After enabling a custom agentic code reviewer system that auto-approves safe PRs (now over 90% are approved like this), PR size dropped in half. I dug into the data and realized it correlate with some some interesting numbers:
1
3
149
Of course this can't be all attributed to the review process: stacks usage got up, and coding agents kept getting better. But knowing that the PR will get reviewed and merged faster, you keep the scope smaller, easier to monitor, and easier to roll back (if needed).
1
18
Yoav Farhi retweeted
We're testing something new in @Base44 - while your app builds, it asks you a few casual questions. Nothing it needs for the build, more like who it's for and how it should feel, to make it yours.
Really loving how its turning out, great work by the team
(Obviously I tested it by building a sandbox startup, will update once it gets to $1M ARR)
Heads up: gpt-image 2.6 is much faster than 2.5, but uses ~4x as many tokens at low/medium for the same image size.
2
151
In case you were wondering - Opus 5.5 can edit (or add weights too).
We have a successful fine-tuned Qwen 3.5 2B running in production for auto-complete. Works great, except occasionally (~2% of requests) it will repeat itself repeat itself repeat itself repeat itself.
🤖 Made with AI
2
1
334
So I asked Opus to take a look - initially it found and tried to turn off the copy heads, but that made overall quality lower. But then it just added three new vectors instead that say: "stop if I'm repeating myself"
1
40
Astra is an incredible model. The first to complete my personal impossible task on Base44. Time to find a new challenge.
4
274
For some reason I've been sitting in this for too long: jst.sh is the cli utility you didn't know you needed. It uses a (free hosted, or local) llm to convert natural language to shell commands.
Designed to just work.
3
5
372
Yoav Farhi retweeted
We just blew past $200M ARR at @Base44
5 months after hitting $100M. We've released Base 1 in between.
i'm proud of the healthy business we're building,
being the only platform to own its entire stack (backend, models, etc) - which also allows us to move fast and be creative.
All credits go to a very special team that i'm having so much fun working with
Yoav Farhi retweeted
Announcing Base 1’s benchmarks:
After rigorous testing in production, attached below Base 1 results for building beautiful, functional and complex web apps inside @Base44.
we released Base 1, our own model, a few weeks ago. we've been measuring it since.
For app building, there are no standard benchmarks. We A/B test in production with millions of builders, who we believe provide the best benchmark for measuring real value across real-world use cases.
Base 1 came close to Fable, while being 4.5x cheaper.
Base 1 has the lowest frustration of anything we've tested. below Sonnet 5, below Opus 4.8, below Fable 5.
Fable 5 still edges it on success rate, 76.4% to 74.9%.
i'll post the full benchmarks in the comments
