Kimi K3 Hits the Frontier | AI Field Notes #71

AnalysisThirteen apps that generate fake nude images of real people drew cease-and-desist letters from San Francisco's city attorney on July 17, eight on Apple's store and five on Google's. David Chiu's notices accuse both companies of aiding the spread of nonconsensual intimate images and profiting from it, arguing they have made millions in fees from tools marketed as harmless face-swappers. Apple said it removed three and is terminating the developer accounts; Google said all five on Play were suspended. The apps often produce images of minors. Enforcement is aimed at the stores, not the developers hiding behind them.

AnalysisServing someone else's open model at production scale turned into a $17.5 billion business this week. Fireworks AI closed a $1.505 billion Series D on July 16, led by Atreides, Index, and TCV, with Nvidia joining, after crossing $1 billion in annualized revenue, up fivefold in a year. Traffic on its platform runs past 40 trillion tokens a day, up from 15 trillion. Founded by seven former Meta engineers, Fireworks lets a company fine-tune an open model on its own data and serve it, with Uber, Shopify, GitLab, and MongoDB among its customers. The money is in inference now, not just training.

AnalysisAn open-weight model from a Beijing lab just won the front-end coding arena outright, with a 76% head-to-head win rate against every leading American system. Moonshot AI released Kimi K3 on July 16, a 2.7-trillion-parameter mixture-of-experts design (a model that activates only a slice of itself per request to cut cost), and priced it at $15 per million output tokens against Anthropic's Fable 5 at roughly $50. Moonshot promises the weights by July 27. The lead America held at the frontier a month ago is now measured in weeks, and in a bill a third the size.

AnalysisA London robot maker crossed into unicorn territory this week without shipping a robot to anyone's home. Humanoid raised $150 million as the first tranche of a Series A on July 16, at a $1.2 billion pre-money valuation, and is chasing another $80 to $100 million by September. The money rides a wave: warehouse and factory robotics drew several nine-figure rounds the same week, from Walden Robotics at $300 million to Germany's Microagi. Investors are betting that physical labor is the next thing software eats. The valuations are arriving well ahead of the working machines.

AnalysisSearch is quietly turning into a place where things get done, not just answered. Google began letting US users connect Instacart, Canva, and YouTube Music inside AI Mode the week of July 16, so a barbecue query can fill a grocery cart, a flyer request can pull Canva templates, and a party plan can build a YouTube Music playlist you save and play. The action happens in the chat, with more partners promised. Google is testing whether people will let its assistant reach into the apps they already pay for. The referral it once sent as a blue link now closes the transaction itself.

AnalysisMicrosoft is close to shipping a security tool that hunts software bugs by picking whichever model fits each job, drawing from its own, OpenAI's, and Anthropic's. Project Perception, reported July 17 and expected this month, aims squarely at Anthropic's Mythos, whose API runs an estimated 100% above Opus and 82% above GPT. A routing layer sends each task to the cheapest capable model, and cost is the pitch. Microsoft has not confirmed pricing or availability, and there is no sign it produced any of this month's patches. The multi-model router is becoming a product category of its own.

AnalysisAn agent can now stop mid-task, ask the user a question, and keep going. Cloudflare shipped support on July 13 for MCP elicitation, the part of the protocol where a tool requests input while it runs. Agents built with its addMcpServer method register handlers for two modes: a form for structured data, and a URL flow for consent before something like a third-party login or a payment. The settings survive hibernation, so a restored agent re-advertises them on reconnect. It is a small primitive with a real consequence: agents that gather what they need instead of failing when a field is missing.

AnalysisA 27-billion-parameter model now fits on a phone in 3.9 gigabytes, small enough to run on an iPhone 17 Pro. PrismML released Bonsai 27B on July 14, a compressed build of Alibaba's Qwen3.6 under an Apache license (free to use and modify, including commercially), using 1-bit weights that store each value in about a bit instead of the usual 16. It keeps a 262,000-token context and reads images, not just text. The catch with aggressive compression is usually quality, and the independent benchmarks that would settle it are still thin. On-device inference just stopped being a lab demo.

AnalysisCompanies that traded workers for AI are opening invoices they did not budget for. Uber burned through its entire 2026 AI spend in four months, one firm ran up $500 million on Claude in a single month after skipping usage limits, and 78% of IT leaders report surprise charges from consumption pricing. The pattern shows up per-seat too: a developer's GitHub Copilot bill jumped from about 67 euros to 966 after usage-based billing arrived June 1. Meta cut 8,000 jobs while moving 7,000 people into AI roles. The savings and the costs are landing on different lines of the same budget.

AnalysisLeaked source code shows an AI music generator logged more than 113,000 hours of audio from YouTube Music alone, routed through a proxy firm to slip past YouTube's defenses. The Suno breach, reported July 15 by 404 Media, traces to a supply-chain worm that stole an employee's credentials last November. The datasets also pull from Deezer, Genius, and stock libraries, and the same access exposed customer emails and Stripe payment details. Record labels already suing Suno now have code suggesting deliberate circumvention, which the copyright law they cite treats as its own violation. The receipts are the story.

AnalysisMore than 200 economists and AI researchers, including 16 Nobel laureates, signed a statement on July 13 warning that AI could reshape the economy faster than the Industrial Revolution and calling for governments to prepare for large-scale job displacement now. Organized by Erik Brynjolfsson and colleagues at Stanford, the letter carries names like Daron Acemoglu and Michael Spence, alongside researchers from OpenAI, Anthropic, and Google. It stops short of specific policy, asking mainly for research and institutions that keep AI complementing human work. A warning signed by the people building the thing carries a particular weight.