🔡🔡🔡 [🤑 Today's AI News 🤑] 🔹Yann LeCun's Comeback! AMI Labs, founded by former Meta Chief AI Scientist Yann LeCun after leaving the company, has raised $1.03 billion in a seed round. The pre-money valuation is $3.5 billion, and investors include Bezos Expeditions, NVIDIA, Samsung, and Toyota Ventures. The initial target was €500 million, but due to overwhelming demand, the funding eventually closed at €890 million. AMI currently researches "world models" based on the JEPA program, rather than the LLM program. This demonstrates LeCun's commitment to his philosophy that "LLMs are not true intelligence." With offices in Paris, New York, Montreal, and Singapore, they plan to release their code and papers as open source. This is noteworthy as a serious European endeavor, proclaiming itself as a "frontier AI lab, neither Chinese nor American." 🔹OpenAI Acquires AI Security Startup Promptfoo Promptfoo is an AI red team and vulnerability assessment tool used by over 25% of Fortune 500 companies, and is a fairly influential project with over 350,000 developers in the open source community. After the acquisition, the project will be integrated into OpenAI Frontier, and the open source project will retain its current license. As AI agents autonomously browse the web, execute code, and make API calls, security issues such as prompt injection and data leakage are becoming increasingly important. OpenAI appears to be intent on incorporating these issues into its platform. 🔹Anthropic files lawsuit against the Pentagon The Trump administration's DOD blacklisted Anthropic as a "supply chain risk," ostensibly because Anthropic demanded that Claude not be used for autonomous weapons or domestic mass surveillance. Interestingly, this dispute backfired. When the news broke, the general public downloaded the app in droves and subscribed, making Claude suddenly much more popular. It was also revealed that Claude Opus 4.6, in collaboration with Mozilla, discovered 22 security vulnerabilities in Firefox, 14 of which were high-risk. This appears to be a good example of AI's practical effectiveness in large-scale vulnerability analysis. 🔹Andrej Karpathy Open-Sources Autoresearch This 630-line Python tool allows AI agents to autonomously run machine learning experiments on a single GPU. It can reportedly run about 100 experiments overnight. In fact, after running it on a depth=12 model for two days, it independently identified about 20 improvements, and applying these improved the nanochat "Time to GPT-2" metric by 11%. It reached 8,000 GitHub stars within a few days of its release, and I think this in itself is an interesting example of the potential of the concept of "autonomous research." 🔹Google Releases Gemini 3.1 Flash-Lite Flash-Lite is very affordable, with an input cost of $0.25/1M tokens, while its response time is 2.5 times faster than Flash 2.5 and its output speed is 45% faster. It reportedly outperformed GPT-5 mini and Claude 4.5 Haiku in six benchmarks. There's also talk on Twitter that "Google is planning additional launches this week." While no official confirmation has been made, it's possible that Gemini 3.1 Ultra or a similar model could be released. Watching today's news, I strongly feel that the AI industry is shifting its focus from a "model performance competition" to a "agent infrastructure + security competition." OpenAI's acquisition of Promptfoo is a symbolic signal of this, and it seems that Anthropic, Google, and OpenAI are all restructuring their competitive landscapes around how to reliably provide agent platforms to enterprise customers. Yann LeCun's AMI Labs is while everyone else is competing for the LLM, but he's going in a completely different direction. It's a high-risk, high-reward bet that could shake things up if successful, or it could be a "LLM was the right choice" if it fails. So, I'm very curious about the future.