ThursdAI - The top AI news from the past week · From Weights & Biases, Join AI Evangelist Alex Volkov and a panel of experts to cover everything important that happened in the world of AI from the past week

📆 ThursdAI - Oct 23: The AI Browser Wars Begin, DeepSeek's OCR Mind-Trick & The Race to Real-Time Video

·1 hr 35 min·5 clips

DeepSeek OCR is framed as a way to shove 10x more context into the same context window.

1. ThursdAI’s October 23 episode focuses on AI OCR releases, browser competition, and real-time video tooling. 2. Alex Volkov hosts the show as an AI evangelist at Weights and Biases, with Nisten, Yam, and LDJ joining the discussion because they react to the week’s models and browser launches. 3. The episode asks what changed this week across OCR, agents, and browsers, then uses two interviews and a news rundown to answer that question. 4. DeepSeek OCR gets the main open-source segment, with the panel describing it as a tiny model that can process screenshots, PDFs, books, and code. 5. The hosts say DeepSeek OCR can rebuild charts as HTML and may enable about 200,000 pages a day on a single GPU. 6. LDJ explains the paper’s compression framing, where text is represented visually rather than tokenized in the usual way. 7. The discussion also compares DeepSeek OCR with Z.ai’s Glyph paper, which LDJ names as “Glyph Scaling Context Windows via Visual Text Compression.” 8. The panel discusses Liquid AI’s LFM2 VL3B, Pokee Research 7B, and Qwen 3 VL releases as part of a crowded week for multimodal models. 9. Qwen 3 VL 2B and 32B get attention because the smaller model is described as runnable on less than two gigabytes of RAM. 10. The hosts also note that the 32B model is being compared with Claude Sonnet and GPT-5 Mini on multiple benchmarks. 11. One memorable point is the claim that DeepSeek OCR may encode 1 million tokens into a 100K vision-token sequence. 12. Another notable moment is LDJ’s caution that the 97% compression figure does not equal standard language-model benchmark accuracy. 13. Nisten says the smaller Qwen model can still hallucinate on terminal-style screenshots even while handling many tasks well. 14. The panel says Qwen’s video support matters because video is expensive to process and small models can make that cheaper. 15. Alex’s browser-war segment ties OpenAI Atlas to his past work on Chromium-based browsers and to a previous browser war he followed for HTML5 releases. 16. The show also covers Microsoft Edge’s Copilot mode, which can navigate pages, manage tabs, and access browser history. 17. The tone is fast, conversational, and opinionated, with the hosts interrupting each other to test claims live. 18. The format mixes TL;DR news, live reactions, and two scheduled guest interviews, with some comments from the live audience. 19. People tracking OCR, AI browsers, and compact multimodal models will get the most value from this episode. 20. People wanting a polished deep-dive on one model or a low-noise recap should probably skip it.

As heard by us

AI browsers, OCR, and agent infrastructure get a focused walkthrough.

Thursday Eye stays on OpenAI's Atlas and DeepSeek's OCR model, while the two interviews with Paul Klein of Browserbase and Quinn Kramer broaden the frame around agent authentication and real-time voice avatars.

Read the full review in PlayNext →

Why you'd press play

If you want the week’s AI browser drama plus the tooling behind it, press play.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from ThursdAI - The top AI news from the past week