Wednesday, October 28, 2026 · Europe/Bucharest · Sources: Hacker News, X, tech press, vendor blogs
Genie 4 generates photorealistic, steerable worlds with sound and interactive dialogue, using Gemini 3.8 Flash as an in-the-loop planner. It's a research preview for trusted testers now, with Google AI Ultra and Gemini API early access next. 825 points on HN.
Six models (text, vision, audio, omni and two decision models) share one 1.7B backbone. The text model fits in about 1 GB at 4-bit, has a 256K context, and is post-trained for agent harnesses; Liquid's own benchmarks put it ahead of Qwen3.5-9B on several agentic tests.
A 560B-parameter MoE with a 1M-token context, post-trained further for long-context and agent work, plus a base checkpoint for fine-tuning. Weights are on Hugging Face and ModelScope.
The widely cited benchmark aggregate has a new methodology version, so expect model rankings to shift in the coming days.
Agents live in a project folder with an instructions.md prompt and TypeScript tools, deploy with one command to Cursor's servers or your own, and come with built-in evals.
It captures full execution traces from long-running agents, runs customizable judges over them, and flags failure patterns. Modal and Fluidstack are named customers.
In a randomized trial, 50 data analysts matched the agent's randomly assigned preferences 86% of the time and later justified those choices as their own. It's a pointed caveat for human-in-the-loop safeguards.
A thoughtful essay on which skills stay essential when agents write most of the code. 166 points on HN.
The group will write a per-minor-version spec for how workloads behave at runtime (security, disruption resilience, resources, networking, storage, observability), plus upstream test tooling. Every requirement must be objectively testable.
A new open-source GPU Fractioning Operator under the KAI Scheduler org sets up MPS per node and enforces per-workload GPU memory limits, with guaranteed minimums and burstable maximums. KAI Scheduler v0.18.0 integrates via its NvFractions mode.
Agent-generated code runs in Wasm isolation via the open-source Spin framework on Akamai Functions, with explicit limits on compute, memory, outbound connections and system capabilities. It shipped alongside a managed Akamai Inference service.
A clear comparison of Firecracker microVMs (E2B, Vercel, Archil), gVisor (Modal's default, claude.ai) and containers, including where gVisor's shared-kernel boundary ends. Useful context now that Wasm is joining the mix.
A 16-year-old buffer overflow lets an attacker on the local network run ring-0 code during PXE boot, before Secure Boot protections apply. The embargo broke early; if you network-boot servers, ask your firmware vendor and consider disabling PXE for now.
Already-authorized users could exceed their local user's permissions. Until you update, Tailscale suggests tightening SSH policies and setting TS_SSH_DISABLE_FORWARDING=1.
Wiz tracks live exploitation attempts and patch uptake for last week's critical PHP code-execution bug; a companion post covers LiteLLM and GitLab patching too.
LWN reports the CPython JIT experiment is being pulled. 480 points and a long HN thread on what it means for Python performance work.
The Stanford Daily reports professors' books were among illegally obtained training material. 500+ points on HN.