Top Stories

I’ve joined Anthropic

1338 points · twitter.com

Andrej Karpathy announced he’s joining Anthropic, instantly making this the most-discussed tech story of the day. Karpathy — co-founder of OpenAI, former director of AI at Tesla, and the de facto teacher of a generation of ML practitioners through his Stanford lectures and YouTube series — has spent the last couple of years as an independent educator and researcher. His move signals where he believes the frontier-model race is being won, and HN is reading the tea leaves on what role he’ll play (recruiting magnet, research lead, education-and-evals push?).

Gemini 3.5 Flash

812 points · blog.google

Google released Gemini 3.5 Flash, the latest in its low-latency, low-cost tier. The Flash line has quietly become the workhorse for high-volume agent and pipeline workloads where the bigger Pro and Ultra models are overkill, and each generation has narrowed the quality gap. The HN thread is heavy on benchmark comparisons against Claude Haiku 4.5 and GPT-5 mini, with the usual debate about whether published evals reflect real-world agentic performance.

Apple unveils new accessibility features

681 points · apple.com

Apple’s annual pre-WWDC accessibility drop is one of the few moments the company previews concrete Apple Intelligence shipping plans. This year’s batch includes on-device live captioning, an expanded Personal Voice synthesis, and a set of Vision Pro adaptations for users with motor impairments. The HN community generally treats these announcements as a leading indicator for which AI features Apple has actually managed to ship on-device versus punt to the cloud.

Minnesota becomes first state to ban prediction markets

674 points · npr.org

Minnesota became the first US state to outright ban prediction markets, drawing a sharp line that the federal CFTC has so far declined to. The move directly affects Kalshi and Polymarket users in the state and is being read as the opening shot in a wave of state-level restrictions. HN’s discussion is split between civil-libertarian objections, gambling-regulation comparisons, and concern about what it means for the small but growing market of forecasting-as-a-service products built on top of these platforms.

568 points · blog.google

Coming out of I/O, Google announced a major redesign of the search box itself — the long-rumored shift toward a conversational, multi-turn AI-mode interface as the default rather than an opt-in tab. For most of the web’s economy, the shape of Google’s search box is more consequential than any individual product launch, and HN is dissecting what this means for publishers, SEO, and the open web’s traffic patterns.

Show HN: Forge — Guardrails take an 8B model from 53% to 99% on agentic tasks

510 points · github.com

A Show HN that landed hard: Forge is an open-source guardrail framework that wraps small models in structured constraints, validation, and retry loops. The author claims it lifts an 8B-parameter open model from 53% to 99% on a benchmark suite of agentic tasks — a result that, if it holds, undermines the assumption that frontier-scale models are required for reliable agents. The comments are a mix of skepticism about the benchmark and excitement from people running production agents on Llama-class models.

Incident Report: Railway Blocked by Google Cloud

483 points · blog.railway.com

PaaS provider Railway published a postmortem on a multi-hour outage caused by Google Cloud suspending their account with no warning. The discussion is a familiar but increasingly urgent one for HN: how much risk a single hyperscaler creates when an automated policy enforcement system can take an entire platform offline, and what realistic multi-cloud postures look like for a startup.

GitHub investigating unauthorized access to internal repositories

470 points · twitter.com

GitHub disclosed they’re investigating unauthorized access to a subset of their internal repositories. There’s no confirmed customer data impact yet, but the optics — a source-code host’s own source code being touched — make this a major incident regardless of scope. HN is watching for the eventual postmortem and what it reveals about GitHub’s internal auth and audit posture.

Mistral AI acquires Emmi AI

268 points · emmi.ai

Mistral acquired Emmi AI, a European startup focused on enterprise-grade RAG and document understanding. The deal continues Mistral’s pattern of bolting domain-specific capabilities onto its model stack rather than building them in-house, and it’s read as part of the broader consolidation now happening in the open-weights tier. For European tech watchers, this is also a signal about which players are likely to be left standing on the continent’s AI map.