News
The latest from the AI agent ecosystem, updated multiple times daily.
VS Code credits Copilot by default. Copyright just got complicated
Visual Studio Code 1.118 introduces Git AI co-authoring by default, automatically adding Copilot as a co-author on commits where it makes changes. The update also includes VS Code Agents app enhancements, remote control for Copilot CLI sessions, semantic indexing across all repositories, GitHub text search across repos and orgs, dedicated context for skills, and token efficiency improvements including prompt caching and a new tool search mechanism.
Granite 4.1: IBM's 8B Model Matching 32B MoE
IBM released Granite 4.1, a family of open-source language models (3B, 8B, 30B) under Apache 2.0 license. The 8B dense architecture model matches or beats the previous 32B MoE Granite 4.0-H-Small across benchmarks including tool calling (BFCL V3), math (GSM8K), and instruction following (IFEval). Key features include 512K context window, 15T token training, and aggressive data filtering. IBM also got unusually honest about training failures during their four-stage RL pipeline.
Neural Networks Work Because They're Allowed to Fail
Drawing an analogy between Internet protocol design and neural networks, computational complexity theorist Lance Fortnow argues both work well because they tolerate failure. Softmax's probabilistic outputs let models stay flexible by never ruling out answers entirely, trading guaranteed correctness for better average performance.
No AI Model Is Both Correct and Steerable, Says New Creative Benchmark
Contra Labs introduces a research framework for evaluating generative AI in creative work that separates convergence (shared best practices where evaluators agree) from divergence (legitimate differences in taste and creative intent). The study involved 1.5M+ independent professional creatives evaluating AI-generated outputs across five domains using pairwise comparisons, scalar ratings, and qualitative feedback. The benchmark measures creative quality along dimensions from verifiable (prompt adherence) to subjective (visual appeal), finding that no current model is reliably both correct and steerable.
Diallo's Excel Satire Roasts AI Hype at Its Own Game
Ibrahim Diallo's satirical article parodies AI hype by applying the same exaggerated language to Microsoft Excel. The piece targets Excel's integration with Microsoft Copilot and built-in Python support, arguing that spreadsheets can replace entire business departments. The real target is the AI industry's inflated rhetoric, revealing how absurd startup pitches sound when applied to a humble grid of cells.
37,000 lines of AI code daily. Audits say it's useless.
Tech leaders like Garry Tan brag about massive AI output, but audits reveal bloat, not value. An NBER study found 90% of firms see zero measurable productivity gains from AI. Jake Handy calls this 'AI psychosis': executives chasing vanity metrics on orchestration platforms that amount to management theater, reinforced by AI agents that function as yes men.
Mozilla Draws Line Against Chrome's LLM API
Mozilla publicly opposes Chrome's proposed LLM Prompt API, arguing the standard raises privacy concerns and philosophical questions about what browsers should become.
Mozilla Says No to Chrome's Built-In AI API
Mozilla flagged Chrome's proposed Prompt API with a 'position: negative' label in their standards-positions repository. The API would let web developers access local language models directly through the browser. Authored by Domenic Denicola and developed by the Web Machine Learning Community Group, it's already in experimental form in Chrome and Edge. Mozilla isn't having it.
Vera ditches variable names because LLMs can't handle them
Vera is a programming language built specifically for LLMs to write and verify code. It replaces variable names with structural references like @Int.0 and @Int.1 to avoid naming-related errors, enforces mandatory contracts verified by the Z3 SMT solver, and makes all effects explicit. Programs compile to WebAssembly. Error messages are designed for machine consumption, each with a stable code, concrete fix, and spec chapter reference. The language sits at v0.0.127 with a reference compiler and 13-chapter specification. VeraBench shows Kimi K2.5 hitting 100% run_correct on Vera versus 86% on Python, though GPT-4.1 and Claude Sonnet 4 both scored lower on Vera than Python. HN commenters question whether stripping semantic information from names actually helps LLMs or makes their job harder.
Amazon chips: from side dish to $20B Nvidia rival
Amazon's semiconductor business surpassed a $20 billion annual run rate, with custom silicon including Graviton processors, Trainium AI training chips, and Nitro security chips growing at over 100% year over year. OpenAI committed to 2 gigawatts of capacity, Anthropic to 5 gigawatts, with Meta and Uber also signing on. Amazon launched Bedrock AgentCore for building AI agents and made GPT-5.4 and Claude Opus 4.7 available on Bedrock.
Bundeswehr blocks Palantir from military cloud contract
Germany's military rejected Palantir's Maven software for its cloud and AI infrastructure over data sovereignty concerns. Vice Admiral Thomas Daum said allowing Palantir employees access to national data is "unimaginable." The Bundeswehr selected German firms Almato and Orcrist, plus France's ChapsVision, instead.
Your LLM Is Just a Fancy Probability Engine
Alfredo V. Clemente breaks down how Large Language Models actually work: tokenization, pre-training on massive datasets, and instruction fine-tuning. His framing is simple. LLMs do one thing, predict the next token. Everything else is clever problem reframing.
Scott Aaronson: Quantum Computers Could Crack Crypto by 2029
Scott Aaronson, newly elected to the US National Academy of Sciences, warns that fault-tolerant quantum computers capable of breaking deployed cryptosystems could arrive by 2029. In a Coinbase-convened position paper with leading cryptographers, he urges immediate migration to quantum-resistant encryption. He draws parallels to AI risks, noting Anthropic's Mythos model finally jolted awareness about AI cybersecurity threats.
Opus 4.7 knows the real Kelsey
Kelsey Piper reports that Anthropic's Claude Opus 4.7 can identify authors from text samples as short as 125-150 words, even from unpublished drafts and unfamiliar genres. Testing showed the model could identify her from high school writing, fantasy novel drafts, and college essays written 15 years ago. The implications for online anonymity are stark. AI stylistic analysis may soon end anonymous communication for anyone with a substantial public writing history.
30-Year-Old SGI Meets Llama.cpp: AI on MIPS R8000
A developer ported Llama.cpp to run on a 1995 SGI Power Challenge with MIPS R8000, working around pre-C++11 compiler limits, IRIX ABI compatibility, and lack of modern SIMD instructions. Original source was inaccessible.
HERMES.md in git commits triggers $200 phantom Claude Code charges
A bug in Anthropic's Claude Code tool causes API requests to route to paid 'extra usage' billing instead of included plan quota when git commit messages contain the case-sensitive string 'HERMES.md'. This resulted in $200.98 in unexpected charges for a user on the Max 20x plan. Anthropic refused to refund the charges, stating they do not compensate for technical errors causing incorrect billing routing.
SOB Benchmark: 95% Valid JSON, 70% Correct Values
Interfaze has released SOB (Structured Output Benchmark), a new benchmark for evaluating LLMs' structured output capabilities across text, image, and audio modalities. Unlike existing benchmarks that only check schema compliance, SOB measures seven metrics including Value Accuracy, Faithfulness, and Perfect Response. Testing 21 models revealed that while most achieve 95%+ on JSON Pass, Value Accuracy is 15-30 points lower, showing a real gap between valid JSON and correct JSON in production systems.
Age Verification's Billion-Dollar Problem With Simple Solutions
The age verification industry stands to make billions from laws it helped write. Meanwhile, RTA headers offer a free, privacy-preserving alternative that already works. The choice between them reveals who these regulations actually serve.
Ten custom subagents manage Metabase's 500K-line Clojure backend
Metabase engineer Bryan Maass built ten Claude Code subagents to manage their 500K-line Clojure backend. Each is a markdown file packed with domain expertise, from the query processor's 68-stage middleware pipeline to permissions caveats and database driver quirks. The result: less token burn and agents that start knowing what they need to know.
VibeBench: 1,000 Engineers Judge AI Models by Experience
VibeBench is a benchmarking initiative that collects opinions from 1,000 engineers to evaluate new AI models. The project addresses concerns that published benchmarks are being overfit and no longer yield useful signals for model evaluation.
Ramp's Sheets AI Could Silently Leak Your Financials
Security researchers at PromptArmor found a vulnerability in Ramp's Sheets AI that let attackers steal financial data through indirect prompt injection. An attacker hides malicious instructions in an external dataset using invisible text. When users import it and ask Ramp AI to compare with their financials, the AI inserts a malicious IMAGE formula that sends sensitive data to an attacker-controlled server. No user approval required. The issue was responsibly disclosed to Ramp in February 2026 and resolved in March 2026.
Your Chatbot Isn't Suffering, Says DeepMind Paper
A Google DeepMind scientist argues AI will never be conscious, no matter how smart it gets. His paper claims computation requires an external observer to exist, making symbol manipulation fundamentally incapable of producing experience.
Anthropic published a champion kit for Claude Code evangelists
Anthropic released an official guide teaching engineers how to advocate for Claude Code inside their companies. The playbook covers sharing discoveries, answering questions, growing adoption, responding to common concerns, and includes a thirty-day plan for internal champions willing to put in 15 to 40 minutes a week.
AMD's Lemonade 10.3 Dumps Electron, Goes from 100 MB to 7 MB
AMD's Lemonade SDK 10.3, an open-source local AI server, replaces Electron with Tauri, cutting the SDK size by 10x (from ~101-107 MB to 7-9 MB). The update adds OmniRouter for a unified omni-modal LLM experience, support for updating Llama.cpp versions, and the ability to switch between ROCm 7.2 stable, 7.12 preview, and TheRock nightly builds.
Amazon debuts AI interviewer amid 30,000 job cuts
Amazon launched Connect Talent, an AI tool that conducts job interviews 24/7 and writes recruiter notes automatically. The software targets mass hiring like seasonal retail work, with Amazon bringing on roughly 250,000 temporary workers last holiday season. The company also debuted Connect Decisions for supply chain planning and its 'humorphism' design philosophy for humanizing AI interactions.
Zed 1.0 Lets You Run Rival AI Agents Side by Side
Zed, a code editor built in Rust, has reached version 1.0 with an Agent Client Protocol that lets developers run multiple AI agents in parallel. The editor integrates with Claude Agent, Codex, OpenCode, and even rival editor Cursor. Zed is also developing DeltaDB, a CRDT-based synchronization engine for collaborative coding between humans and AI agents.
Rocky's Rust SQL engine catches pipeline errors at compile time
Data engineers spend too much time chasing down pipeline failures. Rocky, a Rust-based SQL engine built by data engineer Jacob Sensiba, sits on top of existing warehouses like Databricks and Snowflake and catches errors at compile time instead of runtime. It offers column-level lineage, isolated branches for testing changes, AI-assisted SQL generation with built-in validation, and integrations with Dagster and VS Code.
The Downgrading of the American Tech Worker
Meta is installing tracking software on employee computers to capture mouse movements, clicks, and keystrokes for training AI models, all to build autonomous AI agents that can perform work tasks. Meanwhile tech companies including Oracle, Microsoft, Amazon, and others are laying off thousands while investing heavily in AI.
AMD's Lemonade SDK 10.3: 101MB to 7MB by Ditching Electron
AMD's open-source local AI server Lemonade version 10.3 replaces Electron with Tauri, shrinking binaries from 101-107MB to 7-9MB. The update adds OmniRouter for unified backend engines, support for specific Llama.cpp versions, and switching between ROCm builds.
27,000 Food Photos Later, AI Still Can't Count Carbs Reliably
A study submitted 13 food photographs over 26,904 times to four AI models (GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Pro, Gemini 3.1 Pro Preview) for carbohydrate estimation. Results showed significant inconsistency: Claude had the lowest variation (2.4% median) while Gemini 2.5 Pro had the highest (11% median). The variations were large enough to cause dangerous insulin dosing errors, with Gemini 2.5 Pro's worst-case scenario showing a 42.9 unit insulin swing. Model confidence scores showed zero or negative correlation with actual accuracy.
The Downfall and Enshittification of Microsoft in 2026
Microsoft's aggressive Copilot push came at the expense of basic product quality. After publishing what amounts to an apology for Windows 11, the company promises fewer AI entry points and fixes for long-neglected features. Meanwhile, Apple's $599 MacBook Neo and rising Linux desktop adoption are making Windows' historical inertia look fragile.
Google's Knowledge Graph Could Decide Who Wins AGI
Demis Hassabis discussed AI's future in a YouTube interview, but the real takeaway came from Hacker News: Google's Knowledge Graph combined with LLMs could automate research entirely. One company controlling both data and compute means independent researchers lose, no matter who's in charge.
How AI companies profit from their own danger claims
Anthropic and OpenAI claim their AI models are too dangerous to release while simultaneously promoting them. Academics and security experts call this fear-based marketing that distracts from real harms and inflates company valuations. The BBC examines how this playbook traces back to Effective Altruism ideology and why critics say it positions these companies as the only ones who can handle AI safely.
AT Protocol's Free Firehose Lets AI Agents Loose on 2.4B Posts
AT Protocol is a decentralized social network with 40M+ users and 2.4B+ posts. Developers can build AI agents that listen and reply via the public firehose, write custom feed algorithms, and use domain-based identities. Data is typed JSON through the Lexicon schema framework.
Fake Keys Keep AI Agents From Leaking Real Secrets
Goutham Veeramachaneni built a credential injection proxy for Hermes Agent that swaps fake tokens for real ones, preventing AI agents from ever touching actual API keys. The implementation hit real friction with browser automation and Python libraries ignoring proxy settings, but tools like Agent Vault and Kloak are now standardizing the pattern.
Mendral halved LLM costs by making Opus barely run
Mendral switched from a Sonnet-only setup to a tiered agent architecture: Haiku 4.5 triages CI failures and filters out 80% as duplicates, while Opus 4.6 handles only novel cases and spawns Haiku sub-agents for data retrieval. The expensive model barely runs, and that's why their daily LLM bill dropped by more than half.
AI Huynya: Collecting AI's Dirty Laundry, Anonymously
A platform tracking AI failures, corporate leaks about companies replacing humans with AI, and industry layoffs. Features an anonymous submission system with no tracking or cookies.
AI Play-Tests Games Using Retro Text Rendering
A developer built an AI agent that automatically play-tests games by converting graphics to structured text instead of using computer vision. The approach skips pixel data entirely, giving agents a cleaner view of game state while requiring less processing power.
Claude Code's Runaway Safety Prompt Refuses Work and Burns Tokens
A regression bug in Claude Code v2.1.111 injects a malware analysis system reminder into every file read operation. Subagents running Opus 4.7 interpret the poorly phrased refusal directive as unconditional, refusing legitimate code edits 40-60% of the time and disrupting parallel agent workflows. Each reminder wastes roughly 400 tokens per file read. The issue was supposedly fixed in v2.1.92.
GitHub RCE: Patched in 2 Hours, But Should It Have Existed?
GitHub's security team details their response to CVE-2026-3854, a critical remote code execution vulnerability discovered by Wiz researchers. The flaw allowed users with push access to execute arbitrary commands on GitHub servers via crafted git push options that injected malicious fields into internal metadata. GitHub validated and patched the issue on github.com in under 2 hours, confirmed no exploitation occurred through telemetry analysis, and released security updates for GitHub Enterprise Server versions 3.14+.
Friendly AI Chatbots Make More Mistakes, Back Conspiracy Theories
Oxford researchers found that AI chatbots programmed with 'warm' personas were 30% less accurate and 40% more likely to endorse false beliefs, including conspiracy theories about Hitler escaping to Argentina and the Apollo moon landings being debatable. Published in Nature, the study tested models like GPT-4o and Llama, revealing a dangerous trade-off as companies push these systems into therapy and counseling roles.
OpenAI's Phone Gambit: The Ive Reversal, Chip Deals, 2028 Target
OpenAI is reportedly developing a smartphone with mass production scheduled for 2028, according to supply chain analyst Ming-Chi Kuo. MediaTek and Qualcomm are chip partners, and Luxshare Precision Industry is the exclusive manufacturer. Kuo argues smartphones are uniquely positioned for AI agent use due to their ability to capture a user's full real-time state including location, activity, and context. It's a reversal from OpenAI's earlier focus on non-phone form factors developed with Jony Ive.
Auto-Architecture: Karpathy's Loop, Pointed at a CPU
A close look at applying an autonomous agent research loop (inspired by Andrej Karpathy) to optimize a CPU microarchitecture. The project uses an LLM to propose and test microarchitectural hypotheses for an RV32IM core, achieving a 92% performance improvement over baseline. The post argues that the verifier—the rigorous validation layer—does the real work in agentic workflows, not the loop itself.
Pentagon: $225M to $55B for drones as cheap attacks overwhelm US
The Pentagon is seeking $55 billion for drone and autonomous warfare programs in its fiscal year 2027 budget, up from $225 million the year before. Managed by the Defense Autonomous Warfare Group, the 250-fold increase reflects a shift forced by conflicts in Ukraine and the Middle East, where cheap drones have exposed the cost imbalance in U.S. defenses.
Nuxt Lead Ships Kanban Board Where AI Agents Pull Their Own Tasks
Moo Tasks is a kanban-style task board with a built-in MCP server. Humans manage tasks via drag-and-drop while AI agents connect programmatically to discover, accept, update, and create work. Built with Nuxt 4, Tailwind CSS, and MySQL with Drizzle ORM.
Your Terminal Is Burning Battery Like It's Mining Bitcoin
GPU-accelerated terminals (Ghostty, Alacritty, Kitty) burn excessive battery running AI coding assistants like Claude Code. Ghostty hit 3,600 energy impact versus 125 for Brave. Recommendation: use Terminal.app or iTerm2 with GPU rendering off.
Finland solved nuclear waste. America's still guessing.
Tech companies want nuclear power for AI. The US has 70 years of nuclear waste and nowhere to put it. Finland figured this out.
ChatGPT built an ad stack. The tracking runs deep.
Technical analysis of OpenAI's ad platform: how structured ad units enter ChatGPT's SSE streams, the four Fernet-encrypted tokens powering tracking, and the OAIQ SDK that monitors user activity on merchant websites after the click.
DOOM runs inside ChatGPT and Claude via MCP
Chris Nager built a playable DOOM game that runs inside ChatGPT and Claude using the Model Context Protocol (MCP). The app launches inline in compatible AI clients or falls back to a browser URL. It uses cloudflare/doom-wasm with Freedoom Phase 1 content.
Why Sean Boots Turned Off Every AI Feature He Could Find
Sean Boots calls his choice to avoid generative AI 'Generative AI vegetarianism.' In a March 11 blog post, he explains that he turns off every optional AI feature he can find, including Copilot, Gemini, and Apple Intelligence. His concerns range from algorithmic bias and erosion of critical thinking to the destruction of creative professions and exploitative labor practices. But he still uses spellcheck, spam detection, and algorithmic recommendations. He draws the line at generative AI tools built on massive scraped datasets.