J Rosser
@jrosseruk.bsky.social
Research Scientist Intern at Spotify | DPhil in Machine Learning at the University of Oxford
Quiz time: I shrank my pretrained LLM's VRAM usage from 1TB+ to 50GB with zero drop in performance. What did I do?
1) 🪄Quantization🪄
2) ⚡FlashAttention⚡
3) 🏁Gradient checkpointing🏁
4) 💾KV cache to CPU💾
1) 🪄Quantization🪄
2) ⚡FlashAttention⚡
3) 🏁Gradient checkpointing🏁
4) 💾KV cache to CPU💾
June 27, 2025 at 12:56 PM
Quiz time: I shrank my pretrained LLM's VRAM usage from 1TB+ to 50GB with zero drop in performance. What did I do?
1) 🪄Quantization🪄
2) ⚡FlashAttention⚡
3) 🏁Gradient checkpointing🏁
4) 💾KV cache to CPU💾
1) 🪄Quantization🪄
2) ⚡FlashAttention⚡
3) 🏁Gradient checkpointing🏁
4) 💾KV cache to CPU💾
🎉It's happened! AgentBreeder got its first citation!🎉
arxiv.org/abs/2506.04572
arxiv.org/abs/2506.04572
June 17, 2025 at 10:26 AM
🎉It's happened! AgentBreeder got its first citation!🎉
arxiv.org/abs/2506.04572
arxiv.org/abs/2506.04572
Excited to share that I've joined @spotify.bsky.social as a Research Scientist PhD Intern this summer investigating Mechanistic Interpretability for long context reasoning in LLMs!
#Spotify #InternAtSpotify #LifeAtSpotify
#Spotify #InternAtSpotify #LifeAtSpotify
June 12, 2025 at 8:19 PM
Excited to share that I've joined @spotify.bsky.social as a Research Scientist PhD Intern this summer investigating Mechanistic Interpretability for long context reasoning in LLMs!
#Spotify #InternAtSpotify #LifeAtSpotify
#Spotify #InternAtSpotify #LifeAtSpotify
Had an awesome time presenting mine and @jfoerst.bsky.social's "AgentBreeder" paper at #ICLR2025 to a room of over a hundred researchers! Our approach addresses pre-deployment Al safety risks in multi-agent systems by evolving thousands of "scaffolds" from base LLMs in a multi-objective setting.
June 12, 2025 at 8:11 PM
Had an awesome time presenting mine and @jfoerst.bsky.social's "AgentBreeder" paper at #ICLR2025 to a room of over a hundred researchers! Our approach addresses pre-deployment Al safety risks in multi-agent systems by evolving thousands of "scaffolds" from base LLMs in a multi-objective setting.