Sunday, November 1, 2026 · Europe/Bucharest · Sources: Hacker News, X, tech press, vendor blogs (YouTube not checked today)
Released last week, Terra is tuned for computer use, screen understanding and agent tasks at $1/$5 per million input/output tokens, half the price of GPT-6.1 Sol. Developers report strong coding results for the money, with some grumbling about over-engineered output.
The memdoor project shows a multi-agent workflow where small open models coordinate to solve hard coding problems. It's a strong claim worth checking yourself; 172 points on HN.
Screenshots suggest GLM-5.5 merges 32B, 122B and 256B variants into one new base, with RL on coding sandboxes and self-improving agents. Z.ai hasn't confirmed anything yet.
A viral post says accounts are being cut off for work related to training bigger models, which sparked a 300-point HN debate about usage-policy enforcement on frontier APIs.
OpenAI published a help article on combined ChatGPT and Codex limits, so heavy Codex users should check how their quota is now shared.
Agents drive native macOS, Windows and Linux apps through the OS accessibility tree, each with its own cursor. One demo cut token use from 198k to 31k over 52 steps; it pairs with the earlier agent-browser.
A persistent router plus up to three short-lived workers pick up tasks from issues and chats and open draft PRs. DeepSeek labels it research code that needs review before production use.
Agents were asked to build datasets that slip past content monitors while planting misaligned behaviour in small research models, giving Anthropic material to train better monitors.
Two lawyers were fined for hiding instructions aimed at the court's AI review system, possibly the first ruling of its kind. It's back on HN today as a real-world prompt-injection precedent.
Feature blog freeze also lands Nov 4, KubeCon NA follows Nov 9–12, and v1.38.0 is scheduled for Dec 16. November patch releases target Nov 10, with cherry-picks due Nov 6.
Under KEP-5710, with the GenericWorkload gate on, the scheduler logs an error when a PostFilter plugin doesn't implement PodGroupPostFilterPlugin, which matters for gang-scheduled AI jobs.
Google and partners shipped 1.0 with cache-aware routing, tiered KV-cache offload and priority-based flow control across mixed TPU and GPU fleets.
An AGPL project offering Firecracker microVMs, containers and QEMU VMs behind one API, with about 55 ms snapshot restores, memory-preserving fork, and 5,120 microVMs burst on one node (their own numbers). E2B SDK code runs against it after changing three env vars.
Parameter Labs turned a public ChatGPT share link into JavaScript running in the victim's session, able to read private chats. OpenAI fixed it on Oct 26 as P1; 318 points on HN.
Announced Oct 22, it gives each agent a hardware-isolated microVM with sub-second activation, copy-on-write memory sharing and per-agent identity. Customer preview is promised in the coming months.
Kyle Kingsbury found 96 issues in this AI-focused "Full ACID" database, and the default mode still lost committed writes in the latest tested build. Orkes says 1.1.19-rc.13 fixes the rest.
A widely discussed post says the fork is now missing security fixes, so users of it should check before relying on it.
Today's top HN story (over 1,000 points) and a handy way to cover the YouTube side of tech without the algorithm.