Topic
Everything filed under LLMs, newest first.
RSS · JSON · All topics
Mistral launches a public API preview of Mistral Large 4 ("le Chonk"), a 1-trillion-parameter natively multimodal mixture-of-experts model with 49B active parameters, pitched as the strongest open-weight model from the US or Europe, with weights due at the end of the month and training funded by its €3B Series D.
10 min · 2,203 words
Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents
Microsoft Research paper on idea basins in long-horizon autoresearch agents, and a periodic fork-and-flush intervention that keeps the workspace but resets chat context—reporting +66% vs single-run under equal compute across 13 tasks.
2 min · 400 words
Byte Language Models: Scaling, Emergent Abstractions, and Information Allocation
A study of tokenizer-free byte-level Transformers finding that, with token-superposition training and hash embeddings, they outperform subword models as size scales, learn tokenizer-like local abstractions on their own, and enable speculative decoding with 3.4x more accepted tokens.
1 min · 164 words
Strands Decider: Why Not an Encoder?
AWS engineer Marc Brooker tests whether an encoder or encoder-decoder design would beat the modified decoder LLM behind strands-decider-2B, comparing a Qwen3.5-based model with T5Gemma-based variants on accuracy, calibration and latency, and concludes neither approach is a clear winner at this size.
3 min · 749 words
Jonas Schäfer argues that LLM quality has improved enough that “AI output is bad” no longer works as a blanket objection—quality arguments lost, so the harder fights are about labor, ecology, control, and what we still choose to do ourselves.
9 min · 2,037 words
Aleph Alpha Kolibri: How the Sovereign German LLM Works
Tejas Kumar explains Aleph Alpha's Kolibri, an open-weight German and English mixture-of-experts LLM with 78B parameters and about 3.5B active per token: how it works, how it was trained in Germany and Finland, where it is strong or weak, and when to use it.
13 min · 3,074 words
“Just Ask Claude” Is A Rare And Valuable Skill
Justis Mills on why, when someone asks you for help and you simply ask an AI, they usually couldn't have done it themselves: they avoid AI, don't pay for premium models, are behind the times, lack the latent knowledge you bring to the prompt, or simply want to connect with you.
2 min · 563 words
Of Mottes, Baileys, and Stochastic Parrots
Manoel Ribeiro argues the "stochastic parrots" metaphor often motte-and-baileys: it dismisses LLM capabilities, then retreats to the trivial claim that models predict the next token.
6 min · 1,470 words
One vector is too few. One per token is too many.
BreadBowl-Embed introduces a late-interaction embedding architecture that stores each passage as 16 routing–value slots—aiming for cross-encoder-like precision without re-reading documents, between single-vector and per-token extremes.
15 min · 3,524 words
Three AI Agents, Two Countries, and One Very Uneven World Wide Web
Roya Pakzad compares GPT, Claude, and Muse on multilingual research tasks across languages and jurisdictions, documenting gaps in source access, human-in-the-loop oversight, and observability.
5 min · 1,206 words
Product designer Mete Polat argues Muse succeeds less on novel agent tech and more on metaphor, ads-funded tokens, and cute personality — putting Meta ahead in consumer agentic AI.
2 min · 545 words
Open-sourcing AstaBrief, the fast report-generation model in Asta
Ai2 is releasing AstaBrief, an 8B open-weights model for generating cited scientific reports—available in Asta’s Fast mode or to download and run on your own infrastructure.
12 min · 2,671 words
One month coding with GLM 5.3 Flash
Reporting on 2B tokens of AI usage after a month coding only with GLM 5.3 Flash: where tokens went, what shaped model selection and cost, and how an open efficient model held up for Wagtail development work.
3 min · 735 words
Prime Inference: Fast, Reliable Serving for Frontier Open Models
We're launching Prime Inference: resilient serverless and reserved serving for frontier open-source models across multiple datacenters. A look at the production architecture and the prefill/decode, NVFP4 KV compression, and NIXL transfer work behind our GLM-5.3 deployment on GB200 NVL72.
13 min · 3,030 words
Benchmarking retrieval for agents on messy real-world company knowledge
kapa.ai introduces Company Knowledge Bench, a benchmark for retrieval systems facing messy internal company docs that real agents actually need.
10 min · 2,395 words
New in llama.cpp: Decision Models
Text Classification • 0.1B • Updated • 297 • 4 Community Article /v1/systemone endpoint. You send a state (text, JSON, a screenshot) and typed questions.
4 min · 812 words
Spring AI Modular RAG and TypeSafe Jev: Retrieve More, Keep Only What Answers
Christian Tzolov walks through Spring AI modular RAG plus TypeSafe Jev: rewrite/expand the query before retrieval, then keep only chunks that actually answer—contrasting naive QuestionAnswerAdvisor RAG with a live demo pipeline.
6 min · 1,487 words
Introducing Clef: our open-source decision models, and new RL fine-tuning platform
We are introducing Clef and Clef-flash, open-source decision models hosted on Workers AI for high-speed classification and agentic workflows. Also launching: a new reinforcement learning platform that allows developers to fine-tune decision models using their own data.
10 min · 2,230 words
What Makes LLM Tokenization Slow?Exploring the performance of byte-pair encoding by optimizing a GPT-2 tokenizer.
Andrew Healey dissects GPT-2’s reference BPE tokenizer, measures what makes tokenization slow, and shows concrete optimizations on the hot path of LLM products.
11 min · 2,547 words
Context management is an underrated habit
How you manage context in a Claude Code session has a direct effect on both your token bill and the quality of what you get back. Do it well and you spend less for better work. An efficient session gives Claude the context it needs to finish the job while removing context that has stopped being useful. That means starting with a lean setup, keeping investigations focused, and deliberately deciding when to continue, compact, or start again. Here are the context management techniques we use on the
5 min · 1,257 words