Some LLM takeaways for 2025: reasoning as a signature feature, coding agents were useful, subscriptions hit $200/month, and Chinese open-weight models impressed
It's that time. It's been a hell of a year. — At the start we barely had reasoning models. X: Simon Willison / @simonw : Here's my enormous round-up of everything we learned about LLMs in 2025 - th...
GPT-5.2 models have a 400K context window and 128K max output tokens, like GPT-5 and 5.1, but have an August 31, 2025, knowledge cutoff, vs. September 30, 2024
Simon Willison / Simon Willison's Newsletter :
OpenAI merges ChatGPT's voice mode directly into the main text chat interface by default; users can still switch back to the original, separate voice mode
here's how Aman Kumar / PhoneArena : OpenAI has finally addressed a major problem with ChatGPT's voice mode feature NDTV Profit : ChatGPT Voice Now Built Into Main Interface: What's New For Users? Sar...
Google launches Gemini 3 Pro Image, aka Nano Banana Pro, with more control, improved text rendering, and enhanced world knowledge, for free in the Gemini app
except when it gaslit me Ryan Whitwam / Ars Technica : Google's new Nano Banana Pro uses Gemini 3 power to generate more realistic AI images Robert Hart / The Verge : Google's new AI image creator too...
Nano Banana Pro is great at following instructions, generates interim “thought images”, and makes full infographics with well-rendered text from a short prompt
Simon Willison / Simon Willison's Weblog :
First impressions of ChatGPT Atlas, as browser agents remain confusing, with insurmountable security and privacy risks including prompt injection attacks
a web browser with ChatGPT built in, not bolted on. The browser is the agent now. Tabs are prompts. The search bar is dead. Welcome to the post-URL era. P.S the browser wrote this on its own Arlan / @...
DeepSeek releases DeepSeek-OCR, a vision language model designed for efficient vision-text compression, enabling longer contexts with less compute
the new frontier of OCR from @deepseek_ai , exploring optical context compression for LLMs, is running blazingly fast on vLLM ⚡ (~2500 tokens/s on A100-40G) — powered by vllm==0.8.5 for day-0 model su...
Anthropic announces Claude Skills, a tool with folders of instructions, scripts, and resources that Claude can load to improve how it performs specific tasks
here's how they could supercharge your workflow Robert Brown / Implicator.ai : Claude taps Microsoft 365 as the connector war heats up X: Olivia Moore / @omooretweets : Like many of Anthropic's produc...
Nvidia DGX Spark hands-on: trades performance and bandwidth for 128GB of unified memory, the ecosystem is a big selling point, the design is standard, and more
What It Means for NVDA Stock and AI Crypto Sentiment Maximilian Schreiner / The Decoder : Early reviews suggest Nvidia may have found another way to sell its chips with the DGX Spark Robert Brown / Im...
Andrej Karpathy unveils nanochat, a full-stack training and inference implementation of an LLM in a single, dependency-minimal codebase, deployable in 4 hours
It provides a full ChatGPT-style LLM, including training, inference and a web Ui … X: Clem / @clementdelangue : Am I wrong in sensing a paradigm shift in AI? Feels like we're moving from a world obses...