TLDR AI 2025-12-03

Hear from the creators and curators of MCP at OpenAI and Anthropic, alongside product and engineering leaders from Google, Microsoft, Datadog, Supabase, and WorkOS at MCP Night: The Holiday Special. The event brings the MCP community together at the iconic Regency Ballroom in San Francisco for an evening of learning, networking, and celebration.You'll get an inside look at how MCP is being used in real products today and the direction the protocol is moving next.Register for MCP Night →

OpenAI will delay advertising plans and other products to focus on improving ChatGPT. Google's release of Gemini 3 was very well received, with the service rapidly gaining popularity. Google is quickly catching up to ChatGPT in user numbers. ChatGPT has more than 800 million weekly users - Gemini had 650 million in October.

Mistral announced the release of Mistral 3, which features three new dense models (14B, 8B, and 3B) and Mistral Large 3, a sparse MoE model with 41B active and 675B total parameters. All models are open-sourced under Apache 2.0.

Anthropic is considering an IPO while also exploring private funding. It has reportedly engaged with a law firm and major banks. The IPO could be one of the largest IPOs ever. No decision has been made yet on the timing of the IPO.

Anthropic surveyed 132 of its own engineers and found 27% of Claude-assisted work wouldn't have happened otherwise, but not without costs. Mentorship is declining as Claude is becoming the first stop for questions. Engineers worry that the skills needed to supervise AI are the same ones atrophying from overuse.

The driver for further improvement will be continual learning. Solving continual learning will require incremental improvements, just like solving in-context learning. Labs will likely release something next year that they will call continual learning, but it will just be progress towards continual learning. Human-level continual learning may take another five to ten years of further progress.

There's deja vu, and now there's dejAI vu. Why does every AI tool feel the same? Because they've been designed to generate text, not to emotionally engage your target audience. Opal hits different – it learns your brand, plugs into your workflows and helps your team show up with work that actually lands. Meet Opal

AI coding systems can introduce severe bugs and vulnerabilities, so it is important to check their work. As models become stronger generators, their ability to verify, critique, and support human judgment must scale with them. OpenAI's agentic code reviewer is optimized for low safety tax and high precision to earn user trust. The company's testing shows that the code reviewer can deliver reliable, high-signal feedback without slowing teams down.

RAPTOR (Recursive Autonomous Penetration Testing and Observation Robot) is an autonomous offensive/defensive security research framework based on Claude Code. It combines traditional security tools with agentic automation and analysis to deeply understand code, prove exploitability, and propose patches. The tool empowers security research with agentic workflows and automation. RAPTOR was vibe-coded and is still in an early release stage.

Amazon Nova Forge is a service for building frontier models using Nova. It is the easiest and most cost-effective way to build your own frontier model. Customers can start their development from early model checkpoints, blend their datasets with Amazon Nova-curated training data, and host their custom models securely on AWS. The service is designed for organizations with access to proprietary or industry-specific data who want to build AI that truly understands their domain.

BrowseSafe is a real-time content detection model and benchmark suite designed to protect AI agents from prompt injection in web browsers. The open-source BrowseSafe and BrowseSafe-Bench enable developers to scan HTML for hidden malicious instructions without slowing down performance, offering layered defenses for safer agentic browsing.

Ricursive, a startup founded by ex-Google researchers, aims to build software that can automate the design of cutting-edge chips. This could allow every company to build their own chips from scratch. The company expects to release its first product next year. It recently raised $35 million and is currently valued at $750 million.

Amp is spinning out of Sourcegraph to become an independent research lab. Its goal is to let software builders harness the full power of AI. The company believes that the way we develop software will be challenged fundamentally and drastically. Spinning out into its own company will give it more freedom to focus on the frontier.

75% of leaders believe that AI is overly focused on individual productivity, creating siloes and hurting collaboration. What's the alternative? Get a free copy of the report from Miro

Humanity will have to decide by 2030 whether to risk letting artificial intelligence systems train themselves to become more powerful.

OpenAI's Memory search feature for ChatGPT lets users quickly find stored info via direct queries.

Anthropic announced it has acquired Bun, a fast JavaScript runtime.

Opera has expanded its AI capabilities with Google's Gemini models across its browsers.

The creative process is limited by human scale, but we're getting close to leaving that world.