Sebastian's "from scratch" series keeps delivering. Hard distillation that works with any LLM is exactly the kind of practical resource people need right now.
LLMS
-
NVIDIA Releases Nemotron 3 Super 120B Models
By
–
NVIDIA published a series of new "Nemotron 3 Super 120B" open-source models on Huggingface, designed for agentic use.
-
On-Policy Self-Distillation Improves Reasoning LLM Efficiency
By
–
“On-Policy Self-Distillation for Reasoning Compression” This paper shows that the new bottleneck for reasoning LLMs isn’t “too little reasoning”, it’s that more reasoning tokens often increase mistakes. So they fixed it without RL length penalties or external verifiers, but all
-
OpenRouter Adds Two Stealth Models
By
–
OpenRouter now has 2 new stealth models, Hunter Alpha with a 1M context window for agentic use, and Healer Alpha, an omni-modal model.
-
Claude Poised to Transform Knowledge Work Like Agentic Coding Did
By
–
Claude for all things knowledge work feels like it is on a very similar trajectory to what agentic coding experienced last year.
— Alex Albert (@alexalbert__) 11 mars 2026
I expect entire industries that rely on spreadsheets and powerpoints to begin to be transformed in the next few months. https://t.co/jMxV3iZscaClaude for all things knowledge work feels like it is on a very similar trajectory to what agentic coding experienced last year. I expect entire industries that rely on spreadsheets and powerpoints to begin to be transformed in the next few months.
-
Claude for Excel and PowerPoint now share context
By
–
Anthropic upgraded Claude for Excel and Claude for PowerPoint to enable context sharing between solutions. https://t.co/fgzfPmHpN1 pic.twitter.com/Tcd7EXFMTq
— 🚨 AI News | TestingCatalog (@testingcatalog) 11 mars 2026Anthropic upgraded Claude for Excel and Claude for PowerPoint to enable context sharing between solutions.
-
Harvey AI Launches BLB Research Benchmark for Legal AI
By
–
We partnered with @Harvey_AI on BLB: Research — a benchmark of hard, agentic US case law research problems. We identified a series of US case law research problems that leading models are currently unable to solve — even when provided with search tools like web search. The
-
Research Analysis: Massive Activations and Attention Sinks in Transformers
By
–
Another by Yann LeCun! “The Spike, the Sparse and the Sink” This paper shows that massive activations and attention sinks come from the same pre-norm Transformer pipeline rather than being separate anomalies. Early SwiGLU blocks act like directional quadratic amplifiers,
-
The Silent Trap in RAG Pipelines: Embedding Model Costs
By
–
Most RAG pipelines have a silent trap nobody warns you about. Imagine you built a RAG pipeline with a large embedding model for high retrieval quality, and it ships to production. Six months later, your application traffic and your embedding model costs are soaring while your