Last Week in AI – Episode #210 Summary

Podcast: Last Week in AI
Date: May 26, 2025
Hosts: Andrej Kankov & Jeremy (Gladstone AI)
Theme: A packed week of major AI announcements and releases, spotlighting rapid innovation across model capabilities, consumer products, safety, and the hardware powering the next era of AI.

Episode Overview

This episode covers one of the most eventful weeks in recent AI history, with major updates from Google (I/O 2025), Anthropic (Claude 4), OpenAI (acquisition news & infrastructure), and more. The hosts provide deep-dives, analysis, and industry context on product launches, research breakthroughs, business drama, and AI safety moves.

Highlights and Key Topics

1. Anthropic’s Claude 4: Major Leap in Coding & Workflow Agents

Timestamps: 03:40–09:21

• Product Launch: Claude Opus 4 (large model) and Claude Sonnet 4 (medium model).
• Key Strengths:
  • Big jump in coding ability: “Sonnet 4 hits 80.2% on the Swebench benchmark, up from 72% with OpenAI’s Codex 1. That is a big, big jump.” — Jeremy (05:31)
  • Long workflow execution: Models sustain coherent multi-step agentic tasks.
  • Model flexibility: “You can adjust the reasoning budget for the models—qualitatively not new, but they’re really doubling down on the agentic direction.” — Andrej (03:40)
  • Memory upgrades: Opus 4 excels at creating and using memory files for persistent context.
  • Reduced shortcutting: “Both Opus 4 and Sonnet 4 are 65% less likely to take shortcuts on agentic tasks than Sonnet 3.7.” — Jeremy (07:37)
• Developer Features: Cloud Code tightly integrates with SDKs and dev environments for easier programmatic use.
• System Card: Details saved for a later, deeper technical section.

2. Google I/O 2025: Flurry of Consumer-AI Announcements

Timestamps: 09:21–26:24

• “A week like this hasn't happened in months. Google just went on the attack.” — Andrej (01:26)

a. AI Mode in Google Search

• New tab brings in-depth, conversational AI to search, with follow-up Q&A and sourcing—essentially blurring lines with Perplexity and ChatGPT’s search offerings.

b. Project Mariner & Gemini Agent Mode

• Google’s internet agent can complete up to 10 background tasks in parallel ($250/mo AI Ultra Plan). Agent Mode seems to act as Mariner’s interface in the Gemini app.
• “A difference in kind: Previously a single task ran in your browser; now, many tasks in parallel in the cloud. You’re orchestrating AI workers.” — Jeremy (15:45)

c. VEO 3: Text-to-Video and Audio Generation

• “Mind-blowingly coherent and realistic videos, not just video but video and audio together—and it does a pretty good job.” — Andrej (16:39)
• “A wow moment for text-to-video. 66% win rate over VEO 2 on Meta’s MovieGen benchmark—not a knockout, but remarkably consistent.” — Jeremy (18:16)
• New Flow tool for editing and asset management; competitor to Runway and Sora.
• VEO 2 improved with reference images and all features rolled into FlowTV for browsing output.

d. Imagine 4: Text-to-Image

• Higher realism, prompt fidelity, better at rendering minute details and text, up to 10x faster than Imagine 3.

e. Real-Time Translation and Smart Glasses

• Google Meet gets real-time speech translation (English/Spanish, with more languages on the way).
• “We hit a magic unlock as latency crosses a threshold. Suddenly, this becomes useful.” — Jeremy (24:25)

f. Jules: AI Coding Agent

• Competes directly with GitHub Copilot. Identifies and fixes code errors, preps pull requests.
• “Like every AI agent, it ‘may make mistakes’—we’ll be saying that until we hit superintelligence.” — Jeremy (26:24)

3. Coding Agent Wars: OpenAI, Microsoft, Google, Mistral, GitHub

Timestamps: 26:42–29:52

• Massive competition: Microsoft and OpenAI as frenemies, Google and Mistral entering agentic coding.
• GitHub Copilot for VS Code is now open source to compete with startups like Cursor.

4. OpenAI + Jony Ive’s IO

Timestamps: 29:52–36:10

• OpenAI acquires IO—a little-known startup by Jony Ive—for $5 billion in an unusual, “very strange” deal.
• “Jony Ive gets to just leave and fuck off—this is very esoteric… Normally not how it goes.” — Jeremy (32:38)
• Hints at a hardware play (AI device), but details are scarce and announcement vibes are “weird.”
• Related business maneuvers: OpenAI hires ex-Meta AR hardware lead.

5. OpenAI’s 5GW Abu Dhabi Data Center

Timestamps: 36:10–40:44

• Massive new data center (five times the size of Texas’s 1.2GW Stargate campus) announced in UAE, in partnership with G42.
• National Security Concerns: “Extraordinarily difficult to secure when you can't control the physical land. OpenAI hasn’t been impressive on the security story so far.” — Jeremy (37:09)
• Context: UAE ties, energy availability, and increasing capex for AI compute.

6. Other Noteworthy Business Updates

Timestamps: 41:18–47:26

• LM Arena raises $100M: “What is the profit story for a leaderboard company?” — Andrej (41:18)
• Nvidia chip policy: Adjusting chip supply to China after US restrictions.
• Google Gemini user numbers: Now reportedly at 400M monthly actives, close to ChatGPT scale.

7. Industry Analysis

Timestamps: 47:26–51:33

• Google’s Comeback: “Google is killing it right now… The sleeping giant woke up.” — Jeremy (49:14)
• AI Hardware/Server Glut: J.P. Morgan report shows over a million GPU oversupply—“Marginal costs are collapsing, capex skyrocketing, and compute is doubling yearly.”

8. Open Source and Research

Timestamps: 53:44–90:07

• Meta delays Llama 4 "Behemoth": “A really bad sign for Meta… consistently putting out mid models, open source recruitment strategy faltering.” — Jeremy (54:55)
• Google’s Gemini Diffusion Demo:
  • Diffusion-based LMs generate text near-instantly (up to 2,000 tokens/sec).
  • Potential for new capabilities like non-causal reasoning.
• Chain of Model Paper: Hierarchical sub-networks trained together enable efficient, dynamic scaling and interpretability.
• Test-time Activation Tuning (“Seek in the Dark”): Prompt-engineering-like tweaks at the model’s latent levels.
• Mixture-of-Experts “Reasoning Expert” Discovery: Dialing up cognitive experts in MOE models can improve reasoning—remarkably “obvious in hindsight, but effective.”
• Google Gemini Prompt Injection Defense: Adversarial robustness is an ongoing arms race; novel tricks like "spotlighting" tokens can break up attacker prompts.
• Epic AI on Algorithmic Innovation and the “Intelligence Explosion”:
  • Software-only advances depend heavily on discoveries that scale with compute.
  • Debate: Whether transformative advances (like transformers/MOEs) would be discoverable without hardware progress—hosts give critical perspective on the argument.
• Reinforcement learning fine-tunes small sub-networks: Alignment tends to only update a fraction of model parameters.

9. Policy & Safety

Timestamps: 90:07–102:33

• OpenAI’s Letter to CA Attorney General:
  • Details of OpenAI’s proposed shift to a Public Benefit Corporation; “lots of caveats and contradiction… classic OpenAI PR management.” — Jeremy (92:17)
  • Critics argue new structure weakens nonprofit controls and public benefit obligations.
  • “No Delaware PBC has ever been held liable for failing to pursue its mission”—weak enforcement in practice.
• Anthropic’s AI Safety Level 3 Measures:
  • With Claude 4, Anthropic activates new safety protocols (e.g., jailbreaking resistance, security controls, bug bounty, synthetic data, egress controls).
  • Main concern: “The biorisk side—helping those with undergraduate STEM degrees create bioweapons.” — Jeremy (99:55)
  • “They’re not pretending to defend against nation-state adversaries yet, but this is a major step up.” — Jeremy (101:53)

Notable Quotes & Moments

• On the pace of AI progress:
  “A week like this probably hasn't been seen in a few months… Google just went on the attack.” — Andrej (01:26)
• On Claude 4’s coding leap:
  “Sonnet 4 hits 80.2% now, going from 72 to 80%. That is a big, big jump… There’s not that much left to go.” — Jeremy (05:31)
• On Google’s competitive urgency:
  “If ChatGPT becomes the default for 5% more users, Google’s market cap drops more than 5%… This is a five-alarm fire for Google.” — Jeremy (11:29)
• On OpenAI + Jony Ive:
  “This is a very esoteric kind of deal… Jony gets to just leave and fuck off.” — Jeremy (32:38)
• On AI hardware security:
  “It’s extraordinarily difficult to actually secure something when you can't control the physical land… You’d hope they put in a little bit of effort, but you’d be surprised.” — Jeremy (37:09)
• On Meta’s model struggles:
  “Meta has been consistently pumping out these pretty mid models… Their recruitment/branding play is falling flat.” — Jeremy (54:55)
• On emerging model architectures:
  “Diffusion models can generate entire outputs at once—really changes the user experience, if they can get it to work as well as autoregression.” — Andrej (62:12)
• On AI safety escalation:
  “A big leap, moving to ASL3… This is fundamentally increasing the threat model: organized crime, terrorists, but not yet China.” — Jeremy (101:53)
• “Can an announcement have code smell? Because I feel like that's what this is.” — Jeremy, on the OpenAI/IO news (36:06)

Conclusion

Summary:
This week showcased a pivotal point in the AI landscape, with Google reasserting dominance through aggressive product rollouts, Anthropic pushing the state-of-the-art in code and agentic workflows, and OpenAI making big, if opaque, hardware and business moves. Rapid-fire research advances, policy maneuvering, and escalating safety measures show the field entering a new phase of both capability and risk.

For Listeners:
Stay tuned, as this pace shows no sign of slowing and the competitive, technical, and safety dimensions of AI are more intertwined and consequential than ever.