Darmani

The instruction tuning between 3 and 3.5 was a big deal. If you prompted raw GPT 3 "Tell me a story about a mouse", it was as likely to continue "2) Tell me a story about a cow. 3) Tell me a story about a rabbit" as it was to tell you a story. I tried to use GPT 3 for work and found its only utility was to get over blank page fear by writing something so bad I would fix it in anger.

> Heh, personally I fixed this by just adopting the sleep cycle my body wants of going to bed at 04:00/05:00 and going up at 11:00/12:00, life is much better now when I just accept it. One approach if your life can allow it :)

That was my life in my mid-late 20's.

But as I've gotten older, my sleep schedule has only gotten more messed up. Now I consider it a victory if I manage to go to sleep before the dawn.

> Very much so, obviously prefer something async if possible, just a .patch file could suffice I suppose, but could do a call to have a look if that's the only way :) Reach out to my email from my profile and we can coordinate :)

Cool, let's chat async then. Contacting you now.

Most days I don't ship 20 PRs. But I think my record is 30.

Three things made that possible.

The first, obviously, is having Command Center.

The second is that a lot of those were fixes or UX improvements under 100 lines.

The third is, no joke, not sleeping. I've had quite a few 20+ hour days in the last 6 months. Some of that is work pressure, but also I've considered getting evaluated for a broken circadian rhythm.

> I'd be curious to see some of those PRs if you're saying you've essentially solved the holy paradox of "ship fast = shit code" or "ship slow = good code".

If you're serious, I'll be happy to get on a call and show you.

Hi Fizzadar,

On the one hand, it is true that the website code was pushed 0 minutes before this announcement went up.

On the other hand, I tested just now on two different phones and didn't see any issues. Can you say in more detail what you expected vs. what actually happened?

There was an occlusion issue on some smaller screens, but it's been fixed now.

Hi egamirorrim,

The basic answer is that it runs locally. If you turn telemetry off and don't use our free Gemini credits, it's trivial to verify that no traffic goes to our servers other than a tiny subscription check. For our enterprise customers, we offer a version that doesn't even do that. Everything stays between you and your model providers (and we support custom and local models).

SOC2 is still a work in progress. I'm a former security researcher with work featured in the New York Times, and I know that doing it right (and not going through Delve) takes time. I can tell you that we have passed a compliance check for a company in a highly-regulated space.

I didn't find your contact info, but I'm available at [email protected], and happy to discuss your needs.

Thanks!

The final moments before this launch announcement consisted of me twiddling my thumbs while waiting for our designer to upload any version he could get ready in time that is better than the previous version of our website. So we knew we'd be launching with a lot of imperfections in the visuals. Did test in mobile, but not on iPad.

In fairness, I did most of these interviews last summer, and I know some people have changed. And while I did go a fair bit outside my network to interview people, there are all sorts of hard-to-understand selection effects that come from me being me. A 21 year-old frat boy who tried doing the same kind of interviewing with the people he could find to interview would probably get different results.

But yes, that is indeed what happened. Multiple times, I'd talk to someone that I'd expect to not be reading the code at all (solo founder, mostly nontechnical), then I'd interview him in detail about his workflow and think "Huh, there was absolutely no point in there where he was reading stuff," and then I'd ask "So how much of your time is reading code?" "60, maybe 70%"

We have a referrer program

Doug has not signed up for it.

In honesty, that's not a bad idea, and we hadn't thought of that.

It's pretty expensive to measure even for small programs. It's also more of a relative than an absolute measure, i.e.: it scores two variants of the same codebase, but the raw scores aren't very meaningful on their own. So our goal had been to use this in the benchmark set we're working on when we release a standalone refactoring product.

But the more I think about this suggestion, the more I think: "Hmmm, why not?"

And indeed, I think we're the only agentic coding environment with jj support.

The most difficult code in the 1.0 release is some gymnastics to avoid the appearance of a concurrency conflict with a user running their own jj commands, made at the request of the person who introduced me to jj.