Seldon Lab Batch #2. In SF building the future of programming
Joined February 2023
- Tweets 202
- Following 130
- Followers 314
- Likes 98
62 Photos and videos
3 months ago I joined Seldon Lab as a solo founder.
1 month in, I pivoted. Started building Weft: a programming language designed for AI systems, where LLMs, humans, and infrastructure are first-class primitives.
2 months of heads-down building. Yesterday morning, I shipped.
Since then:
- 2,000 users in 48 hours
- 700 GitHub stars
- 10 paying customers
- Demo day today, went great
Still beta. Production-ready in ~2 months if I keep focus on the product.
Huge thanks to Seldon Lab, @EsbenKC and @MetzFinn81065 for betting on a solo founder.
And Sorhan for the visibility push!
If you want to build real AI systems on Weft, please reach out, I'll start onboarding business in a month or so
12
453
I'm open sourcing Weft today. A programming language where LLMs, humans, databases, and APIs are base ingredients, not libraries you import. You wire them together. The compiler checks the architecture. You get a visual graph of your program, automatically.
Why I built it ↓
2
1
6
527
The bet: don't steer the model harder, change the language under it.
1
170
Two months of solo work. Very early.
The catalog is small, the shape is still settling, and I'm building in public. If any of this resonates, I'd love your eyes on it.
📖 Blog post: weavemind.ai/blog/future-of-…
💻 GitHub (stars help more than you think): github.com/WeaveMindAI/weft
🚀 Try it: app.weavemind.ai
1
6
209
You won't be able to see this post — after this
1
2
159
Github: github.com/qfeuilla/emnuke
Chrome: chromewebstore.google.com/de…
Firefox: addons.mozilla.org/en-US/fir…
1
104
Well... things got out of hand. I made a pre-launch video on instagram and it went viral. What I expected even less when I woke up the next day was @Instagram banning me. If you discovered me there please contact me here in the meantime while I try to figure out an appeal. I'll give more details about the programming language I am building here next week 🚀
@Meta @instagram
My fresh account got banned (1 day). I just launched with an influencer on insta and I think this is because I got a spike of activity (>80k view), I tried every possible way to contact you but each account I made to do so were banned. My account is quentin_101010 is there any way to appeal on your end, or at least unblock my IP so I can create a new account. This is quite urgent I had important conversation going on. I can send more informations in PM
1
5
300
Quentin Feuillade--Montixi retweeted
@Meta @instagram
My fresh account got banned (1 day). I just launched with an influencer on insta and I think this is because I got a spike of activity (>80k view), I tried every possible way to contact you but each account I made to do so were banned. My account is quentin_101010 is there any way to appeal on your end, or at least unblock my IP so I can create a new account. This is quite urgent I had important conversation going on. I can send more informations in PM
23
1
7
565
I just thought of a big vulnerability for AI coding agent.
The agent needs contents:write to create branches and PRs. But that same permission lets it create files in .github/workflows/. If it pushes a branch with a new workflow triggered on on: push, GitHub runs it from the pushed commit, not from main.
Even if you reject the PR, the workflow already ran with all your Github secrets and elevated autorization.
1
2
135
I am only aware of one mitigation that actually prevents this while still letting the agent push code normally: a Push rulesets. It restrict who can push to .github/**/* while allowing pushes everywhere else. The agent can still create branches and PRs, it just can't touch workflow files. That's it. Everything else I am aware of is partial. And this is paid or OS, on the free plan, there's no way to let an agent push code while preventing it from pushing workflow files.
1
1
99
The permission to create PRs and the permission to create workflows are the same permission. You can't give an agent one without the other, unless you set up push rulesets.
Most people treat PR review as the security boundary. But on: push runs before review.
1
92
New post: Three Properties for Alignment (and Why We're Not Training for Them)
I think "alignment" means almost nothing without concrete properties. So I tried to define three: Red Lines, Embodiment, and Resilience.
2
2
13
1,213
My claim: we should train Embodiment and Resilience explicitly: "deeply absorb this configuration and resist external override". Only then add Red Lines to prevent specific configurations. We'd get models that are more controllable, more diverse, and harder to break.
1
3
196
Also: I think an open source model with strong Embodiment and Resilience, deployed in a well-configured system, can be more aligned than current closed source models that collapse everything into one persona for billions of users.
Full post: lesswrong.com/posts/sJph46Ea…
1
3
194