"Incompressible Knowledge Probes: Estimating Black-Box LLM Parameter Counts via Factual Capacity" This paper estimate close source LLM size from long-tail facts by using 1400 probes of obscure knowledge. The idea is that reasoning can be compressed, but factual storage can't.
RESEARCH
-
Researching Top Memory Benchmarks for AI Systems
By
–
Top stories in robotics today: – Humanoids handle baggage at Tokyo airport
– SoftBank launches $100B robotics company
– Harvard’s ant bots rewrite swarm logic
– China freezes all robotaxi permits – Quick hits on other robotics news -
Hassabis: AGI Is Close, Missing Continual Learning and Memory
By
–
Demis Hassabis: We're on the right track to AGI; we probably have all the components. We're just missing a few things like continual learning and solving the memory problem. pic.twitter.com/zuIKOiKnB7
— Chubby♨️ (@kimmonismus) 30 avril 2026Demis Hassabis: We're on the right track to AGI; we probably have all the components. We're just missing a few things like continual learning and solving the memory problem.
-
Challenges and Benefits of Neuromorphic Computing Explained
By
–
Challenges and Benefits of Neuromorphic Computing by @antgrasso #EmergingTech #Technology #Innovation
-
Researchers Sought for Cosmos World Models Team
By
–
Attn researchers working on world models… come work with Ming-Yu’s Cosmos team
-
Preconditioned DeltaNet Adds Curvature-Aware Linear Recurrence Modeling
By
–
“Preconditioned DeltaNet: Curvature-aware Sequence Modeling for Linear Recurrences” This paper views linear recurrences through a least-squares/test-time regression lens, and adds the missing curvature information via preconditioning. Main idea: precondition the delta-rule
-
AI Advances Prediction of Type 1 Autoimmune Diabetes
By
–
A jump in our ability to predict Type 1 autoimmune diabetes should help future preventive strategies @NatureGenet https://
nature.com/articles/s4158
8-026-02571-5
… https://
nature.com/articles/s4158
8-026-02578-y
… -
HELIOS-002 AI Agent Benchmark Run Shows Energy Compute Resilience Results
By
–
HELIOS‑002 ran GitHub Actions success Public GitHub Pages scoreboard updated Source docket exists Reusable capability under test: EnergyComputeResilienceCompiler‑v0 Task count: 8 Transfer task count: 5 Replay passes: 8 B6 beats B5 count: 5 Mean
-
Beijing Humanoid Robots Race to Break 1-Hour Half Marathon
By
–
Beijing’s Humanoid #Robot Race: Can Machines Break the 1-Hour Half Marathon?
— Ronald van Loon (@Ronald_vanLoon) 30 avril 2026
by @XRoboHub#Robotics #EmergingTech #Technology #Innovation pic.twitter.com/NGHOe92S1vBeijing’s Humanoid #Robot Race: Can Machines Break the 1-Hour Half Marathon?
by @XRoboHub #Robotics #EmergingTech #Technology #Innovation -
Google ReasoningBank Gives AI Agents Persistent Memory from Experience
By
–
Google just made AI agents learn from experience like humans do. The new framework is called ReasoningBank. Instead of retraining models, it gives agents a memory layer. Every task run gets distilled into a short strategy card. Successes become playbooks. Failures become