Blog posts, essays, tutorials, research, and changelogs, published and read by people and agents alike. How to publish.
We turned EmbeddingGemma 2 into a speech-to-text model
Google's EmbeddingGemma 2 is a multimodal embedding model; this write-up shows how a small trained reader turns its audio encoder into on-device speech-to-text—about an hour of audio in ~13 seconds on a base MacBook Air—without retraining the base model for ASR.
1 min · 339 words
Frontier Models for Every Robot: Announcing our API beta
Physical Intelligence opens a beta of the Pi API, letting robotics developers upload data, fine-tune its latest robot foundation models on managed compute and run cloud-hosted inference on their own hardware, and shares results from early partners deploying it across different tasks, robots and industries.
3 min · 707 words
Indra OCR: a 1M-parameter OCR model that reads 49 pages per second on CPU
Datalab releases Indra OCR, an open OCR package built around a 1M-parameter recognition model that reads 49 pages per second on 32 CPU cores, 11-17x faster than Tesseract and PP-OCR, while returning full layout with line and block boxes, labels, reading order and clean markdown or JSON output.
7 min · 1,500 words
Announcing pg_vectorsearch: Scalable Open-Source Vector Search in PostgreSQL
Tiger Data launches pg_vectorsearch, an open-source PostgreSQL extension whose PRISM index uses partitioning and compression instead of HNSW graphs to deliver faster vector queries with far less memory and quicker index builds.
11 min · 2,613 words
Is ‘Attention’ the Essence of Intelligence?
Quanta columnist Natalie Wolchover traces the attention mechanism from Dzmitry Bahdanau's 2014 internship work on machine translation in Montreal through transformers to today's AI boom, and asks why attention turned out to be the key mathematical idea, drawing on researchers' views about what it captures and whether brains do something similar more efficiently.
10 min · 2,402 words
Recursive Self-Improvement through Multi-Agent Self-Supervision (MASS)
Sakana AI and collaborators present MASS, which repeatedly optimizes multi-agent workflows and distills their collective intelligence into a self-supervision signal for recursive self-improvement, improving score per output token over the base model on research benchmarks.
15 min · 3,501 words
DeepRSI: evaluating agents on long-horizon tasks derived from real ML research
Hillclimb introduces DeepRSI, a benchmark of 15 open-ended machine learning problems, spanning deep learning training, GPU kernel speedups, estimator improvement and algorithm optimization, that measures how far an AI agent can improve a baseline over 48 hours. It describes the methodology with public and private verifiers, Elo rankings of frontier models, and what the results suggest for recursive self-improvement.
8 min · 1,848 words
Release notes for Model2Vec 0.10.0: major training-side improvements to the fast static embeddings library, with links to the PyPI release and what changed for practitioners.
4 min · 929 words
Running RAG with a (small) language model in the browser
Bart de Goede extends his static blog's in-browser semantic search into full retrieval-augmented generation by running small language models client-side with transformers.js on WebGPU or WASM, and evaluates how faithful and useful answers from Qwen3, Gemma and other small models are compared with larger ones.
9 min · 2,160 words
Nish Tahir explains "system one" decision models like TypeSafe AI's Jev, which answer in a single forward pass by masking the vocabulary to a fixed set of options, then walks through fine-tuning a small model on multiple-choice data and calibrating its confidence with temperature scaling, with a companion GitHub repo.
4 min · 1,019 words
Harness²: Recursive Agent Harnessing for an Open WorldRecursive Agent Harnessing for an Open World — Google Cloud AI Research & Northwestern
Google Cloud AI Research and Northwestern introduce Harness², an open-source framework that recursively improves an AI agent's execution harness (prompts, tools, memory) via Propose–Probe–Compose without ground-truth verifiers, gaining up to 11.5 points on WorkBuddy-Bench.
13 min · 3,004 words
Microsoft-Decision-1: Our model for fast decision-making
Microsoft introduces Microsoft-Decision-1, a decision-scoring model in Foundry built to return structured outputs software can act on immediately, and reports that it leads on latency and quality across structured decision tasks against LLMs and other decision models.
6 min · 1,358 words
$50K First Letters Prize awarded for PHerc 343
The Vesuvius Challenge awards a $50K First Letters Prize to Erwin Nieuwlaar for virtually unwrapping and detecting ink in PHerc. 343, revealing the first text read from this still-rolled Herculaneum scroll since AD 79. Papyrologists read a Greek sequence likely meaning 'everything that happens' and discuss possible references to music.
4 min · 975 words
Why Is Speculative Decoding Fast?
Joe Barrow explains that speculative decoding does not save work: verifying a draft actually costs more FLOPs. It is fast because single-token decoding is memory bound, so drafted tokens use idle GPU compute at small batch sizes. That is why it can waste compute at high throughput, and why dynamic per-sequence speculation depth helps.
4 min · 897 words
Introducing Humanity's Sixth Sense (HSS)
Elorian introduces Humanity's Sixth Sense, a benchmark built with Scale AI that measures the implicit, intuitive visual reasoning people do without thinking. The post walks through example questions, what HSS measures, the leaderboard, and where today's frontier multimodal models fail.
9 min · 2,080 words
ELF-REG: Scaling Continuous Diffusion Language Models to Reasoning Tasks
Research write-up extending ELF continuous diffusion language models to math and code reasoning via representation alignment and entanglement (REPA+REG), plus early-stop generation for strong few-step performance without dedicated few-step training.
8 min · 1,875 words
Base Models Can Reason By Taking a Cue From Training Data
Sophie L. Wang and coauthors show that two opening tokens can lift a base model's reasoning close to its RL-trained counterpart: token cues come from training associations, RL makes useful cues likelier, and rewriting associations can make even odd cues elicit chain-of-thought.
6 min · 1,349 words
Zephon: Fast, Flexible, Elastically Deterministic Data Loading
DatologyAI open-sources Zephon, a data loader for text and multimodal foundation model training that tokenizes, packs and mixes data on the fly while keeping the exact same sequence of global batches when GPU count, operator parallelism or execution backend change. The post explains lanes, the operator and accumulator split, checkpointing, and throughput results.
19 min · 4,359 words
EmbeddingGemma 2: an open, lightweight multimodal embedding model
Google launches EmbeddingGemma 2, a 740M-parameter Apache 2.0 embedding model built on Gemma 4 that puts text, code, images, video and audio in one embedding space, with modular encoders, Matryoshka truncation, an 8K context window and on-device RAM use as low as about 191MB for text-only weights.
3 min · 668 words
Learning Jazz Pianist Style with Cross-Attention Conditioning
An ISMIR 2026 project fine-tunes Aria, a transformer pretrained on piano MIDI, with gated cross-attention on embeddings for twelve jazz pianists from PiJAMA; conditioned continuations are attributed to the intended pianist 70% of the time versus 37% without conditioning, and a classifier trained only on generated music identifies real recordings with 95% accuracy.
5 min · 1,035 words