Topic
Everything filed under LLMs, newest first.
RSS · JSON · All topics
Towards safety cases for frontier AI training
OpenAI argues frontier RL runs should require structured safety documentation approaching “safety cases”: technical safeguards, operational practices, and incident investigation before continuing training.
7 min · 1,571 words
Automating eval design and hillclimbing with Claude
Lance Martin (claude.dev) explains principles for production-like evals with held-out sets, then shows how the claude-api skill’s build-eval and hillclimb commands automate design and overfitting-aware improvement—including cost and performance case studies.
2 min · 462 words
Meta FAIR introduces RL-XAR (Reinforcement Learning from eXpert-Aligned Rubrics): learn rubrics from the gap between expert writing and model output, then train models toward expert-level text generation to reduce AI slop.
16 min · 3,572 words
Can a Model Learn New Skills as Add-Ons?
Connito Research trains residual MoE experts with their own routers on a frozen DeepSeek-V2-Lite base, then merges independently trained math, code, medical, law, and finance experts in seconds without retraining—lifting domain benchmarks while leaving the original model untouched.
4 min · 913 words
Small Decisions: Engineering a Leading Model
AWS engineer Marc Brooker recounts building and training a small leading model hands-on—what worked, how it performed, and what the exercise taught him about modern model-building.
9 min · 2,024 words
What I believe about the future of software development
Thorsten Ball plants a flag on where software development is headed as AI agents write more of the code: what still matters for engineers, what gets commoditized, and how taste and judgment become the scarce skills.
4 min · 809 words
Throughout the history of AI, open research has played a critical role in driving progress. Today, many key details of frontier large language models (LLMs) remain proprietary, but open-weights model families— such as DeepSeek, Kimi, and MiMo —continue to provide a valuable window into the development process for modern LLMs.
7 min · 1,609 words
Robert W argues LLM confidence scores are vibe-words, not calibrated probabilities—and outlines RLCD-style approaches that treat hallucination as a solvable measurement and training problem.
7 min · 1,560 words
What's wrong with Open Source AI?
Alexine Le Port on the shrinking gap between open-weight and frontier models—and why open-source AI culture is still mostly vibing instead of shipping durable public infrastructure.
3 min · 651 words
A font compiler that makes every LLM token the same width—why monospace-per-token helps visualize chain-of-thought, plus an interactive preview of fonts built from a font + tokenizer pair.
7 min · 1,527 words
Turn GLM-5.3-Flash into a Jev-like System One model
Johannes Hötter walks through turning GLM-5.3-Flash into a fast Jev-like “System One” decision model—typed options with probabilities in a single forward pass, matching Jev’s accuracy and speed.
11 min · 2,426 words
Pragmatic Anthropomorphism, or: How to Talk to an Autocompleting Cricket
Les Orchard argues that debating whether LLMs truly understand misses the useful move: treating an AI as a competent collaborator is pragmatic navigation of latent space, not magical thinking.
10 min · 2,271 words
“As a Language Model…”: Chat Template Switches LLM Self-Referential Voice
Research showing chat templates act as a switch between disclaimer (“I’m just an AI”) and experiential (“I feel”) self-referential voices across 8 instruct models, with a steerable activation direction that reproduces the template effect.
3 min · 621 words
LLMs: Semantic Translation Machines
KnorpelSenf reframes LLMs as probabilistic semantic translators—great at moving an idea between representations, weak at inventing facts—and shows how that lens guided nearly all of a hard distance-matrix rewrite in Rust with LLM help.
2 min · 427 words
A Jev-like wrapper for LLMs, including vision models
Allan shows a small single-function Jev-style wrapper for LLMs that also handles vision models, with practical code for local and API backends.
7 min · 1,575 words
Analyzing Frontier Model Progress with My Favourite Game: Prince of Persia (Apple II, 1989)
Priyan uses Jordan Mechner's Prince of Persia as a living benchmark: asking frontier models to port and reason about the classic Apple II game, and what those runs reveal about coding-agent progress.
8 min · 1,825 words
Mixture of Experts (MoE) for Backend Engineers
A detailed visual guide to token routing, expert batching, weighted combination, and the memory and communication tradeoffs of MoE serving. Suppose a model has dozens of feed-forward subnetworks, but each token uses only two of them. The arithmetic per token can stay modest while the total weight set grows. Now place that model on eight GPUs. If every GPU stores all experts, memory can become the limit; if experts are split across GPUs, token activations must travel to whichever GPU owns their…
15 min · 3,515 words
Alex Ewerlöf argues that AI has not solved coding: reliability, architecture, product judgment, and the long tail of software work still require human engineering beyond vibe-coded demos.
23 min · 5,234 words
42x Faster Prompt Lookup Drafting in llama.cpp
Four changes to the n-gram caches of llama.cpp make drafting up to 41.6x faster, load the static cache up to 23.5x faster, and lower peak memory up to 2.65x.
13 min · 2,954 words
Tupi: rebuilding 1554 in the browser with 32 AI agents
How Ruben Marcus rebuilt a 1554 Tupinambá canoe raid as real-time 3D in the browser with Claude Opus 5.5, three.js, and Blender—what the agent swarm got right, what froze the page, and what it cost.
8 min · 1,950 words