what if we stopped betting everything on one agent rollout? "The Unreasonable Effectiveness of Scaling Agents for Computer Use" Generates multiple trajectories in parallel & selects the best using "behavior narratives" 69.9% on OSWorld, nearly matching human-level 72%
MACHINE LEARNING
-
Humans and Machines Coevolve in Dynamic Synergistic Existence
By
–
Humachineology and humachinekind studies show that as we have been making our machines, our machines have been making us. Humans and machines have been evolving together, and they will continue to further evolve to maintain their dynamic and synergistic existence.
-
Recent Progress in Reasoning Methods Research Synthesis
By
–
It's actually thriving. I probably have bookmarked at >100 interesting papers on reasoning-related methods in just the last few months. I will share some thoughts / synthesis of all this in the upcoming weeks.
-
Humans and Machines Evolving Together in Synergistic Relationship
By
–
Humachineology and humachinekind studies show that humans and machines have been evolving together, and they will continue to further evolve to maintain their dynamic and synergistic existence.
-
Relaxing fundamental challenges in LLM training and metalearning
By
–
But 2 and 4 can be easily relaxed. They are not fundamental challenges. For 2, note that PAQ8 LMs were trained in continual style. We could easily do the same with LLMs, but indeed what we’re doing here is accelerating evolution in the outer metalearning loop. 4 is often
-
Apple’s MoE Scaling Breakthrough: RoE Hyper-Parallel Inference
By
–
An intriguing paper from Apple. MoEs Are Stronger than You Think: Hyper-Parallel Inference Scaling with RoE Paper: https://
arxiv.org/abs/2509.17238 -
Unified Architecture Different Parameters Activations
By
–
Unify in architecture, not in details. The parameters/activations will be different.
-
Missing GPT-1 Base Baseline in Comparison Analysis
By
–
Ok, so my criticism here is not the scores. But the right baseline is missing.
Should have a "GPT-1 base" column. -
Model-Tool Cotraining Achieves Better Results Than Larger Models
By
–
err i commented too quickly and people might get overexcited. i was just saying model-tool specific cotraining is achieves same or better results than using a 10-100x larger model without cotraining. this research doesn’t touch on memory/context compression, which is yet another