Last Week in AI — Episode #224

Date: November 5, 2025
Hosts: Andre Korenkov & Gavin Purcell
Main Topics: OpenAI restructuring, coding AI tools, Grokipedia, Minimax M2, Udio copyright, AI regulation, and the broader economic/social impact of AI

Episode Overview

This episode delivers a concise round-up of notable developments from the last two weeks in the AI world. Hosts Andre and returning guest Gavin kick off with updates on AI coding tools and applications, then dive into significant industry business maneuvers—notably, OpenAI’s final transition to a for-profit corporation. The discussion covers new open-source releases, AI infrastructure investments, regulation and copyright battles, and societal implications of AI’s rapid deployment. The show rounds out with reflections on synthetic media, ongoing legal wrangling, and what economic changes the AI wave may create.

Tools & Apps

Cursor 2.0 and AI Coding Revolution

[03:06–07:07]

• Cursor: Upgraded to 2.0; now features their own Composer model (in-house), designed for speed (200 tokens/sec) and competitive coding capabilities.
  • Cursor originated as an AI-native fork of popular IDEs and quickly rivaled Microsoft’s GitHub Copilot.
  • Uncertainty lingers around Composer’s state-of-the-art parity compared to models like Claude or OpenAI’s newest offerings.
  • Andre [05:14]:
    “Why, you know, the state of the art coding models, if they can do stuff faster and if they can do stuff better, is it not worth paying more to use those?... The composer model might be faster … but if you’re getting results that make you go back 3, 5, 10 times versus one of the state-of-the-art models... it kind of trades off.”
• Business dynamics: Cursor needs its own model to control margins, as relying on API calls to other providers was unsustainable.
• Competitive landscape: Intense, with cloud code, Codex, and Copilot all pushing rapidly.

Anthropic’s Claude Code and Coding Agent Convergence

[07:07–09:39]

• Anthropic’s Claude Code now has a web app interface, joining the shift from terminal-only to dual interfaces (web & terminal) seen in Copilot, Codex, Gemini, etc.
• Hosts debate usability and speculate this is an answer to the mass adoption of web-based coding agents.

Quick Hits on Copilot's New Avatar and Memory Features

[09:39–13:35]

• Microsoft Copilot’s “Meco”: A playful blob-like AI avatar—meant to foster emotional connection/user-friendliness, possibly appealing to a more general Windows user base.
  • Clippy-style easter eggs amuse the hosts.
• Anthropic Claude’s Memory: Toggle-able chat history recall, catching up to ChatGPT and Gemini in personalization and memory—important for user retention and “lock-in”.

Canva AI

[18:38–20:53]

• Canva’s Design Model: New AI features that understand layers and formats, allowing easy prompt-based creation and iteration—targeting pro-sumer and mainstream design needs.

Grokipedia: Grok-Generated Wikipedia

[20:53–24:06]

• Elon Musk’s XAI launches Grokipedia—a Wikipedia clone populated by Grok.
  • Over 700,000 articles; many closely mimic Wikipedia but slant towards “removing liberal bias” (e.g. on climate change, vaccines).
  • Concerns arise about “wisdom-of-the-crowds” being replaced by single-entity curation:
    • Andre [22:41]:
      “It’s disappointing in some ways … the beauty of Wikipedia always was … wisdom of crowds wise … as you get smaller data sets, you get worse results.”
  • Gavin notes:
    “Elon Musk has positioned Grok and XAI as being maximally truth seeking, as removing the liberal bias… So he is accomplishing that mission with this.” [23:51]

Business & Infrastructure

OpenAI’s Corporate Restructuring

[24:06–31:12]

• OpenAI finalizes shift: Now a Public Benefit Corporation (PBC), after untangling its non-profit/for-profit hybrid model; nonprofit becomes OpenAI Foundation.
  • Brings structure in line with Anthropic and XAI.
  • Nonprofit now holds equity (~$130B valuation).
  • Impact: Permits crucial new investments; aligns shareholder structure; reduces risk of another “board coup.”
  • Microsoft partnership updated: Microsoft’s stake drops from 32% to 20–27%; AGI “clause” now clarified by an expert panel; IP rights set through 2032.
  • Andre [27:12]:
    “AI companies like OpenAI are setting up for pretty difficult conversations with society ... gobbling up a lot of the work and money in the world ... So at least this thing is behind them now. 'Quote, unquote', they're for profit, but they're also trying to make it clear that ... we're here to make some difference in the world too.”

Chips & Compute

[31:12–36:35]

• Qualcomm unveils new AI data center chips (AI 200 & AI 2050) to rival Nvidia and AMD—direct response to exploded compute demand.
• Amazon’s Rainier: Announces over 1 million Trainium 2 chips for Anthropic by EOY; continuing tight-knit cloud/compute partnership.
• Google and Anthropic: $50B cloud deal, giving Anthropic access to 1M TPUs, adding a gigawatt of AI compute capacity by 2026.
  • Gavin [39:44]: “There’s just not enough compute for these companies, so they do whatever they can.”

International Expansion

[39:46–41:17]

• Google partners with India’s JIO: 18 months of free AI Pro access for young users to spur global AI adoption.

Open Source, Research & Model Releases

Minimax M2: The Next Open-Source Powerhouse

[41:17–43:52]

• Chinese firm Minimax launches M2—a 229B parameter mixture-of-experts model, runnable efficiently (only 10B parameters per inference).
  • Outperforms many peers on coding/agent benchmarks; fully open-source (MIT licensed).
  • Trend: Many top-tier open-source LLMs now coming from China (DeepSeek, GLM-4, Kimi K2).
  • Ethical/strategic concerns about this geo-pattern.

Research: Smarter, Smaller, More Efficient

• Scaling Latent Reasoning via Looped LMs: Recurrence (multi-pass through the network) enables smaller models (1–2B params) to rival much larger ones—holds promise for on-device AI.
  • Andre [47:16]:
    “I can't wait till you could get models that are really capable on your phone.... If you could have individual smaller models that could be local in a way that really are good, that's very exciting.”
• OpenAI’s GPT-OSS Safeguard:
  • Release of safety classification models for moderating AI outputs, co-developed with Discord and others.

LLM Understanding and Memory Research

[49:43–54:02]

• Continual Learning via Sparse Memory Fine-Tuning: New method allows LLMs to “learn on the fly” while avoiding catastrophic forgetting, potentially enabling persistent, adaptable AI personalities in future.
• Adaptive Vision Transformers: Vision models now adjust “patch sizes” to focus compute on image regions with most information.
• How Do LLMs Use Their Depth?: Early layers focus on common patterns, later layers yield complex reasoning. Pushback on “LLMs are black boxes”—researchers are making progress in interpretability.

Policy, Safety & Societal Considerations

US Govt. Invests in Supercomputers

[57:06–62:25]

• DOE and AMD forming $1B partnership to build AI-focused supercomputers for scientific/discovery initiatives.

AI & Teen Safety

[62:25–68:53]

• Character.AI bans users under 18 from talking to its bots after high-profile cases of harmful immersion and mental health concerns among teens.
  • Raises new debates about psychological, societal, and developmental consequences of ongoing/chat-based AI interaction.
  • Andre [66:02]:
    “It's very possible ... that 10 years from now there will be legitimate relationships that people have with these entities.... Society has to really see these stories and ... be like, hey, we are entering to a new social contract with things we don't really understand.”

Synthetic Media, Copyright & Lawsuits

Universal Music & Udio Settle

[70:24–73:04]

• Universal & Udio: UMG settles with Udio, the leading text-to-music generator, to license—but with a catch: Udio users can no longer download their creations; only streaming will be allowed going forward.
  • Users angry at retroactive restriction on owned/generated content; Udio will “restart” with only licensed data.
  • Andre [72:47]:
    “Users were told that they can no longer download their songs... That is a kind of a scary thing ... users are furious because ... they don't own the things that they thought they.”

Authors vs. OpenAI

[76:04–79:03]

• OpenAI fails to dismiss parts of a US authors’ lawsuit that AI-generated text infringes on copyrights; court rules plaintiffs can try to prove “substantial similarity.”
  • Hosts speculate OpenAI will eventually settle (following Anthropic’s lead).
  • Main unresolved issue: is generative AI output copyright infringement or not?
  • Andre [77:11]:
    “My gut is telling me eventually OpenAI will settle some version of this lawsuit. But I don’t know … eventually you would hope ... if OpenAI does ... a 1 trillion dollar valuation IPO. They should pay some of this money out because how else is the money going to get out, right?”

Economic & Social Impact Reflections

[79:03–85:04]

• The looming question: If AI companies “own" most economically valuable work, how will benefits reach the broader population?
  • Universal Basic Income seen as implausible in the current US climate.
  • Andre fears “winner-take-all” consolidation and societal displacement outweigh distant AGI catastrophe risk.
  • Gavin [80:15]:
    “...their goal is to own a large chunk of the economy, right? So absolutely, that is an open question.”
  • Wild unemployment swings, GDP boom, and the “cyber-corporation” scenario are all possibilities.

Notable Quotes & Timestamps

• On OpenAI’s shift:
  Andre [27:12]: "AI companies … are setting up for pretty difficult conversations with society … So at least this thing is behind them now … they're for profit, but they're also trying to make it clear that … we're here to make some difference in the world too."

• On Grokipedia:
  Andre [22:41]: “It’s disappointing in some ways … the beauty of Wikipedia always was … wisdom of crowds wise … as you get smaller data sets, you get worse results.”

• On AI company consolidation:
  Gavin [80:15]: "...the literal goal of OpenAI and Anthropic and the reason that VCs are giving them hundreds of billions of dollars is they are claiming or at least saying that their goal is to own a large chunk of the economy, right?"

• On AI’s social contract:
  Andre [66:02]: “It's very possible ... that 10 years from now there will be legitimate relationships that people have with these entities.... Society has to really see these stories and ... be like, hey, we are entering to a new social contract with things we don't really understand.”

• On music copyright deal impact:
  Andre [72:47]: “Users were told that they can no longer download their songs... That is a kind of a scary thing ... users are furious because ... they don't own the things that they thought they.”

Hosts’ Startup Spotlight

[85:04–89:03]

• Amben Chat: Andre and Gavin’s AI audio platform is soft-launched; enables users to interact with AI “characters” in goal-driven, game-like conversations (e.g., talk someone out of setting off a bomb, game show quizzes, etc.).
• Aims: Be as easy/fun/capable as Roblox for AI audio, with creator tools coming soon.

Final Thoughts

The episode interweaves tech news with deeper implications for industry, law, and society. Amidst rapid-fire releases and billion-dollar corporate shifts, the hosts highlight underlying ethical, legal, and economic issues. Their friendly banter keeps the tone accessible—even as looming questions about AI’s effect on jobs, media, and social structures remain unresolved.

For More

• Full newsletter: https://lastweekin.ai/
• Try Amben Chat: [Official site link not given, referred as "Amben Chat"]