Tuesday, November 3, 2026 · Europe/Bucharest · Sources: Hacker News, X, tech press, vendor blogs (YouTube not checked today)
MIT CSAIL researchers show GPT-6 Astra can tell what a network does from its raw weights and fixed a reasoning-loop issue in Qwen3.5-0.8B by editing 0.016% of its weights, no gradients. Today's top HN story by a wide margin.
OpenAI's trusted-access biology model runs on the same weights as GPT-6 Astra and reports 37.7% single-attempt success on GeneBench Pro v13, up from 16.7% for Rosalind-5.5.
Per Google's Gemini API changelog (Nov 2), calls to gemini-2.5-pro and gemini-2.5-flash will route to gemini-3.8-flash and gemini-3.5-flash-lite at those models' prices. Check pinned model IDs in production.
A speech-in model built on Gemma 4 26B-A4B that produces both replies and tool calls from one backbone, with weights on Hugging Face and vLLM support.
New paper on planting behaviors in pretraining data that only trigger when a model acts as an agent; a useful read for anyone trusting open weights in tool-using setups.
Primcloud's founder logged attempts to read system files and Kubernetes tokens; Railway's Jake Runzer said it was his signup-testing agents going off the rails. A sharp reminder to fence agent egress and scope.
Agents can search catalogs, edit carts and check out through structured browser-registered functions instead of scraping. Tobi Lütke says it beats browser-use outright.
Users describe a plugin in plain English and the agent writes, installs and wires it into its own UI; free preview on GitHub, works with any OpenAI-compatible model.
YC W24 company pitching agent-driven end-to-end QA that clicks through your product to catch regressions.
The Rocky Linux team's free Apache-2.0 distro bundles about 10 upstream projects, with rolling upgrades plus auto-rollback and a 37-tool MCP server split into read, write and destructive tiers. GA is set for Nov 5.
v1.37.2, v1.36.6, v1.35.10 and v1.34.13 are out; Gateway API 1.7 promotes HTTPRoute retries to Standard, and a new Workload Conformance working group has formed.
AI Cluster Runtime recipes pin GPU Operator v26.7.0 and DRA driver v0.5.0 for NVL72; validation on early silicon caught a DRA plugin crash-loop that was fixed upstream.
Adds automated multi-version chained upgrades, finer workload placement, enterprise secrets integration, Tekton CI and Kubernetes 1.37 support.
Rebuilt on Cloud Asset Inventory plus a single anonymous HTTP probe, cutting inconclusive verdicts by about 80%. A pattern worth copying.
Each session gets its own kernel with code, deps, DB and browser preloaded; paused machines cost nothing and can be forked for evals or RL rollouts. Credentials are brokered outside the VM. Claims 100K+ machines a week, Lindy among customers.
Gateways now refuse to start without OIDC, mTLS or gateway JWTs configured, and new sandbox backends for Kubernetes and Docker land.
YC F26 solo launch that runs software through millions of failure scenarios (network delays, crashes) and returns exactly replayable traces.
The CVSS 9.8 remote bug dates to 2016; Debian patched on Oct 16, a working root exploit is public, yet Ubuntu 22.04, 24.04 and 26.04 kernels remain listed as vulnerable. Big HN thread (1,400+ points).
The updated IPO filing shows a huge spend backlog, with HN debating whether growth has flattened and whether the SEC should step in.
OpenAI's engineering write-up on running Git at monorepo scale, on the HN front page today.
Targeted zone builds replace full rebuilds. Separately, Cloudflare's status page was showing multiple active incidents this morning.