Stanislas Polu
@spolu
Followers
18K
Following
3K
Media
433
Statuses
9K
_co-founder+engineer(https://t.co/SXBR0l9TrF @dusthq), _alumni(https://t.co/8jAnpFAkp1, https://t.co/e99AaHzlA0, https://t.co/4jg6knqi2S, https://t.co/kXE6PNf8xH)
Paris
Joined November 2007
Inverse scaling laws were fundamental in the exploration of pre training scaling laws so it’s super exciting to see them emerge for test time compute scaling. They eventually faded away (because scale won or because model size started being something we don’t readily have access.
New Anthropic Research: “Inverse Scaling in Test-Time Compute”. We found cases where longer reasoning leads to lower accuracy. Our findings suggest that naïve scaling of test-time compute may inadvertently reinforce problematic reasoning patterns. 🧵
0
2
11
The secret to a healthy startup is an healthy cofounder relationship and the secret to an healthy cofounder relationship is to take as often as possible time to think and build from first principle as if it was the first day. I think this also applies to any team. There is.
1
1
24
Disparate platforms (from Github to Notion or Google Drive), one unified filesystem-like representation for agents to navigate company data. We’ve built it for agents now I want to use it for work. `ˋ`.cat /work/notion/Engineering/AdvancedSearch > dust -a ProjectPlanner.ˋ`ˋ.
We built synthetic filesystems that map disparate data sources into navigable Unix-inspired structures. This transforms AI agents from search engines into knowledge workers capable of both structural exploration and semantic investigation across company data.
2
1
14
Every year we take an intern from high school for 1-2 weeks. It has been mind blowing to witness the evolution of what you can make them do with 0 prior knowledge in CS over such a short period. 2 years ago: ~nothing, a bit of JS. 1 year ago: a fairly complex self contained.
we've seen nothing yet! hosted a 9-13 yo vibe-coding event w. @robertkeus this w-e (h/t @antonosika @LovableBuild). takeaway? AI is unleashing a generation of wildly creative builders beyond anything I'd have imagined. and they grow up *knowing* they can build anything!
2
1
22
Interconnected computers augmented and accelerated us but it also increased the number of burnouts at work in the last decade. No surprises here as interconnected computers also drastically increased the volume of bits we're exposed to an ingesting at work, yet the information.
The current trope on entry-level white-collar jobs being wiped out is interestingly reminiscent of Meta making us believe that we would all work from home in the Metaverse: projection of a few people desires for a technology onto a world view that is seriously disconnected from.
2
3
21
The current trope on entry-level white-collar jobs being wiped out is interestingly reminiscent of Meta making us believe that we would all work from home in the Metaverse: projection of a few people desires for a technology onto a world view that is seriously disconnected from.
7
4
35
Often asked the question why we're positioning Dust as an horizontal product?. Horizontality makes Go To Market motions more complex and therefore harder to scale. We're ready to pay that price to build upon the vision of custom agents built by users for users that operate not
1
0
13
When the time ends and our branch of the simulation gets to present itself to the rest of the group:. « Fun fact, in my branch, artificial intelligence generalizes before IPv6 ».
0
0
11
When joining a project I generally start by exercising my craft. I'm in FPS mode (first-person shooter). I write code, I do research, I talk to users, I spin up instances. There are tasks that need to be done to make the project a reality, and I shoot them down as fast as.
2
0
18
Reflecting on @DarioAmodei's 'Machines of Loving Grace', there’s a phrase he used that I’m still chewing on: the "marginal returns to intelligence". He references economists' "factors of production": typically labor land, and capital.”. If you break these factors down
10
4
47
MiniF2F about to get saturated 🤯.
We believe formal math is the future. 🔥Introducing Kimina-Prover Preview, a Numina &.@Kimi_Moonshot collaboration, the first large formal reasoning model for Lean 4, achieving 80.78% miniF2F.
1
0
12
The reasoning <> knowledge superposition hypothesis.
You cannot separate reasoning and knowledge as cleanly as you think. If you'd asked me what I care about in 2020/2021, I'd have said it was “decoupling the capacity that language models have for understanding text from how they store knowledge” (quote from link below this.
0
0
5
Last night I heard tales about someone using Claude code in the most crazy/futuristic way. They basically have 10 local checkouts of their code base and launch in parallel Claude code on each of them doing various stuff or just exploring the space of approaches concurrently.
7
3
32
Feature request for X: filter out tweets based on regex. I would use /this changes everything/gi right away.
1
0
5