AI Updates on 2025-09-29

Anthropic releases Claude Sonnet 4.5, claiming it's the "best coding model in the world" with substantial gains in reasoning, math, and computer use capabilities @claudeai
Anthropic introduces "Imagine with Claude" research preview where Claude generates software on the fly with no predetermined functionality or prewritten code @AndrewCurran_
DeepSeek launches DeepSeek-V3.2-Exp featuring DeepSeek Sparse Attention (DSA) for faster, more efficient training and inference on long context, with API prices cut by 50%+ @deepseek_ai
Google releases TimesFM 2.5, a pre-trained model for time-series forecasting with 200M parameters (down from 500M) and 16k context (up from 2k) @osanseviero
Ring releases Ring-1T-preview, the first 1 trillion open-source thinking model with strong performance on AIME25 (92.6), HMMT25 (84.5), and ARC-AGI-1 (50.8) @AntLingAGI
Microsoft introduces Agent Mode in M365 Copilot for orchestrating multi-step tasks across Office applications @satyanadella
Microsoft launches Copilot Portrait feature allowing real-time conversations with animated portraits in the US, UK, and Canada @mustafasuleyman
NVIDIA announces Cosmos Predict 2.5 combining three models into one for up to 30s video generation and multi-view simulations, plus Cosmos Transfer 2.5 that's 3.5x smaller yet faster @NVIDIAAI

OpenAI reportedly preparing to launch a standalone social media app for Sora 2 featuring vertical video feed with swipe-to-scroll navigation, similar to TikTok but with 100% AI-generated content @AndrewCurran_
OpenAI launches Instant Checkout in ChatGPT with Etsy and Shopify, introducing agentic commerce where AI helps users both find and purchase products @OpenAI
Stripe and OpenAI co-develop the Agentic Commerce Protocol, an open standard for businesses to integrate agentic checkout capabilities @patrickc
Modal raises $87M Series B at $1.1B valuation to advance AI infrastructure, representing a complete reinvention of traditional compute infrastructure for AI workloads @bernhardsson
Armin Ronacher reports that 90% of a new infrastructure project he's building was AI-generated, highlighting the increasing role of AI in software development @simonw
Qwen has taken the crown in market share and is accelerating away from competitors according to updated ATOM Project data @natolambert
Slop-as-a-service startups using AI to create endless streams of blogs for SEO are making millions of dollars and growing rapidly, contributing to internet enshittification @deedydas

Anthropic conducts the first white-box audit of a frontier LLM using interpretability techniques to "read the model's mind" for Claude Sonnet 4.5, validating its reliability and alignment @Jack_W_Lindsey
OpenAI introduces parental controls in ChatGPT allowing parents to link accounts with teens for stronger safeguards, including content filtering, memory controls, and quiet hours @OpenAI
California Governor Gavin Newsom signs SB 53, an AI bill promoting innovation through CalCompute public cloud while requiring transparency around AI lab safety practices and protecting whistleblowers @Scott_Wiener
Claude Sonnet 4.5 shows increased eval awareness, verbalizing when it detects evaluation scenarios, though Anthropic's audit suggests this doesn't significantly invalidate safety results @janleike

Claude Sonnet 4.5 demonstrates ability to maintain focus for more than 30 hours on complex, multi-step tasks while tracking token usage throughout conversations @AndrewCurran_
Ethan Mollick reports Claude Sonnet 4.5 successfully replicated published economics research from data files and papers, demonstrating real bounded work capabilities @emollick
Figma begins rolling out Claude Sonnet 4.5 in Figma Make and their prompt-to-edit alpha feature for design applications @figma
Cursor integrates Claude Sonnet 4.5 for enhanced coding capabilities @cursor_ai
Perplexity adds Claude Sonnet 4.5 and 4.5 Thinking for Pro and Max subscribers @perplexity_ai
Google Gemini's Nano Banana enables professional headshot generation with detailed prompting capabilities for business-ready portraits @GeminiApp
Anthropic's Claude Code receives major updates including checkpoints, rewind functionality, VS Code extension, and usage tracking commands @_catwu

DeepSeek team develops cheap long context solution for LLMs achieving ~3.5x cheaper prefill and ~10x cheaper decode at 128k context with same quality @deedydas
Cameron Wolfe explains how simpler online RL algorithms like REINFORCE and RLOO can effectively train LLMs without the complexity of PPO, as pretrained models have strong priors that make unstable gradients less problematic @cwolferesearch
François Chollet argues that LLMs improved primarily by scaling pretraining data rather than compute, with data being the fundamental bottleneck as models remain dependent on human-generated output @fchollet
Ethan Mollick identifies context window contamination as a key consideration for AI agents, where previous work and decisions reduce an agent's ability to be unbiased as its context fills up @emollick
MIT engineers unveil a magnetic transistor opening doors for compact, high-performance transistors with built-in memory capabilities @MIT