Top Stories

Gemma 4 12B: A unified, encoder-free multimodal model

855 points · blog.google

Google’s latest open-weight model drops the separate vision encoder entirely, folding image understanding directly into a single unified architecture. That’s a meaningful design shift — instead of bolting a vision tower onto a language model, Gemma 4 treats everything as one stream, which tends to simplify training and deployment.

The HN crowd cares because Gemma has become the go-to open model for people who want capable multimodal inference they can actually run locally, and a 12B “encoder-free” design is the kind of architectural bet that could ripple across the whole open-model ecosystem.

Elixir v1.20: Now a gradually typed language

792 points · elixir-lang.org

The long-running effort to bring a type system to Elixir reaches a major milestone: v1.20 ships gradual typing as a first-class feature, letting developers add types incrementally without abandoning the language’s dynamic roots. It’s the payoff of years of research work led by José Valim and the core team.

This is catnip for HN — gradual typing in a BEAM language touches the perennial static-vs-dynamic debate, and Elixir’s approach (set-theoretic types, no annotations required to benefit) is genuinely novel rather than a Python-style afterthought.

They’re made out of weights

759 points · maxleiter.com

A reflective essay arguing that we keep anthropomorphizing language models when, fundamentally, “they’re made out of weights” — matrices of numbers, not minds. The piece pushes back on the instinct to treat model behavior as intention or understanding.

It resonated because it lands in the middle of an ongoing fight over how to talk about LLMs honestly, and it pairs neatly with the Ted Chiang piece also on the front page today. Expect strong opinions on both sides in the comments.

I was recently diagnosed with anti-NMDA receptor encephalitis

631 points · burntsushi.net

Andrew Gallant — better known as BurntSushi, author of ripgrep and a fixture of the Rust ecosystem — shares a deeply personal account of being diagnosed with a rare autoimmune brain condition. It’s a candid, sobering read from a developer whose tools millions rely on.

HN rarely rallies around health stories, but when a beloved open-source maintainer writes one this honestly, the community shows up — both to learn about a poorly understood disease and to wish a respected contributor well.

Uber’s $1,500/month AI limit is a useful signal for AI tool pricing

500 points · simonwillison.net

Simon Willison unpacks news that Uber capped internal AI tool spending at $1,500 per engineer per month, and argues the number is a revealing data point about where real-world AI coding costs are landing. When a company this size sets a ceiling, it tells you something about both consumption and the economics vendors are betting on.

Developers are watching AI tool pricing nervously, and a concrete corporate cap gives the community something tangible to argue over instead of vibes.

Artificial intelligence is not conscious — Ted Chiang

464 points · theatlantic.com

Sci-fi author Ted Chiang, long one of the most thoughtful skeptics of AI hype, makes the case that current systems are not conscious and that conflating capability with sentience is a category error. As always, his argument is more careful than the headline suggests.

Chiang carries enormous credibility on HN, and dropping a consciousness essay the same day as “They’re made out of weights” guarantees a sprawling, philosophical comment thread.

Failing grades soar as professors see greater AI usage at UC Berkeley

357 points · dailycal.org

Berkeley CS professors report a jump in failing grades alongside heavier AI use and what they describe as eroding fundamental math skills among students. The story crystallizes a fear many educators have been voicing: leaning on AI to get through assignments can hollow out the underlying competence.

This hits a nerve for a community full of engineers who learned the hard way, sparking the recurring debate over whether AI is a crutch, a tool, or both.

A Post-Quantum Future for Let’s Encrypt

267 points · letsencrypt.org

Let’s Encrypt lays out its roadmap for issuing post-quantum certificates, preparing the web’s largest free CA for a world where quantum computers could break today’s public-key crypto. Given that Let’s Encrypt secures a huge fraction of the internet, its timeline effectively sets the pace for everyone.

Security-minded readers care because PQC migration is moving from theory to logistics, and the operational details — performance, certificate size, client support — are where the hard problems live.

Mathematicians issue warning as AI rapidly gains ground

246 points · science.org

Science reports on mathematicians grappling with AI systems that are now solving research-level problems faster than expected, prompting both excitement and unease about the field’s future. The “warning” is less doom than a call to figure out what mathematical work looks like when machines can do real proofs.

It’s a natural HN flashpoint: a domain long considered uniquely human-creative facing the same disruption developers are already living through.