Last Week in AI – Episode #214 Summary

Podcast: Last Week in AI
Hosts: Andrei Karenkov, Jeremy Harris
Date: July 4, 2025
Episode Title: Gemini CLI, io drama, AlphaGenome, copyright rulings

Episode Overview

Andrei and Jeremy discuss the latest developments in the AI industry from June 2025, focusing on new tools like Google’s Gemini CLI, drama and legal disputes in the AI hardware space, significant hardware and energy news, recent research advances from DeepMind and others, evolving AI company structures, talent wars, and the latest critical rulings on AI and copyright.
The tone is witty, conversational, and informed—a blend of lighthearted banter and in-depth technical analysis.

Key Discussion Points & Insights

1. AI-Generated Podcasts and Human Touch

• The hosts joke about the rise of AI-generated podcasts and the challenges of replicating their unique banter, especially when life events (like Jeremy’s daughter’s teething) interfere.
  • Quote: “But dude, what AI could compete with starting 30 minutes late because its daughter's teething? ... You're not going to find an AI that can pull that off.” (03:51, Jeremy)

2. Tools & Apps

Gemini CLI Launch

• What: Google announced Gemini CLI, a command line LLM tool akin to Anthropic’s Claude Code, offering developer-oriented AI assistance from the terminal.
• Features: Powered by Gemini 2.5 Pro; generous free tier (60 requests/min, 1000/day); designed for agentic workflows with interactive turns.
• Competitive Landscape: Slightly less capable at code tasks than Claude Code, but rapidly improving; Google’s competitive advantage is massive compute resources and scaled down costs.
  • Quote: “Google is being pretty aggressive, giving away a lot of usage... There's also a lot of usage for free without having to pay... the cap is very high.” (07:13, Andrei)
• Systemic Trend: Shift from “one big model” thinking to complex, agentic software systems combining models and backend logic.

Anthropic Artifacts & App Building

• What: Anthropic adds the ability to publish "artifacts" (mini interactive web apps) built with Claude; can now be shared and can query Claude in-app.
• User Impact: Easier for non-technical users to develop and share apps; strong threat to platforms like Replit.
  • Quote: “You could see OpenAI, you could see Anthropic launching a kind of app store... Not actually that crazy sounding today.” (13:33, Jeremy)
• Industry Convergence: All major LLM providers are evolving towards integrated app-building and tool-connecting frameworks.

3. Applications, Business & Drama

OpenAI’s IO vs. EO Trademark Dispute

• What: Lawsuit between OpenAI (IO) and hardware startup EO (IYO), alleging trademark infringement and idea theft over similar names and wearable AI hardware concepts.
• OpenAI’s Response: Sam Altman published email receipts to counter EO’s claims.
  • Quote: “Here's the evidence, here's the receipts of your emails. I'm not too sure if what you're saying is legit.” (16:11, Andrei)
• Legal Outcome: EO obtained a temporary restraining order preventing OpenAI from using the IO name for now.
• Bigger Picture: More public, sensational legal drama at OpenAI compared to competitors—a recurring pattern of handling disputes publicly.

4. Hardware & Infrastructure

China’s Chip Progress: “Still at 7nm”

• What: TechInsights confirms Huawei’s Kirin X90 SoC still uses SMIC’s 7nm process, not the rumored 5nm breakthrough.
• Strategic Implications: China remains ~2 years behind the West in advanced semiconductor manufacturing; heat and efficiency issues persist with older nodes.
  • Quote: “...the decisive quashing of that rumor.” (23:25, Jeremy)

AMD’s Ultra Ethernet Network Card

• What: AMD unveils industry’s first ultra ethernet-compliant NIC—the Pensando Polara (400 Gbps)—to compete against Nvidia’s InfiniBand in connecting AI superclusters.
• Competitive Shift: Open, industry-backed ultra ethernet standards aim to break Nvidia’s monopoly, cut costs, and support massive GPU clusters.
• Deployed By: Oracle Cloud, in forthcoming Zetta-scale clusters.
  • Quote: “This is actually a pretty seismic shift in the industry.” (28:12, Jeremy)

Nuclear Power for Data Centers

• Amazon: Buys 1.9 GW from Talon Energy’s Susquehanna plant (PA) to power data centers through 2042—first deal was blocked for ‘off-grid’ intent, new deal rerouted properly through the grid.
• Nvidia & Bill Gates: Invest in TerraPower, a company building small modular nuclear reactors (SMRs) for powering future data centers, though commercial impact likely post-2030.
• SMRs: Seen as the next major innovation for data center energy supplies in the US.

5. AI Talent Wars & Company Structure

Thinking Machines Lab (Mira Murati)

• What: Former OpenAI CTO Mira Murati raises $2B at a $10B valuation (seed round) for new AGI startup with a team of AI all-stars—despite not disclosing the specific plan to investors.
  • Quote: “...hundreds and hundreds of millions of dollars just on the basis of like, yeah, you know, Mira is a serious fucking person.” (37:25, Jeremy)
• Governance: Murati holds majority board voting power—echoes of post-OpenAI governance experiments.

Meta Poaches OpenAI Talent, Restructures AI Division

• What: Meta aggressively hires away OpenAI researchers for its new “Superintelligence” team, led by high-profile figures (including Scale AI’s Alex Wang).
• Talent Strategy: Zuckerberg personally recruiting, offering impressive (allegedly not $100M, per Lucas Baer) packages.
• Cultural Conflict: Meta’s historic AI culture, shaped by Yann LeCun’s scaling skepticism, hampered progress; new hires represent a hard pivot to “scale-pilled” alignment and product focus.
  • Quote: "Essentially Mark Zuckerberg is being forced to pay the Yann LeCun tax right now." (46:24, Jeremy)

6. Research & Advancements

AlphaGenome (DeepMind)

• What: DeepMind releases AlphaGenome, a model for predicting gene function and effects of genetic mutations, analogous to AlphaFold but focused on causative genomics.
• Output: Model and API publicly available for non-commercial research.
  • Quote: “…beats all existing techniques out of a park on almost every single benchmark...” (50:08, Andrei)

Direct Reasoning Optimization (DRO)

• What: New LLM training algorithm for better open-ended reasoning—refines how models learn from their own reasoning traces, overcoming flaws in previous reward averaging.
• Key Innovation: “If your model's reasoning was, ‘I think it was Cristiano Ronaldo,’ and the actual answer was Lionel Messi ... this is just a way that they use to kind of detect good reasoning and then they feed that in ..." (59:58, Jeremy)

Farseer – Refined Scaling Law

• What: New empirical scaling law that more accurately predicts LLM loss for different model/data sizes, outperforming the “Chinchilla” law, critical for expensive AI training investments.

LLM First Search

• What: New search algorithm letting LLMs dynamically choose how to branch/explore in reasoning or problem-solving tasks (e.g., math games)—outperforms classical search methods like Monte Carlo Tree Search.

7. Policy, Safety & AI Supervision

Unsupervised Elicitation of LLMs (Anthropic)

• What: New Anthropic-backed method for “eliciting” correct behavior from LLMs without human-labeled data, useful for when AIs surpass human performance. Trained a Claude 3.5-based assistant to outperform its human-supervised cousin via internal coherence maximization.
  • Quote: “...once you get to superhuman AI, well, maybe humans can't actually see what it does and kind of give it to labels…” (74:20, Andrei)

Taiwan Expands Export Bans

• Taiwan blacklists Huawei and SMIC for a wide array of tech components, broadening US restrictions on semiconductors and manufacturing inputs by closing loopholes.

8. Societal Impact Studies

ChatGPT and Cognitive “Debt”

• Study showed heavy use of AI assistants for writing (vs. tasking the brain or using Google) led to measurably less mental effort, lower recall, and a weaker sense of ownership over one’s output (based on EEG/follow-up tests).
  • Quote: “If you don't put in effort doing something, you're not going to get better at it. That's already something we know.” (83:01, Andrei)
• Criticized for overblown conclusions (small sample, short timeline), but illustrates the ongoing question of digital tools’ impact on cognition.

9. Synthetic Media, Copyright & Legal Rulings

Meta Copyright Win

• Judge dismisses a lawsuit by 13 authors (incl. Sarah Silverman, Junot Díaz) alleging Meta’s use of their books to train Llama infringed copyright—ruled as “fair use” as the outputs are transformative—not direct copies, and book sales didn’t drop post-model release.
  • Quote: “The outputs of LLAMA are transformative. So you’re not infringing on copyright.” (84:44, Andrei)
• Emphasizes these rulings are case-specific and do not set universal precedent.

Getty Drops Key Claims Against Stability AI

• Getty Images drops some major copyright claims against Stability AI in the US, citing procedural weaknesses; trademark and secondary infringement claims continue. Seen as a partial win for the AI image generation sector.

Notable Quotes & Memorable Moments

• On AI Replacing Human Hosts:
  “Can they be so lacking in wit and thought as we can be? Sometimes, that's a challenge.” (04:45, Andrei)

• On Replit Facing Platform Disruption:
  “If I'm Replit, I'm getting pretty nervous looking at this.” (13:33, Jeremy)

• On the Nature of Tech Legal Battles:
  “This is probably not a huge deal ... more than anything, just another thing to track with OpenAI, right?” (20:25, Andrei)

• On US-China Chip Race:
  "We're still not on the 5 nanometer node yet. That's really, really interesting." (23:39, Jeremy)

• On Amazon Buying Nuclear Power:
  “Imagine trying to… predict power needs through 2042.” (33:09, Jeremy)

• On Meta’s AI Culture Shift:
  “Essentially Mark Zuckerberg is being forced to pay the Yann LeCun tax right now.” (46:24, Jeremy)

• On Scaling Laws:
  “Bigger models just fundamentally learn from data in a different way.” (64:58, Jeremy)

• On the Tricky Impact of Generative AI on Learning:
  “I think... good writers are good thinkers because...When you are forced to... write something, I don't really understand something until I've written something about it with intent.” (80:29, Jeremy)

Timestamps for Important Segments

• Gemini CLI Launch and Analysis: 05:55–11:32
• Anthropic Artifacts & App Store Trend: 11:54–14:47
• OpenAI vs. EO Trademark Drama: 16:11–21:35
• China/Huawei-SMIC Chip Update: 23:42–25:39
• AMD & Ultra Ethernet/Nvidia Competition: 27:08–29:54
• Nuclear Power for Data Centers: 31:50–36:03
• Thinking Machines Lab Fundraising: 36:10–42:54
• Meta’s Talent Poaching & Org Changes: 42:54–49:01
• AlphaGenome DeepMind Research: 49:40–55:14
• Direct Reasoning Optimization Paper: 55:14–61:53
• Scaling Laws – Farseer: 61:54–66:05
• LLM-First Search Paper: 66:05–71:14
• Unsupervised Elicitation (Anthropic): 71:14–76:48
• Taiwan Tech Export Controls: 76:48–78:22
• ChatGPT & Cognitive Debt Paper: 78:22–83:01
• Meta Copyright Ruling: 83:01–88:35
• Getty Drops Claims Against Stability AI: 88:35–End

Note: For further details, papers, and coverage, consult episode links and source notes as referenced by the hosts.