the open weights shipped, the talent got poached, and RL went looking for the whole internet

An open padlock made of circuitry leaking glowing cables toward a distant server tower, beside stacked luminous cubes and a self-overlapping loop of light.

Week ending July 18, 2025; archive coverage: July 14.

Two things happened this week that rhyme. An open-weight Chinese model showed up claiming to beat the paid incumbents at exactly the task everyone's monetizing—coding—and one of the hottest coding startups got its brain surgically removed and shipped to Google. The models are commoditizing faster than the business models can adapt, and the people who know how to build them are the scarce asset now.

Meanwhile the unglamorous middle layer kept improving quietly: sandboxes for computer-use agents, tighter control loops for anything predicting actions in chunks, and a genuinely provocative argument that reinforcement learning is the next thing to scale—if we can figure out what it even means to scale it.

The loud stuff was the launches. The story, as usual, was the plumbing and the org charts.

the open coder that actually acts

Kimi K2 is a trillion-parameter mixture-of-experts model with 32B active, open-weighted, and tuned specifically for tool use rather than sounding pleasant. Moonshot claims 65.8% on SWE-bench Verified and better-than-GPT-4.1 numbers on LiveCodeBench and MATH-500.

Read the benchmarks with the usual suspicion—everybody's numbers are cherry-picked and MATH-500 is nearly saturated anyway. But the positioning is the real signal. This isn't another chatbot; it's an agentic-coding model you can self-host, priced to undercut the API incumbents and free to download. The MuonClip optimizer claim—stable trillion-parameter training with "zero instability"—is the kind of thing that either generalizes and matters enormously or quietly doesn't. We won't know for a while. But an open model that's competent at multi-step tool orchestration compresses everyone's coding-tool margins, and that's not hypothetical.

the acquihire that wasn't an acquisition

Which brings us to Windsurf. OpenAI's $3B deal collapsed, and Google DeepMind promptly hired CEO Varun Mohan, co-founder Douglas Chen, and top researchers—paying a reported $2.4B for a nonexclusive tech license and the people, while taking no stake and no control.

This is the reverse-acquihire playbook again, same as Character.AI and Inflection: grab the talent and the IP license, leave the corporate husk behind, and skip the antitrust review that a real acquisition would trigger. The ~250 remaining employees keep the enterprise product; history (Scale, Inflection) suggests that's a rough road once the founders leave. The tell here is that even the biggest labs would rather buy fifteen brains than build a coding tool from scratch. Talent is the moat now, not architecture.

the boring layer got useful

ScreenEnv is the kind of thing I actually want: reproducible Ubuntu desktops in Docker, spun up in under ten seconds, with both a direct sandbox API and Model Context Protocol support. Computer-use agents have been demoed on brittle VM rigs forever; giving them isolated, disposable, typed-tool environments is how you go from impressive video to something you can benchmark and deploy without it clobbering your host.

And async robot inference is a lovely little systems result that travels well beyond robots. Chunky policies leave the robot idle while the next action chunk computes. Decouple prediction from execution across a gRPC client/server, keep executing the current queue while the next one bakes, aggregate on overlap—roughly 2x faster task completion with comparable success. Any system that predicts action batches and then sits waiting can steal this idea. Overlap your compute with your execution; stop letting the loop stall.

RL wants the whole internet

The most interesting read was Jack Morris arguing how to scale RL to 10^26 FLOPs. His claim: today's reasoning models are in their GPT-3 moment—trained for a few thousand steps on narrow verifiable domains like math and code because those happen to be easy to auto-score. The proposal is to notice that next-token prediction is itself verifiable, and turn all of web text into RL training signal, letting models learn to reason on anything rather than just what we can unit-test.

He's honest that the specifics are unsolved—what's the reward, how often to insert reasoning, whether it beats plain pretraining. A recent preprint titlegrabbed the idea and only finetuned on one math dataset. But the first-principles pull is strong: if something makes sense, keep hammering until it works. I buy the framing more than any single implementation.

what I’d do Monday

Actually test Kimi K2 on your coding-agent harness before believing or dismissing the numbers. Self-hostable open weights change your cost and compliance math.
Move computer-use experiments into disposable sandboxes now. ScreenEnv or equivalent; stop debugging agents against your own machine.
Audit any inference loop for idle time. If you predict in batches, overlap execution with the next prediction.
Don't over-invest in a single startup's tool as strategy—this week showed how fast the people, and the leverage, can walk out the door.

vocabulary inflation

Reverse-acquihire: hire a startup's leaders and license its tech without buying the company—conveniently below the regulatory radar. MuonClip: Moonshot's optimizer tweak (rescaling query/key projections) they credit for stable trillion-parameter training. RNTP / reasoning via next-token prediction: Morris's pitch for RL-ing on general web text instead of hand-built verifiable environments.