NEW Research from Meta Superintelligence Labs and collaborators. The default approach to improving LLM reasoning today remains extending chain-of-thought sequences. Longer reasoning traces aren't always better. Longer traces conflate reasoning depth with sequence length and
MACHINE LEARNING
-
MIT’s Loop-Closure Grasping: Balancing Versatility and Strength in Robotics
By
–
Robotics just took a meaningful step forward by rethinking the shape of grasping, not just the force behind it. MIT researchers have introduced loop-closure grasping, a concept that elegantly solves one of robotics’ oldest tradeoffs: Versatile grasps are rarely strong.
-
New Research on Bounding Hallucinations in RAG Systems
By
–
This paper explains why RAG works in demos and breaks in production. The paper is “Bounding Hallucinations: Information-Theoretic Guarantees for RAG Systems via Merlin–Arthur Protocols.” And it attacks the hallucination problem from a direction almost nobody in applied AI is
-
DeepCode: Open Agentic Coding System by Li et al.
By
–
DeepCode: Open Agentic Coding Li et al.: https://
arxiv.org/abs/2512.07921 #ArtificialIntelligence #DeepLearning #MachineLearning -
Macaron focuses on smarter training over benchmark chasing
By
–
Most AI companies chase benchmarks nobody cares about. Macaron chases real intelligence that learns from actual experience. Research that ships: https://
macaron.im/mindlab
Product that works: https://
macaron.im Bigger models won't save you. Smarter training will. -
RL Efficiency Gains in Production Apps
By
–
Mini-apps went from 20 minutes to 2 minutes. That's not optimization. That's demolition. The RL efficiency isn't hiding in benchmarks. It's crushing it in production. Stop reading papers. Start using products that actually ship the research.
-
Mind Lab’s Technical Innovations in AI Training and Memory Architectures
By
–
Mind Lab's unfair advantages: 1T-scale LoRA RL. Train 10x cheaper. Everyone else is bleeding cash.
Memory Diffusion. Forgets the noise. Keeps what counts.
OpenAI, DeepMind, MIT, Tsinghua refugees. 200+ papers. 30k citations. These people actually built shit that matters. Not a -
Macaron achieves 1T parameter RL training at 10% cost
By
–
Every AI lab is pissing away millions on compute. Macaron just ran RL on a 1 trillion parameter model at 10% cost.
Then open sourced the whole damn thing. Already integrated into NVIDIA Megatron and ByteDance Seed Verl while most companies are still writing Medium posts about -
CuspAI hiring: join generative models team for materials research
By
–
🔥🔥Amazing opportunity 🔥🔥 Work with my awesome and wonderful colleagues @johannbrehmer @wellingmax and @lonepair on generative models for materials here at @cusp_ai! Fun and exciting research on real-world problems that actually matter! #MachineLearning #Materials Johann Brehmer (@johannbrehmer) Come work with us at @cusp_ai in the generative model team! Excited about flow / diffusion models and chemistry? Looking for impact? jobs.ashbyhq.com/cuspai/b810… Join a great team lead by @wellingmax & Aron Walsh, work in Amsterdam / Cambridge / London / Berlin. 1/2 — https://nitter.net/johannbrehmer/status/2000488276029813139#m
→ View original post on X — @wellingmax, 2025-12-15 09:08 UTC
-
CuspAI hiring generative model researchers in multiple European cities
By
–
Come work with us at @cusp_ai in the generative model team! Excited about flow / diffusion models and chemistry? Looking for impact? jobs.ashbyhq.com/cuspai/b810… Join a great team lead by @wellingmax & Aron Walsh, work in Amsterdam / Cambridge / London / Berlin. 1/2
→ View original post on X — @wellingmax, 2025-12-15 08:49 UTC