Nicolas Joseph

CTO, Co Founder @514hq - The views expressed here are my own. ex @dotswoosh, ex @datalogue

[Nicolas Joseph @thatsFrScience]

Agents now review each other's PRs until the reviewer finds nothing. Then I get one card: review rounds, CI, Bugbot, approve or not.

Typical one last week: zero findings after 4 rounds.

Trying https://t.co/GSphldAcei next to @cursor_ai Bugbot now. The reviewer's context is the knob I tune most.

[Nicolas Joseph @thatsFrScience]

@bcantrill Congrats! This is a really awesome milestone, hardware is soo hard that speaks to the talent of your team.

I secretly hope I am someday in a position where I can order some of those babies 👶

[Nicolas Joseph @thatsFrScience]

When someone's stuck in a doubt I've already lived through, my reflex is to hand them the conclusion.

It rarely works.

What works: sharing the doubt itself. How it felt, what kept me stuck, what got me out.

The conclusion says where to go. Your story shows they can get there.

[Nicolas Joseph @thatsFrScience]

@Docker's new docker-agent pitch: teams of specialized agents. We just tested that.

1 research subagent vs 4 parallel specialists, same model:
• invented claims: 25 → 13
• coverage: equal or better
• speed: same or faster
• tokens: 2.2×

Narrower context, fewer made-up facts.

[Nicolas Joseph @thatsFrScience]

@QuinnyPig @rauchg @vercel 😹

[Nicolas Joseph @thatsFrScience]

Hey @rauchg we are hitting an issue with env variables size for a project and it doesn't look like you have a way to raise that limit.

This is core to leveraging @vercel for AI-native development; we use previews to create an isolated dev environment for agents, tied to @supabase branches.

We are throttled in our ability to spin up agents and work because the preview envs are not scaling.

It looks like nothing is committed to the roadmap to remediate this; can you fix it?

[Nicolas Joseph @thatsFrScience]

@rauchg @vercel envs_size_too_large ("Unable to save environment variables. Insufficient space.")

[Nicolas Joseph @thatsFrScience]

The “wrong path” problem isn’t new. Managers always burned weeks if the shot on goal was wrong.

What’s new: every IC is now a manager of agents. Waste is less calendar time, more tokens — full workflows compress into days.

Anyone else coaching ICs through that shift?

[Nicolas Joseph @thatsFrScience]

Hey @artman, could you share updated numbers on engineering velocity in your org, with AI agents at the core of your build workflow?

I'm a big Linear user and a fan of the quality you guys deliver. I'd like to get a sense of the throughput benchmark.

[Nicolas Joseph @thatsFrScience]

Something clicked: the unlock isn’t one smarter agent — or even one manager agent.

It’s several agent managers in parallel, one per work track. Same instinct as Cursor Projects: scoped context, own workers, clear seams. I review across tracks instead of living in every ticket.

Bigger unlock for me than any single model bump.

[Nicolas Joseph @thatsFrScience]

@buildcanada @harleyf @cbcwatcher When build Canada - Montreal?

[Curtis G. Northcutt @cgnorthcutt]

Today we establish that AI is as good as expert human tutors for immediate GRE learning gains (p=.015, n=2,383 students, study conducted July-Sep 2026).

One hour with an expert GRE tutor: $75.
One equivalent hour with an AI tutor: 7 cents .
918x cheaper, equivalent learning (p=.044, n = 140).

In 5 of 7 academic topics, the top AI tutor beat the expert human tutors on average.

How we measured it: pre-test -> 1 hour with (AI or human or no) tutor -> post-test.
There are no LLM judges. Learning gains are post-pre. Everything is real students.

Why we did this:
> 1 trillion USD is being spent per year on machines making machines better. StudentBench shows us how to shift these resources so machines also make humans better.

Why I did this:
I grew up in rural Kentucky. My dad, grandad, and great grandad were all mailmen. Getting into a good college changed my life. Now we have done a rigorous study to help AI labs evaluate LLMs and help show how bright AI can be for students of all incomes.

I believe in good science that leads to the democratization of opportunity.

The research paper behind StudentBench: https://t.co/JfkDDllP4i
Data, code, details: in this thread

If you're wondering what I'm selling, sorry to disappoint: we released all study data for free on Hugging Face, and made https://t.co/5G9L8yAxrt free to use. Enjoy.

lots more in the video / thread!!

[Nicolas Joseph @thatsFrScience]

Hey @mitchellh how are you guys thinking about alpha? I am very happy to test early software the versions I have seen so far seems right smack into what I need. The server multiplexer and the web lib. I believe I would replace kittyjs and tmux with this.

[NVIDIA HPC Developer @NVIDIAHPCDev]

Introducing CUDA Rust!

CUDA Rust lets you write GPU kernels natively in Rust, not just launch them from it.

Two paths: cuda-oxide for SIMT kernels compiled to PTX, and cutile-rs for Tile-based programming on stable Rust. Both can catch aliasing errors at compile time.

Technical blog: https://t.co/5mYSRnQkBB

[Nicolas Joseph @thatsFrScience]

Hey @karrisaarinen, @linear agent should be able to take in feedback for the Linear team:

>Can you send feedback to the Linear team saying we need to have project description history with agents?

I can’t send feedback directly to Linear. Please send this to [email protected]

^ not very agentic

[Nicolas Joseph @thatsFrScience]

Hey @nikitabase, did you guys stop developing mooncake? Or do you have something else in the works for neon and columnar?

[Nicolas Joseph @thatsFrScience]

@adamshuaib Hey Adam, I am curious how I would score! Did you publish anywhere the questionnaire / algorithm?

[Nicolas Joseph @thatsFrScience]

@mitchellh @sudomateo @Tailscale I am assuming @mitchellh that what you are building won’t cover the networking layer.

[Nicolas Joseph @thatsFrScience]

@mitchellh @sudomateo Hey @Tailscale would be great to have tailscale be natively compatible with superlogical when it comes out. Like you do for ssh

[Nicolas Joseph @thatsFrScience]

@Codie_Sanchez Yes