p(doom) (@prob_doom) profile picture

Turning sand into intelligence. One bitter lesson at a time.

Joined June 2024
Photo by p(doom) (@prob_doom)Photo by p(doom) (@prob_doom)Photo by p(doom) (@prob_doom)Photo by p(doom) (@prob_doom)Photo by p(doom) (@prob_doom)Photo by p(doom) (@prob_doom)Photo by p(doom) (@prob_doom)Photo by p(doom) (@prob_doom)Photo by p(doom) (@prob_doom)Photo by p(doom) (@prob_doom)
If you want models to work on tasks for months, you have to train on months-long trajectories. Today we release an inverse dynamics model for screencasts. It watches a screencast and recovers the actions behind the pixels: keys, clicks, cursor movement and scrolls. 🧵(1/5)
2
6
8
47
9,352
Grifter SF startup openly published 11h of screencasts and calls our dataset 'less open-ended' We have 15,000 hours of screencasts of top-tier PhD students working over 3-12 months. Their longest trajectory is 12h. Ours is 12 months.
🤖 Made with AI
We're releasing RealCUA, the first computer-use dataset of open-ended work in evolving environments. 600+ hours (and growing!) of real multi-objective work with interleaved tasks across thousands of steps in realistic desktops. 11-hour sample & access: data.fig.inc/realcua
1
8
1
21
852
Replying to @figbrains
Your claim is not true. 7 months ago, we openly released 600+ hours of the 15,000 hours we have internally (of top-tier PhD students working over 3-12 months). Also, our cross-platform capture tooling is open source. teapawt.net/prob_doom/status/20372…
You really think we're going to scale data labelers to AGI? Today, we release the largest public long-horizon dataset of human digital work. 600h of long-horizon AGI research across 3 months. 🧵(1/n)
2
3
1
24
1,094
We're spending $250,000 to record cracked people working. The problem: finding them. So we printed 10,000 flyers and started hitting Europe’s AI hubs.
4
1
3
16
3,408
The model was trained only on macOS data, but transfers to Windows and Linux without adaptation. Mouse movement and clicks generalize best. The remaining cross-platform errors are mostly hotkey confusion. (4/5)
1
10
432
TRY US
Replying to @menhguin @prob_doom
200k nukes, these traces are never going to be on-policy, my policy cant be imitated teapawt.net/ComplexMusic/status/98…
1
1
2
1,508
We’re p(doom), an AGI research lab. We’ll pay you $300/month to record your screen while working. If your work is open-source and involves research, engineering, design, editing, or similar long-horizon digital work, fill out the form: forms.gle/Um8pzes644tL661F8
17
10
13
201
555,473
You really think we're going to scale data labelers to AGI? Today, we release the largest public long-horizon dataset of human digital work. 600h of long-horizon AGI research across 3 months. 🧵(1/n)
3
22
6
168
24,105
As with crowd-code and AGI-CAST, we will publish the cleaned dataset under the most permissive Creative Commons license (CC0). (8/n)
1
8
1,046
If you openly publish your work, consider participating in crowd-sourcing by installing crowd-cast. (n/n) 🖥️ crowd-cast: github.com/p-doom/crowd-cast… 🌐 Blog: pdoom.org/crowd_cast.html 📈 Dashboard: pdoom.org/crowd_cast_dashboa… 🤗 Dataset: huggingface.co/datasets/p-do…
14
1,018