They're almost always much worse than frontier models (no matter what anyone/benchmarks say). The last few releases have been very solid though (but I suspect frontier will pull far ahead again soon, so this won't last long).
RESEARCH
-
LLMs Struggle With Mature Scholarship Beyond Mathematical Problems
By
–
LLMs are “ good at math style problems, where you tell them A, B and C are true, and then ask them to figure out D [but]
extremely bad at anything involving what I would call mature scholarship .. [something] like naive undergrads” –
@kareem_carr -
Jensen Huang on AI Diffusion, Policy and Real-World Ecosystems
By
–
The Jensen + @dwarkesh_sp podcast was fantastic.
— Soumith Chintala (@soumithchintala) 20 avril 2026
Jensen is someone who understood how ecosystems work and someone who understands real-world trade, policy and controls work. And in some deeper sense how AI will actually diffuse into the world.
In this podcast, Dwarkesh came off… https://t.co/A2T9WcKQiBThe Jensen + @dwarkesh_sp podcast was fantastic. Jensen is someone who understood how ecosystems work and someone who understands real-world trade, policy and controls work. And in some deeper sense how AI will actually diffuse into the world.
In this podcast, Dwarkesh came off -
AI Transformers Cut Cancer Trial Failure Rate to 5%
By
–
🔬 Training Transformers to solve 95% failure rate of Cancer Trials
— Latent.Space (@latentspacepod) 20 avril 2026
the AI for Science pod is back with @RonAlfa, CEO of @NOETIK_ai, and Daniel Bear, VP Research at Noetik, explaining exactly how their team of top AI x Bio researchers and engineers (shoutout @owl_posting) will… pic.twitter.com/xdyYWraTkUTraining Transformers to solve 95% failure rate of Cancer Trials the AI for Science pod is back with @RonAlfa
, CEO of @NOETIK_ai
, and Daniel Bear, VP Research at Noetik, explaining exactly how their team of top AI x Bio researchers and engineers (shoutout @owl_posting
) will -
Kimi K2.6 Open-Source Model Beats GPT-5 Claude and Gemini
By
–
An open-source model just topped GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro on some of the hardest benchmarks in AI. Moonshot AI just released Kimi K2.6. What it's good at: > Long-horizon coding (12+ hour autonomous runs)
> Coordinating hundreds of AI agents in parallel > -
Japanese Cardboard Drone Achieves 120 km/h Speed Record
By
–
Cardboard #Drone from Japan Hits 120 km/h—Built in Just 5 Minutes
— Ronald van Loon (@Ronald_vanLoon) 20 avril 2026
by @FurkanGozukara#EmergingTech #Technology #Innovation pic.twitter.com/oxaz3D2G1pCardboard #Drone from Japan Hits 120 km/h—Built in Just 5 Minutes
by @FurkanGozukara #EmergingTech #Technology #Innovation -
NVIDIA’s Self-Evolving Logic Synthesis Framework with Multi-Agent LLMs
By
–
NEW paper from NVIDIA. EDA tools like ABC have been hand-tuned by humans for decades. New research from NVIDIA shows they can evolve themselves. The work introduces the first self-evolving logic synthesis framework: multi-agent LLMs autonomously refine the entire ABC codebase,
-
Kimi K2.6 becomes open-source SOTA on HLE
By
–
Kimi K2.6 from @Kimi_Moonshot is a new open-source SOTA on HLE with tools, SWE Bench Pro, and other benchmarks!
– HLE w/ tools – 54.0
– SWE-Bench Pro – 58.6
– SWE-bench Multilingual – 76.7 Looks like it is testing time now -
Gut Microbiome Signature Predicts Parkinson’s Disease Risk
By
–
We've known the gut-brain axis is a key underpinning of Parkinson's disease. Today, for the 1st time, a gut microbiome signature denoting risk found in healthy individuals with genetic predisposition @NatureMedicine https://
nature.com/articles/s4159
1-026-04318-5
… -
Simple Prompting Trick Improves LLM Randomness and Output Diversity
By
–
Getting LLMs to simulate “true” randomness or generate diverse outputs is surprisingly difficult. We found a simple prompting trick that solves this by having the model generate and manipulate a random string. To be presented at #ICLR2026 this week!
— hardmaru (@hardmaru) 20 avril 2026
Blog: https://t.co/CyevqqJ5ej https://t.co/dN0yZZ5MijGetting LLMs to simulate “true” randomness or generate diverse outputs is surprisingly difficult. We found a simple prompting trick that solves this by having the model generate and manipulate a random string. To be presented at #ICLR2026 this week! Blog: https://
pub.sakana.ai/ssot