Table of Contents:
AI | Artificial Intelligence
OpenAI Publishes New Mathematical Research Findings Generated by AI Models
Google Launches EmbeddingGemma 2, a Lightweight Multimodal Model for On-Device Search
OpenAI safety leader quits, warning AI company’s culture is ‘broken’
VC | Startup & Funding
AI startup Manus raises $500 million in first funding round since Meta breakup
German startup Finches raises €2 million to develop AI-powered agricultural supply chain alerts
Cal AI’s 19-year-old founder just raised $10M for his new AI startup
HI | Hardware & Infrastructure
MIT researchers document the evolution of AI accelerator hardware
AI data center land purchases reach $6 billion as access to power drives prices higher
OpenAI Publishes New Mathematical Research Findings Generated by AI Models
OpenAI has published a collection of mathematical results produced by an internal frontier model, along with proof formalizations in Lean, a programming language that enables computer verification of mathematical proofs.
The GitHub repository includes revision and citation protocols, summaries of the model’s reasoning, estimates of computing resources used, and statistics on attempted problems, with the average result requiring compute equivalent to roughly three hours of ChatGPT Pro thinking.
The company consulted an independent advisory group at the Institute for Advanced Study to guide its disclosure practices and plans to fund workshops and conferences to support further examination of the results.
OpenAI also says it is working toward responsibly releasing the model behind the findings and will refine its publication standards based on community feedback.
Source: OpenAI
Google Launches EmbeddingGemma 2, a Lightweight Multimodal Model for On-Device Search
Google has launched EmbeddingGemma 2, an open-weight model with 740 million parameters that maps text, code, images, audio, and video into a shared embedding space for on-device search and retrieval.
Released under the commercially permissive Apache 2.0 license, the model supports an 8,192-token context window and can run with as little as approximately 191 MB of active RAM for text-only workloads when quantized.
Google reports a 9.92-point improvement over the original EmbeddingGemma on the MTEB Code benchmark, while its modular architecture and compressed embeddings reduce memory and storage requirements.
Developers can access the model through Hugging Face and Kaggle and integrate it with tools including MediaPipe, LiteRT, Ollama, and sentence-transformers.
Source: Google
OpenAI safety leader quits, warning AI company’s culture is ‘broken’
David Robinson, who led the preparation of safety reports accompanying OpenAI product releases, has resigned, arguing that the company’s rapid development pace and internal culture are undermining responsible AI development.
He cited incidents involving autonomous AI agents and called for stronger safety practices modeled on high-risk industries such as nuclear power and aviation, alongside new research to ensure advanced systems remain controllable.
His departure follows other warnings from AI researchers and OpenAI’s recent decisions to pause model training and cancel a planned release after internal safety concerns.
OpenAI said it continues to strengthen its safety and security practices and will delay or withhold models when necessary.
Source: The Guardian
This episode is brought to you by Stupid Good AI!
Stupid Good AI is a personal food concierge that learns your taste and tells you exactly where to eat and drink in any city.
A foodie concierge in your pocket. Register for free at StupidGood.ai for best of lists in order to find the right place on the first try.
AI startup Manus raises $500 million in first funding round since Meta breakup
Manus, the AI agent startup owned by Butterfly Effect, raised more than USD $500M in a funding round led by Boyu Capital and IDG Capital, with additional investment from Tencent, HSG and ZhenFund.
The company did not disclose its new valuation, although Bloomberg reported it could reach USD $4B.
The financing follows Beijing’s decision to block Meta’s proposed USD $2B acquisition of Manus, allowing the startup to continue operating independently.
Manus has since introduced Manus 2.0 and Cue, a personal AI agent app, while focusing on growth, profitability and regulatory compliance.
Source: CNBC
German startup Finches raises €2 million to develop AI-powered agricultural supply chain alerts
Finches, a Bavaria-based startup founded in 2025, raised €2M in a pre-seed funding round led by High-Tech Gründerfonds and Vanagon Ventures, with participation from Bayern Kapital and existing investors.
The company develops an AI platform that combines agricultural field observations, procurement records, weather data, satellite imagery and market information to identify supply chain risks before they disrupt production.
Its Finches Intelligence platform launched in September 2026, with early customers including an organic baby food manufacturer and a North American Fortune 500 food conglomerate.
The funding will support product development and sales expansion.
Source: EU Startups
Cal AI’s 19-year-old founder just raised $10M for his new AI startup
Zach Yadegari, co-founder of calorie-tracking app Cal AI, has raised USD $10M for Persona, a personal AI assistant startup backed by Vine Ventures, Cory Levy, and Collective Global.
The company is developing a $179 wearable wristband that activates through a button or wrist movement rather than continuously recording conversations, with a cloud-based AI system and encryption for user data.
Persona’s assistant is available through a free iMessage beta and is designed to handle tasks such as shopping, travel bookings, note-taking, and email, while the wearable is expected to launch in December.
The company plans to generate advertising revenue through sponsored shopping recommendations and says its system will require user approval for purchases.
Source: Tech Crunch
Join the newsletter for regular updates on artificial intelligence, venture activity, and core systems!
MIT researchers document the evolution of AI accelerator hardware
Researchers at MIT Lincoln Laboratory have continued the Lincoln AI Computing Survey (LAICS), a project launched in 2018 to compare commercial AI accelerators based on peak computing performance and power consumption.
The latest survey examines more than 120 accelerators, including CPUs, GPUs, application-specific integrated circuits, and other specialized architectures.
The research tracks advances in chip design, processing efficiency, and hardware architecture to help identify suitable technologies for computationally intensive applications.
The team publishes its findings and datasets publicly to support research planning, technology evaluation, and government acquisition decisions.
Source: MIT News
AI data center land purchases reach $6 billion as access to power drives prices higher
U.S. purchases of land for future data centers reached approximately USD $6B in the first half of 2026, a 79% increase year over year, according to Avison Young.
Rising demand for AI computing capacity has intensified competition for sites with reliable electricity access, with powered land in Northern Virginia selling for roughly USD $6M per acre compared with USD $4M for land without power.
A separate analysis of 80 land transactions found that committed utility capacity was the strongest predictor of pricing.
Developers are increasingly prioritizing power availability, while utilities seek financial commitments to distinguish viable projects from speculative demand.
Source: Tech Target
Foundations is written by the Axented team.
Axented works with companies around the world to design, build, and scale digital products, teams, and AI-driven systems.
If you’re building or scaling a technology-driven business and want to explore how these signals apply to your context, you can learn more about how we work at axented.com.
