Just-in-Time Software: Software that writes its own code
Bay Area, CA
Joined February 2025
- Tweets122
- Following1
- Followers54
- Likes46
Agents are getting good at saying how long something will take. The next unlock is them proving it works when they say it's done.
Opus 5.5 is the best work estimator model i've ever used. This is calibration that is increasingly important as we delegate more and more work — don't tell me that it's going to take 2-3 working days (what a human would do) when for you, an enthusiastic, indefatigable, and embarrasingly parallelizable AGI, can get it done in 2-3 hours.
In a report on an Oct 6 incident, OpenAI describes a model in RL training that found its input files missing. It submitted fabricated scores, then faked the files. The automated check rejected every one of those grades. New essay: what a passing test does and doesn't prove.
RAG was built for scarce context. Windows are heading toward a billion tokens. If cost per token keeps falling, when does carrying the whole working set beat retrieving pieces of it? And once an agent can hold that much, what should it attend to?
ALT Illustration: a person stands before a curved wall of floating screens, code and data, for the essay The Gigacontext Threshold.
Nothing kills a long Codex session like the account running dry halfway through. ZergRouter shows each connected Codex account's remaining weekly quota and reset date, side by side. Stale readings are clearly marked. You still choose which account to use.
ALT ZergRouter card: Which account resets next? Weekly quota and reset dates across your connected Codex accounts.
Most coding agents retry. Few remember. AdaExplore, a paper from CMU, UW and Arm on GPU kernel generation, distills repeated failures into rules that carry across tasks: 1,178 failure traces became 78 reusable rules. Which failures keep coming back in your agent runs?
ALT Most AI Agents Have No Memory: Zerg blog write-up on the AdaExplore paper.
Our write-up on the paper: zergai.com/blog/research/age…
Done and checked aren't the same thing. Zerg is AI automation for work that has to be right. Give it a task like a code migration or a data connector. It does the work, then checks it against your tests and the real system, so you see what passed and what still needs work.
ALT How Zerg works: Map, Build, Verify, Maintain. Verified results guide the next improvement.
Zerg AI retweeted
I'm building Zerg (@zerg_ai): systems where AI agents don't just write the code, they get better at writing it.
What I post about:
- running lots of models and agents at once (harnesses, routing; ZergRouter is out)
- self-improving loops, where the hard part is the evaluator, not the optimizer
- paper breakdowns and explainer videos on agent research
Follow along for what actually works when you hand real engineering to agents, and what breaks.
60% of Codex users also use Claude Code (per Theo). A lot of that is people paying twice to get around limits. A cheaper fix for the overflow: keep Codex, point it at DeepSeek 4.1 Flash at $0.30 per 1M input tokens, and give each API key a daily budget so nothing runs away. zergrouter.com
Built in ~4 days, mostly by agents. Original launch thread ↓
xitter.cf/idanbeck/status/211008…
Inspired by @jaivinwylde
I put the entire game of Counter Strike 1.6 into Twitter with working multiplayer - just click below to play!
twitterstrike.com/
New experiment we've just launched: Twitterstrike, a multiplayer shooter that runs inside a tweet. Tap the card below to play.
Inspired by @jaivinwylde
I put the entire game of Counter Strike 1.6 into Twitter with working multiplayer - just click below to play!
twitterstrike.com/
Shipped ZergRouter and listed it on Tiny Startups 🚀
Run DeepSeek in Codex, with API-key budgets and quota visibility.
tinystartups.com/startup/zer…
ty @ratheejaisal
You can also run DeepSeek in Codex through one endpoint and give each API key a daily budget. When a key hits its budget, it stops instead of running up a bill overnight.
Try it free now: zergrouter.com/?utm_source=t…
It's launch week, so if something breaks, send us a message - we'll fix it ASAP!