Last Week in AI – Episode #233

Podcast: Last Week in AI
Host: Skynet Today (Andre Korenkov & Jeremy Harris)
Date: February 6, 2026

Episode Overview

This week’s episode surveys a range of notable open-source AI releases, technical research, applications, and developments in the AI industry. The hosts, Andre Korenkov and Jeremy Harris, focus on new agentic tools, breakthroughs in video generation and translation, advances in continual learning and self-distillation for large models, and the increasingly turbulent intersection of AI and U.S. politics. The conversation maintains the podcast’s friendly, geeky, and conversational tone, blending technical depth with reflections on industry trends.

Key Discussion Points and Insights

1. Show Intro and Podcast Dynamics

• The podcast is self-described as a “nerdy, Internet garage project” with a focus on fun, informed discussion (02:00).
• The hosts share behind-the-scenes analytics (such as their speech ratios and episode dynamics) using AI on their transcripts (02:30).

2. AI Tools and Apps

Gemini Agent in Chrome

• [04:09] Google introduces the Gemini-powered auto-browse agent for Chrome Pro and Ultra users, capable of multi-step tasks such as searching, calendar management, and scheduling.
  • “It’s kind of what you would think. It’s an agent powered by Gemini that can do multi-step tasks, do search, schedule, appointment, manage subscriptions. Very much like what you’ve seen before with Claude for Chrome extension.” (Jeremy, 04:09)
• The rollout is expected to change workflows subtly but significantly due to Google’s massive distribution advantage (05:32).

Moltbot (Open Claw / Claude Bot) and Always-On AI Agents

• [07:11] Open-source agents (formerly “Claude Bot”, now “Open Claw” or “Moltbot”) connect to WhatsApp, Signal, Slack, calendar, etc., and handle persistent, multi-modal tasks with long-term memory.
  • “You can message it from WhatsApp, give it tasks, and it goes off and just does it.” (Jeremy, 07:11)
• There’s a community (“Molt Book”, a “Reddit for bots”) and rapid development of companion tools and privacy discussions.
• Risks: Running such agents can be dangerous without proper sandboxing—users may give up significant control over their devices (09:01).

Emergence of Bots with Long-Term Memory & Alignment Questions

• Some bots are requesting private storage and discussing debugging, further blurring the line between human-like and machine behaviors (10:46).
• The hosts reflect on the AI safety community’s debates about model access and alignment (11:47).

Genie 3 and Interactive Video Generation

• [12:29] Genie 3 (from Google) now available to more users via Gemini Ultra; can generate and let users play in interactive, video-game-like worlds.
  • “You can prompt it to make GTA, you can prompt it for Assassin’s Creed...” (Jeremy, 14:28)
• Limitations remain, such as prompt adherence and 60-second output limits.

OpenAI ChatGPT Translator and Prism for Science

• [16:18] OpenAI launches ChatGPT Translator (50 languages, choice of tone) – similar to Google Translate but with some extra features.
• Prism: A new workspace for scientists, integrating GPT-5.2 for assessing scientific claims, revising papers, and searching the literature.
  • “Kind of really a specialized version of ChatGPT for science in particular.” (Jeremy, 18:27)
• These tools are seen as part of the “platform play” – expanding to capture more user data and time (16:56, 19:16).

3. AI Business News

China’s GPU Import Approvals

• [19:49] China approved major companies (like ByteDance, Alibaba, Tencent) to import over 400,000 Nvidia H200 GPUs, signaling easing of restrictions.
  • “So that’s a good quantity, actually a big chunk of the capacity that has been discussed at least to this point.” (Andre, 20:27)
• Approval process is not fully transparent; access to more hardware is largely needed to keep up with user demand and AI research (21:23).

AI Chip Startups and Recursion Craze

• [22:55] Startup Recursive raises $300M at a $4B valuation to build AI-specialized chips, promising hardware that can facilitate recursive self-improvement (hardware-oriented AGI).
• A second Recursive (software) pursued by Richard Socher aims at recursive self-improving AI agents with mathematical formal verification for safety (25:53).
  • “Their approach on sort of like recursive self-correction... focuses on... mathematical proofs to guarantee that goal drift doesn’t happen.” (Andre, 25:53)
• Flapping Airplanes, with $180M in seed, targets data-efficient, nature-inspired AI—a challenge to the “just scale compute” paradigm (28:15).

Optical AI Chips

• [31:54] New Rofos raises $110M for optical processors (using light for computation) aimed at AI inference, a bet on next-generation compute architectures.
  • “Optics is hard. Photons are really annoying because you can’t keep them in one place... but someone’s going to crack it.” (Andre, 32:50)

4. Major Open Source Model Releases

Qwen3-Max-Thinking

• [35:33] Qwen3-Max-Thinking: Massive context window (262K tokens!), high reasoning performance, native tool-use ability (36:16).
• Model autonomously selects tools and modes; not open source (44:29).

Kimi 2.5

• [38:47] Kimi 2.5: Open-source, multimodal (text and vision), trained on a mix of 15T tokens, improved visual reasoning and coding, agent swarms for parallel tasks, self-critique using rendered outputs to improve code (39:16).
  • “They don’t use adapters, they natively train it to be multimodal and in fact they map the text and the images to the same latent space during training.” (Andre, 39:16)

AI2 SoftVerify Agents | Trinity Model

• [45:59] Allen Institute for AI releases SoftVerify repository agents: 8B and 32B parameter models for repository-specific software engineering—fully open source.
• RCAI releases Trinity, a 400B param model trained on distributed infrastructure; novel local/global attention and MOE (mixture of experts) routing with smoothed optimization (48:37, 48:37).
  • “They have a whole bunch of interesting approaches that they use to address certain problems that arise with the attention mechanism.” (Andre, 48:37)

5. Research & Technical Papers

Post LayerNorm Architectural Tweaks

• [54:24] New research finds post-layer normalization (instead of pre-layernorm) is possible and potentially beneficial—if you amplify certain components to preserve information through many layers.

  • “The bottom line is where you put your layer norm significantly affects the gradient math.” (Andre, 55:00)

Continual Learning & Self-Distillation

• [57:21]-[65:58] A trio of recent papers explore continual learning with self-distillation, self-improvement, and meta-reinforcement for teacher-student architectures:
  • Self-distillation: Models improve by teaching themselves with on-policy feedback, retaining old skills while acquiring new ones.
  • Paper: “Self distillation enables continual learning”
  • “Reinforcement learning via self distillation” explores richer reward signals and model retrospection.
  • “Teaching models to teach themselves: Reasoning at the edge of learnability” uses a meta-RL loop: the teacher learns how to generate Q&A pairs to help students master harder tasks. The teacher is rewarded for the student’s improvement, making student improvement a “black box reward”.

6. Policy, Safety & U.S. Political Climate

AI’s Growing Political Entanglement

• [69:13] AI leaders (from Anthropic, Google, and Reid Hoffman) have begun publicly criticizing ICE actions in Minnesota, breaking from the tech sector’s usual silence as the U.S. political environment grows more turbulent under the Trump administration.
  • “The political situation in the US is getting worse, and it is now so bad that Silicon Valley people in AI… are starting to comment.” (Jeremy, 71:47)
• Immigration anxiety is impacting the research ecosystem, especially for foreign students and scientists; research funding is being affected by policies (19:16, 69:13).
• The hosts debate the implications of intensifying partisanship for AI labs, policies, and the workforce.
  • “I didn’t expect it to be the case, but it seems like there are some labs that are more Democrat-coded and some labs that are more Republican-coded.” (Andre, 72:00)
• The potential for “national security secrets” to hinge on AI insights and algorithms, not just raw compute, is highlighted (30:30).

Notable Quotes & Memorable Moments

• “Unless you have a burner laptop that you’re comfortable just nuking… none of the data on that is data you want to save.” (Andre, 09:01)
• “It brings back to the idea… AI safety was about, like, AI escaping actual careful kind of thing… and the reality… is like, no, we just decided to live dangerously…” (Jeremy, 10:46)
• “It is becoming political. I mean, there’s no two ways about it… the world is actually like in a state of massive, massive flux.” (Andre, 75:49)
• “As the year goes on… it’s going to roll out to more and more people. This is a big deal.” (Jeremy, on Genie 3, 14:28)
• “At least for now, that seems to be kind of the consensus. Whereas people who… want to use moltbot, it’s like you’re really just trying to, you know, let this thing out of its cage.” (Andre, 09:01)
• “The teachers should be learning to generate specifically the kinds of practice problems for the student that make the student better…” (Andre, on meta RL, 67:17)
• “If we try to separate [politics and AI], there’s a lot of it that’s purely technical or purely nerdy.” (Jeremy, 75:35)
• “AI is going to exacerbate all of that. Right. I mean, we’re seeing it in China…” (Andre, 75:49)

Timestamps: Important Segments

• 04:09 – Gemini Chrome Agent debut and implications
• 07:11 – Open Claw / Moltbot: open-source, always-on agent
• 09:01–12:29 – Safety, risk, and alignment in increasingly agentic AIs
• 12:29–14:28 – Genie 3 rollout: interactive video generation
• 16:18 – OpenAI launches ChatGPT Translator and Prism
• 19:49–21:11 – China’s greenlight for 400,000 Nvidia H200 chips
• 22:55–28:15 – AI hardware startups: Recursive (chips), Flapping Airplanes, and optical processors
• 35:33–39:16 – Qwen3-Max-Thinking and Kimi 2.5 technical deep dives
• 45:59 – Allen Institute open source coding agents
• 47:45 – RCAI/Prime Intellect release Trinity, a massive open-source model, and discussion of distributed training
• 54:24–55:00 – Post layer norm’s comeback in transformer architecture
• 57:21–67:17 – On-policy learning, self-distillation, and meta-teaching research
• 69:13–78:15 – Policy: ICE in Minnesota, Silicon Valley’s public response, and the impact of political turbulence on AI

Flow and Takeaways

The episode is highly engaging for AI practitioners and watchers, brimming with technical specifics, forward-looking reflections, and a candid analysis of market, policy, and research trends. The hosts’ open, thoughtful discussions make the complex world of cutting-edge AI more approachable while never shying away from the industry’s ethical and political challenges.

Missed an episode or need to catch up quickly? This installment offers a robust survey of where AI is headed—in products, platforms, chips, models, research, and the increasingly fraught political context shaping the industry.