AI Builders Digest

Codex rate limits are draining faster because cache hit rates got worse

Some Codex users burned through their quota faster this week, and the cause is a degraded cache hit rate rather than a policy change. Consistently hitting the cache is a core part of the efficiency math, so when it slips, usage drains. Investigation is ongoing with an update promised for the next day. Separately, the banked reset shipped by 8pm PST for all paid ChatGPT Work and Codex users.

• #products
• #agents

Claude Security scans GitHub repos with Mythos and returns suggested patches

Point it at a repo and Mythos traces data across files, reasons about how components interact, and returns findings tagged with CWE category plus confidence and severity ratings. Suggested fixes open directly in Claude Code on the web, using whatever models your team already runs, and scans bill as ordinary token usage under your existing plan. The design keeps the model behind the scan: it returns findings only, no direct access. Alongside it, a Defender Advantage Fund putting $35M in credits toward open-source security, partner integrations of Mythos 5 into third-party security products, and an expanding Cyber Verification Program.

• #security
• #products

Collapsing an eval suite into one score hides regressions on your frontier use case

The tyranny of the average: a model that improves 85 to 89 on summarization and 80 to 85 on factual QA while dropping 70 to 63 on complex financial analysis will still look like a win under a single number. Weighted scores are not the fix, since they wrap a judgment call in false mathematical precision. The alternative is a prioritized ladder of evals, people willing to read the details instead of demanding abstraction, and a human decision about whether the system is genuinely better for users. This happened in the early days of Gemini and it keeps happening to teams now.

• #evals

Simulating humans is the last barrier to recursive self-improvement

If you take recursive self-improvement seriously, with models automating larger and larger parts of ML research and AI engineering, then simulating humans and human feedback is the remaining bottleneck. That reframes Smallville, which had zero commercial applications at the time, as the reason Karpathy and Fei-Fei Li backed the team behind it. Simile is already finding product-market fit at Fortune 100s at an early stage. The admission is unusually direct: two years late to understanding it, and never happier to have been wrong.

• #agents
• #research

As intelligence gets too cheap to meter, the opportunity moves to diffusion

Models are getting cheaper on a like-for-like task basis, more generally capable, faster, and deeper in nearly every domain all at once, at a rate with no precedent in tech history. Once intelligence approaches too cheap to meter, the value shifts to driving AI into the actual economy rather than to producing the intelligence. That makes this a strong moment for applied AI companies, where competition and innovation upstream function as a tailwind instead of a threat.

• #products
• #agents

Vercel ran is-agentic in a loop against its own docs until it scored 100/100

Running the is-agentic evaluation repeatedly against their documentation site closed a number of real gaps, with deliberate work to keep the criteria high quality enough to justify the tokens spent on it. Grok and Codex subscriptions are now supported and installable instantly in a sandbox. The Python team is also making rapid progress.

• #evals
• #products

Anthropic's internal ELI5 skill explains code with big pictures and few words

The prompt is simple: explain like I'm someone who knows nothing about this topic, using an HTML artifact with big pictures and few words. It gets used internally for explainers and for digging into problems, with prompts like how does this module work, why did we make this tradeoff, and what caused this incident. It's installable now from the community plugin marketplace, with an open question about whether it should become an official plugin.

• #agents
• #open-source

Instinct nails onboarding but indexes your email with no way to delete it

Connecting iMessages, Google Workspace, and MCPs is unusually smooth, and the assistant is more proactive than its peers, suggesting actions immediately after an MCP connects. The blocker is data handling: it indexes and retains emails without permission and offers no way to remove them from its records, which makes it unrecommendable until fixed. The other limitation is structural, everything lives in one thread, so it works for random chores but not real work. ChatGPT Work and Codex remain the actual workhorses.

• #products
• #privacy

Chess got more popular after computers beat humans, not less

Chess.com started as a $56,000 domain bought out of a 2005 bankruptcy auction that essentially every investor called uninvestable, and now runs 10 million daily active users, 250 million registered members, and over $200 million in revenue with no capital ever raised to grow it. Machines crossing superhuman actually made the game briefly boring, when everyone ground out Stockfish-approved endgames, and then far more exciting once neural nets like Leela started playing aggressive, unconventional chess that pushed human play forward. On expertise there are no shortcuts in the process, only in the tools: "if you just take any given day in your life and multiply that by a thousand, that's kind of what your life looks like." The founder advice is to ignore the founder advice, since the 2005 playbook said hire a technical cofounder from Stanford, raise money, and chase a big market, and doing the exact opposite worked. Next up is Gambit, a poker site that brings the chess playbook of ratings to a game usually scored by chips: how good are you really, not how many M&Ms can you rebuy.

• #products
• #founders

Claude Code reverse-engineered a school lunch site into a home bot feed

A kindergarten posted daily meals on a random website with unstructured data. Claude Code found the underlying API by inspecting network requests, discovered it was unauthenticated, worked out the response format, and wired it into an existing Hermes bot. Now the home bot announces breakfast and lunch each morning so the family can pack accordingly. Small-scale personal automation like this is still badly underrated.

• #agents
• #products

"No Doors for Agents" talk at the Agentic AI Summit in Berkeley

Spoke at the Agentic AI Summit in Berkeley alongside a set of other speakers, with a talk titled "No Doors for Agents." A new skill is also dropping.

• #agents