AI Action Summit
The leaders of both governments and AI research labs convened in Paris this week for the AI Action Summit.
The big picture:
AI regulation is losing steam. Some Europeans view the EU AI Act as misguided, while the US and UK refused to sign a new international agreement.
China has kickstarted the next wave in an AI arms race, as CEOs push for more resources and governments eagerly provide them.
And while the politicians blustered, AI leaders keep pointing to incredibly short timelines: DeepMind's CEO suggested AGI was merely five years away, while Anthropic's CEO called out "profound economic, societal, and security" impacts in the next two years.
Elsewhere in AI geopolitics:
The UK has renamed its AI Safety Institute to AI Security Institute while exploring Claude's use in public services.
Apple will partner with Alibaba to power its AI features in the Chinese market, though, like in the US, it is working with multiple providers simultaneously.
And Sam Altman reportedly won the President's favor by suggesting that AGI would be achieved during his term.
Elsewhere in frontier models:
Google's Gemini Advanced subscribers can now access chat memory features that enable referencing past conversations for more personalized assistance.
Baidu announced its AI chatbot Ernie will become free for all users starting April 1st, citing technological improvements and lower costs.
DeepMind's AlphaGeometry2 outperformed average gold medalists by solving 84% of selected International Math Olympiad geometry problems.
And Hugging Face created an open-source AI research agent that aims to match OpenAI's Deep Research capabilities.
Reuters v. Ross
Thomson Reuters won the first major US AI copyright case against legal research AI startup Ross Intelligence (though it did not cover generative AI).
Why it matters:
Back in 2020, Ross Intelligence was building an AI-powered legal research tool, and needed training data. It tried to license data from Westlaw (a competitor owned by Thompson Reuters) but was refused - instead, it paid a third party with Westlaw access to generate the training data.
Legal experts suggest this ruling could severely undermine AI companies' fair use arguments, with one professor noting that much of the case law companies currently cite may be "irrelevant."
Even though Ross Intelligence shut down in 2021 (due to litigation costs), new AI copyright lawsuits crop up every week - including a new one from several media outlets against Cohere.
Elsewhere in AI anxiety:
A Microsoft study warns that AI tools are causing human cognitive atrophy, leaving people unprepared for complex tasks.
Anthropic's new Economic Index tracks AI’s impacts on the labor market and economy.
OpenAI has removed some warning messages from ChatGPT that previously flagged potential terms of service violations.
And a security researcher has demonstrated how prompt injection can manipulate Gemini's long-term memory across multiple user sessions.
GPT-501(c)(3)
Elon Musk launched a $97.4 billion offer to acquire OpenAI's nonprofit arm, escalating his ongoing attempt to block the company's for-profit conversion.
Between the lines:
To be clear, Musk is almost certainly not going to buy OpenAI - he's already said he'll withdraw the bid if the for-profit conversion is halted. But the offer throws a significant wrench in OpenAI's plans.
To convert to a for-profit organization, OpenAI must value its nonprofit assets at some number. This valuation directly impacts the number of shares Microsoft (and other investors) will get after the conversion.
Based on reported upcoming investments, OpenAI is looking to value itself at $260B, roughly making its nonprofit stake worth about $65B (the nonprofit is slated to get a 25% stake in the for-profit).
But Musk's $97B offer implies a "fair market" valuation of well over $65B, meaning either a) Microsoft et al will get a sweetheart deal (bad for getting regulatory approval) or b) Microsoft et al will get far fewer shares than they're currently planning (bad for them and future investors).
Elsewhere in OpenAI:
Sam Altman provided some major roadmap updates: GPT-4.5 (codenamed Orion) will be OpenAI's last "traditional" GPT model, while GPT-5 will incorporate reasoning capabilities (a la o1 and o3).
OpenAI significantly expands its Model Spec with a focus on customizability and transparency.
And the company is preparing to finalize its first in-house chip design to reduce dependency on Nvidia.
Things happen
AI agents are everywhere and nowhere. Anthropic may be launching a new hybrid AI model. Google CEO says quantum computers are 5-10 years away. The revenge of GPT wrappers. Meta's brain-typing AI system. Sam Altman's observations on AI and society. Arm CEO on DeepSeek and AI's "Sputnik moment." Adobe launches Generate Video in public beta. AI Startup School. Tech giants' capex spending soars to $246B in 2024. Saudi Arabia commits $1.5B to Groq and a $5B AI data center. Meta in talks to acquire FuriosaAI. Ilya Sutskever's SSI valued at $20B+. Christie's plans first AI art auction. Three principles for AI policymaking. "True Crime" series features AI-generated murders. DeepSeek R1 susceptible to jailbreaking. Malaysia to give 445,000 civil servants Google AI tools. Snowflake and Anthropic partner on AI integration. France pledges nuclear power for AI. Google edits Gemini's response in Super Bowl ad. Firing programmers for AI is a mistake. Open-source AI tool helps diagnose autoimmune disease. Building a personal AI computer on a budget. Meta launches AI Demos website. Indigenous engineers use AI to preserve culture. Lawyers cite AI-hallucinated cases. Building an AI company to save open-source. AI might erase what enabled its creation. Why AI chatbots struggle to say "I don't know". Britons support AI safety laws. YouTube Shorts adds AI video generation.
