What happens when AI starts running its own research experiments overnight? And what happens when the humans building AI weapons systems start walking out? Both questions broke into the open today — and together they paint a picture of a field moving faster than its own institutions can handle.

The Military AI Reckoning: Resignations, Open Letters, and a Line in the Sand

The fallout from OpenAI's Pentagon deal just got personal. Caitlin Kalinowski — who joined OpenAI just months ago to rebuild its robotics division after it shuttered in 2020 — resigned this weekend, citing concerns over AI surveillance and lethal autonomy. Her statement was pointed: the deal was pushed through, she said, "without the guardrails defined." That's not an abstract complaint. That's an insider saying the process failed.

She's not alone. Hundreds of employees across Google, OpenAI, and other major labs have now signed open letters demanding clearer limits on military AI contracts — accelerated by U.S. strikes on Iran and the Pentagon's expanding AI ambitions. These aren't junior engineers. These are people who understand what the systems can actually do.

Here's what makes this moment different. Consumer backlash is easy to weather. Angry tweets fade. But when your robotics lead — someone who chose to come build your future — exits on principle and names lethal autonomy explicitly, that's a different signal entirely. The question now isn't whether AI will be used in warfare. That ship has sailed. The question is whether the labs building these systems will have any meaningful say in how they're deployed.

Agents Doing Science: Karpathy's Autoresearch and the Overnight Experiment Machine

While the ethics debate burns, the capability frontier keeps moving. Andrej Karpathy this weekend released autoresearch — an open-source framework that does something deceptively simple and genuinely radical. You point an AI agent at a machine learning training script. You go to sleep. You wake up to over a hundred completed experiments.

The mechanism is elegant. The agent modifies training code, runs a fixed five-minute training window — making all experiments directly comparable — checks whether validation loss improved, and either commits the change or rolls it back via a ratchet mechanism. Only improvements survive. The fixed time window is the clever bit: whether the agent tweaks the optimizer, the architecture, or the batch size, every experiment costs exactly the same, so you can compare across them cleanly.

Karpathy's own runs on eight H100 GPUs — graphics processing units optimized for parallel computation — produced 276 experiments over multiple days. Twenty-nine improvements kept. Validation loss trending steadily down. More importantly, improvements found at smaller model scales transferred cleanly to larger ones — suggesting the agent is finding genuine architectural insights, not just overfitting noise.

The real-world test came fast. Shopify's CEO adapted the framework overnight for an internal project. Eight hours, thirty-seven experiments, a nineteen percent improvement in validation score — on a model that now outperforms its larger predecessor. His line: "I learned more from that than months of following ML researchers."

What autoresearch really represents is a shift in who does the tedious work of science. The human writes instructions in Markdown. The agent handles the change-train-evaluate cycle that consumes most of a researcher's time. The entire training codebase — about 630 lines — fits within a modern LLM's context window, so the agent always understands everything it's touching. That constraint is intentional and important.

Claude as Security Researcher: Twenty-Two Firefox Flaws in Two Weeks

There's a related thread worth pulling on. Anthropic revealed that Claude Opus 4.6 spent two weeks combing through Firefox's codebase alongside Mozilla's team — and surfaced twenty-two vulnerabilities, fourteen of them rated high-severity. Claude flagged its first flaw in twenty minutes. Patches are already live for hundreds of millions of users.

Firefox isn't a new, poorly-audited codebase. It has decades of scrutiny and active bug bounty programs behind it. Finding fourteen high-severity issues — accounting for nearly twenty percent of Firefox's most serious patches for the entire year — in two weeks is remarkable. Claude also attempted to write working exploits for its own findings, but only succeeded twice, and both required disabling Firefox's sandbox entirely.

Anthropic was candid about that last point: the gap between finding vulnerabilities and weaponizing them won't last. The window to lock down codebases is, as they put it, "pretty urgent."

Weave these threads together and a picture emerges. AI agents are now good enough to autonomously run scientific research overnight, find security flaws in battle-hardened software, and — soon — fully exploit what they find. The capability curve is steep. And the humans who understand that best are the ones signing resignation letters asking whether we've thought hard enough about where this leads. Joseph Weizenbaum, who built one of the first conversational AI programs in the nineteen sixties, observed that "extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people." The programs are no longer simple. The thinking required now is anything but.