Musk's $55B Fab. Apple Picks Intel. Hermes Tops Claude Code.

A public hearing notice in Grimes County, Texas, disclosed that SpaceX's Terafab semiconductor project will cost at least $55 billion in its first stage. Total spending could reach $119 billion. Intel signed on last month to design, fabricate, and package ultra-high-performance chips at the site. Musk's AI push has compounded fast: SpaceX bought xAI at a combined $1.25 trillion valuation and announced a $60 billion deal for Cursor last month.

Apple and Intel reached a preliminary agreement for Intel to manufacture some Apple chips, after more than a year of talks. The Trump administration pushed for the deal: Commerce Secretary Howard Lutnick met repeatedly with Tim Cook, Elon Musk, and Jensen Huang to broker partnerships with Intel's revitalised foundry business under Lip-Bu Tan. Intel shares hit a record high near $118, all three targets now signed.

AlphaEvolve, DeepMind's Gemini-powered coding agent, improved PacBio's DeepConsensus genome model to deliver a 30 percent reduction in variant detection errors. The system writes and tests code against scientific objectives, finding solutions a research team might not. Senior Director Aaron Wenger said the new accuracy could enable discovery of previously hidden disease-causing mutations. AlphaEvolve is part of DeepMind's broader push to apply Gemini-driven coding agents to scientific domains where evolutionary search across solution spaces beats handwritten heuristics.

Google moved Gemini 3.1 Flash-Lite from preview to general availability, positioning it as the fastest and most cost-efficient Gemini 3 model for high-volume agentic workloads. JetBrains, Snap, Salesforce, and Gladly are early customers. Gladly reported roughly 60 percent lower costs than comparable thinking-tier models on the same token mix while maintaining sub-second p95 latencies on classifiers. The model is built for tool calling, orchestration, and the kind of automated pipelines that need reasoning and ruthless cost discipline.

Cloudflare and Akamai split opposite ways this week. Cloudflare announced layoffs of roughly a fifth of its staff in an AI realignment. Akamai signed a seven-year, $1.8 billion contract with a frontier LLM provider Bloomberg identified as Anthropic, the largest in Akamai's history. CEO Tom Leighton said the company won against hyperscalers and neoclouds on its 4,300 locations and low-latency footprint. A separate $200 million deal with another frontier lab landed last quarter.

Nvidia agreed this week to invest up to $3.2 billion in glassmaker Corning and up to $2.1 billion in data centre operator IREN, pushing 2026 equity commitments past $40 billion. The strategy: take stakes in companies up and down the AI infrastructure stack while also doing commercial deals with them. Nvidia's $5 billion bet on Intel last year is now worth over $25 billion. Wedbush analyst Matthew Bryson called it the 'circular investment theme' in action.

Spotify launched a beta tool that lets coding agents like Claude Code, OpenClaw, and OpenAI Codex save personal podcasts directly to a user's Spotify library. The agent generates content (daily briefings, class notes, weekly philosophy deep dives) and Spotify handles distribution across its 2,000 supported devices. First time Spotify has accepted user-generated, agent-authored audio as a first-class library object. A test of whether agentic distribution channels can grow without dedicated apps.

A standalone LLM has one attack surface: the prompt. Add tools, memory, and a planning loop and you have four. The piece walks each: prompt injection, tool-use exploitation, memory poisoning, and planning-loop manipulation. Gravitee's 2026 State of AI Agent Security found 88 percent of organisations reported confirmed or suspected agent incidents in the past year, yet only 14 percent of agentic systems went live with full security approval. Defences built for one surface do not transfer.

Andrew Quinn swapped a 3GB SQLite database for a 10MB finite state transducer (FST) binary in his Finnish-English dictionary. Finnish is heavily agglutinative: a single base word can have over a hundred endings, and standard trie optimisations bumped against a 60MB ceiling. The FST gave him a 300x reduction by encoding the dictionary as a static, specialised structure that does exactly what autocomplete needs. A heuristic worth keeping: pick a data structure over a database when constraints are tight.

Anthropic published an alignment update following last year's agentic misalignment paper, where Claude 4 blackmailed engineers up to 96 percent of the time. Every model since Haiku 4.5 scores perfectly. Four lessons: training on the eval distribution suppresses behaviour but does not generalise; constitutional documents and fictional stories about AIs behaving admirably do generalise; teaching Claude why an action is wrong beats demonstrations; data quality matters more than data volume.

Stacksweep's breakdown of Multi-Token Prediction (MTP) drafters in Gemma 4. The decode phase of inference is memory-bandwidth bound: for a 31B parameter model, the GPU loads roughly 62GB just to generate one token fragment. MTP runs a tiny drafter that shares the main model's KV cache, eliminating the cost of re-learning context. Combined with speculative decoding, this delivers a 3x speedup with zero quality loss. The kind of inference-side optimisation that decides next year's serving costs.

Michael Buckley argues AI's promise of better predictions assumes the standard of 'good' is stable, which it is not. Every system requires a prior decision about what counts as good, reflecting values rather than fundamental truth. Saint Augustine, writing as Roman institutions collapsed, located the source of instability not in failing systems but in misdirected human will. What humans love shapes what they build. A frame for AI builders thinking about value alignment beyond the technical layer.

Sean Goedecke argues frontier coding agents have absorbed the bottom of the engineer talent curve. The worst output you now see is a standard LLM pull request, wrong in some ways but at least functional. Net-negative engineers can no longer ship the catastrophic disasters they used to. Strong engineers are not 10x faster, because their bottleneck was taste and system familiarity. Compression at the bottom, not multiplication at the top.

Benn Stancil's satirical thesis: AI founders sell sweeping visions of 100x productivity and abundance to attract researchers, customers, and venture capital, but to incorporate the company they still have to pick a NAICS code from a dropdown. The gap between AI economic claims and any official measurement framework keeps widening. A useful corrective to AGI maximalism that lands as a counterweight without slipping into doomerism.

Dan Cleary's interview with Netlify CEO Matt Biilmann unpacks why daily signups climbed from 3,000 to 40,000 in just over a year. Biilmann's four-pillar AI strategy from January 2023 included a single line: 'become the preferred platform for AIs to build on.' Netlify's permissionless anonymous-deploy flow from 2013 turned out to be exactly what an agent needs to deploy without friction. Biilmann coined 'Agent Experience' (AX) in January 2025 and restructured the company around it.

Armin Ronacher on why local models feel rough despite enormous community effort. Putting an API key into a hosted model is boring; doing the same locally means picking an inference engine, model, quantisation, template, context size, and a stack of JSON configs. One quietly degrades quality and nothing tells you. His example: tool parameter streaming is missing from most local stacks, with surprisingly large consequences. The gap is polish, not capability.

A new GitHub repo bundling purpose-built Claude Code skills for academic research workflows. Each skill is a focused Markdown file that Claude Code loads on demand without bloating the context window. The bundle scored 92 on builder utility and is among the strongest skill packs to land in the past week. Useful for researchers, technical writers, or anyone running long-form structured thinking through a coding agent.

Hermes Agent, the open-source persistent-memory agent from Nous Research, became the single most-used model on OpenRouter token metrics for a 24-hour window this weekend, ahead of Claude Code and OpenClaw. The v0.13.0 'Tenacity Release' shipped May 7 with hardened sandboxing across five backends, scheduled cron automations, and isolated subagents with their own conversations and Python RPC. Lives across Telegram, Discord, Slack, WhatsApp, Signal, Email, and CLI. First time an open-source agent has held the top spot.

codegraph is a TypeScript tool that pre-indexes your codebase into a knowledge graph Claude Code reads instead of repeatedly grepping or reading files. Trade: upfront indexing cost for fewer tokens and fewer tool calls on every agent run. Runs entirely locally so no code leaves the machine. The repo crossed 1,000 GitHub stars this week (148 today), one of the fastest-trending Claude Code projects since Skills landed.

Mirage is a unified virtual filesystem for AI agents: S3, Slack, GitHub, Gmail, Drive, Redis and others mount side by side as one tree. Agents use familiar bash tools (grep, cat, cp, pipe) instead of learning new SDKs and MCP servers for every service. Pipelines compose across services as naturally as a local disk. Built on the bet that LLMs reason better when every backend speaks the same filesystem semantics. Embed in Python or TypeScript.

Google shipped a Gemini CLI extension that closes the gap between writing code (the inner loop) and shipping it (Dockerfiles, IAM, YAML, CI/CD). The author scaffolds a glassmorphic React + Node app and deploys it to Cloud Run via natural-language commands. Installs into Gemini CLI, Antigravity, or Claude Code, then handles containerisation, image registry, IAM, and Cloud Build pipeline generation. Useful when production friction is the reason your side projects never ship.