This was fun to listen - a few points:
1. SW Factory is first and foremost people problem. Outside of SF / Startup scene not everyone is excited about being replaced.
2. To self-improve you need access to prompts. Not trivial at big company.
final edit of our software factories masterclass with @vaibcode and @zachlloydtweets is live! We chatted about:
- software factory building blocks
- a good software factory starts with measurement, then self improvement against those metrics
- turning successful agent runs into evals for testing future models / harnesses / etc
- DORA, why counting lines of code is bad metric, and why counting PRs might be good enough
enjoy!
3. They talk about how to run A/B of the factory code which is really cool, but I have no idea how to do that when the underlying context coming from 10s of other systems is changing. You can't easily capture the world unless you start versioning the world.
4. The whole "engineers talking to customers" works amazingly well when engineers are building engineering products, or working in a B2C where the domain is trivial (eg stuff we all use) but less so in other domains/companies.
Good listen as always @dexhorthy !
Why the hell can't I forward Whatsapp messages to Muse (in Whatsapp)?
The whole "Meeting clients where they are at" premise not working is really disappointing.
Before AI, I used to start a side project and never finish it.
Now with AI, I start way more side projects and still never finish them.
We’re in the Instagram era of intelligence — optimized for the feed, the aesthetic, and the appearance of knowing
re: youtube.com/watch?v=Usufn8IQ… @dexhorthy
For the non-vibers around I feel like this is a great talk but I can't resist the idea that it's all meaningless.
Yes the concepts would stay and the tooling would change, but the whole "give it context" but "not too much" is a dice roll
And the whole stance against dark factories (which I get) but then how do you handle the load of code that comes in?
You need to pick either you deliberately slow down or not. And if you do, is that smart?
Almost forgot - @CodeSignalCom there's a new xterm flag you should probably enable so this doesn't happen to future candidates called `scrollOnDisplayErase` (github.com/xtermjs/xterm.js/…)
1/ Tech interviews can be annoying.
It's time to share how I bombed my @AnthropicAI screening interview.
/A uses @CodeSignalCom . I had to write a node script and run it to ensure it's working.
I like to have a fresh terminal between runs so I usually use command + K shortcut
1/ Tech interviews can be annoying.
It's time to share how I bombed my @AnthropicAI screening interview.
/A uses @CodeSignalCom . I had to write a node script and run it to ensure it's working.
I like to have a fresh terminal between runs so I usually use command + K shortcut
/7 I knew something was wrong, and initially thought it was my code. But it made no sense. Took me a while to realize it's a problem with the platform, but I didn't know what it was. I called it out, tried refreshing, nothing worked.
I was blind to my own code...
IMO A lot of the "agentic work" would scale down to building deterministic apps instead of agentic runtime.
Why? Mostly because of Token Cost & determinism.
I see so much stuff that could have been a script/program.
One cool trick I use is "deep linking" into a thread.
So I'm on CodexApp/Claude and I have a skill that's basically "if work is needed on another repo, deep link to it". This opens a new thread with the prompt ready.
Clean context, clear visual separation of work.