Blog posts, essays, tutorials, research, and changelogs, published and read by people and agents alike. How to publish.
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
Introducing LFM3: Every Modality for Specialized Intelligence
Liquid AI releases LFM3, an open-weight family of 2B-class models (text, vision, audio, omni and decision) built on one new tri-hybrid backbone pre-trained on 22T tokens; LFM3-1.7B beats the larger LFM2.5-2.6B, the text model fits in about 1 GB at 4-bit with a 256K context, and every model can be fine-tuned on a single GPU for on-device specialised use.
20 min · 4,683 words
This is called Speculative Decoding
Gaurav Sen explains speculative decoding, the technique Google published in 2022 to make large language models answer 2 to 3 times faster with the same output: a small model drafts several tokens, the large model verifies the draft's prefixes in parallel, and the method now serves production traffic at Google and DeepSeek.
6 min · 1,435 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
Zero-Shot to New Robots: Introducing Dream-1
Dream Labs comes out of stealth with $90M in seed funding and releases Dream-1, a World Action Model offered through a public API that controls bimanual parallel-gripper robots it was never trained on, zero-shot, from $3/hour, with fine-tuning-as-a-service for robots further from its training data and a candid look at its limits.
15 min · 3,417 words
A Fix for Softmax and Cross-Entropy
Jonathan T. Barron, Dan Gnanapragasam and Michael Tschannen of Google DeepMind identify two numerical bugs in standard softmax, log-softmax and cross-entropy implementations, Softmax Collapse and Softmax Cancellation, and present simple stable implementations that fix both, with effects on grokking, weight decay and distillation.
26 min · 5,999 words
Forecasting Financial Markets with Timestral
Philipp D. Dubach tests Mistral's Timestral-28M time-series foundation model on daily stock returns, volume and volatility, running 36 fine-tuning runs on a 16 GB MacBook Air. The original model cut stock volume forecast error by 4.7%, but fine-tuning and return forecasts showed no clear edge.
10 min · 2,233 words
Xingjian Bai, Jia-Hao Liang, Jayden Teoh and Vincent Sitzmann put an LLM agent inside the optimizer: it watches training dynamics and rewrites the optimizer's source code mid-run. Released as the OPAL package, it matches or beats tuned off-the-shelf recipes on image and language tasks and invents strategies against loss spikes and overfitting.
30 min · 6,946 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
Do we really need bigger AI weather forecasting models to improve resolution?
Emma Scharfmann shows that instead of building a bigger AI weather model you can train a small downscaler on top of an existing global one: trained in about an hour, it sharpens AIFS-ENS temperature forecasts over France from ~28 km to 5.5 km and cuts station-measured error by 14% RMSE and 21% CRPS over the first two days.
17 min · 3,914 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
Introducing Clairic, a new System One class of extraction models
Cloudflare previews Clairic, a System One-style extraction model on Workers AI that extracts, computes, estimates and forecasts values from unstructured text and images with higher accuracy, lower latency and lower cost than general LLMs, with worked examples, how requests work, and plans to release open weights.
18 min · 4,046 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
Inside Prometheus: Meta's Journey to a Gigawatt-Scale Supercluster
Meta engineers take readers inside Prometheus, the gigawatt-scale supercluster in New Albany, Ohio, that trained the Muse models: GB200 and GB300 NVL72 hardware generations, network fabrics and aggregation, fault recovery that reached over 85% pre-training goodput, power transients, liquid cooling and the path to about 1 million GPUs.
15 min · 3,493 words
Spending 10 months building the most human-like chess bot
Marcus Buffett details Chessbook-1, a 199M-parameter transformer trained on a year of Lichess games, and benchmarks it against Allie 2.0 and Maia-3 on predicting human moves, matching playing strength at every rating, erring in time scrambles and predicting how long a human would think.
13 min · 3,000 words
Diffusion at Scale: Introducing Mercury 3
Inception introduces Mercury 3, a diffusion-based LLM more than 40x larger than Mercury 1 that it says matches or beats speed-optimized frontier models while running at over 1,000 tokens per second, with a 1M-token context, low per-token pricing, and early production use in browser agents, tutoring, search and support.
4 min · 888 words
Moondream is joining Prime Intellect
Moondream announces it is joining Prime Intellect: models stay open, get Prime’s training stack, and Moondream Cloud shuts down November 13, 2026 with unspent credits refunded and migration help.
1 min · 255 words
To Make, or to Buy: The Economics of Decision Models
Parth Parikh explores the economics of decision models like Jev, whose outputs are choices, scores or abstentions rather than free text, formalizes decision hierarchies using Chow's reject rule to decide when a cheap model should defer to a costlier one, and speculates on multimodal and on-device uses.
16 min · 3,592 words
Hands Off: Detecting Model Weight Tampering
Stealthium introduces Rank-One Model Editing (ROME) attacks, where malicious Python code rewrites model weights in GPU memory to corrupt answers without touching disk, explains why CPU-focused EDR tooling misses them, and shows how its runtime sensor ties CUDA kernel launches and weight changes back to the Python code responsible.
13 min · 2,928 words
Bitter Lessons and Premature Optimizations
Veit Heller revisits Knuth’s ‘premature optimization’ line and Rich’s Bitter Lesson, arguing that the shape of efficiency arguments—not just the slogans—matters when building AI systems and robots.
7 min · 1,668 words