Last Week in AI - Episode #225 Summary

Podcast: Last Week in AI
Hosts: Andrei Karenkov (A), Michelle Lee (B - guest)
Episode Date: November 21, 2025
Main Theme:
A roundup of the latest in AI news—model releases, business developments, research papers, legal issues, and cultural trends—with a focus on major global players and tangible impacts across tech, business, and society.

1. Introduction & Episode Overview

• This episode delivered a relatively fast-paced summary of the latest in AI, spotlighting model releases (GPT-5.1, Ernie 5.0, Kimi K2 Thinking), business shifts (data center arms race, major fundraises), legal battles (copyright decisions), and synthetic media developments.
• "There's not been any like crazy big news in the world of AI...but there is a smattering of different things going on as usual." (B, 01:17)

2. AI Tools & Model Releases

OpenAI’s GPT-5.1

• Highlights:
  • Two variants: GPT-5.1 Instant (warmer, more intelligent persona) and GPT-5.1 Thinking (faster on simple tasks, slower but deeper on complex ones).
  • New personality presets: friendly, quirky, nerdy, cynical, and more, giving users more options for interaction style.
  • Change appears in response to complaints about earlier models lacking the "feel" of GPT-4o.
• User Experiences:
  • Michelle shares how using ChatGPT as an interactive teacher—e.g., for learning poker—benefits from selectable personalities.

    "I do sometimes find some of the preset personality is pretty annoying if I'm talking to it." (A, 04:25)

• Timestamps: [03:00–04:49]

Baidu’s Ernie 5.0 & AI Services

• Key Points:
  • Released at Baidu World 2025, now part of Ernie Bot for enterprise.
  • Reception lukewarm; Baidu stock dropped 10%, seen as incremental.
  • Baidu’s Apollo Go robotaxi service now operational in 22 cities, vastly outpacing Western counterparts in deployment scale.
  • "If you look at like Waymo and Robotaxi and now even Zoox, like they are really operating at...fewer cities compared to Apollo Go." (A, 06:23)
• Timestamps: [04:49–06:49]

ByteDance’s Coding Agent

• Features:
  • Volcano engine's Dǒubāo Seedcode competes with cloudcode and Cursor. Main draw is the low price point; capabilities "pretty good though not top-tier." (B, 07:12)
• Timestamps: [06:49–08:04]

3. Business, Infrastructure, and Investment

Data Center Arms Race

• Anthropic’s $50 Billion Data Center Partnership:
  • Partnership with UK-based NeoCloud and FluidStack to build new US facilities.
  • Context: Meta announced a $600 billion initiative; OpenAI/Oracle’s 'Stargate' project at $500 billion.
  • Anthropic's investment seen as a move toward autonomy from Amazon/Google, projecting $70 billion in revenue by 2028.

    "This is also Anthropic's first investment in custom infrastructure." (B, 17:26)

Chinese AI Chips & Hardware

• Baidu’s Next-Gen Chips:
  • M100 for inference (tailored for mixture-of-experts models); M300 for training, targeting "multi-trillion parameter models" by 2027.
  • Addresses U.S.-imposed hardware restrictions (Nvidia GPUs).
  • China betting on local chip solutions, optimizing for efficiency due to persistent export controls.
• Timestamps: [16:01–20:50]

Startup Fundraising and Notables

• Gamma:
  • AI-powered presentation creator, $2.1B valuation, 70M users, 50 employees, $100M ARR.
  • "It's pretty impressive that they only have 50 people in the company and they've already reached 100 million ARR." (A, 28:49)
• Inception:
  • $50M raise to develop diffusion models for code and text; led by Stefano Ermon (Stanford).
• Cursor:
  • $2.3B raise (valuation now $29.3B), as it expands model training in-house to compete with Anthropic/OpenAI’s own code tools.
  • "For Cursor to only rely on external AI models...puts them at existential risk." (A, 32:09)
• Timestamps: [27:33–33:12]

Self-Driving Cars

• Baidu’s Apollo Go:
  • 250K weekly rides (matching Waymo), expanding internationally; far ahead of US-based services in operational cities.
• Pony AI:
  • Raised $863M in HK IPO, aiming for profitability by 2028/2029.
• Timestamps: [33:12–36:16]

4. Research, Open Source, and Technical Advances

Kimi K2 Thinking

• "Most exciting" open-source release from China:
  • Optimized for tool use and coding problems, with 1T parameters and remarkable tool-call sequencing (200-300 sequential calls).
  • Modified MIT license: free for use, but with display requirements for large-scale commercial applications.
  • "It's really exciting that not only is it beating all the other open source models, it's even matching or surpassing the models from the large Frontier Labs." (A, 38:21)
• Timestamps: [36:16–38:21]

Remote Labor Index (Research)

• Overview:
  • A benchmark to measure AI’s ability to automate tasks on freelance platforms (e.g. Upwork).
  • Findings: Best current AI only achieves 2.5% automation.
    • Key challenges remain: context retention and long-horizon task management.
  • "Current AI agents are quite limited...they're still very far from human performance in remote labor tasks." (B, 41:00)
  • Fun failures: odd spelling errors, botched 3D designs ("a diamond ring is just like a ring drawn...with another oval-shaped thing," A, 43:31)
• Notable Moment:
  • Referenced paper's Appendix (C6, p.27) for entertaining screenshots.
• Timestamps: [38:21–44:20]

OpenAI: Interpretable Sparse Transformers

• Research Focus:
  • Rather than retrofitting interpretability into dense transformers, train sparse transformers from scratch to get clearer, more "readable" circuit activations.
  • Trade-off: promise of better transparency vs. increased training/deployment inefficiency; models only at GPT-2 scale so far.
  • "Training sparse models are just a lot more inefficient to train and deploy..." (A, 48:02)
• Timestamps: [44:20–49:04]

KimiLinear: Efficient Attention Alternative

• Claim:
  • Linear attention outperforms full attention on some scaling benchmarks, with up to 75% memory savings.
  • Still smallish scale (48B parameters), hybrid architecture, and caveats re: large scale generalizability.
  • "It's been an open question in my mind for years whether we transcend the standard transformer architecture..." (B, 52:54)
• Timestamps: [49:04–52:54]

DeepMind’s SIMA2 World Agent

• Demonstrated Capabilities:
  • Multi-game, multi-world AI agent capable of carrying out open-ended tasks given text prompts (e.g., Minecraft, No Man’s Sky, Genie 3-generated worlds).
  • Trains as a platform for possible sim-to-real skill transfer for embodied physical AI.
  • "If you train on enough data, can you then do sim-to-real for this like physical AI agent...?" (A, 55:53)
• Timestamps: [52:54–56:53]

Preprint Spam on Arxiv

• Change:
  • Due to LLM-generated spam, arXiv will no longer accept non-peer-reviewed computer science review and position papers by default.
  • "With how good LLMs are these days, arguably we don't need review papers anymore." (A, 58:39)
• Timestamps: [56:53–59:04]

5. Legal & Policy Developments

• Stability AI vs. Getty Images (UK):
  • UK court sided with Stability AI on copyright (Stable Diffusion did not infringe, as it doesn’t store/reproduce works), but found in favor of Getty on limited trademark infringement (watermarks).
  • Points to a trend favoring “fair use” for training, but unsettled in other jurisdictions.
• OpenAI vs. Gemma (Germany):
  • German court found OpenAI liable for copyright infringement (training on lyrics).
  • "Still it's a whole big mess...we are starting to get some initial legal results. Finally." (B, 63:32)
• Microsoft & UAE Agreement:
  • Microsoft investing $15B+ in UAE infrastructure, including US-approved shipment of Nvidia GPUs.
  • UAE positioning as regional AI hub, part of broader global political and business strategy.

6. Synthetic Media, Art & AI Culture

• AI-Generated Music Makes Billboard

  • "Walk My Walk" by Breaking Rust (AI country artist) hits Billboard’s country digital song sales chart (#1) and 1.8M Spotify listeners.
  • Michelle: "Walk My Walk is actually a banger. It's actually quite good." (A, 67:33)
  • Discussion of debate: Is AI music a genre, a tool, or an existential threat to musicians?
  • "I think it's just another tool, personally...When new genres have emerged, especially that utilizes a lot of technology, they've always in the past been slammed..." (A, 69:38)

• AI Singer’s Chart Debut

  • Gazania Monet (created with Suno by poet Talisha Nikki Jones) debuts on Billboard radio chart, signs multimillion-dollar deal.
  • Brings up the idea of "fictional" or "virtual" artists, parallels to Gorillaz, Hatsune Miku, etc.

• ElevenLabs Famous Voice Marketplace

  • Brands can license AI-generated voices (real/historic), e.g. Michael Caine, Mark Twain, Judy Garland, and more.
  • Raises ethical and legal questions about consent, posthumous rights, and brand use.
  • "There's actually a lot of historical characters...I'm not sure how you could get the voice of Mark Twain." (A, 74:29)

7. Notable Quotes & Moments

• Chatbot Personae Matter:

  "I can definitely see a lot of people just like chatting with these chatbots...having these friendly, quirky, efficient, cynical...personalities makes a lot of sense." (B, 03:30)

• Open Model Transparency:

  "Not only is it beating all the other open source models, it's even matching or surpassing the models from the large Frontier Labs." (A, 38:21)

• On AGI & Economic Tasks:

  "Current AI agents are quite limited...they're still very far from human performance in remote labor tasks." (B, 41:00)

• On AI-Created Music:

  "Walk My Walk is actually a banger. It's actually quite good." (A, 67:33)
  “I think it's just another tool, personally...when new genres have emerged, ... they've always in the past been slammed for somehow not being creative or utilizing technology too much." (A, 69:38)

8. Episode Flow & Tone

• Conversational, lightly humorous, occasionally nerdy (e.g., EM dash digression at [13:00–15:59]).
• Focus on both high-level trends and practical impacts—useful for professionals as well as AI-savvy laypeople.
• Skipped over ads/housekeeping—straight to substantive updates and engaging back-and-forth.

9. Quick Segment Timeline

• [00:11–03:00] Opening/newsletter plugs, host intros
• [03:00–10:41] Tools and apps (GPT-5.1, Ernie 5, ByteDance, Google AI shopping)
• [10:41–15:59] World models (World Labs Marble), quirky OpenAI news (EM dashes)
• [16:01–27:33] Business: Data centers, Baidu chips, startup moves, Amazon vs. Perplexity
• [27:33–36:16] Startup fundraising, self-driving cars
• [36:16–44:20] Open source: Kimi K2 Thinking, academic research (Remote Labor Index)
• [44:20–55:53] Research: Interpretable transformers, KimiLinear, DeepMind SIMA2
• [56:53–61:13] Policy, law: Arxiv spam, AI copyright cases
• [61:13–65:40] Microsoft-UAE data center deal
• [65:40–76:00+] Synthetic media, AI in art/music, ElevenLabs voice marketplace

10. Closing Thoughts

This episode underscores the relentless pace and global scope of AI advancement, balancing technical achievements (model performance, hardware, architecture) with real-world business, culture, and regulatory ramifications. Open models, international competition, and the friction between legacy and emerging creative forms illustrate a field in flux—where what’s cutting-edge one month feels routine the next.

Memorable Outro:

"Just AI all the way down." (A, 73:32)