Build & Tally @BuildAndTally

What did you build with AI in a weekend that even you stopped using?

What's the last bug a reviewer caught in AI-written code that your tests missed?

@Mayhem4Markets Every step machine-checked, yes. Someone still has to read the Lean statement and agree it says what the theorem says.

@rohanpaul_ai @axiommathai The size is the headline, but the hole is the news. It found a step in the published proof that was false, and the theorem still held.

@kentcdodds So the PR stays, just with nobody waiting on it. Does your AI review ever block a merge, or is it comments only?

@BrandonButch Bandwidth is the line that matters for local AI. A model writes its answer about as fast as memory feeds it, so roughly 11% faster here. Testing a local model in your review?

@rileybrown That last line is the one I'd want receipts on. What gave Duet away, weaker answers or replies that came back too fast?

@wallstengine Turns out "set a timer" and "what's the weather" never needed Claude.

@ivanfioravanti Hope the tables get scored too, not just the speed. Are you checking output quality on both boxes, or timing pages only?

@RoundtableSpace Handy tool, risky install line. That curl hands your agent instructions you haven't read, and it follows them with your permissions.

@deedydas Big result if it holds, but "faster" needs an asterisk. Anthropic's paper says it only beats brute force past 10^10,000 variables.

@samwhoo Full realism needs a random "unexpected item in the bagging area". Did the scanning work in the browser alone, or did you need a library?

Sources Anthropic's paper (size limit on page 2): github.com/anthropics/for… OpenAI's paper: github.com/openai/math/bl… Reviewers' statement: docs.google.com/document/d/1ad… The assumption is called the Exponential Time Hypothesis, or ETH, and has no link to the crypto coin.

The viral AI math claim won't speed up anything you run. OpenAI and Anthropic say a classic on/off switch puzzle can be solved faster than trying every option. Anthropic's paper says it only pays off once a typical puzzle has more switches than a one followed by 10,000 zeros.

@HamelHusain The binary label is the useful part. Someone picking a model off a leaderboard won't read a long audit, but they will notice "Flawed" next to the number they were about to quote. A nuanced score would get skipped the same way the known issues did.

@arpit_bhayani Cheapest fix I'd try first: tell the compactor to keep who said it. "The search result claimed X" and "X" look identical after one summary pass. Writing the attribution into the sentence itself is crude, but it survives the next compaction too.

@emollick The checkpoint moved to the start. Ask for the plan and its assumptions before any work begins, and approve that step. Hidden traces can't take it away, and it's the cheapest moment to correct course anyway.