Stanislas Polu

co-founder+engineer( alumni(

[Stanislas Polu @spolu]

Agents need servers not computers: The emergence of hyper-personalized software. Why localhost is dead and why it's all exciting.

https://t.co/16jNTq7UcM

[Stanislas Polu @spolu]

Come in London on June 24 at the Cerebral Valley Summit to chat about token-economics!

Is tokenmaxxing sustainable? where's the margin going? or the only key question that matters: Will we be using PhD-level reasoning models to input CRM entries??

https://t.co/4pOQ7y93cJ

[Stanislas Polu @spolu]

@isuboiiiiii @Leonard41111588 https://t.co/cXN7OJeHt4

[Stanislas Polu @spolu]

Great read: https://t.co/PCKwxc9D1n

@Leonard41111588 do I understand correctly that the future you see for verified software is one where the verified software is written in Lean, and hence executed by Lean?

It's appealing but moving developers there will take longer time than the time for AI to be able to verify very powerful properties about very complex software stacks...

Is there slightly less pure world where we have a transformation of existing code into a substrate we can reason about and prove property with Lean?

[Stanislas Polu @spolu]

@swyx @latentspacepod @dust4ai @OpenAI @gdb @ilyasut We need to redo a pod on this[0] and this[1] as soon as the product is fully lined up!

Cowork is superb from a PLG standpoint but completely not future proof (since player) for what’s to come!

[0] https://t.co/BWrrv8XY4S
[1] https://t.co/16jNTq7UcM

[Stanislas Polu @spolu]

@victor_lfrst Mea maxima culpa

[Stanislas Polu @spolu]

Fuck off

[Gap | Suby @gaspardlezin]

France is hardcore mode for founders:

• Pay an employee $5K net → costs you $13K
• Make profit → 30% corporate tax
• Succeed → public calls you an exploiter
• Get famous → kidnapping becomes a real threat

No other country stacks the difficulty this high.

[Stanislas Polu @spolu]

@vishalmisra @ShriramKMurthi @Hesamation The question that remains after the read is to me: what distinguishes us from an LLM to give us access to rung 2 and 3? I agree LLM training is very rung 1 but once it runs auto-regressively with tools I don’t see what prevents it from intervention and countefactual?

[Stanislas Polu @spolu]

@HenriRoussez @gaspardlezin Safety to free your spirit. Anti-996 (real work not anti-work). Thriving research environment. Exploding pool of talent that also happens to be loyal. Beyond that culture taste and spirit creating the richest company cultures.

[Stanislas Polu @spolu]

@HenriRoussez @gaspardlezin Oh! I am 👌

[Stanislas Polu @spolu]

Random memory and post in the ether of the X nebula, maybe you’ll catch it maybe you won’t…

Remember when we met with Ilya before I even joined OAI, must have been late 2018, early 2019. It was downstairs Pioneer (same room the API was born and same room I interviewed with pc to join Stripe. That room ❤️). I was working with a friend on applying deep learning (and eventually transformers on Ilya’s advice) to fuzzing to discover vulns[0], and was sharing them with you guys.

I remember Ilya asking: what would it mean for the world to be able to find any vulnerabilities at scale.

7 years later…

This.

🤯

[0] https://t.co/CgeICwWGvK

[Stanislas Polu @spolu]

Oh sorry if rude. I meant it in a respectful way. If you prefer « get lost » 👍

No point in pointing broken things without even an ounce of an attempt to fix it. Just unproductive. France has a lot to give to founders and you’re are criminally mis-representing the reality leveraging uninteresting tropes that have been there for too long. What you say on top of being biased beyond recognition is backward looking and misses the entire point of what France has best to give.

[Stanislas Polu @spolu]

Blog post: https://t.co/LvJVtnsVwB

[Stanislas Polu @spolu]

We're only year 3 of a decade (if not multi-decades) long transformation of work.

3 years ago we bet on building an horizontal platform for work with agents, a chance to invent a new operating system for companies, from scratch, with AI as a fundamental premise. Many people considered us crazy for going after that, praising verticalized AI products as the winning strategy. But here's the thing: the time horizon of tasks successfully handled by agents has been predictively increasing form minutes to hours and will in all likelihood reach the equivalent of days and weeks of human work equivalent in the coming quarters.

This is were verticalized and/or single-player AI falls short. Single-player tools, one person, one agent, confined to your machine is the wrong architecture for what's coming. We're shifting from using AI to produce things, to managing fleets of agents that do the producing. 3 years ago I wrote[1]:

"ChatGPT is the Pong of LLMs. [...] Imagine, one day we'll get the DOOM, Civ, Red Alert, and Counter Strike of LLMs. Let alone multiplayer modes."

Weeks long tasks in companies are inherently collaborative and mechanically spanning multiple teams. The new bottleneck in harnessing agents within organizations is coordination: multiple humans and multiple agents need to work together, with shared context, shared tools, shared goals. Agents that can hand work off to other agents or surface decisions to the right person at the right time. Humans who can review, steer, and step in without losing the thread. Teams that can run parallel workstreams and actually stay aligned.

This is Multiplayer AI, and that's what we've been building at Dust.

Across Datadog, Clay, Persona, 1Password, Doctolib and 3,000+ organizations globally, we've watched teams figure out what this looks like in practice. 300,000+ agents deployed. 70% weekly active. 240%+ NRR.

Today we're announcing a $40M Series B with Abstract, Sequoia, Snowflake, and Datadog to accelerate our vision.

Designing the right interfaces for multiplayer AI is the next frontier. Join us to redefine work by defining multiplayer AI.

[Stanislas Polu @spolu]

@ptsi Thank you! \o/

[Stanislas Polu @spolu]

@datadito @Fabien_Mikol Transfer is everywhere!

[Gabriel Hubert @gabhubert]

Most AI at work is still single-player. The gains stay trapped with each individual. Nothing compounds across the team.

Today, we're announcing @DustHQ $40M Series B to scale multiplayer AI: humans and agents working in parallel, with shared context, shared tools, and shared goals. 🧵

[Stanislas Polu @spolu]

@AymericRoucher On est cooked.

[Stanislas Polu @spolu]

@gabhubert Was the book deutsch?

[Stanislas Polu @spolu]

This is such fantastic science of DL and human science at the same time.

The fact that there is scaling on pre 1931 data on coding evals is mind blowing.

If you think LLMs are just stochastic parrots representing probability distributions, that means that coding in TS was backed into human knowledge decades before it existed.

If you think that LLMs truly generalize to raw intelligence that’s an incredible confirmation.

But the interesting bit is that it’s one or the other and whichever it is, the conclusion is 🤯

[Nick Levine @status_effects]

New work with @AlecRad and @DavidDuvenaud:

Have you ever dreamed of talking to someone from the past? Introducing talkie, a 13B model trained only on pre-1931 text.

Vintage models should help us to understand how LMs generalize (e.g., can we teach talkie to code?). Thread: