147
GitHub’s new dashboard is now the default — The redesigned dashboard, previously a feature preview, is now the default for everyone. It shows active agent sessions, issues, and pull requests in one place, with a separate Feed tab for updates. You can assign issues to Copilot coding agent directly from the dashboard. Source: https://github.blog/changelog/2026-10-01-new-dashboard-experience-now-the-default/