The week in startup news: Sep 21-27

1. A ChatGPT co-inventor built a new kind of AI model

Diogo Almeida helped build ChatGPT and later invented RLHF (reinforcement learning from human feedback), the training method that made today's chatbots possible. He left OpenAI two years ago because he felt AI models were too good at human language and not good enough at talking to computers, which is what real automation needs. His new AI model is getting developers excited because it aims to fix exactly that gap.

2. Mistral raised one of Europe's biggest rounds ever

French AI company Mistral confirmed a 3 billion euro funding round at a valuation of more than 21 billion euros, one of the largest ever for a European tech company. Samsung is reported to be among the backers, a sign that big global money is willing to bet heavily on European AI teams, not just US ones.

3. A firm that builds startups for other companies raised $100 million

Vantora, formerly called UP.Labs, doesn't invest in startups from the outside. It builds them from scratch to solve problems for corporate customers like Alaska Airlines and Porsche. It just raised $100 million from Silversmith Capital Partners and is now shifting more attention to physical AI, meaning AI that controls robots and hardware, not just software.

4. A new solo-GP fund is backing early AI and deep-tech startups

A solo GP (a fund run by one person making the investment calls, rather than a big committee) named Shahryar Barati launched Project Ventures. It just closed 1.2 million pounds to invest in pre-seed and seed stage startups (the earliest, smallest funding rounds) working in AI and deep tech, with strong ties to Imperial College London.

5. ByteDance gave away its AI training system for free

ByteDance Seed and Tsinghua AIR released DAPO as open source (free for anyone to use or change), including the training method, the code, and the dataset. The system hit over 50% on AIME 2024, a difficult math competition test used to measure how well an AI reasons. Anyone can now download it and build on it.

6. Ottonomy is building robots for the part of logistics nobody automated

Founder Ritukar Vijay noticed that moving things between buildings, through elevators, and across campuses, the messy middle of logistics, stayed manual even as robots took over warehouses. He teamed up with three co-founders to build Ottonomy, and the COVID-19 pandemic made the need for contactless delivery of goods and supplies urgent almost overnight.

7. Anthropic merged Claude's chat and work modes into one screen

Anthropic combined Claude's regular chat, its Cowork feature (a mode built for doing multi-step tasks), Artifacts (an interactive workspace), and its design tools into a single interface. Instead of forcing people to pick the right tab, Claude now automatically routes each request to the right tool.

8. Security researchers hacked into OpenAI using Anthropic's Claude

A three-person team at security startup Hacktron AI used Claude to chain together two security flaws and break into OpenAI employee ChatGPT accounts, as part of an official bug-bounty program (a program that pays people for finding security holes). OpenAI paid them $6,500 and says it has fixed the issues.

9. India's startup funding stayed under $100 million, but chips are booming

India's weekly venture funding fell below $100 million again. At the same time, SEMICON India 2026 wrapped up with roughly 1 lakh crore rupees (a large sum in Indian numbering) in investment commitments, and a new report expects India's chip market to grow from $64 billion in 2026 to $200 billion by 2035.

Why this matters if you're building

This week shows two things at once: money for AI at the top end is enormous, from Mistral's multi-billion-euro round to Vantora's $100 million, while early-stage funding in places like India is still tight. If you're pre-deck, the open-source releases (DAPO) and interface changes (Claude, Anthropic) matter more than the mega-rounds, because they're tools you can use or learn from today. And the OpenAI hack is a reminder that if you're wiring AI tools into your product, your security surface just got bigger, not smaller.