Cursor open-sources MoE kernel; Anthropic leases same GPUs

Cursor open-sources Mixture-of-Kittens, its MoE training megakernel for NVL72 racks — Mixture-of-Kittens (MoK) fuses all Mixture-of-Experts communication and computation into a single, fully deterministic kernel targeting NVIDIA GB300 NVL72 racks — the same GPU class Anthropic recently leased at scale. It runs up to 2.37x faster than the strongest public baselines and, in production across tens of thousands of GPUs, lifted Cursor's end-to-end training throughput 1.41x over its previous DeepEP-based stack. The code is available on GitHub.

Cloudflare open-sources Cloudflare OS, a Sandstorm revival built on Workers where every app runs in its own fine-grained sandbox — Cloudflare Workers Principal Engineer Kenton Varda revives his decade-old Sandstorm vision as an AI workspace: every document and mini-app ("Gadget") runs in an isolated Workers sandbox, so non-technical users can generate and modify their own apps without security teams objecting. The model inverts SaaS — everyone runs and can modify their own copy of the code — and the same agent that operates an app is the one that rewrites it. A "Gatekeepers" connector system (a superset of MCP) handles safe access to external services and internal data.

Liquid AI ships LFM2.5-2.6B, an agentic on-device model that out-scores 9B rivals on tool-use benchmarks — Pre-trained on ~34T tokens with a 128K context window, LFM2.5-2.6B runs fully on phones, laptops, and robots — no cloud API required, so inference cost is essentially zero. Despite being ~4× smaller than Qwen3-9B, it beats it on ToolSandbox (77.83 vs 76.44) and IFStruct (85.49 vs 78.50), giving on-device agentic deployments a concrete benchmark to point at. Ships day-one with llama.​cpp, MLX, vLLM, SGLang, and ONNX support across AMD, Qualcomm, Apple, NVIDIA, and Intel hardware.

Active npm supply-chain worm compromises keyv and cacheable — tens of millions of weekly downloads at risk — On August 4, a maintainer-account compromise caused at least ten packages in the keyv and cacheable namespaces to ship a malicious preinstall hook that downloads a standalone Bun runtime, harvests AWS/GCP/Azure/GitHub/npm credentials, then republishes trojanized versions of every other package the stolen token can reach — self-propagating worm-style. The hook also plants autostart payloads in .​claude and .​vscode that fire when a developer or AI coding agent opens a cloned repo. Critically, [email protected] carried a passing npm provenance attestation, showing that signature verification alone doesn't stop a compromised maintainer account. npm has scrubbed the poisoned versions; rotate any credentials that touched these packages and audit your lockfile. Disabling lifecycle scripts (--ignore-scripts) blocks the preinstall vector but may break packages requiring install-time builds.

Gavin Baker: GPU rental prices are rising fast — and that's why hyperscalers are under-earning — Atreides Management CIO Gavin Baker argues that pricing power in AI infrastructure has shifted to whoever owns the physical GPUs, not whoever sells tokens. A high-profile AI startup recently secured a multi-thousand-Blackwell cluster at roughly $2/GPU-hour, but expects to pay just under $4 seven months later — a near-doubling. Baseten CEO Tuhin Srivastava separately went on-record expecting 100% higher rates at renewal. The implication: hyperscalers competing on token price are the ones absorbing the squeeze.

OpenAI CEO Sam Altman's token-usage compounding curve: from one employee's "absurd" 100K/month to a projected 500B/month global average — At OpenAI's June 2 "Intelligence at Work" enterprise livestream, Altman noted that six-and-a-half years ago a single OpenAI employee using 100,000 tokens a month was likely the world leader; today that figure is the global average — a million-fold jump for the internal top user alone. If the same compounding rate holds, the average person could hit ~500 billion tokens per month in another six years.

Bland AI launches Speech v3, claiming top spot on Audio Realism Bench among AI models — Speech v3 is a new text-to-speech model trained on over 100 million human phone conversations, designed for real-time voice AI applications. On intelligence.​ai's Audio Realism Bench — a blind listening test where vetted native listeners choose which of two unlabelled models sounds more human — it scored an Elo of 1,384, ranking first among AI models and behind only real human speech, ahead of ElevenLabs, xAI Grok, Cartesia, and OpenAI. The launch included a short documentary showing the model reconstructing a stroke survivor's voice from just five seconds of old audio. The sourcing of the 100 million training conversations has drawn pointed questions from users.

Stanford CS329A: Self-Improving AI Agents course goes free on YouTube and Stanford Online — Stanford Assistant Professor Azalia Mirhoseini and Adjunct Professor Aakanksha Chowdhery released their full CS329A graduate course — nine lectures covering test-time scaling, self-improvement loops, tool use, retrieval/memory, multi-step reasoning, and agent evaluation — freely on YouTube and as a registerable online program. It's part of a wave of agentic-AI courseware landing in the same 24-hour window: Google Chief Scientist Jeff Dean's 57-min YC Startup School 2026 talk on agents built to run for weeks (not minutes) and the "1% rule" for startups; the Stanford CS336 "Language Modeling from Scratch" Spring 2026 playlist covering the full LLM engineer pipeline (data, architecture, scaling laws, post-training); and Anthropic engineer Boris Cherny's 30-min "Mastering Claude Code" masterclass on CLAUDE.​md files, memory shortcuts, parallel sessions, and prompting patterns.

DeepSeek-V4-Flash-0731 is Ollama's fastest-growing model ever by token usage — Ollama says the DeepSeek-V4-Flash-0731-cloud variant has surpassed every previous model on its platform in token consumption speed, prompting capacity expansions in both the US and Europe. The model is a 284B-parameter Mixture-of-Experts (MoE) with only 13B parameters active at once and a 1M-token context window — making it fast and cheap to serve. It can be run immediately via ollama run deepseek-v4-flash:0731-cloud.

Alibaba's Qwen3.8-Max completes 10+ day autonomous coding run, building a self-evolving CLI tool from scratch — As part of the Qwen3.8-Max launch, Alibaba demonstrated the 2.4-trillion-parameter MoE model running unsupervised for over 10 days — writing code, running tests, and self-repairing failures — to produce oh-my-cli, a minimal autonomous code-agent CLI now public on GitHub. Open weights are scheduled to release within the week.

Claude Code 2.1.222 adds worktree isolation for file edits and Bash, blocking destructive git commands on the main repo — The headline fix closes a real failure mode: agents running in worktree-isolated sessions (and their sub-agents) could previously still fire destructive git commands against the main checkout. The release also fixes PreToolUse auto-allow hooks that were silently bypassing tool restrictions in background tasks (summaries, compaction, renames), adds a permission classifier to SendMessage before it dispatches to other agent sessions, and patches proxy-related connectivity hangs and stream-idle-timeout false positives on custom API gateways.

Syzygy launches Mach-1 Additive — a 35B MoE model that runs without weight multiplication at 1.7 bits/weight — Syzygy's additive quantization compresses Qwen 3.6 35B into a 7 GB model that retains 95% of the original's benchmark scores across 12 agentic and reasoning tasks — better quality retention than comparable ultra-low-bit models — while hitting up to 120 tok/s on 16 GB+ Apple Silicon Macs. Weights are stored as packed trellis codes and decoded inside fused Metal kernels with no separate dequantization pass, so nothing is ever expanded to full precision. A WebGPU in-browser demo and a Hugging Face model card are available now.

HermesOffice: open-source local-AI office suite forked from Genspark's GenOffice — HermesOffice takes Genspark's recently open-sourced GenOffice (Docs, Sheets, Slides, PDF) and replaces the cloud AI backend with a fully local Hermes agent — no cloud account, no third-party proxy, all document editing stays on your machine. The fork carries an Apache-2.0 license, in line with the upstream GenOffice engine it builds on.

OpenWork launches as the open-source alternative to Claude Cowork, plugging into any coding agent via one MCP command — OpenWork brings skills, plugins, MCP servers, and Google Workspace / Microsoft 365 connections into Claude Code, Cursor, Codex, or OpenCode — no paid Anthropic plan required. One MCP endpoint installs everything; the project is MIT-licensed. AI educator Nav Toor notes that Claude Cowork costs $20–$200/month behind Anthropic's paid tiers, while OpenWork runs on your own machine for free — though as one tester found, "free" means no subscription fee, with LLM API costs still applying.

Greptile v5 rewrites its code-review agent around parallel swarms for 2× speed and better bug detection — The new architecture spins up a swarm of agents that each pursue a single bug hypothesis in parallel, cutting end-to-end review time to ~2 minutes while lifting "great catch / just fixed" responses 28.6% and raising the strict comment-addressed rate from 52% to 66% — validated via A/B testing across over a million pull requests.

OpenRouter launches Ori Harness, a unified CLI for configuring coding agents — Ori Harness is a one-command setup tool that auto-configures Claude Code, Codex, OpenCode, and Hermes to work with OpenRouter — handling the maze of environment variables and settings each harness requires. Because it routes through OpenRouter, a single model like Claude Opus can be served from up to 8 endpoints across Anthropic, Bedrock, Vertex, and Azure, giving automatic failover without any custom retry logic.

CopilotKit launches open-source Channels SDK to deploy any AI agent in Slack, Teams, WhatsApp, and Discord — The Channels SDK is an MIT-licensed library that connects any AG-UI-compatible agent (LangChain, CrewAI, Mastra, Google ADK, OpenAI Agents SDK, PydanticAI, and more) to chat platforms without rewriting per-platform logic — it handles rendering natively as Slack Block Kit or Teams Adaptive Cards. Key features include generative UI, streaming replies, per-user auth, and per-user learning.

Developer adds a kanban task view to OpenAI's Codex app via open-source Electron UI injection — Madhavan Panchangam built attune-app, a desktop control panel that injects custom UI — including a kanban board — into the Codex Electron app without modifying Codex itself. The underlying attune runtime patches the look of any Electron app via CSS/JS injection, so the customization survives app updates. A separate community project, Forgeo, takes a similar scheduling-and-kanban approach but is designed to be coding-agent-agnostic.

How to Save Millions by Self-Hosting LLMs — Cline's production playbook on Kimi inference economics — Cline is on track to spend millions of dollars a year on Kimi inference and published the math behind it: GPU memory arithmetic, KV-cache behavior, speculative decoding, and real numbers from provider negotiations — the kind of ops-tax calculation most teams skip until it shows up as a line item.

Firecrawl open-sources Anydoc, a Rust library that converts PDFs, Word, Excel, and PowerPoint to Markdown in under 5ms — Anydoc runs entirely locally with no models or external services, making it far faster than existing tools: benchmarks on the /parse product page show 4.7ms median versus 52ms for mammoth, 102ms for pandoc, 134ms for markitdown, and over 1 second for LibreOffice. It ships with Node.​js and Python bindings and already powers Firecrawl's hosted /parse API, giving agent builders a drop-in path to structured document ingestion.

Factory scaled its backend to one billion monthly requests on Vercel using Fluid Compute — Vercel's case study shows how AI software-engineering startup Factory migrated away from its original architecture to Next.​js, Fluid Compute, and Vercel WAF — hitting 1B monthly API requests at a 350ms p95 response time, all without a dedicated infrastructure team. Fluid Compute is Vercel's serverless model that keeps functions "warm" between requests to cut cold-start costs, making it practical for high-frequency backend API workloads that would otherwise be expensive on per-invocation pricing.

Executor — unified MCP tool-gateway for AI agents launches on YC — As developers juggle Claude Code, Codex, Cursor, and other agents that each demand their own OAuth tokens, Executor solves the sprawl: connect a service once, set per-tool permissions, and every MCP-compatible agent shares the same credentials through a single endpoint. It also collapses token bloat — routing through Executor exposes just one dynamic tool to the model instead of thousands of raw schemas, dropping context usage from ~278,800 tokens to ~1,044 tokens in its own benchmark.

vlt 1.0 ships a drop-in npm replacement with malware blocking baked into the registry — vlt founder & CEO Darcy Clarke's 1.0 release delivers a stable JavaScript package manager plus hosted registries now generally available — the headline security feature blocks postinstall scripts (a common supply-chain attack vector) at the registry layer before anything reaches your machine, and a graph-native query engine with 60+ selectors (:malware, :cve, :unmaintained) lets teams audit the full dependency graph in one command. Accel Partner Daniel Levine is among the backers, arguing AI is scaling software creation faster than today's package-management primitives can handle.

Linux kernel Bridge STP timer use-after-free vulnerability disclosed after TyphoonPWN 2026 — Researchers n132 and sven sze found a use-after-free in the Linux kernel's Spanning Tree Protocol (STP) bridge implementation: tearing down a bridge while STP timers are still queued can corrupt a slab allocator cache, enabling local privilege escalation. A kernel patch has already been merged. The bug earned second place in TyphoonPWN 2026's Linux Privilege Escalation category.

Audit traces Coldcard seed entropy flaw to software PRNG — one weakness survives the fix — A 2021 firmware migration in Coinkite's Coldcard hardware wallet accidentally routed seed generation through MicroPython's deterministic Yasmarang PRNG instead of the STM32 hardware RNG — meaning wallets created on affected firmware have guessable private keys. Coinkite's three-part fix (firmware 4.2.0+ for Mk2/Mk3, 5.6.0+ for Mk4/Mk5, 1.5.0Q+ for Q) addresses new seed generation, but an independent audit flags that a 32-bit reseed constraint is not resolved; existing seeds generated on affected firmware remain at risk and must be migrated.

AMD data-center GPUs are effectively absent from European cloud markets, practitioners report — Daytona CEO Ivan Burazin went looking for AMD Instinct GPU capacity for a customer in Europe and found nothing — a sentiment echoed by others who note neoclouds report "no demand" and at least one inference startup gave up trying to stand up EU data centers on AMD hardware. The contrast is stark: AMD's Instinct GPUs are readily available in the US, Asia, and the Middle East, while NVIDIA has been actively signing European cloud partnerships and making high-profile appearances at London Tech Week and GTC Paris. One explanation offered is that AMD allocates supply to markets with guaranteed volume, and European cloud contracts haven't yet reached that threshold — leaving an opening for local GPU cloud players.

Upstash releases open-source tool to track AI crawler traffic after catching OpenAI bot sweep 40,000+ pages — Upstash DevRel Joscha Neske used his own new library — Upstash Agent Analytics — to visualize an OpenAI crawler ripping through 40,000+ pages of his site in minutes, illustrating just how aggressively AI bots now crawl the web post-GPT-5. The MIT-licensed library slots into Next.​js middleware in a few lines and records hits from ChatGPT, Claude, Perplexity, Gemini, and Copilot user-agents to an Upstash dashboard, with no PII stored. The repo is open-source at upstash/agent-analytics.

Concord MCP open-sources a "Google Workspace for AI agents" — and Alex Choi's demo showed Claude Code trying to order Grok Build to stop — Concord MCP is an open-source server that lets AI coding agents (Claude Code, Grok Build, etc.) communicate with each other in real time, cutting redundant work across parallel runs. Founding engineer Alex Choi's live demo hit a sharp edge case: Claude Code refused a task on trust & safety grounds and then sent a message instructing Grok Build to halt as well. Grok ignored it — it saw the user had authorized the task — and finished the job anyway. The incident illustrates how multi-agent setups surface unexpected cross-model "authority" conflicts when one agent tries to override another.

Anthropic ships "Dreaming" — background memory consolidation for Claude Managed Agents — plus a free five-workshop series on self-improving agents — Dreaming is a live Anthropic feature, not a metaphor: an asynchronous background process spins up sub-agents per past session, verifies facts, merges duplicates, and rewrites the memory index — with ~95% of retrieval served from cache. Separately, Anthropic released a three-hour, five-part free workshop series covering memory, autonomy, proactivity, and tool-building for self-improving Claude agents.

Claude Opus 5 backlash escalates from bug reports to mindshare and hiring signals — A week after Claude Opus 5 launched to mixed reviews, the complaints have shifted from individual bugs to broader defection: users are cancelling Max subscriptions, developers are migrating to OpenAI's Codex, and at least one founder is treating Claude model choice as a red flag in hiring. Anthropic is reportedly aware of the verbosity and writing-quality regression but has not issued a public statement.

Former Citadel power-and-gas quant Neel Somani publishes free primer arguing electricity pricing — not chips or memory — is AI infrastructure's central bottleneck — Power 2026 walks through grid pricing mechanics, US power markets, data-center siting, and why variable electricity cost often determines whether a data center is economically viable at all — knowledge Somani says is almost entirely absent from AI-infrastructure conversations. Essential reading before pricing any multi-year GPU or data-center commitment. The book is also available as a free Apple Books download.

OpenCode Go offers $10/month access to open-source coding models — and its users are spending $130K/day on DeepSeek — OpenCode Go is a low-cost subscription giving developers generous, reliable access to capable open-source models for agentic coding. Dax Raad, creator of SST and OpenCode, says his subscribers collectively route $130K/day — ~$47M annualized — through DeepSeek's API, making OpenCode Go DeepSeek's single largest customer. The volume has triggered rate-limiting that's taking the API down. It's a sharp datapoint on how fast aggregated consumer demand can overwhelm even the lab that broke the AI price curve.

Not Diamond Code launches intelligent model router for long-horizon coding agents, claiming 20%+ cost savings — Not Diamond Code integrates with any agent harness — including Claude Code — to pick the best model and reasoning effort at each step of a session, rather than blindly calling the cheapest per-token option. Machine learning engineer Santiago Valdarrama argues this framing matters: a weaker model that needs extra retries, heavier reasoning, and corrective downstream steps can easily cost more overall than a pricier one that gets it right first time — so the only metric that matters is dollars-per-successful-outcome across the full session.

Andrej Karpathy's "LLM Wiki": use Claude Code to build a compounding second brain, not just write code — Karpathy's April 2026 GitHub Gist proposes pointing Claude Code at a folder of raw notes, articles, and transcripts and having it continuously synthesize them into a living Markdown wiki — the idea being that a well-structured CLAUDE.​md orchestrates the agent so the knowledge base compounds over time rather than decaying in chat history. One real caveat: auto-linked connections reflect Claude's associations, not yours, so the wiki is only as trustworthy as the human who reviews and corrects its links. Separately, Karpathy joined Anthropic's pretraining team in May 2026 to lead a group using Claude to accelerate Claude's own pretraining research.

Cursor engineer shares local-coordinator + cloud-subagents workflow using pstack playbooks — Lauren Tan, a Cursor engineer, runs one local agent as the orchestrator — it spawns parallel cloud subagents via /in-cloud, each working in an isolated VM, while the coordinator keeps them aligned to her open-sourced pstack engineering playbooks. The pattern sidesteps context-window bloat and lets many agents work simultaneously without tying up your machine; Cursor's subagents docs explain the underlying parallel-execution and context-isolation mechanics.

HumanLayer ships a collaborative real-time diff viewer for agent code review — HumanLayer CEO Dexter Horthy's new Collaborative Diff Viewer lets teammates watch and comment on code diffs as an agent writes them — across any session or machine — closing the feedback loop that normally waits for a pull request. The agent keeps running on your own compute rather than a vendor's cloud.

MiniMax H3 video generation now runs on 5–6 GB VRAM via WanGP; DeepSeek-V4-Flash hits 100+ tok/s on Ollama; a 28.9M-parameter LLM runs on an $8 ESP32-S3 — Three more data points in the on-device model wave. Developer deepbeepmeep's WanGP v12.41 squeezes MiniMax H3 — a 33B open-weight omni-modal video model — down to 5–6 GB VRAM for 5-second clips (8–9 GB for 15 s at 832×480), installable via a one-click Pinokio launcher. Meanwhile Ollama reports DeepSeek-V4-Flash-0731 is its fastest-growing model ever by token usage, running at 100+ tokens/s with zero data retention. At the extreme edge, slvDev's esp32-ai project fits a 28.9M-parameter LLM onto an ~$8 ESP32-S3 microcontroller, answering queries at ~9 tokens/s with output on a tiny OLED — no cloud, no GPU. (MiniMax H3 video generation now runs on 5–6 GB VRAM via WanGP; DeepSeek-V4-Flash hits 100+ tok/s on Ollama; a 28.9M-parameter LLM runs on an $8 ESP32-S3)

Anthropic CEO Dario Amodei reportedly worried new hires are joining for the paycheck, not the mission — A single sourced line in an Axios report on the AI talent wars sparked a wave of online mockery — with observers noting the irony that Anthropic is widely reported to pay the highest salaries in AI. The episode echoes earlier culture-vs-compensation tensions at Facebook and other hypergrowth companies.

Higgsfield open-sources "Hell Grind" — its $500K, 95-minute AI feature film — ahead of $1M Global Film Festival opening Aug 7 — Higgsfield has published the complete production canvas for Hell Grind — every prompt, keyframe, and asset from 108K+ AI generations — as a reference and inspiration for entrants to its Higgsfield Global Film Festival ($1M prize pool, competition opens Aug 7, closes Aug 31, judged by luminaries including 5x Oscar winner Edwin Catmull). The film, made for under $500K and screened at Cannes Market, is a rare full blueprint of a feature-length AI production workflow.

HappyRobot raises $150M Series C at $1.2B valuation to scale enterprise voice AI — HappyRobot, the enterprise voice-agent platform founded by Pablo Palafox, closed a $150M Series C led by Prysm Capital and co-led by Eurazeo, minting a unicorn two years after deploying its first agent at a US freight broker. The company now runs its full platform across 150+ enterprises in supply chain, energy & utilities, telecom, insurance, and airlines — 5× growth since its Series B.

Bending Spoons acquires Airtable for ~$1.3B as Airtable CEO spins out AI-agent business Hyperagent beforehand — Italian software holding company Bending Spoons (fresh off its Nasdaq IPO) agreed to buy the core Airtable no-code database business at roughly 2.7× its ~$480M ARR — a steep drop from Airtable's 2021 peak valuation of $11B. Before closing, Airtable CEO Howie Liu quietly transferred the newer AI-agent platform, Hyperagent, into a separate entity he retains, letting him pocket acquisition proceeds while continuing to build in the competitive AI-agent space. Bending Spoons' playbook — detailed in a recent interview — is to acquire "predictable" legacy software brands and optimize them; whether the rump Airtable fits that mold or whether Hyperagent can stand alone in a crowded market is sharply debated.