Good morning. OpenAI is teeing up another Codex-bound model release, a Dartmouth study is making bold claims about AI tutoring that the internet is happily poking holes in, and Zuckerberg — following yesterday’s town hall confession — is now on the record with Reuters saying agents aren’t accelerating. Also: Mechanical Turk is finally winding down, and Jim Keller wants to mass-produce chip fabs.
GPT-5.6 Sol Ultra is coming to Codex. OpenAI’s Thomas Sottiaux announced that GPT-5.6 Sol Ultra, which uses subagents to push past single-agent limits, will land in Codex. The timing lines up with reporting from The Information that OpenAI has cut inference costs in half, and one commenter on the HN thread noted they already have access via a corporate account — though their employer’s tune has shifted from “use more tokens” to “please use fewer.” No word yet on public availability or how this stacks against Pro.
A Dartmouth AI tutor claims huge effect sizes, with caveats. A new paper reports 0.71–1.30 SD improvements in a statistics course using a Claude Sonnet-graded practice quiz platform. The HN discussion is not buying it: there was no randomized control, roughly 90% of students opted in voluntarily, and the “tutor” is really an autograder wrapped around a quiz engine. As one commenter put it, “Engaged students score 0.71 – 1.30 SD better in tests” would be a much simpler headline.
Zuckerberg, in public this time, says agents aren’t accelerating. Following Wednesday’s leaked town hall, Zuckerberg told Reuters that “the trajectory of agentic development over at least the last four months hasn’t really accelerated” as expected, though he still expects meaningful returns in three to six months. The HN crowd was unsurprised, with the best framing coming from one commenter: “A chatbot can be wrong 10% of the time and still help you. An agent that’s wrong 10% of the time is sending bad emails and making wrong API calls with no one checking.”
Does clean code help coding agents? Sort of. A 660-trial study using Claude Code found that code cleanliness doesn’t move task completion rates but cuts token usage by 7–8% and file revisits by 34%. Practitioners on HN mostly disagreed with the null result on completion, arguing dead code, leaking abstractions, and half-baked patterns clearly hurt agents in real projects. The methodology took heavier fire: the “clean” and “degraded” versions of each repo were themselves generated by an LLM, which one commenter called “sufficient to condemn the entire study.”
Canada’s AI sovereignty pitch, minus the sovereignty. Al Vigier argues in The Line that Canada’s new “AI for All” strategy — which positions Ottawa as a “strategic anchor customer” for domestic AI — is undermined by roughly $46.8M in largely undisclosed Palantir contracts, no procurement mandates for Canadian vendors, and a preference for equity stakes over actual purchases. HN commenters were split between “Canada shouldn’t touch Palantir at all” and “Canadian critical infrastructure is 100% on US clouds anyway, so what exactly are we protecting?”
Mechanical Turk is winding down. Amazon will stop accepting new Mechanical Turk customers on July 30, with no new features planned for existing users. It’s a fitting end for a service that a 2023 study found was already 33–46% LLM-powered on the worker side — the human-labeled data pipeline for AI was itself being quietly automated by AI.
EU Council fast-tracks Chat Control 1.0. The Council pushed through a renewal of the voluntary CSAM-scanning permission for messaging providers, which had recently expired. Commenters on HN were careful to distinguish this from the far more dangerous Chat Control 2.0, which would break end-to-end encryption — but noted the direction of travel is clear enough.
LangChain’s OpenWiki generates docs for your agents. LangChain released OpenWiki, a CLI that generates and maintains agent-readable documentation and appends pointers to AGENTS.md and CLAUDE.md. The HN reaction was tepid: several commenters called it a thin TypeScript wrapper around prompts that could have been a Claude Skill, and questioned what it offers over just asking an agent to write docs.
Jim Keller’s chip startup wants to build a fab factory. Atomic Semi has rebranded to Fab2 and moved to Texas, where it’s building a facility to mass-produce small, software-defined semiconductor fabs using electron-beam lithography — trading throughput for prototyping speed. Co-founder Sam Zeloof famously fabricated chips in his garage as a teenager. One semiconductor engineer on HN was skeptical no fundamental manufacturing advances have actually been demonstrated, but others called it the most interesting non-AI hardware project going.
That’s the round-up. If GPT-5.6 Sol Ultra shows up in your Codex sidebar this weekend, you’ll know what it is — assuming it can finish a thought without hitting a 512-token boundary.