TLDR AI 2025-11-03
Anthropic $300B rumor 💰, Siri’s private Gemini model 📱, Ilya’s deposition 🧑⚖️
Sam Altman says 'enough' to questions about OpenAI's revenue (3 minute read)
Sam Altman grew testy when pressed about OpenAI's ability to fund $1.4 trillion in committed spending, snapping at interviewer Brad Gerstner with, "If you want to sell your shares, I'll find you a buyer". Altman said OpenAI is doing "well more" than the previously reported $13 billion in annual revenue and suggested the company could reach $100 billion by 2027, but denied rumors of an IPO in 2026.
From Ilya's deposition (2 minute read)
Ilya Sutskever plotted for over a year with Mira Murati to remove Sam Altman as OpenAI CEO. Dario Amondei wanted Greg Brockman fired and himself in charge of all research. Murati told Sutskever that Altman had pitted her against Daniela Amondei. Sutskever wrote a 52-page-long memo to get Altman fired and also a separate document on Brockman. OpenAI is paying Sutskever's legal bills.
New Version of Siri to 'Lean' on Google Gemini (2 minute read)
A revamped version of Siri that uses a custom Gemini-based model is expected to roll out in March next year. Google is reportedly developing the model, which will run on Apple's Private Cloud Compute servers to power Siri. Apple is reportedly running into problems with the launch of Apple Intelligence in China. It is being mired by regulatory issues, so the launch is now a rolling target.
OCR Models Explained (4 minute read)
Recent advancements in OCR models, such as Chandra and Deepseek, eliminate traditional scripting by using vision transformers that convert raw pixels into semantic tokens. These models improve accuracy by understanding and compressing visual data into concise tokens before decoding them into text, allowing context-aware text recognition. Deepseek-ocr achieves this efficiently with ~10x compression, enabling flexible adaptation to different fonts and handwriting styles without retraining.
Why do AI models use so many em-dashes? (13 minute read)
Language models use em-dashes so much that real humans who like using them have stopped out of fear of being confused with AI. It's surprisingly hard to prompt models to avoid using em-dashes. Researchers still don't really know why language models use them so much. This article looks at some possible explanations.
Agents Rule of Two for AI Security (15 minute read)
Meta has proposed the "Agents Rule of Two" to reduce prompt injection risks in AI agents by limiting them to no more than two of three capabilities in a session: handling untrusted inputs, accessing private data, or taking external actions.
👨💻
Engineering & Research
Getting AI-ready: real-world lessons from Atlassian and Udemy (Sponsor)
Tongyi DeepResearch: A New Era of Open-Source AI Researchers (12 minute read)
Tongyi DeepResearch is a full open-source Web Agent that achieves performance on par with OpenAI's DeepResearch across a comprehensive suite of benchmarks. It is specifically designed for long-horizon, deep information-seeking tasks. This post shares a complete and battle-tested methodology for creating advanced agents like Tongyi DeepResearch.
Ghosts in the Codex Machine (7 minute read)
After weeks of user reports claiming Codex had gotten worse, echoing a perennial concern that frontier labs secretly downgrade their models, OpenAI assembled a squad to investigate every possible culprit and publish their findings. The team found no smoking gun, but instead uncovered a catalog of small bugs: older hardware subtly dragging down performance, a constrained sampling bug that made the model switch to Korean mid-sentence, and context compaction that quietly degraded accuracy in long conversations.
Get the most interesting AI stories and breakthroughs delivered in a free daily email.
Join 920,000 readers for
one daily email