Isomorphic Labs raises $2.1B to accelerate AI drug discovery
Google DeepMind co-founder Demis Hassabis founded Isomorphic Labs completes $2.1 billion funding round for AI protein structure prediction and molecular design.
Isomorphic Labs, the AI-first drug discovery company founded by Google DeepMind co-founder Demis Hassabis, has raised $2.1 billion in new funding. The company builds on DeepMind's groundbreaking AlphaFold protein structure prediction technology to design novel therapeutic molecules and dramatically shorten the drug development timeline. For years, Hassabis and his team have been building powerful AI systems that can predict protein structures with atomic accuracy. This latest funding round signals growing investor confidence in the convergence of artificial intelligence and pharmaceutical research. The capital will accelerate Isomorphic's efforts to bring AI-designed therapeutics to clinical trials, potentially transforming how new medicines are discovered and brought to patients worldwide.
Paper proposes unified scaling method, trains Olympiad gold-level reasoning model
Researchers propose a general method to transform post-training reasoning models into Olympiad-level problem solvers. Using reverse perplexity curriculum SFT and two-stage reinforcement learning on a 30B-A3B backbone, the resulting SU-01 model achieves gold-medal performance at IMO 2025, USAMO 2026, and IPhO 2024/2025, with stable reasoning trajectories exceeding 100,000 tokens and generalization to other scientific domains.
Codex integrated into ChatGPT app, enables remote Mac control from phone
OpenAI announces Codex is now available directly inside the ChatGPT app, giving users full access to programming capabilities anywhere. A standout feature allows users to control their Mac remotely from an iPhone, effectively turning the phone into a mobile development terminal and blurring the line between desktop and mobile coding workflows. The integration marks a meaningful expansion of Codex beyond its original developer-facing interface.
X platform open-sources its recommendation algorithm on GitHub
Elon Musk announced that the X platform's "For You" recommendation algorithm has been published to GitHub under the xai-org/x-algorithm repository. The open-sourcing of the algorithm powering the For You feed represents a significant move toward transparency in social media recommendation systems, inviting developer contributions and independent external audits of the ranking logic.
OpenAI merges ChatGPT and Codex into unified product team
OpenAI announced a major internal reorganization: ChatGPT, Codex, and the developer API business have been integrated under a single core product team led by co-founder and president Greg Brockman. In an internal memo, Brockman stated that the company will merge ChatGPT and Codex into a unified experience, signaling a strategic consolidation of OpenAI's consumer and developer product lines under one cohesive strategy.
The difference between Grok foundation model 8 and 9 is gigantic.
@teortaxesTex
Gemini Deep Think powers Aletheia to autonomously solve math problems
The Aletheia system, powered by Gemini Deep Think, autonomously solves complex mathematical problems, marking a new milestone in AI mathematical reasoning.
NVIDIA DC reveals how to serve agentic workloads on trillion-parameter models
NVIDIA DC demonstrates serving agent workloads on trillion-parameter models at 400 tokens per second per user, balancing quality and inference performance.
MiniMax M2.7 model now available on OrcaRouter via OpenAI-compatible API
MiniMax's M2.7 model goes live on OrcaRouter, accessible through a single OpenAI-compatible API endpoint for seamless integration.
vLLM-Omni achieves production-grade omni-modal inference at Baseten
The vLLM-Omni extended framework now runs in production at Baseten, supporting streaming multimodal inference and real-time TTS workloads previously dominated by closed-source APIs.
v0 launches Browser Use: autonomously test and fix web applications
v0 can now open the apps it builds, use them, critique designs, debug complex flows, and fix things proactively, all while sending screenshots of what it sees.
Vercel's v0 design-to-code platform takes a significant step toward autonomous development with the launch of Browser Use. The AI can now open the web applications it builds, interact with them, critique visual designs, debug complex interaction flows, and proactively fix issues without human intervention. While v0 works, it sends screenshots of its browser view to the user, creating a transparent feedback loop. This feature blurs the line between AI-assisted design and autonomous quality assurance, pushing the platform closer to full-stack agentic development.
Grok subscription integrates with Nous Research Hermes Agent
xAI announces Grok subscriptions can now be integrated into Nous Research's Hermes Agent, allowing users to call relevant functions via their Grok account within the open-source self-improving agent framework.
vLLM delivers Day-0 support for Intern-S2-Preview science model
Intern-S2-Preview is an open-source scientific multimodal foundation model (36B total, 3B active) supporting material crystal structure generation. vLLM achieved Day-0 compatibility, with the model featuring 262K context length and MTP-accelerated inference.
Runway Agent turns product shots into finished ads in one session
Runway introduces Agent mode, allowing users to go from a product shot and an idea to a fully finished advertising video in a single session, streamlining the creative production pipeline.
LTX-2.3 audio finetuning achieves SOTA emotion-controlled TTS
Researchers extracted the audio component of LTX-2.3, fine-tuned it for text-to-speech, and achieved state-of-the-art emotional control in synthesized speech, now available for trial on Hugging Face.
Microsoft releases 3.8B parameter text-to-image model Lens
Microsoft's Lens, a 3.8 billion parameter text-to-image generation model, lands on Hugging Face with efficient training characteristics and high output quality.
Microsoft unveils Fara-7B, optimized for browser OS environments
Microsoft releases Fara-7B on Hugging Face, a model purpose-built for browserOS environments that has received strong praise from early developer adopters for its performance.
Luma releases Creative Agents: generate origami-style works from ideas
LumaAI launches Creative Agents supporting planning, generation, iteration, and refinement with full-context continuity. A demo titled "Origami" showcases two birds and one sword in the signature style.
Ollama significantly expands Blackwell GPU capacity for GLM-5.1
Ollama adds a large number of NVIDIA Blackwell GPUs to better serve the GLM-5.1 model on its cloud, with daily capacity expansion continuing for other models, while also supporting Claude Code and Codex app integrations.
Ollama now supports Codex app, open models selectable
Starting from version 0.24, Ollama supports the Codex app. Users can run `ollama launch codex-app` and choose from available open models to use with the coding agent.
ChatGPT rolls out personal finance management for Pro users
US Pro users can now securely connect bank, investment, and credit card accounts inside ChatGPT, view a unified asset dashboard, and ask questions based on real financial data.
SWE-ZERO-12M trajectory dataset released, largest open agent traces
SWE-ZERO-12M contains 12 million agent trajectories, making it 5.7 times larger than the previous largest open-source agentic trace dataset, fully open for research and training.
Key pitfalls when evaluating diffusion language models
A blog post highlights pitfalls such as entropy differences across inference settings and non-comparable ELBO when evaluating diffusion language models, proposing a principled KL-divergence-based evaluation framework.
AI generates interactive 3D worlds from a single image in Cursor
Upload an image and the AI agent automatically generates a 3D environment with object meshes, collision bodies, and sound effects.
Open-ended coding training data no longer a bottleneck; AI can self-expand
Research shows AI can scale open-ended coding tasks autonomously, even surpassing human-annotated data quality.
Recraft V4.1 versus Midjourney V8.1: same prompts, divergent visual styles
Recraft compares V4.1 against Midjourney V8.1, showing distinct creative choices across composition, mood, texture, and detail.
Hermes Agent runs fully locally on DGX Spark via Ollama
NVIDIA AI PC publishes a tutorial showing how to run Nous Research's Hermes Agent entirely locally on DGX Spark through Ollama.
MiniMax models used in open-multi-agent framework with DAG task decomposition
MiniMax models automatically break goals into DAG tasks and execute them in parallel within an open multi-agent framework.
ChatGPT subscriptions now usable inside the Zed agent editor
Users can leverage their ChatGPT subscription within the Zed agent, with the same usage and rate limits as in Codex direct.
Vercel Ship London to feature Google DeepMind and ElevenLabs agents
Vercel Ship 26 in London on June 17 will showcase video agents and real-time voice agents, with production stories from Telegraph, Currys, and easyJet.
Vibecon NYC brings AI and code to the stage with Spike Jonze and Refik Anadol
Vibecon takes over New York June 17-18, featuring performances at the intersection of software and stage, with talks, workshops, and installations.
Runway Agent mode praised for building complex stories from short text
Early reviewers describe Runway's new Agent mode as impressive, capable of constructing complex visual stories from brief text descriptions.
Replit mobile app upgraded to Agent 4 with multi-task control
Replit's mobile application upgrades to Agent 4, adding support for controlling multiple tasks simultaneously on mobile devices.
Replit resolves Apple dispute, app returns to App Store after four months
Replit reached an agreement with Apple and relaunched its app with multiple updates, thanking the community for their support.
Kling AI Conference at Cannes unveils featured filmmaker speaker lineup
Three world-class filmmakers gather at Cannes to present works empowered by Kling AI technology, showcasing AI's growing role in cinema.
MiniMax powers AI-voiced Italian drama at Cannes Film Festival
MiniMax Speech 2.8 powers the Italian drama Il Cinese with immersive AI-voiced storytelling, presented at the Cannes Next panel on May 16.
SWE-ZERO and AgentTrove agent trajectory datasets gaining traction
SWE-ZERO and AgentTrove agent trajectory datasets are trending on Hugging Face, reflecting growing interest in agentic training resources.