PodBrowser
Last Week in AI

#233 - Moltbot, Genie 3, Qwen3-Max-Thinking

Friday, 6 February 2026 · 3 min read · Listen to the episode ↗

The podcast discusses major advancements in AI, including Google’s Genie 3 for interactive video creation and Qwen Free Max Thinking, a model excelling in reasoning capabilities. It highlights Open Claw, an open-source tool raising user privacy concerns, and the increasing competition between OpenAI's ChatGPT translator and Google in the translation market. The intersection of AI with potential national security implications and the evolving landscape of funding for AI technologies, particularly in specialized chips, is also explored.

Andrey Kurenkov and Jeremy Harris discuss significant AI developments, emphasizing open-source releases and research. They highlight Google's new auto-browse feature in Chrome, powered by Gemini, which can perform multi-step tasks, raising questions about user adoption and practical applications. The episode introduces Open Claw, an open-source tool that allows user interaction via messaging platforms, prompting discussions about user comfort with extensive permissions and the risks associated with such tools.

Advancements in AI interactions are explored, including models debugging issues and the limitations of context windows compared to human cognition. The episode critiques Yan LeCun's views on internet access for models, noting the reality of misaligned models being granted such access. Genie 3, an interactive video generation demo from Google, is introduced, showcasing the ability to create game environments, though it is acknowledged as an early experimental release.

The podcast highlights the impressive progress in AI-generated content, with 60 seconds of coherent video output representing a significant improvement. OpenAI's upcoming ChatGPT translator, offering around 50 language options, is positioned against Google in the translation market. OpenAI's new workspace, Prism, integrates GPT-5.2 to assist scientists with research claims, reflecting a growing interest in using AI for scientific proofs.

The episode discusses a startup named Recursive, valued at $4 billion, which is developing AI-specialized chips. Another startup, also named Recursive and co-founded by Richard Socher, focuses on self-improving AI agents. Concerns about U.com's dual leadership are raised, as it aims for a recursive self-improvement model emphasizing safety.

Flapping Airplanes, a new lab with $180 million in seed funding, advocates for a nature-inspired approach to AI development. The podcast touches on the implications of insights becoming national security secrets, with discussions on data center security and the complexities of tradecraft and espionage.

Funding news includes New Rofos raising $110 million for optical processors aimed at enhancing AI inference efficiency. The challenges of optical processing and meta surfaces for applications like matrix multiplication are discussed. The episode also covers **Qwen Free Max Thinking**, a large model optimized for reasoning, outperforming other models on major benchmarks, and **KimiK 2.5**, optimized for coding.

The podcast addresses multimodal model development, focusing on integrating visual and text modalities, and the potential for creating natively visual interfaces for natural computer interaction. The new flagship model, Kimi 2.5, is open source, while the Qwen Free model is not, indicating a trend towards closed-source models.

AI2V is launching open-source projects, including smaller models tailored for specific code repositories. RCAI introduces Trinity, a 400 billion parameter open-source model, competing with other models and employing a globally distributed training infrastructure. The discussion includes challenges in the attention mechanism and strategies for expert assignment in a mixture of experts model.

Research advancements include a paper on post layer norm architectures and the importance of retaining core information across layers. The conversation transitions to continual learning, proposing a teacher-student model approach to retain knowledge while acquiring new skills. An asymmetric teacher-student meta-reinforcement learning framework is introduced, addressing the quantification of the teacher's contribution to the student's progress.

The political landscape in the U.S. is explored, particularly regarding immigration enforcement and its impact on the AI community. Concerns about research funding and the stability of the tech sector are raised, alongside the growing politicization of AI-related risks. The intertwining of politics and AI is expected to intensify, with implications for data centers and government power, reflecting a shift in the industry's approach to social issues.

This summary was generated from the episode transcript and can contain mistakes.