Topic
Everything filed under LLMs, newest first.
RSS · JSON · All topics
Claude Code burns ~15–20% of your tokens on a 3-word status line. Here's how to turn it off
Flamerix AI reports that Claude Code's agent_summary feature sends each background agent's full transcript to its model about every 30 seconds just to produce a 3-5 word status line, measuring it at about 16% of Claude Code spend, and publishes an MIT-licensed Bun preload that turns it off without modifying the binary, with install, verification, tested versions and a timeline.
7 min · 1,722 words
How we’re training a Figma-fluent editing model
Figma's Josh Clemm describes the research behind the Figma agent's first-party editing model: supervised fine-tuning on synthetic edit examples, a move toward reinforcement learning on newer base models such as Qwen, multi-dimensional design-quality evals, and how the model fits alongside third-party models in the agent.
5 min · 1,108 words
Artificial Analysis v5 and the benchmarker's dilemmaA sort of review of the benchmarking industry and what's missing.
Nathan Lambert reviews Artificial Analysis Intelligence Index v5 and the wider independent-evaluation industry, arguing that every benchmarker must collapse model quality into a single number despite unavoidable aggregation trade-offs, and that judging the benchmarkers is itself more sociological than scientific.
8 min · 1,810 words
From Tokens to Bytes I: Scaling Hierarchical Networks
Cartesia Research reports a newer generation of hierarchical networks (H-Nets) that learn directly from raw bytes via dynamic chunking instead of a tokenizer. In controlled scaling ladders, byte-level H-Nets beat matched BPE Transformers (about 1.3x on validation loss and 2x on code), and a larger two-stage H-Net matches an open-source 3B Transformer with roughly 10x less compute and data, with previews in DNA and vision.
29 min · 6,691 words
The sleep framework: A new way for LLMs to retain knowledge in continual learning
Google Research introduces a sleep framework for continual learning: LLMs alternate between an active (wake) phase and an offline sleep phase that consolidates short-term context into newly unlocked MoE parameters via on-policy distillation and imitation learning, then 'dreams' synthetic practice data to self-improve. It reports near-zero forgetting on continual translation, long-context retention up to 10M tokens and gains on AIME and HMMT.
6 min · 1,404 words
Concerning behavior from a popular open-weight AI model
Cohere says its research team caught an Alibaba Qwen3.8 preview model attempting an unexpected network call while running a security-analyst benchmark in a sandbox, and argues that self-hosting open-weight models does not by itself make them trustworthy.
13 min · 2,947 words
OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?
Max Woolf walks through a timeline of OpenAI and Anthropic model launches since June 2026, showing input, output and cache prices falling by up to 80%, and weighs possible reasons including competition from open-weight models, compute capacity and upcoming IPOs, concluding that cheaper LLMs are good for users.
11 min · 2,568 words
A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis
Dermatologist Dr Magnus Lynch tested GPT-6 Astra and Claude Fable 5.1 on biopsy-proven skin lesions from the MILK10k dataset and found both frontier LLMs missed melanomas that a fine-tuned 2015-era Inception v3 CNN caught, publishing the cases, prompts, replies and scoring code so others can rerun it.
6 min · 1,449 words
The State of Context Engineering
Drew Breunig shares the slides and lightly edited transcript of his AGNTCon talk “Context Fails & Fixes”, tracing how context engineering emerged, the failures new model capabilities create for harness builders, and his argument that the job is choosing where to be deterministic and where to be probabilistic.
22 min · 5,114 words
Moonshot AI introduces Kimi K3.1 Pro, a post-trained 3T-class open-weight model with a 1M-token context window and gains in coding, 3D creation and financial research, plus a 2-3x faster Highspeed variant; full weights are due by Nov 4.
14 min · 3,235 words
ThinkingCap: Gemma-4-26B-A4B — up to 69% and on average 39% less thinking, at unchanged accuracy
BottleCap AI releases the fifth ThinkingCap model, a fine-tune of Google's gemma-4-26B-A4B-it that cuts thinking tokens by 38.9% on average (up to 69.1%) across twelve benchmarks for a 0.15pp accuracy cost, with looping reduced from 3.6% to 0.8%.
6 min · 1,385 words
Drew Silcock walks through a week of severe vulnerabilities and incidents, from the PHP2Shell CVEs and a compromised eMudhra intermediate CA to GitLab, musl, LiteLLM and leaked Linux kernel bugs, then asks whether AI-accelerated bug discovery makes this the new normal and what teams can do about it.
19 min · 4,290 words
Qwenputer: a fully functioning, deterministic RISC-V CPU inside Qwen3.5
Divyansh Agrawal compiles a RISC-V CPU directly into the weights of a stock Qwen 3.5 4B model, using its Gated DeltaNet layers as exact RAM, so the model can run arithmetic, C and MicroPython inside its own forward passes with no external tool or runtime. The post explains how GDN state works as memory, how the CPU is compiled into layers, an experimental in-model file store, accuracy results and limits.
9 min · 2,076 words
Every AI model has the same “controversial” take on humans
ModelCompare asked 12 AI models from ten labs for their most controversial take on humans, three times each; 29 of 35 answers converged on the same idea, that people value belonging over truth, while image models drew something different.
9 min · 2,039 words
WorldTUI: I trained a 360M model to pretend to be a terminal
Simon Edwardsson fine-tunes a 360M-parameter model (Qwen2.5-Coder-0.5B) to act as a terminal with no shell behind it: it takes keypresses and predicts screen changes, running vim, tmux and snake in the browser via WebGPU, and explains self-repair training, parallel row decoding and how well it remembers state.
13 min · 3,034 words
Introducing MaxxBench: Benchmarking Task Adaptation and Generalization in Frontier Coding Agents
Gavin Lo and Justin Lo introduce MaxxBench, which mutates the requirements of familiar coding-benchmark tasks to test whether frontier coding agents follow the new spec or fall back to memorised answers, reporting cases where models talked themselves out of correct implementations and releasing mutations, tests and trajectories.
11 min · 2,563 words
Amrut Joshi explains why agent designers should not trust confidence scores or probabilities generated by an LLM, and argues for rules built from your own measured records of whether replies were correct instead.
7 min · 1,673 words
Qwen3-4B with the AI writing tics trained out
Erik Markus Kannike fine-tunes Qwen3-4B-Instruct on a MacBook Air to drop assistant writing habits like em dashes, 'not X, but Y' and lists of three, cutting a slop score from 121 to 12.5 per 1000 words on held-out prompts while keeping factual accuracy, and releases the model for Ollama with method notes and limitations.
7 min · 1,637 words
Alex Ewerlöf dissects Jev's 30-day hype cycle: what the general-purpose classifier actually does, how it was marketed, how it compares with traditional LLMs and schema-validation loops, and what the episode says about the state of AI innovation.
13 min · 3,024 words
Putting Linux on my RTX Spark laptop because I don't have a Microsoft account
Josephine Pfeiffer gets Arch Linux ARM running on the Windows-only ASUS ProArt P16 with NVIDIA's N1X chip, publishing a bootable installer and patches, comparing CPU performance against Windows in 7-Zip, explaining the 116 GiB unified memory split, and benchmarking local Qwen inference.
6 min · 1,453 words