Theo pledges one meaningful T3 Code improvement every day
The search rebuild that ran 18 hours straight is now day 21 of a shipping streak he says he will keep up as long as he can.
Theo has turned a month of steady shipping into a public commitment. After noticing that T3 Code had landed at least one solid improvement every day of the month, roughly matching the pace of what he calls the "28 days of improvements to Codex", he made it official and said he will ship one meaningful improvement to T3 Code every day for however long he is capable. The search overhaul counted as day 21.
Gonna ship one meaningful improvement to T3 Code every day for however long I'm capable
That overhaul is the one that ran for 18 hours straight, covered earlier. The loop had four steps. He cloned his real data from the T3 Code production databases, started from a set of example queries that had not returned the results he wanted, synthesized 50 more examples based on his own usage, then hill climbed with Opus, Sol and Wonder Full until 53 of 54 cases passed.
Hill climbed with Opus, Sol and Wonder Full until I hit 53/54 cases
The result is in nightly for now, with stable shipping in the next one to two days by his estimate. He says results should be more relevant, more forgiving, more recent and much clearer, and that the PR is a funny read.
What transfers
The interesting part is not which models did the climbing. It is that he had 54 concrete query cases before letting any model touch relevance ranking. Search quality is the classic place where a model can confidently make things worse and nobody notices for a week, because "better results" is a feeling until you write it down as cases that either pass or fail. A production data clone plus a set of queries you already know are bad gives you a score to climb, and a 53 of 54 claim means something specific rather than a vibe.
It is also worth being clear about what this is. One engineer, one product, one overhaul, with the numbers coming from the person who shipped it and no outside check. The daily pledge is a pledge, not a track record, though he does point at a month of improvements behind it. Tomorrow is going to be a banger, he says, possibly dropping at midnight.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.