In September 2026 I also outlined a European Transformative AI strategy. The plan calls for a three‑year political compute pre‑commitment of 16 GW, an early operational footprint of 50‑100 staff with hundreds of MW of compute, and a 99 %/1 % split of compute between research/training and inference. Independent technical reviews are scheduled after about 18 months and then at increasing frequency. The overarching goal is to accelerate the learning loop by shipping competitive models every few months and releasing them as open‑weight checkpoints.
The political compute pre‑commitment is 16 GW for the first three years.
Compute allocation is 99 % to research/training and 1 % to inference.
Independent technical reviews are planned after ~18 months for the first review and then more frequently.
In September 2026 I posted a series of updates on X about my move to San Francisco, the launch of decision model d1, collaborations with Qualcomm, chess‑engine experiments, and infrastructure improvements.
"should_try_d1": {
"noul": 0.9999
}
output_tokens: 0
price: $0
It’s the first model to outperform Jev on @huggingface's Decision Index.
I had to stop training it because it was destroying Opus 5.
The Fly model achieved a 4‑1‑5 win‑loss‑draw record against Jev in a 10‑game match, but further training was halted for safety.
Fixed‑depth search contributed roughly 400 Elo points to our chess engine’s strength, outperforming more complex MCTS approaches.
before: ~100 GB on disk after: ~20 GB
Provides up‑to‑2.8× throughput of stock TRL
Our partnership with Qualcomm enabled Liquid Context to run efficiently on Snapdragon processors and the Hexagon NPU.