Alec Dewulf

Introducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.

Automating AI research is the next major step in AI We let Claude Code (Opus 4.7) and Codex (GPT 5.5) run autonomously on the nanoGPT speedrun optimizer track using our idle compute. ~10k runs, ~14k H200 hours Opus now holds the record at 2930 steps vs the 2990 human baseline

Automating AI research is the next major step in AI We let Claude Code (Opus 4.7) and Codex (GPT 5.5) run autonomously on the nanoGPT speedrun optimizer track using our idle compute. ~10k runs, ~14k H200 hours Opus now holds the record at 2930 steps vs the 2990 human baseline

I started a vibe run using the aurora optimizer. model is of param size 4B and is trained on 32k sequences. HPs are not carefully tuned (used the same lr & wd as muon baseline). i do observe some loss improvement over baseline early in training (see fig, -0.0093 at step 3k).

Ok directly relevant for ongoing work (on memorization): avoiding a "huge percentage of neurons to effectively die early in training (…) so that many parameters no longer meaningfully contribute to network outputs". This optimizer is going to see some SYNTH data.

Introducing Aurora, a new optimizer for training frontier-scale models. We train Aurora-1.1B, which achieves 100x data efficiency on open-source internet data. Despite having 25% fewer parameters, 2 orders of magnitude fewer training tokens, and using fully open-source internet-only data, Aurora matches Qwen3-1.7B on several benchmarks. Aurora was developed after identifying a major failure mode that can occur under Muon, an increasingly popular optimizer that has shown strong gains over Adam(W). We find that Muon can cause a huge percentage of neurons to effectively die early in training, reducing effective network capacity so that many parameters no longer meaningfully contribute to network outputs. By redistributing update energy more uniformly across neurons while preserving Muon’s stability properties, Aurora prevents neuron death and recovers substantial model capacity. What makes this work especially exciting is that it points toward a broader direction for ML research: better optimizers may not come purely from elegant mathematical abstractions, but from understanding and addressing the concrete dynamics and pathologies that emerge inside real training systems.

Introducing Aurora, a new optimizer for training frontier-scale models. We train Aurora-1.1B, which achieves 100x data efficiency on open-source internet data. Despite having 25% fewer parameters, 2 orders of magnitude fewer training tokens, and using fully open-source internet-only data, Aurora matches Qwen3-1.7B on several benchmarks. Aurora was developed after identifying a major failure mode that can occur under Muon, an increasingly popular optimizer that has shown strong gains over Adam(W). We find that Muon can cause a huge percentage of neurons to effectively die early in training, reducing effective network capacity so that many parameters no longer meaningfully contribute to network outputs. By redistributing update energy more uniformly across neurons while preserving Muon’s stability properties, Aurora prevents neuron death and recovers substantial model capacity. What makes this work especially exciting is that it points toward a broader direction for ML research: better optimizers may not come purely from elegant mathematical abstractions, but from understanding and addressing the concrete dynamics and pathologies that emerge inside real training systems.

Introducing Aurora, a new optimizer for training frontier-scale models. We train Aurora-1.1B, which achieves 100x data efficiency on open-source internet data. Despite having 25% fewer parameters, 2 orders of magnitude fewer training tokens, and using fully open-source internet-only data, Aurora matches Qwen3-1.7B on several benchmarks. Aurora was developed after identifying a major failure mode that can occur under Muon, an increasingly popular optimizer that has shown strong gains over Adam(W). We find that Muon can cause a huge percentage of neurons to effectively die early in training, reducing effective network capacity so that many parameters no longer meaningfully contribute to network outputs. By redistributing update energy more uniformly across neurons while preserving Muon’s stability properties, Aurora prevents neuron death and recovers substantial model capacity. What makes this work especially exciting is that it points toward a broader direction for ML research: better optimizers may not come purely from elegant mathematical abstractions, but from understanding and addressing the concrete dynamics and pathologies that emerge inside real training systems.

👀 Aurora dropping tomorrow. 3175 steps → beating NanoGPT Track 3 SOTA by 50 steps. And it scales 🚀

Distillation (especially on-policy) has become a pivotal component of the post-training stack. ☕ To dramatically accelerate distillation at scale, we open-source Nitrobrew, a communication-efficient, fused strategy for logit distillation. It’s built for both on- and off-policy distillation with: 100x faster loss computation 50% peak memory savings 3x faster on-policy distillation and more! A 🧵 (1/8)

Qwen3.5 2B and 0.8B Benchmarks almost seems like progress on tiny models has come to a stop