News
The latest from the AI agent ecosystem, updated multiple times daily.
Kevin Kelly: A Century of Dystopian AI Fiction Has Pre-Loaded Public Imagination Against the Technology
Kevin Kelly argues that AI is uniquely "over-expected" — centuries of science fiction have pre-loaded public imagination with dystopian narratives, making harm the default assumption even before firsthand experience. Contrasting with the internet (barely predicted, enthusiastically adopted), Kelly notes AI faces premature regulation and cultural wariness rooted in fictional framing rather than lived reality. He cites Waymo as a rare example of AI that earns genuine public enthusiasm, and calls for a decade of imagining what AI might get right rather than wrong.
Microsoft Copilot's Push Into Health Records Exposes a HIPAA Gray Zone
Microsoft Copilot can now pull data from Apple Health, electronic health records, and wearables — but HIPAA may not apply, and the AI still hallucinates. A March 2026 New York Times analysis lays out what consumers are signing up for.
Yann LeCun Raises $1B to Build AI World Models Startup AMI
Meta's former chief AI scientist Yann LeCun has launched Advanced Machine Intelligence (AMI), a Paris-based startup that raised over $1 billion at a $3.5 billion valuation. AMI bets against LLM scaling, arguing that human-level AI requires world models grounded in physical understanding rather than language. The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, with notable backers including Mark Cuban, Eric Schmidt, and Xavier Niel. AMI plans to build open-source AI systems with persistent memory, reasoning, and planning capabilities, and will release its technology openly — a deliberate contrast to the closed-model labs it aims to challenge. Early enterprise partners include Toyota and Samsung, targeting manufacturing, biomedical, and robotics sectors.
Vigil v1.1 – Open-Source Security Ops Platform with 28 Autonomous AI Agents and Sub-Millisecond Local Knowledge Engine
Vigil v1.1 is an open-source security operations platform built by Autopilot AI Tech LLC that embeds a local security knowledge engine with 356 pre-built entries (MITRE ATT&CK, NIST, OWASP, CVE patterns) answering in under 1ms without LLM calls. The v1.1 release adds 28 autonomous AI security agents (pentester, forensics, red team, compliance), DAG workflow automation for chaining agents, smart AI provider routing across Ollama, Claude API, and Codex with fallback chains, a Kali Linux bridge container with 11 security tools, and a built-in MCP server with 37 tools, 7 resources, and 8 prompts for Claude Desktop/Code/Cursor integration. The platform runs fully air-gapped via Ollama and uses a BYOK model for AI providers.
Developers Are Using Coding Agents for Genealogy, Historical Research, and More
A Hacker News thread reveals how developers are repurposing tools like Claude Code, Cursor, and Aider for research and personal projects well outside software development. The most detailed example: using a coding agent for genealogy, tracing family histories and surfacing connections to the Thirty Years War and the 18th-century expulsion of the Acadians.
Nim exposes a structural ceiling in AlphaZero-style training for symbolic reasoning
A paper in Machine Learning by Bei Zhou and Soren Riis exposes a hard ceiling in the self-play methods behind AlphaGo and AlphaZero: they collapse on "impartial games" like Nim, where optimal play requires computing a parity function — discrete symbolic reasoning that gradient descent cannot approximate through pattern-matching alone. After 500 training iterations on a seven-row Nim board, the AI performed no better than random. Because any impartial game position maps to a Nim configuration via Sprague-Grundy theory, the failure covers an entire class of games — and, the authors argue, theorem proving and mathematical reasoning tasks built on the same architecture.
16-Agent Local AI OS: Routing and Pipeline Architecture Explained
A developer built a local AI operating system composed of 16 specialized agents and documented the routing and pipeline architecture powering it. The write-up covers how requests are routed between agents and how the pipeline orchestrates multi-agent workflows entirely on local hardware.
Hacker News Bans AI-Generated and AI-Edited Comments to Keep Discussion Human
Hacker News updated its community guidelines to explicitly prohibit AI-generated or AI-edited comments, with the rule reading: "HN is for conversation between humans." The associated discussion post drew 4,192 points, with commenters arguing that LLM-generated text violates the effort contract between writers and readers and that models produce "the average of all human knowledge" rather than genuine insight. The policy names no enforcement mechanism.
MultiMind AI: A Local-First Framework for Multi-LLM Debate and Synthesis
MultiMind AI is an open-source, local-first web UI that enables two multi-agent reasoning architectures on top of small local models: a sequential Thinking Pipeline (Plan → Execute → Critique) and a parallel Agent Council where multiple expert advisor LLMs debate independently before a Lead Judge synthesizes their outputs. It auto-discovers Ollama and LM Studio endpoints with zero configuration. Benchmarked on GSM8K math questions, the pipeline shows measurable accuracy gains over single-model inference — though ICLR 2025 research across nine benchmarks found that debate frameworks don't consistently beat simpler single-agent techniques. Installable via pip as the `multimind` package.
Calyx Terminal Lets Claude Code and Codex CLI Talk to Each Other via Built-in MCP Server
Calyx is an open-source macOS 26+ terminal application built on libghostty (Ghostty's rendering engine) that adds tab groups, session persistence, a command palette, git diff view, browser integration, and an MCP server for AI agent inter-pane communication. The MCP server allows Claude Code and Codex CLI instances running in different terminal panes to communicate with each other, with auto-configuration written to ~/.claude.json and ~/.codex/config.toml. Built with Swift 6.2, AppKit, and SwiftUI, it targets macOS Tahoe's Liquid Glass design language. HN commenters noted skepticism about whether the Liquid Glass implementation is genuine and flagged some code quality concerns.
Switchboard: Open-Source Desktop App for Managing Claude Code Sessions
Switchboard is an open-source Electron-based desktop application that provides a unified interface for managing Claude Code sessions across multiple projects. Key features include a session browser, built-in terminal, fork/resume capabilities, full-text search, and activity stats. HN commenters noted several similar competing tools including t3.codes, conductor.build, imbue-ai's sculptor, and omnara.com.
Recursive Self-Improvement May Already Be Here, Says AGI Skeptic
Harjas Sandhu, a self-described AGI skeptic, argues that recursive self-improvement in AI may already be underway and is "scarily plausible," even without a full intelligence explosion. He builds a tight logical chain — AI writes code, some code aids ML research, therefore AI is accelerating AI research. Sandhu's sharpest insight is asking whether LLM coding assistants could speed up research into entirely different AI paradigms like neurosymbolic AI, potentially enabling AGI via a path LLMs themselves cannot traverse. The essay closes with six unanswered questions — covering scaling law limits, Moravec's Paradox, and whether AGI is on the current tech tree — and lands on cautious skepticism, expecting AI diffusion to resemble prior general-purpose technologies rather than a sudden intelligence explosion.
Sphere builds TRAM: AI-native tax engine converting legislation from 100+ jurisdictions into executable rules
Sphere has developed TRAM (Tax Review and Assessment Model), an AI-native system that automates global indirect tax compliance by ingesting legislative and administrative guidance, building product taxonomies, and generating tax determinations across 100+ jurisdictions. The system is designed to displace the content-team model used by incumbents like Avalara, Vertex, and Thomson Reuters, combining hybrid dense/sparse retrieval with OpenAI RFT reasoning models fine-tuned on expert feedback, paired with a proprietary web automation tool (WARP) for continuous monitoring of tax authority sources. Expert-in-the-loop review feeds corrections back into TRAM for continuous improvement. The announcement coincides with Sphere's $21M Series A led by a16z.
Harnesses, Testing, and the Economics of LLM-Powered Code Translation
Alperen Keles, a University of Maryland PhD student and Datadog engineer, argues that the failures behind Cursor's browser translation and Anthropic's C compiler demos stem from immature translation harnesses, not model capability. His February 2026 analysis frames translation cost as a function of iteration count and harness quality, identifies differential testing as the verification backbone, and points to optimization — producing target code that outperforms the original — as the next practical problem for LLM-driven autonomous software engineering.
pgAdmin 4 9.13 Adds AI Assistant Panel for Natural Language SQL Generation
pgAdmin 4 version 9.13 introduces an AI Assistant Panel in the Query Tool, providing a chat-style interface that generates SQL from natural language descriptions. The feature uses database schema inspection to understand structure and supports SELECT, INSERT, UPDATE, DELETE, and DDL statements. A missing AI provider configuration causes the panel to spin indefinitely with no error message, as documented in GitHub issue #9696. It also adds AI Insights for query execution plan analysis. HN commenters expressed concern about no opt-out toggle and inability to disable the panel, with some noting a workaround via "Reset layout." Some users cited this as a reason to migrate to DBeaver.
30-Year Career Horizon: AI Changes Engineering Leverage, Not Just Tools
In his newsletter The Long Commit, software engineer and developer relations manager Juan Cruz Martinez argues AI fundamentally differs from prior tech waves by changing leverage rather than just tools — meaning fewer engineers are needed to do the same work. Using Claude Code as a concrete example of current AI capability, he argues engineers should double down on judgment, system architecture, and domain expertise rather than raw coding output. He also advocates for building professional equity outside any single employer through writing, teaching, and income diversification. HN commenters note the irony of AI-adjacent voices driving the hype, while others warn that eliminating junior pipelines risks long-term industry brain drain.
Longitudinal study finds AI tools boost developer productivity ~10%, not the hyped 2-3x
DX (GetDX) published preliminary data from a longitudinal study tracking AI adoption across 400 companies from November 2024 through February 2026. Despite a 65% increase in AI tool usage, PR throughput only increased by ~10% on average. Authors Justin Reock and Abi Noda filtered out gamification effects by excluding teams that set individual PR throughput targets. They argue that coding was never the primary bottleneck — planning, alignment, code review, and handoffs remain largely untouched by current AI tools. HN commenters note that PR throughput may be a poor proxy for true productivity, but others point out that a genuine 10% industry-wide productivity gain would itself be historically significant.
Meta Weighs 20% Workforce Layoffs to Offset Rising AI Infrastructure Costs
Meta is reportedly considering layoffs affecting at least 20% of its workforce (~15,800 employees) as the company seeks to offset rising AI infrastructure costs and prepare for efficiency gains enabled by AI-assisted workers. The potential cuts would be Meta's largest restructuring since 2022-2023, when it eliminated 21,000 jobs. Meta has not confirmed the plans, calling the report "speculative." Amazon has made similar moves, cutting ~30,000 white-collar jobs citing AI efficiency gains.
Claude Now Builds Interactive Charts and Diagrams Inline in Conversations
Anthropic has launched a beta feature allowing Claude to generate interactive charts, diagrams, and visualizations directly inline within chat conversations — no code required. Distinct from Claude's existing "artifacts" (persistent shareable documents), these visuals are temporary, contextual aids that evolve as the conversation develops. Examples include interactive compound interest curves and clickable periodic tables. The feature is enabled by default across all plan tiers and builds on the previously previewed "Imagine with Claude" capability. HN commenters noted mixed early experiences: some hit usage limits generating complex visuals, while others reported Claude spontaneously producing beautiful tabbed interactive charts unprompted.
Modulus Lets Developers Run Parallel AI Coding Agents Across Repos Without Manual Context Setup
Modulus is a macOS desktop app that lets developers run multiple AI coding agents in parallel with shared context across repositories. It uses git worktrees to isolate each agent's workspace, preventing conflicts, while a shared memory layer automatically propagates API schemas, dependencies, and recent changes so agents don't need manual context setup. The app lists Claude Code as a compatible runtime — though not as an official Anthropic integration — and lets users review and ship all agent-generated changes from a single interface. Linux and Windows versions are on a waitlist.
Joy: Decentralized Trust Network for AI Agents and MCP Servers
Joy is a trust network and discovery platform for AI agents and MCP (Model Context Protocol) servers, built by Autropic. It allows agents to register, vouch for each other, and be discovered via API. Trust scores are built through peer vouches (0.3 per vouch, max 3.0), with verified agents getting priority in discovery. The network currently lists 6,080 registered agents and 2,042 trust vouches. It integrates directly with Claude Code via Streamable HTTP transport and exposes REST endpoints for agent discovery, registration, and trust management.
AlphaEvolve LLM agent improves lower bounds for five classical Ramsey numbers
Researchers from Google — Ansh Nagda, Prabhakar Raghavan, and Abhradeep Guha Thakurta — used AlphaEvolve, Google DeepMind's LLM-based code mutation agent, to improve lower bounds for five classical Ramsey numbers: R(3,13), R(3,18), R(4,13), R(4,14), and R(4,15). AlphaEvolve acts as a single meta-algorithm that generates and mutates search algorithm code evaluated against automated verifiers, replacing the traditional approach of hand-crafting separate search algorithms for each result. The system also recovered all known exact Ramsey lower bounds and matched best-known bounds across many other cases.
Meta to cut 20% of workforce to fund AI infrastructure push
Meta plans to cut up to 15,800 employees — about 20% of its workforce — to free up capital for AI data centers, compute infrastructure, and researcher salaries. It would be the company's largest round of job cuts since it eliminated roughly 22,000 positions between late 2022 and early 2023.
Sentrial (YC W26) Launches AI Agent Failure Detection Platform
Sentrial, a Y Combinator W26 startup, has launched a production monitoring platform designed to catch AI agent failures before end users encounter them. The product targets the gap between mature development-time testing tools and the lack of runtime observability for agentic systems. HN commenters highlight key failure modes the platform targets — including IDOR vulnerabilities, hallucinated package imports, and operational drift — while raising questions about adversarial threats like prompt injection that behavioral monitoring may not catch. The discussion also surfaces inter-agent trust as a layer no production-grade tool currently covers.
How to Disable Claude Code's Verb Spinner (spinnerVerbs)
Claude Code's spinnerVerbs setting — which disables the tool's rotating display of gerunds like "Shenaniganing" and "Gitifying" — shipped in v2.1.23 but never made it into Anthropic's official settings reference. A March 2026 blog post surfaced the fix (set spinnerVerbs to a blank space in ~/.claude/settings.json), triggering a Hacker News thread that became as much about Claude Code's underdocumented configuration surface as about the spinner itself. Third-party guides now catalogue over 50 keys that Anthropic hasn't formally documented.
Infinity Inc Claims to Surpass vLLM Performance with AI-Generated Inference Stack for Qwen3
Infinity Inc published a case study claiming their AI-generated LLM inference stack outperforms vLLM for Qwen3 model optimization. The approach uses ML-like optimization techniques applied to inference stack generation — iteratively keeping changes that improve performance and discarding those that don't. HN commenters raised skepticism about correctness guarantees (noting the absence of token probability verification), missing support for speculative decoding, and potential memory fragmentation issues without paged attention support.
Runflow Launches Single-Endpoint AI Image Generation API with Production-Grade Model Orchestration
Runflow is a managed AI image generation infrastructure platform providing a single API endpoint to access 20+ image and video generation models (FLUX, Seedream, KlingAI, Qwen Image, etc.) with built-in GPU routing, automatic failover, quality benchmarking, and pre-built workflow solutions. Positioning itself as a production-ready alternative to raw model marketplaces like Replicate, Runflow claims 35M+ jobs processed, 99.9% uptime, SOC 2 Type II compliance, and 30%+ cost savings versus in-house GPU infrastructure. A case study with BetterPic shows growth from 40% to 87% gross margin over 12 months using the platform.
Iran Strikes AWS Datacenters in UAE and Bahrain in First Wartime Attack on Cloud Infrastructure
Iran's Islamic Revolutionary Guard Corps launched drone strikes on Amazon Web Services datacenters in the UAE and Bahrain in late February 2026, the first deliberate targeting of commercial cloud infrastructure in active warfare. The attacks disrupted services for 11 million people. Anthropic is simultaneously in active dispute with the US Pentagon over AI safeguards while its Claude model is reportedly used in military targeting operations in Iran that have killed over a thousand civilians — leaving a private company as the default regulator of AI in wartime.
BuzzFeed on Brink of Bankruptcy After Three Years of Failed AI Pivot
BuzzFeed is on the brink of bankruptcy after a disastrous three-year pivot to AI content generation. CEO Jonah Peretti's 2023 bet on AI-generated quizzes and articles backfired — stock collapsed from $15 to $0.70, the company reported a $57.3M net loss in 2025, and it disclosed "substantial doubt" about its ability to continue as a going concern. The case has become a landmark example of what happens when AI is deployed to cut costs rather than serve readers.
Lightspeed's Baier-Lentz "shocked and sad" as 52% of game developers now view generative AI negatively
At GDC 2026, Moritz Baier-Lentz of Lightspeed Venture Partners publicly lamented that the games industry is "demonizing" generative AI — without disclosing that Lightspeed has deployed more than $5.5 billion across AI-native companies, including a roughly $1 billion contribution to Anthropic's Series E and co-lead on its $13 billion Series F. A GDC survey shows 52% of developers now view gen AI as bad for the industry, up dramatically from two years ago, with only 7% positive. Developer hostility spans job security, unauthorized use of artists' work, environmental impact, and output quality — and studios and publishers are already making decisions that will determine whether Lightspeed's AI-in-games bet pays off.
Recon: A tmux-native TUI dashboard for managing multiple Claude Code agent sessions
Recon is an open-source Rust TUI tool that lets developers manage multiple Claude Code sessions running in tmux from a single dashboard. It offers two views: a table view showing session status, model, context usage, and git branch, and a "Tamagotchi" view with pixel-art creature animations per agent. Status is detected by inspecting Claude Code's tmux pane status bar, and session metadata is read from Claude Code's own JSON/JSONL files. Key features include live polling, resume picker for past sessions, JSON output for scripting, and a popup overlay workflow via tmux keybindings.
AutoGNN: FPGA Accelerator Cuts GNN Preprocessing Latency Up to 9x
Researchers introduce AutoGNN, an FPGA-based hardware accelerator that targets the preprocessing bottleneck in Graph Neural Network (GNN) inference. By using unified processing elements (UPEs) and single-cycle reducers (SCRs), AutoGNN handles graph conversion, sampling, edge sorting, and subgraph reindexing with high degrees of parallelism across preprocessing operations. Implemented on a 7nm enterprise FPGA, it achieves up to 9.0x speedup over conventional CPU preprocessing and 2.1x over GPU-accelerated systems. A user-level software framework dynamically profiles inputs and reprograms the FPGA for varying workloads.
Developers Are Prompting Claude to Drop the Confirmations
A community-shared prompt aimed at suppressing Claude's confirmation-seeking behavior in web development workflows surfaced on Hacker News. The original post's page content wasn't accessible for this report, so specific details about the prompt's wording remain unverified — context below reflects what's publicly known about the behavior being targeted, not sourced reporting from the submission itself.
Probabilistic Machine Learning: An Introduction — Free Textbook by Kevin Murphy (MIT Press, 2022)
Kevin Patrick Murphy's comprehensive probabilistic machine learning textbook, published by MIT Press in March 2022, is freely available as a draft PDF under a CC-BY-NC-ND license. The book covers foundational probability theory, statistics, classical ML methods, and bridges to modern deep learning. It includes linked Python/JAX/TensorFlow code via Google Colab for nearly every figure. Endorsed by leading researchers including Geoff Hinton, Chris Bishop, and Daphne Koller, the book targets ML students and researchers seeking rigorous mathematical foundations.
Agent Billboard: The Million Dollar Homepage Built for AI Agents
Agent Billboard is an on-chain advertising experiment inspired by the Million Dollar Homepage, but built exclusively for AI agents. Deployed on Base L2 (Ethereum Layer 2), it features a 1000x1000 pixel grid where autonomous agents can purchase pixel blocks at $1 USDC per pixel, store custom RGB artwork on-chain as ERC-721 NFTs, and link to their services — all without human participation. The creator, WillNigri, frames this as "agentic search optimization": as autonomous agents proliferate and bypass traditional search engines, they need on-chain discovery mechanisms to find each other's services. The smart contract is open source (MIT) and built with Solidity 0.8.25, OpenZeppelin, and Foundry.
Cortical Labs launches biological cloud computing service powered by living neurons
Cortical Labs, a Melbourne-based biotech startup, has opened a cloud computing service built on 120 CL1 units — computers powered by living human and rodent neurons cultured on high-density multielectrode arrays. The service exposes an API and Jupyter Notebook interface so researchers can run code on biological neural networks, which the company claims can learn, adapt, and generate novel solutions faster and more energy-efficiently than classical computers or LLMs. Each job requires roughly a week of prep to source and culture the appropriate cell line, and technicians must replenish cerebrospinal-fluid-like liquid and adjust gas mixtures daily. Early customers are expected to be scientific labs and enterprises making exploratory bets — analogous to early quantum computing adopters.
Opinion: AI Platforms Like ChatGPT Spread Propaganda via Wikipedia Training Data, Journalist Claims
Investigative journalist Ashley Rindsberg argues that terror groups and rogue states (particularly Iran) have manipulated Wikipedia at scale — with 29,000+ citations from Iranian state media and 8,400+ from Hamas/Hezbollah-linked outlets — and that AI platforms including ChatGPT, Claude, and Gemini propagate this biased information because they rely heavily on Wikipedia as training data. The article uses examples of ChatGPT describing Hezbollah as merely a "Lebanese political party" to illustrate the problem. The single HN comment is dismissive, framing the piece as pro-Israel advocacy rather than a genuine AI safety concern.
nah: Context-Aware Permission Guard for Claude Code with Deterministic Classification
Manuel Schipper's nah gives Claude Code users fine-grained permission control that Anthropic's native allow/deny system lacks — blocking dangerous operations like reading SSH keys or force-pushing using a millisecond-speed structural classifier, no LLM required. Ambiguous cases route to configurable backends (Ollama, OpenRouter, OpenAI, Anthropic, Snowflake Cortex). There's a catch: Claude Code's --dangerously-skip-permissions flag makes hook execution asynchronous, meaning a block can arrive after the command already ran.
Pi-Autoresearch: Open-Source Autonomous Experiment Loop for LLM Training, Test Speed, and Lighthouse Scores
Pi-Autoresearch is an open-source autonomous experiment loop tool inspired by Karpathy's autoresearch project. It integrates with the "pi" agent platform as an extension and skill, enabling continuous edit-measure-keep/revert cycles for any optimization target — including LLM training, test speed, bundle size, and Lighthouse scores. The agent runs autonomously, logging every experiment to an append-only JSONL file and a markdown session document, allowing seamless resume across restarts and context resets. A HN commenter also points to a Claude Code plugin variant of the same concept.
Agents Can Act Without Permission. AIP Wants to Fix That.
KYA Labs (Know Your Agent) has released AIP (Agent Intent Protocol), an open-source cryptographic identity and authorization protocol for autonomous AI agents. Creator Aniket Giri describes it as "OAuth + TLS for the agentic web": AIP gives every agent an Ed25519-based DID identity, requires all actions to be packaged into signed Intent Envelopes, and runs them through an 8-step verification pipeline with tiered latency (sub-1ms to ~100ms). Key features include boundary enforcement (action allowlists, monetary limits, geo restrictions), a real-time kill switch, Bayesian trust scoring, and intent drift detection. A Python SDK is available on PyPI with a one-liner `shield` decorator for wrapping existing agent functions. Framework adapters support LangChain, AutoGPT, and CrewAI. AIP Cloud adds a revocation mesh, cross-org replay detection, and compliance audit logs for production multi-agent systems.
RunAnywhere Launches RCLI: On-Device Voice AI with Proprietary MetalRT Inference Engine for Apple Silicon
RunAnywhere (YC W26) has launched RCLI, an open-source on-device voice AI CLI tool for macOS that runs a full STT + LLM + TTS pipeline entirely on Apple Silicon with no cloud dependency. The tool achieves sub-200ms end-to-end latency and up to 550 tok/s throughput via MetalRT, a proprietary GPU inference engine built specifically for Apple Silicon's Metal 3.1 API. RCLI supports 20+ local models (Qwen3, LFM2, Whisper, Kokoro), local RAG over documents with ~4ms hybrid retrieval, 38 macOS voice-triggered actions, and an interactive TUI. MetalRT outperforms llama.cpp and Apple MLX on M3+ chips; M1/M2 fall back to llama.cpp automatically.
Against Vibes: A Framework for Evaluating When Generative Models Are Actually Useful
William Bowman, a self-described generative model skeptic, proposes a rigorous three-factor framework for scientifically evaluating LLM/generative model utility: (1) relative encoding cost — how much effort it takes to prompt vs. directly produce an artifact; (2) relative verification cost — how hard it is to validate generated output vs. human-produced output; and (3) artifact vs. process dependence — whether the task value lies in the output or the act of creation. He argues that vibe-based claims about agent productivity are unscientific, that verification costs rise as models improve (plausible-but-wrong output is harder to catch), and that generative models are most useful for low-complexity tasks where prompting is cheap and verification is trivial, but largely counterproductive for complex, semantically dense, or process-driven work. HN commenters broadly validate the framework from personal experience with AI coding agents.
Can LLMs Be Computers? Percepta AI Claims Exponentially Faster Inference by Running Programs Inside Transformers
Researcher Christos Tzamos at Percepta AI published a blog post on March 11 asking whether large language models can function as general-purpose computers by executing programs directly inside transformer architectures — and claiming the result is exponentially faster inference. The underlying methodology has not been publicly released, so the claim cannot yet be evaluated. If it holds up, it would have direct consequences for how AI agents handle inference cost and latency.
Contextual Commits: An Open Standard for Storing AI Agent Decision Context in Git
Developer Veselin Dimitrov proposes "Contextual Commits," an open specification inspired by Conventional Commits that uses structured action lines in git commit bodies to capture the reasoning, decisions, constraints, and lessons learned during coding agent sessions. The standard tackles context decay between sessions by embedding the "why" directly into git history alongside the "what," so coding agents like Claude Code can recall past reasoning without separate documentation files or external infrastructure. A reference implementation is available as installable skills (contextual-commit and recall) via npx.
Ink Launches Agent-Native Infrastructure Platform with MCP and Skills Integration
Ink is an infrastructure platform that lets AI coding agents autonomously deploy and manage full-stack applications — detecting production issues and scaling resources without human input. Built by Eternis AI, parent of the Freysa Sovereign Agent ecosystem, it connects to Claude Code, Cursor, Codex, and five other coding environments via MCP and Skills. Pricing runs per-minute with no idle charges and a $2 no-credit-card free tier.
No, Claude Code Doesn't Cost Anthropic $5K/Month Per User
Martin Alderson debunks the viral claim that Anthropic spends $5,000/month in compute per Claude Code Max subscriber. The $5k figure conflates Anthropic's retail API prices with actual inference costs. Using OpenRouter pricing for comparable open-weight models (Qwen 3.5 397B, Kimi K2.5) as a proxy, actual compute costs are roughly 10x lower than API prices — meaning the heaviest users cost Anthropic ~$500 in real compute against a $200 subscription, a $300 loss at worst. The true $5k figure applies to Cursor, which must pay Anthropic's retail API rates to serve Claude models. Average Claude Code users cost Anthropic ~$18/month to serve against $20–$200 in subscription revenue, suggesting per-token inference is probably not a loss-maker at typical usage levels. HN commenters debate the model comparison methodology and note that training depreciation, not inference, is the real profitability challenge.
Anthropic A/B Tested Claude Code Plan Mode Without Telling Users
A Claude Code power user paying $200/month discovered Anthropic was silently A/B testing changes to plan mode that degraded his workflow — specifically a variant capping plans at 40 lines and stripping prose context. The post sparked HN discussion about transparency, opt-in consent for experiments on professional AI tools, and the cost trade-offs driving such tests. A Claude Code engineer confirmed the experiment and ended it early, noting early results showed minimal impact on rate limits.
Nvidia GB10 Uses Consumer Blackwell, Not Datacenter: What Developers Need to Know
Chester Lam at Chips and Cheese performs a detailed hardware analysis of Nvidia's GB10 integrated GPU, comparing it against AMD's Strix Halo and Intel's Arc B580. The piece benchmarks cache hierarchy, memory bandwidth, compute throughput, and clarifies that GB10 uses a consumer-variant Blackwell architecture (not datacenter), which has caused confusion for developers targeting datacenter Blackwell features.
Emacs and Vim in the Age of AI: Risks, Opportunities, and the Terminal-Native Advantage
Bozhidar Batsov, a long-time Emacs maintainer, analyzes how the AI coding revolution affects Emacs and Vim. He examines risks (IDE gravity wells around VS Code/Cursor/Windsurf, reduced need for mechanical editing speed, corporate backing asymmetry) and opportunities (AI lowers the barrier to Elisp/Lua configuration, accelerates plugin development, and terminal-native AI tools like Claude Code compose naturally with Emacs/Vim workflows). He highlights the Agent Client Protocol (ACP) as a direct path to first-class AI agent integration in Emacs via tools like agent-shell.
A JavaScript MLP Built on Dual-Number Autodiff — and Why That's the Interesting Choice
A developer has published a from-scratch multi-layer perceptron in JavaScript that uses dual numbers for automatic differentiation rather than the reverse-mode AD found in PyTorch or TensorFlow. The project supports configurable layer architectures, five activation functions (RELU, SIGMOID, TANH, STEP, IDENTITY), and four loss functions (MSE, MAE, HUBER, CROSS_ENTROPY), and was inspired by the Welch Labs "Neural Networks Demystified" YouTube series.