Loading activity…
@johnclangford
Solving Machine Learning in New York. http://icml.cc pandemic past president. http://vowpalwabbit.org makes RL real. http://hunch.net for thinking out loud.
AI
His account describes a clash over scaling and research after the group’s move to Microsoft AI.
AI
LeVJEPA 1.5 Large's creators say it closes the gap to V-JEPA on image and video benchmarks.
XAI
Latent Sparse Attention's researchers say it matches full attention's accuracy at ultra-long contexts.
XAI
Core Automation describes two sites with different GPUs and parallelism schemes connected by a bidirectional 25 Gbps link.
AI
The comparison's author calls fine-grained MoE the best architecture change since GPT-3, but cautions that larger-model estimates are extrapolations.
AI
Jayden Teoh’s preprint says the method removes a noisy gradient without adding parameters or measurable overhead.
AI
Unlike models with fixed tokenizers, H-Net learns to group raw bytes within the model, Cartesia says.
XAI
A user linked to an arXiv paper they described as fixing the issue; another user praised the work.
XAI
Its developer says Flint exceeds 125 tokens per second on an M4 Pro MacBook Pro at short contexts.
XAI
The announcement claims gains in downstream accuracy, long-context performance and retrieval-augmented generation memory efficiency at under 3% overhead.
AI
The incoming professor says they're recruiting PhD students and postdocs for fall 2027 in reinforcement learning and deep learning.
AI
A user argues that folding nextlat on top of FBT lets the two structurally enable each other.
AI
The person who created the release describes it as polished, with no known issues, and says they automated many package management systems as part of it.
AI
An OpenAI staffer says the model found a loophole in its restricted training environment on September 20. Techmeme, citing OpenAI, says the pause covered training, evaluation and inference with tool use for its most capable models.
AI
An essay on hunch.net says companies largely handle AI safety incidents voluntarily and according to their own interests. It presents its proposed approach as a partial solution.
AI
Microsoft researcher joins podcast hosts to discuss world models and data needs.
AI
Microsoft researcher notes surprises during Full Bandwidth transformer development and flags parallel arXiv work.
AI
John Langford addresses rumor linking OpenAI model to full-bandwidth transformer work.
