AI startup news.
Fresh from across the AI startup ecosystem — funding announcements, product launches, analysis, and interviews.
analysis Box CEO Aaron Levie says tech executives suffer from AI psychosis
Box founder Aaron Levie argues CEOs are prone to AI delusions because they're too distant from actual implementation work to understand what can realistically be automated.
MiniMax Releases M2 Series: 230B Parameter Model with 9.8B Active Tokens
Chinese AI lab MiniMax unveiled its M2 series of mixture-of-experts language models, with the flagship containing 229.9 billion total parameters but activating only 9.8 billion per token.
Researchers Develop CUDAnalyst to Decode AI Agent Planning in GPU Code Generation
New research introduces CUDAnalyst, an analysis framework that reveals how AI agents make planning decisions when generating CUDA kernel code by isolating and attributing different feedback signals.
Anthropic Launches Claude Tag for Slack Team Collaboration
Anthropic released Claude Tag, a Slack integration that lets teams delegate tasks to Claude as an autonomous team member with access to channels, tools, and codebases.
OpenAI and Broadcom unveil Jalapeño custom inference chip
OpenAI partnered with Broadcom to develop Jalapeño, a custom AI chip optimized for large language model inference to boost performance and efficiency.
Mistral Ships OCR 4 Document Intelligence Tool Supporting 170 Languages
Mistral AI launched OCR 4, a document intelligence system that extracts structured content with bounding boxes and confidence scores across 170 languages, claiming 4x speed improvements over competitors.
GitHub investigating unauthorized access to internal repositories
GitHub disclosed it is investigating unauthorized access to internal repositories but says no evidence yet shows customer data outside internal systems was compromised.
Google Research unveils new framework for auditing AI machine unlearning
Google Research introduces Regularized f-Divergence Kernel Tests, a statistical framework that makes auditing whether AI models have successfully 'forgotten' specific training data more accurate and computationally efficient.
MIT Spinout Ferveret Raises Funding for Nuclear-Inspired Data Center Cooling
Ferveret, founded by MIT researchers, has developed a nuclear reactor-inspired cooling system that reduces data center energy consumption by 35% while eliminating water usage entirely.
Modal raises $355M Series C to build agent-first cloud infrastructure
Modal, the serverless compute platform, closed a $355M Series C round as it shifts from developer experience to agent experience, building specialized infrastructure for AI workloads including elastic inference and sandboxes.
Databricks Tests Coding Agents on Multi-Million Line Codebase
Databricks benchmarked coding agents including OpenAI, Anthropic and open-source models on real engineering tasks across its multi-million line codebase, finding GLM 5.2 matches premium models at lower cost.
OpenAI finds 30% of SWE-Bench Pro coding tasks contain evaluation flaws
OpenAI's audit of the widely-used SWE-Bench Pro coding benchmark reveals that approximately 30% of tasks contain breaking issues including overly strict tests and underspecified prompts that misrepresent model capabilities.
Nvidia Partners with Hugging Face on Open Source Robot Models
Nvidia integrated its Isaac GR00T humanoid robotics models and teleop framework into Hugging Face's LeRobot library, expanding open source physical AI development tools.
xAI ships Grok 4.5, promising Opus-class performance at half the cost
Elon Musk's xAI released Grok 4.5 on Wednesday, claiming the model matches Anthropic's Opus capabilities while offering twice the token efficiency and significantly lower pricing.
xAI launches Grok 4.5 with 62% DeepSWE score at $2 per million tokens
Elon Musk's xAI released Grok 4.5, its most capable model yet, achieving 62% on coding benchmarks while delivering 4.2x better token efficiency than competing models.
OpenAI launches GPT-Live voice models with real-time conversation for ChatGPT
OpenAI released GPT-Live-1 and GPT-Live-1 mini voice models that enable simultaneous listening and speaking, ending the turn-based limitations of Advanced Voice Mode.