AI Updates on 2025-10-06

OpenAI announces GPT-5 Pro and Sora 2 are both available in the API starting today at DevDay @AndrewCurran_
OpenAI launches AgentKit, a complete set of building blocks for developers to build, deploy and optimize agent workflows with visual builder, evals, and guardrails @gdb
OpenAI introduces Apps in ChatGPT, allowing users to chat with apps like Canva, Booking.com, Spotify, and Figma directly within conversations @OpenAI
OpenAI makes Codex generally available with new SDK and enterprise features, demonstrated with live vibe coding including voice interface @gdb
Anthropic releases Petri, an open-source automated auditing tool for testing AI models across diverse scenarios for behaviors like sycophancy and deception @AnthropicAI
Google DeepMind announces CodeMender, an AI agent using Gemini Deep Think that automatically patches critical software vulnerabilities, having already submitted 72 high-quality fixes to major open-source projects @GoogleDeepMind
Microsoft updates Copilot memory to allow users to add, modify, and delete what Copilot knows about them, with the ability to direct both remembering and forgetting @Copilot

ChatGPT reaches 800 million weekly active users and OpenAI's API processes over 6 billion tokens per minute, with 4 million developers now building with OpenAI tools @AndrewCurran_
Private AI startups raised $377 billion in H1 2025, more than any full year in history, with 2x the capital per company averaging $36M @deedydas
OpenAI partners with AMD to deploy 6GW of AMD GPUs, beginning with a 1GW deployment in the second half of 2026, as part of scaling next-gen AI infrastructure @OpenAINewsroom
Perplexity expands internationally by opening an office in Berlin, Germany, with 4 MTS onboarded @AravSrinivas
Engineering leaders interviewing for AI product positions often lack actual AI knowledge beyond using ChatGPT, according to a recruiter at a publicly traded tech company @GergelyOrosz
AI infrastructure spending may be driven partly by lack of market exposure options to transformative AI, with data centers being one of the few ways to get "AGI" hedges in portfolios @emollick
2026 is expected to be when recent massive AI infrastructure investments start becoming available as usable compute @natolambert

Microsoft researchers reveal a confidential research effort exploring how open-source AI tools could bypass biosecurity checks, helping create fixes now influencing global standards @MSFTResearch
Concerns raised about the trajectory of open AI models in America, with debates about potential bans on open weights models despite practical implementation challenges @natolambert
Discussion of whether interacting with AIs might actually be better for human flourishing in some cases, challenging assumptions about AI interaction being inherently negative @jeffclune

Figma launches integration with ChatGPT allowing users to create FigJam diagrams through natural language prompts @figma
Mattel uses Sora 2 for instant sketch to toy concept generation, demonstrating AI video applications in product design @gdb
Comet browser introduces new addiction pattern where users open long YouTube videos and use AI assistant to navigate to specific timestamps based on questions rather than linear viewing @AravSrinivas
AI-assisted online shopping continues booming according to new U.S. holiday e-commerce forecasts @TechCrunch
Stanford introduces MedAgentBench, a virtual environment to test whether AI agents can handle complex clinical workflows like retrieving patient data, ordering tests, and prescribing medications @StanfordHAI

GPT-5 Pro achieves breakthrough results in mathematics, solving a problem previously unsolved by LLMs and only solved by 60 humans, plus solving an open problem in real analysis @deedydas
Research shows small Transformers perform better at multiplication when trained to stop relying on explicit Chain-of-Thought steps, suggesting hidden-thought circuits might emerge spontaneously in frontier-scale training @davidad
A 7B model fine-tuned for forms and documents beats GPT-4.1 on 1,000 extraction tasks, trained for only $196 using synthetic training data and LoRA with Group Relative Policy Optimization @rohanpaul_ai
GLM-4.6 becomes the new #1 top open model on Hugging Face Arena, ranking #4 overall and surpassing DeepSeek R1 which had been champion for months @arena
Research confirms LoRA rank=1 closely matches full fine-tuning performance on many RL fine-tuning problems, with successful reproductions showing significant parameter efficiency @johnschulman2
New lightweight open-source text-to-speech model kani-tts-370m released with 370M parameters, achieving natural and expressive voice with real-time inference on RTX 3060 @Tu7uruu
Science systems are breaking under flood of human-created knowledge, with concerns about how to handle potential flood of AI-generated discoveries and translate them into streams of inquiry and practice @emollick