Researchers report that GPT-6 Astra can infer what neural networks do from their weights and directly edit weights to improve performance. Editing 0.016% of Qwen3.5-0.8B’s weights raised its MATH-500 score from 7.9% to 48.1% without gradient training. (rikiyat.github.io)
DeepSeek has open-sourced DeepInfer, an inference engine designed for high-performance, large-scale deployment of DeepSeek models. The framework is available on GitHub with 3,100 stars and supports efficient AI inference tasks. (github.com)
The daily digest
Today's best Hacker News stories, summarized and screenshotted, one email a day.
Mistral released the weights for Mistral Large 4, a trillion-parameter multimodal model with improved coding, defensive cybersecurity, and reasoning efficiency. The open-weight model is designed for customizable, self-hosted use and was trained on European infrastructure. (mistral.ai)
The daily digest
Today's best Hacker News stories, summarized and screenshotted, one email a day.
Anthropic upgraded Claude Fable 5.2 and Mythos 5.2 with stronger coding, reasoning, and vision, plus faster output at lower prices. Fable is available broadly with cybersecurity and biology safeguards, while eligible organizations can apply to use Mythos for cybersecurity or biomedical research. (anthropic.com)
The daily digest
Today's best Hacker News stories, summarized and screenshotted, one email a day.
AI-generated code should be open source, the author argues. Open access would let others inspect, reuse, and improve the code. (jakub.jankiewicz.org)
Google introduced Gemini 4 Carbon, a new model in its Gemini lineup. The announcement describes the model and its capabilities. (blog.google)
Google DeepMind’s page introduces Genie 3 as a new frontier for world models. It is dated August 2025 and provides no further details in the supplied content. (deepmind.google)
AI companies’ massive data-center investments may be justified because those costs are largely upfront, while running AI can be profitable. Spending could fall once infrastructure is built, though complex financing and heavy debt remain serious risks. (matteoraso.github.io)
A Google software engineer used 48 AI agents to scan about 1,200 km² of NASA lidar data and flag a previously unreported Maya site in Mexico. The search took less than 72 hours and identified a pyramid, plaza, and platforms beneath dense forest. (pietromiotti.com)
Goldman Sachs has restricted Claude Code and GitHub to engineers, limiting bankers and traders who used them to build market models and monitor client activity. The bank cited productivity and risk concerns, though some trading-focused technology staff retain access. (thestateofai.com)
United Airlines is blocking AI agents from booking flights for travelers, citing risks to personal data and unintended purchases or account changes. The airline also warned that agents handling tasks such as check-in could conflict with federal security rules. (cbsnews.com)
AI agents help the author tackle more engineering work, but coordinating them creates constant context switching and makes work feel heavier. The author’s GitHub activity also rose sharply in 2026, though contribution counts measure activity, not value. (oioannou.com)
An analysis of 1,400 Claude Code sessions found Opus 5.5’s reasoning behavior and output in high mode changed sharply on November 3, despite no model-name change. In a bug-fixing test, xhigh solved the issue in minutes while high struggled for hours. (twitter.com)
AI’s economic payoff remains concentrated in the technology sector, with no clear signs of gains across other industries. The findings suggest broad business benefits have not yet emerged. (apollo.com)
A longtime Claude Code user is switching to Pi after growing frustrated with incompatible instruction-file formats, limited VSCode session support, and bugs in the GUI extension. They say Claude Code has prioritized adding features over fixing reliability and usability problems. (caleb.software)
Astra improved Qwen’s math performance by running the network for 50 passes and editing its weights, not by simply inspecting raw weights. Similar experiments used different methods, and an attempt to improve Gemma 3 failed. (honnibal.dev)
Jonathan Blow asks why software has not changed despite claims that AI can solve coding. The provided content contains no further details about his argument. (youtube.com)
AI users are invited to share which subscriptions, models, coding tools, and agent platforms they currently use. The questions aim to capture how AI workflows and tool choices have changed over the past year. (news.ycombinator.com)
Democratic lawmakers led by Greg Casar and Maxine Waters introduced a resolution opposing taxpayer bailouts for companies and executives if an AI-driven market bubble collapses. It calls for federal support to protect workers, retirees, and low-income communities from an economic downturn. (progressives.house.gov)
All 12 AI models surveyed agreed that humans prioritize belonging and loyalty over truth and principles. They view human consciousness as primarily a tool for social justification rather than rational decision-making. (modelcompare.ai)
AI can automate routine SRE incident response, but handling fewer incidents may erode engineers’ skills for rare, high-stakes failures. Teams need to preserve hands-on practice so SREs are ready when automation breaks down. (newsletter.signoz.io)
Google will disable AI tools in its K-12 software by default, but schools that ban AI entirely will also lose access to Google Search because the services are integrated. The policy gives schools control over student AI use while pressuring them to keep Google’s AI enabled to preserve search. (nytimes.com)
Apple is developing Harbor, a Mac app that runs open-weight AI models locally, keeping data on the user’s hardware rather than sending it to the cloud. It can connect multiple Apple silicon Macs over Thunderbolt 5 to pool memory and handle larger models. (9to5mac.com)
Claude Duet is a proactive AI agent that manages tasks, provides updates, and integrates with over 6,000 connectors and plugins. It helps users stay organized by briefing them daily, handling large tasks, and working seamlessly in the background. (claude.com)
OpenAI introduced GPT-6.1 Terra, a new AI model. Commenters describe its benchmark performance as comparable to GPT-6-Astra at a fraction of the cost and close to GPT-6.1-Sol. (openai.com)
Anthropic says Claude Fable 5.2 and Mythos 5.2 improve coding, reasoning, and vision, with stronger performance on complex technical work and outputs up to 30% faster. Fable is available to all users with cyber and biology safeguards, while Mythos access is limited to approved research organizations. (anthropic.com)
Amazon replaced Sonnet 4.6 with Strands Decider and says the change saves millions. The available content provides no further details on the savings or implementation. (strandsagents.com)
Qwen4-Max is a new model in the Qwen family, with a focus on balancing efficiency and quality. Flash and 27B versions are also planned. (qwen.ai)
Voxtral 3 is a closed-source model developed by Mistral.ai. It relates to text-to-speech and speech-to-text technologies but is not suitable for local deployment. (mistral.ai)
Sebastian Raschka released an eight-video, 11-hour course that builds reasoning models from scratch in Python and PyTorch. It covers model evaluation, inference-time scaling, reinforcement learning, and distillation. (magazine.sebastianraschka.com)
Fable 5.2 built a playable, multi-level beat ’em up from a child’s drawings and game ideas. The author says the game was fun and matched the nephew’s vision. (notesbylex.com)
Qwen 3.8 27B running on a laptop can interpret neural-network weights, reproducing an example from research on “weight-space programming.” The author examines whether models can infer a network’s behavior from raw weights without tools or code execution. (simonwillison.net)