Top Stories

Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows

822 points · research.meta.ai

Meta is back in the open-weights conversation in a big way. Muse Glimmer is a 30B-parameter model tuned specifically for always-on, local agentic workflows — the kind of thing you’d run on your own hardware to power a persistent assistant rather than call out to a hosted API. The HN crowd is paying attention because a capable agent-focused model you can actually run locally hits the sweet spot of privacy, cost, and latency that developers keep asking for.

The size is telling: 30B is small enough to run on a beefy workstation or a single high-memory GPU, but large enough to handle real tool-calling and multi-step reasoning. Coming from Meta’s research arm, it signals the company doubling down on open models as a strategic wedge against closed frontier labs.

Docker Sandboxes – Disposable, isolated sandboxes for AI agents

545 points · docker.com

As AI agents get more autonomous, the obvious question is: where do you let them run code without blowing up your machine? Docker’s answer is Sandboxes — disposable, isolated environments purpose-built for letting agents execute arbitrary code, install packages, and poke at a filesystem without touching anything you care about.

Developers on HN are treating this as a natural and overdue extension of Docker’s core value proposition. The rise of coding agents has made “give the model a safe playground” a first-class need, and Docker is positioning its existing container tech as the sandbox layer for the agent era.

tl;dv: Over 180k meetings left wide open

408 points · bobdahacker.com

A researcher found that meeting-recording service tl;dv left more than 180,000 meetings exposed, with recordings and transcripts accessible to anyone who knew where to look. Given how much sensitive material flows through AI meeting-notetakers — internal strategy, hiring conversations, customer calls — this is exactly the kind of leak that makes security-minded folks wince.

The thread is a cautionary tale about the exploding category of AI transcription tools that quietly sit in on your meetings. The convenience is real, but so is the attack surface, and access-control bugs at this scale turn an entire company’s internal conversations into a searchable archive for whoever finds the hole.

Mark Zuckerberg attacks ‘closed’ AI rivals as Meta returns to open models

114 points · ft.com

Paired with the Muse Glimmer launch, Zuckerberg is making the open-vs-closed argument loudly, framing Meta as the champion of open AI against rivals that lock their models behind APIs. It’s both a genuine philosophical stance and a shrewd competitive move — if you’re behind on frontier models, commoditizing the layer is a great way to undercut the leaders.

HN loves this debate because it cuts to the heart of who controls AI. Commenters are split between cheering the open-weights momentum and noting that “open” from a company Meta’s size is never purely altruistic.

Mistral patent for “Code implemented tool calls”

151 points · patentsgazette.uspto.gov

Mistral has been granted a patent covering “code implemented tool calls” — the increasingly standard technique of having an LLM emit executable code to invoke tools rather than structured JSON. The pattern is everywhere now, which is precisely why a patent on it is setting off alarm bells in the comments.

The worry is familiar: a foundational technique that much of the ecosystem already depends on, locked up by one lab. Whether the claims are actually broad and enforceable is up for debate, but the mere existence of the patent has developers nervous about the direction of IP in the agent tooling space.

Show HN: Ante, a coding agent in a single binary that runs offline

80 points · github.com

Ante is a coding agent packaged as a single binary that runs entirely offline — no cloud dependency, no telemetry, just a self-contained tool you can drop onto a machine and use. It rides the same wave as Muse Glimmer: developers increasingly want capable agents that don’t phone home.

The single-binary distribution is a big part of the appeal for the HN crowd, who consistently reward tools that are easy to install and easy to reason about. Offline-first coding agents are a small but growing niche, and Ante is a clean example of the pattern.

Learning more about Claude’s mathematical capabilities

83 points · anthropic.com

Anthropic published research probing how well Claude handles serious mathematics, using work related to the Riemann zeta function as a test bed. It’s part of an ongoing effort to understand where frontier models actually reason versus pattern-match, and where they break down on hard, structured problems.

Math is a favorite yardstick on HN because it’s unforgiving — you can’t hand-wave a proof. The discussion digs into what genuine mathematical capability in an LLM would even look like and how much of the performance is real reasoning versus sophisticated recall.

Tail-call optimization in C is relatively recent (2025)

96 points · lwn.net

An LWN piece explains that guaranteed tail-call optimization in C — long taken for granted in functional languages — only landed relatively recently via a compiler attribute. It’s a satisfying deep-dive into why a seemingly basic optimization took so long to become dependable in the C world.

This is peak HN comfort food: a careful look at compiler internals, standards-committee inertia, and the practical implications for people writing interpreters and state machines where tail calls matter for correctness, not just performance.