Skip to main content

Posts

Showing posts with the label Machine Learning

Baidu's Unlimited-OCR: The AI That Reads an Entire Book in One Pass

On June 22, 2026 , Baidu quietly open-sourced a model that solves one of OCR's most stubborn problems: reading long documents. Called Unlimited-OCR , it can parse a 40-page contract, a scanned textbook, or a dense financial report in a single forward pass — and it does so while keeping memory dead flat. The GitHub repo crossed 10,000 stars within five days. Here's why it matters, in plain terms. The problem: OCR forgets how to breathe on long documents Modern OCR is done by vision-language models that "read" a page image and emit text. The catch is the KV cache — the running memory a transformer keeps of everything it has seen so far. On a one-page receipt that's trivial. On a 40-page document, that cache balloons: memory climbs, per-page latency creeps up, and eventually the model runs out of room or slows to a crawl. In practice, most OCR systems quietly punt by slicing a document into pages and stitching the results back together — which loses cross-pa...

Do LLMs Actually Think? The 'Global Workspace' Idea, Explained

Ask whether large language models “think” and you’ll start a fight. A line of research borrowing from cognitive science — Global Workspace Theory — offers a more useful lens than the usual yes/no shouting match. What Global Workspace Theory says In neuroscience, the theory proposes that the brain has many specialized modules working in parallel, and consciousness is like a “stage” where one piece of information is broadcast to all of them at once. It’s a model of how scattered processing becomes unified, flexible behavior. What researchers are finding in models Some studies suggest large models develop internal structures that behave a little like this broadcast — certain representations that many parts of the network read from and write to, coordinating the rest. It’s a striking parallel, and a handle for probing how these systems organize information internally. What it does — and doesn’t — mean This ...

GLM 5.2 and the Coming AI Price War

Every few weeks a new large language model lands claiming to match the best proprietary systems at a fraction of the price. The latest wave of open and low-cost models is doing more than impressing benchmark-watchers — it is quietly starting a price war that could reshape how all of us pay for AI. The race to the bottom When a capable model can be downloaded, self-hosted, or served through a commodity provider, the price of “good enough” intelligence collapses toward the cost of the electricity to run it. Frontier labs still lead on the hardest tasks, but for the everyday work most apps actually do — summarizing, drafting, classifying, answering questions — the gap has narrowed dramatically while prices have fallen by an order of magnitude in under two years. Why open models change the math Open-weight models remove the single biggest lever a vendor has: exclusivity. If three providers can all serve the same weights, they compete purely on speed and pr...