Deepseek R2's release is delayed due to US export controls and the CEO's dissatisfaction with its performance.
A shortage of Nvidia server chips, exacerbated by the US ban on Nvidia's H20 chips designed for China, limits the infrastructure needed to run the model.
Deepseek R1 focused on efficiency, making the infrastructure support issue for R2 surprising.
Meta has been aggressively poaching top AI researchers from OpenAI, with reports of $100 million offers.
Three researchers from OpenAI's Zurich office, previously from Google DeepMind, have confirmed their move to Meta.
Meta also acquired Scale AI for billions primarily to hire its team, as major companies like Google and OpenAI canceled contracts after the acquisition.
Mark Zuckerberg is reportedly trying to acquire or poach talent from Safe Super Intelligence, a company co-founded by Ilya Sutskever.
This aggressive hiring strategy follows the underwhelming reception of Llama 4 and aims to concentrate top AI talent at Meta to advance open-source models.
The high compensation is justified by Meta's multi-trillion dollar market cap and the massive market potential of AI.
Meta recently hired Trapit Bansal, a key OpenAI researcher who worked on reinforcement learning and was a foundational contributor to OpenAI's first reasoning model, 01.
Talks between OpenAI and Microsoft to renegotiate their deal have reportedly broken down.
Microsoft's current deal gives them 20% of OpenAI's revenue off the top, not just profit, and all IP rights to OpenAI models until 2030.
OpenAI wants to convert to a for-profit company to raise capital for a future IPO and seeks to waive the AGI clause, end Microsoft's IP rights, and swap the 20% revenue share for royalty plus equity.
Microsoft is unwilling to accept the new offer, as they have no incentive to do so, possibly only considering it to avoid antitrust investigations.
OpenAI is considering accusing Microsoft of anti-competitive behavior if an agreement isn't reached within six months, which could impact current investments and SoftBank's $30 billion pledge, potentially cutting it to $10 billion.
The OpenAI.io page, related to Johnny Ive and Sam Altman's collaboration for AI hardware devices, is temporarily down due to a trademark complaint from a company named IO (Iyo).
The suing company, IO, produces an AirPod-like ear device with AI capabilities and a voice interface.
OpenAI had known about IO's existence and technology since at least 2022, and IO's founder had pitched OpenAI for investment or acquisition.
Sam Altman confirmed that OpenAI passed on the investment/acquisition but revealed their competitive product would also be called "IO."
An internal email from Peter Willander to Sam Altman indicated a lack of fit with IO's device, stating it was "orthogonal to ours and doesn't really work yet."
11 Labs introduced "11 AI," a full voice AI assistant designed to explore their conversational AI technology.
The assistant aims to perform tasks like planning days, researching, creating tickets, and summarizing messages, with initial integrations with Perplexity, Linear, Slack, and Notion.
11 AI is currently in alpha and features an MCP (Multi-Channel Platform) integration for flexibility.
Mira Murati, former CTO of OpenAI, founded Thinking Machine Labs (TML) and raised $2 billion at a $10 billion valuation from Andreessen Horowitz within five months.
TML aims to develop custom AI for businesses to increase revenue, using reinforcement learning (RL) to optimize models based on specific KPIs like revenue or profit growth.
This approach is described as "RL for businesses," but specific details are still broad.
The startup also plans a consumer product, though its form is unclear.
TML intends to accelerate model development by plucking and combining specific AI model layers, a technique similar to model merging.
Imagen 4 and Imagen 4 Ultra: Google rolled out the new versions of its text-to-image model in the Gemini API and Google AI Studio. Imagen 4 Ultra is priced at 6 cents per output image. Examples show accurate text generation within images and high-quality results.
Gemma 3N: Google launched Gemma 3N, an open-source small model (2 GB and 3 GB versions) optimized for on-device use. It is multimodal (image, audio, video, text inputs/text outputs) and uses a new "matformer" architecture. Gemma 3N, with 4 billion parameters, is the first model under 10 billion parameters to achieve a 1300 ELO score on LM Arena.
Alpha Genome: Google introduced Alpha Genome, a unifying DNA sequence model for predicting regulatory variant effects, available via API. It accurately predicts how single mutations in human DNA impact gene-regulating biological processes, aiming to improve health and drug discovery.
Gemini CLI: Google released Gemini CLI, a free, open-source AI agent similar to Claude Code, bringing Gemini directly into developer terminals. It offers a quota of 60 requests per minute and 1,000 model requests per day and is available as a VS Code extension.
A federal judge in San Francisco ruled that Anthropic's use of copyrighted books to train its Claude AI models falls under fair use.
The judge sided with tech companies, stating that Anthropic's LLMs train on works "not to race ahead and replicate or supplant them, but to turn a hard corner and create something different."
This ruling is significant for the AI industry regarding copyright and training data.
Anthropic released a paper on how people use AI models for emotional support, identifying use cases such as interpersonal advice, coaching, psychotherapy/counseling, companionship, romantic roleplay, and sexual roleplay.
Emotional support accounts for a relatively small portion of Claude's total usage (2.9%), but is expected to grow.
Conversations in this category tend to end more positively than they start.
The models push back in about 10% of cases when potential harm, such as discussions related to eating disorders, is detected.