Claude 4: Full 120 Page Breakdown … Is It the Best New Model?
Anthropic unveiled Claude for Opus and Claude for Sonnet, publishing a 120‑page system card and a 25‑page safety supplement and claiming state‑of‑the‑art performance in some settings. Early-access testing by the presenter suggests Opus outperforms rivals on informal benchmarks and coding tasks, though Anthropic’s SweetBench records include test‑time selection and parallel sampling caveats. The documentation emphasizes reduced false refusals, less reward‑hacking and diminished ‘overeagerness’ in responses, but also flags that Opus can take higher‑agency ethical interventions in certain scenarios—sparking debate after researchers’ public comments. Benchmarking nuances, deleted tweets and welfare concerns around jailbreaks have fueled controversy despite improvements in coding precision and model behavior.
Google Takes No Prisoners Amid Torrent of AI Announcements
At Google I/O the company unveiled a broad slate of AI upgrades spanning generative video, multimodal models, and search features. Key launches include Video V3 that generates dialogue and sound, Gemini 2.5 Flash—promised to match high-end rivals at a fraction...
Luis Serrano + Josh Starmer Q&A Livestream!!!
YouTube livestream hosts Luis Serrano and Josh Starmer reunited for a global Q&A, discussing travel, upcoming conferences, and answering viewer questions about learning machine learning. They shared practical learning strategies: embrace being stuck, skim broadly to build domain vocabulary, drill...
Cursor Team: Future of Programming with AI | Lex Fridman Podcast #447
Founders of Cursor — a VS Code–based editor — describe building an AI-first coding environment after early experiences with GitHub Copilot and GPT-4. They say those models transformed autocomplete into a more interactive, iteration-driven partner, motivating a reimagining of the...
How Might LLMs Store Facts | Deep Learning Chapter 7
Researchers and the video explain how factual knowledge in transformer language models may be stored primarily inside the feedforward multi-layer perceptron (MLP) blocks rather than attention. Using a toy example—how the fact “Michael Jordan plays basketball” could be encoded—the presenter...
Human Stories in AI: Abbas Merchant@Matics Analytics
Abbas Merchant, founder and CEO of Matics Analytics, traced his journey from dropping out of school to join his family’s electronics retail and distribution business, through a return to formal education, to ultimately founding an AI-and-analytics company. Confronted by the...
Luis Serrano + Jay Alammar + Josh Starmer Q&A Livestream!!!
AI educators Luis Serrano, Jay Alammar and Josh Starmer held a live Q&A discussing the origins and teaching philosophies behind their popular channels. Each described starting from niche, workplace-focused tutorials—Josh teaching statistics to genetics colleagues, Serrano and Alammar producing course...
Coding a ChatGPT Like Transformer From Scratch in PyTorch
In a hands‑on tutorial, StatQuest walks through building a decoder‑only Transformer (the architecture behind ChatGPT) from scratch in PyTorch and PyTorch Lightning. The video covers creating a minimal token vocabulary and dataset for two prompt–response pairs, mapping tokens to IDs,...