Romeo Dean (@romeovdean) / X

Romeo Dean

135 posts

Romeo Dean

@romeovdean

Researcher @AI_Futures_. AI 2027

Berkeley, CA

Joined April 2025

Romeo Dean
@romeovdean
May 1, 2025
We’re not the only one's confused by OpenAI model names. Which reasoning models come from which base models? Why is the new o3 different? What is the plan for GPT-5? How good will it be? We dive into our current best answers to all these questions and more in a new @AI_Futures_
15K
Romeo Dean
@romeovdean
Aug 7, 2025
52.8 > 69.1 = 30.8
57K
Romeo Dean
@romeovdean
Apr 8, 2025
AI 2027 features 100M H100-equivalents, 10^28 FLOP models (1000xGPT-4), and 300K AI copies running at 50x human speed by the end of 2027. Takeaways from the compute forecast behind these predictions. 1/7🧵
17K
Romeo Dean
@romeovdean
Apr 24, 2025
In response to concerns that we underestimated China in AI 2027, @slatestarcodex and I wrote up Why America Wins. In a word: compute.
5.6K
Romeo Dean
@romeovdean
Apr 5, 2025
Replying to @DaveShapi
Thanks for engaging David, we appreciate disagreements and that's part of the reason for laying something out with this much detail! Here's a partial reply that might address some of your concerns: 2. ai-2027.com/research/timel… covers some of the empirical data / evidence for the
Timelines Forecast — AI 2027
From ai-2027.com
2.8K
Romeo Dean
@romeovdean
Apr 23, 2025
Awesome, comprehensive work! It is nice to see how closely this data matches my estimates for the global distribution of chips: US / China / RoW (Dec 2024) 74% / 12% / 14% (AI 2027 compute forecast estimate) 74% / 14% / 12% (Konstantin & Epoch dataset)
Konstantin Pilz
@KonstantinPilz
Apr 23, 2025
After 2 years tracking GPU clusters, we're releasing our dataset of 700+ AI supercomputers. The US clearly dominates countries, and xAI’s Colossus is leading system: 200k AI chips, $7B price tag, and power needs of a medium-sized city. Here are my personal highlights🧵
6.1K
Romeo Dean
@romeovdean
Oct 29, 2025
Replying to @elonmusk
Neat. And while your car is parked at home charging, your household robot can operate off of the car’s chip, giving it way more low latency compute.
677
Romeo Dean
@romeovdean
Apr 27, 2025
Replying to @1a3orn
Data is definitely an input, but it’s mostly either freely available (so not worth mentioning because both sides have ~equal access) or synthetic data (that basically can be thought of as part of algorithms).
5.3K
Romeo Dean
@romeovdean
Apr 8, 2025
Replying to @romeovdean
There's also a large spreadsheet you can check out that is behind the compute production model. See ai-2027.com/research/compu… for the full report and ai-2027.com for the scenario it informs. 7/7
1.2K
Romeo Dean
@romeovdean
Oct 24, 2025
The amount of electricity generated by solar has been doubling every 3 years. In the US solar has also been growing with a 3 year doubling time. Wind has doubled every 6 years and gas has doubled in 16 years. Other major sources like Coal and Hydro have decreased, and Nuclear is
Dr Paul Dorfman
@dorfman_p
Oct 24, 2025
4K
Romeo Dean
@romeovdean
Oct 28, 2025
Damn. Nvidia is miles ahead in robotics. Look how lifelike!
Shay Boloor
@StockSavvyShay
Oct 28, 2025
$NVDA SAYS HUMANOID ROBOTS WILL BECOME THE LARGEST ELECTRONICS MARKET
3.2K
Romeo Dean
@romeovdean
Apr 8, 2025
Replying to @romeovdean
Compute production breaks down the chip supply chain to answer how much compute will be available? Our forecast is 100M H100-equivalents by the end of 2027, which is 10x more than exists today, as there's plenty of fab capacity for more raw logic wafers to shift to AI. 2/7
1.1K
Romeo Dean
@romeovdean
Apr 8, 2025
Replying to @romeovdean
Compute distribution projects a plausible breakdown of compute between AI companies in 2027, with leading companies expected to have 10-20M H100-equivalents each. 3/7
987
Romeo Dean
@romeovdean
Apr 5, 2025
Replying to @DaveShapi
2. ai-2027.com/research/compu… argues why 30:1 should be on the limit 5. We do say the weights will constantly be updated for new versions of the model, haven’t heard of the term ‘liquid,’ but guessing it’s a similar idea. 6. sorry might have been the wrong link?
Compute Forecast — AI 2027
From ai-2027.com
665