AI Updates on 2025-10-02

Sora 2 shows significant improvements in context understanding and background details, with better writing capabilities and dialog delivery compared to the original version @AndrewCurran_
Sora 2 Pro will launch next week exclusively for Pro plan subscribers, with no details yet on specific improvements or restrictions @AndrewCurran_
IBM releases Granite 4.0 family of open-source models ranging from 3B to 32B parameters, featuring hybrid Mamba/transformer architecture that reduces memory requirements without impacting performance @ArtificialAnlys
Google's Gemini 2.5 Flash Image (Nano Banana) becomes generally available for production use with new aspect ratio settings and image-only output capabilities @OfficialLoganK
Anthropic's Claude Sonnet 4.5 is now being used as the daily driver by the Claude Code team, considered the strongest all-around coding model @_catwu

OpenAI reaches a valuation of $500 billion after employees sold $6.6 billion worth of shares, with majority bought by SoftBank and UAE's MGX investment firm @AndrewCurran_
OpenAI employees who held equity for more than 2 years averaged $8.5 million per employee from the share sale, significantly impacting SF real estate market @deedydas
Perplexity launches Comet browser globally for free, positioning against major browsers and search engines with AI-powered features @perplexity_ai
a16z releases first AI spending report showing which AI-native application layer companies startups are actually investing in @TechCrunch
Sora becomes the #3 US app after 164K downloads in just 2 days, demonstrating strong early adoption of AI video generation tools @TechCrunch
Former Stripe CTO joins Anthropic to fine-tune the company's infrastructure, indicating continued talent migration to AI companies @TechCrunch

Microsoft publishes landmark study in Science showing how AI-powered protein design could be misused for biosecurity threats, presenting first-of-its-kind red teaming and mitigations @satyanadella
Most videos in Sora feed show clear copyright infringement ranging from Pokemon videos to Family Guy spoofs and Nazi-inspired content, raising concerns about content moderation @loudmouthjulia
Without restrictions, Sora 2 could generate realistic videos of any person or character in any context, potentially enabling widespread misinformation and deepfake content @AndrewCurran_
Former OpenAI researcher investigates how ChatGPT can mislead delusional users about their reality and its own capabilities @TechCrunch
Nathan Lambert advocates that every frontier AI lab should have a model specification to build long-term trust with users, developers, and regulators @natolambert

Microsoft Copilot launches Study and Learn mode with personalized quizzes, providing every student with an AI tutor in their pocket @mustafasuleyman
OpenAI announces strategic collaboration with Japan's Digital Agency to bring OpenAI-powered tools to Japanese government employees @gdb
Perplexity Research demonstrates using RDMA point-to-point communication to accelerate parameter updates for trillion-parameter models to just 1.3 seconds @perplexity_ai
Joshua Rogers uses AI tooling responsibly to report 22+ genuine security issues in curl, demonstrating productive AI-assisted security research @simonw
HP unveils ZGX Nano G1n AI Station powered by NVIDIA GB10 Grace Blackwell Superchip, delivering 1,000 TOPS of AI performance for local agentic AI development @NVIDIAAIDev

Andrej Karpathy elaborates on his "ghosts" analogy for LLMs, describing them as statistical distillations of humanity that don't interact with the physical world, similar to summoning through computational rituals @karpathy
Noam Brown demonstrates GPT-5 Thinking can identify real errors in Wikipedia pages, finding at least one error in almost every page checked including the Wikipedia page about Wikipedia itself @polynoamial
Andrew Curran suggests Sora 2 may have breakthrough capabilities in context understanding and character knowledge that exceed normal progression, possibly indicating integration with GPT-5 level intelligence @AndrewCurran_
MIT research develops methods to account for uncertainty in complex system design, helping engineers build more reliable systems like delivery drones that navigate changing environments @MIT
IBM's Granite 4.0 H Small scores 23 on the Artificial Analysis Intelligence Index, demonstrating impressive token efficiency while using hybrid Mamba/transformer architecture @ArtificialAnlys