Nick Lothian

My Gemma4-in-the-browser agent autonomously decided to use a sub agent on its own for the first time. Gemma4 E4B is great!

Levels and Locks as Agent Orchestration primitives nicklothian.com/blog/2026/02...

Llama 3.1 8B running at 16,000 tokens per second on taalas.com custom hardware. umm yeah ok.

@timkellogg.me you might find x.com/AJakkli/stat... interesting: "What happens when you leave two copies of the same model talking to each other? They have different attractor states: Grok devolves into gibberish while GPT-5.2 starts writing code and editing imaginary spreadsheets"

Nemotron is a *great* model and the 3B and 12B models are free on OpenRouter for people to try out.

LMNotebook generates great infographics of WTF code does if you rename python files to .txt and dump them in there.

Gaslighting from Opus. Grr

Kimi K2.5 looks very impressive. 100 agent co-ordination, 4.5x speedup on tasks. It's only January and we have a half order of magnitude improvement in real-world performance, and yet people think progress is slowing down? Really, we are running out of benchmarks.

Tired: Sprint point estimates Wired: Number of tokens to implement feature estimates

It shocks me how many people think LLM inference providers are somehow running at a loss. Do the math! If you can hire a cloud H100 and run it above 50% capacity for something approaching the per token price inference providers are charging then it's easy to see they aren't.

OpenAI drama making crypto bros look like the sensible ones.

🚀Happy to announce my recent open source project- LightPHE: A Lightweight Partially Homomorphic Encryption Library for #Python. It allows computations on encrypted data without decryption. It wraps many algorithms such as RSA, ElGamal, Paillier and more! github.com/serengil/Lig...

We evaluate the combined architecture on two challenging planning tasks: -Graph traversal -Tower of Hanoi Graph traversal: significant improvements over zero-shot prompting or in-context learning. Notably the architecture lowers hallucination of invalid moves to 0% (from ~20% for 4-step paths!) 4/n

This seems like an appropriate first post.