AI Updates on 2025-06-03

OpenAI rolls out Codex to ChatGPT Plus users with internet access capabilities and user control over HTTP methods and domains @OpenAI @gdb
Anthropic announces Research and Integrations features are now available on their Pro plan, allowing Claude to search across web, Google Workspace, and connected tools @AnthropicAI
Hugging Face releases SmolVLA, a 450M parameter Vision-Language-Action model for robotics with best-in-class performance and inference speed @huggingface
H Company open-sources Holo-1 3B and 7B parameter Action Models achieving 92.2% SOTA on WebVoyager benchmark @huggingface
Shisa AI releases Shisa V2 405B, described as "the highest-performing LLM ever developed in Japan," trained on top of Llama 3.1 405B @simonw

Meta signs 20-year nuclear power deal with Constellation Energy for 1121 MW from Clinton Clean Energy Center to power AI operations @AndrewCurran_
Survey reveals 43.2% of US workers now use generative AI at work for 1/3 of their tasks, reporting tripled productivity on those tasks, though gains aren't passed to organizations @emollick
Builder AI, a $1.5B company, declares bankruptcy after being caught in loan fraud and money laundering cases, with auditors slashing revenue by 75% @deedydas
Amazon reportedly making a movie about OpenAI's 2023 board events with Sam Altman potentially played by Andrew Garfield @AndrewCurran_
a16z publishes thesis on AI disrupting the $140B market research industry, replacing human surveys with AI-moderated interviews and synthetic agent societies @a16z

AI Now Institute releases "Artificial Power" report examining fallout from AI hype cycle and warning about tech companies pushing AI into social, political, and economic systems @AINowInstitute
UK tech secretary Peter Kyle's ChatGPT logs obtained through FOI request, revealing questions about AI adoption in UK small business community to a model with outdated training data @simonw
Yoshua Bengio launches LawZero, a nonprofit AI safety lab focused on existential risk, with Jeff Clune joining as scientific advisor @TechCrunch @jeffclune
Christopher Manning warns against "democracy-washing" in OpenAI's country-specific initiatives, suggesting US government benefits most from such programs @chrmanning

Perplexity CEO highlights rapid growth in agentic commerce applications and superior travel search capabilities @AravSrinivas
Andrew Ng advocates for universal coding with AI assistance, reporting that everyone at AI Fund can now code using AI tools for enhanced creativity and productivity @AndrewYNg
Claire Vo demonstrates shift from task execution to system building, using Zapier agent for automatic email categorization and reply drafting @clairevo
Soleio showcases AI-generated company wiki that automatically updates from meeting conversations, noting benefits of real-time accuracy but challenges with manual additions @soleio
Gergely Orosz creates "vibe coding" survey application to study differences between developer and non-developer AI coding approaches @GergelyOrosz

Meta researchers publish findings that GPT-style language models memorize 3.6 bits per parameter, using Shannon theory to compute total memorization capacity @AndrewCurran_
Berkeley AI Research introduces FeelTheForce (FTF), enabling robots to learn force-sensitive manipulation from human interaction videos @berkeley_ai
Nathan Lambert discusses DeepSeek's strategy of using synthetic data from top API models to overcome GPU limitations while having cash resources @natolambert
Jeff Clune proposes AGI definition as "something that passes a good version of the Turing Test" and explores AI systems that can evolve autonomously beyond human-designed constraints @jeffclune
Hugging Face releases Video-XL-2 model capable of handling 10,000+ frames on single GPU with 2048-frame encoding in 12 seconds @huggingface