#230 - 2025 Retrospective, Nvidia buys Groq, GLM 4.7, METR | Last Week in AI

The Last Week in AI podcast kicks off 2026 by recapping major AI trends of 2025, including reasoning models, agentic AI, vibe coding, and image editing advancements, while also looking ahead to 2026. The discussion covers the realignment around AI alignment strategies, hardware developments like the HBM market shift with Micron's emergence, and pseudo corporate acquisitions driven by antitrust concerns. Data center security and the rise of Chinese open-source models are also highlighted. The hosts debate whether AI will achieve automated AI research in 2026, with considerations for jagged intelligence and the challenges of smoothing out AI capabilities. They also analyze the rising costs of AI agents, the implications of METR evaluation results, and policy updates like New York's RAISE Act.

Outlines

Part 1: 2025 Year in Review

Part 2: Research, Alignment, and Interpretability

Part 3: 2026 Predictions and Metrics

Part 4: Audio Modality and Hardware Wars

Part 5: Industry Acquisitions and Global Strategy

Part 6: Benchmarks and Scientific Models

Part 7: Agent Economics and Performance

Part 8: Safety, Regulation, and Technical Monitoring

Part 9: Ethics and Conclusion

Sign in to continue reading, translating and more.

Continue

#230 - 2025 Retrospective, Nvidia buys Groq, GLM 4.7, METR

Last Week in AI

Part 1: 2025 Year in Review

00:10AI in 2025: Reasoning Models, Agentic AI, Vibe Coding, and World Models

AI in 2025: Reasoning Models, Agentic AI, Vibe Coding, and World Models

02:31Alignment Realignment, Hardware Developments, and ROI Concerns in AI

Alignment Realignment, Hardware Developments, and ROI Concerns in AI

05:22HBM Market Shifts, Groq Acquisition, and Microsoft's Inflection Play

HBM Market Shifts, Groq Acquisition, and Microsoft's Inflection Play

07:10Data Centers, AI Bubble Worries, and Data Center Security Concerns

Data Centers, AI Bubble Worries, and Data Center Security Concerns

Part 2: Research, Alignment, and Interpretability

09:58Persona Vectors, Valuigi Effect, and Activation Steering in Alignment Research

Persona Vectors, Valuigi Effect, and Activation Steering in Alignment Research

12:38Tracing Thoughts in LLMs: Anthropic's Mechanistic Interpretability and Google's Shift

Tracing Thoughts in LLMs: Anthropic's Mechanistic Interpretability and Google's Shift

15:13Distributed Training's Absence and the Rise of Chinese Open Source Models

Distributed Training's Absence and the Rise of Chinese Open Source Models

Part 3: 2026 Predictions and Metrics

17:04On-Device Models, Video Advancements, and the Future of Transformer Architecture

On-Device Models, Video Advancements, and the Future of Transformer Architecture

19:48Automated AI Research: A Pivotal Metric for 2026

Automated AI Research: A Pivotal Metric for 2026

21:23Jagged Intelligence, Continual Learning, and the Magic Sauce for AI Progress

Jagged Intelligence, Continual Learning, and the Magic Sauce for AI Progress

Part 4: Audio Modality and Hardware Wars

24:07OpenAI Bets Big on Audio: Consolidating Teams for Voice-First Personal Devices

OpenAI Bets Big on Audio: Consolidating Teams for Voice-First Personal Devices

25:55Audio as a Key Modality and Nvidia's Acquisition of Groq

Audio as a Key Modality and Nvidia's Acquisition of Groq

27:00Nvidia's Acquisition of Groq: Acquihire, IP Licensing, and Antitrust Concerns

Nvidia's Acquisition of Groq: Acquihire, IP Licensing, and Antitrust Concerns

29:39Groq's SRAM Advantage and Nvidia's Vera Rubin Architecture

Groq's SRAM Advantage and Nvidia's Vera Rubin Architecture

31:20Groq's Inference Focus, Air-Cooled Racks, and Nvidia's Market Dominance

Groq's Inference Focus, Air-Cooled Racks, and Nvidia's Market Dominance

Part 5: Industry Acquisitions and Global Strategy

33:25Meta Acquires Manus: A Talent Acquisition and Subscription Model Play

Meta Acquires Manus: A Talent Acquisition and Subscription Model Play

35:43Meta's Agent Ambitions and the Yan Likun Tax

Meta's Agent Ambitions and the Yan Likun Tax

37:22Cursor Acquires Graphite: AI-Assisted Coding and Code Review

Cursor Acquires Graphite: AI-Assisted Coding and Code Review

39:15Micron's HBM Leapfrog: Hikes CapEx to $20 Billion

Micron's HBM Leapfrog: Hikes CapEx to $20 Billion

41:50Chinese Fabs Upgrade Older ASML Machines: A Strategy of Circumvention

Chinese Fabs Upgrade Older ASML Machines: A Strategy of Circumvention

46:08China's Hardware Independence: Power Abundance and Networking Prowess

China's Hardware Independence: Power Abundance and Networking Prowess

Part 6: Benchmarks and Scientific Models

47:52Z.ai Launches GLM 4.7: A New Open Source Model for Coding

Z.ai Launches GLM 4.7: A New Open Source Model for Coding

50:11Frontier Science: A New Benchmark for Expert-Level Scientific Capabilities

Frontier Science: A New Benchmark for Expert-Level Scientific Capabilities

54:29Democritus: Building Large Causal Models from Large Language Models

Democritus: Building Large Causal Models from Large Language Models

57:33Universally Converging Representations of Matter Across Scientific Foundation Models

Universally Converging Representations of Matter Across Scientific Foundation Models

Part 7: Agent Economics and Performance

1:02:12MetaRL Induces Exploration in Language Agents: The Lamer Framework

MetaRL Induces Exploration in Language Agents: The Lamer Framework

1:06:26Are the Costs of AI Agents Also Rising Exponentially?

Are the Costs of AI Agents Also Rising Exponentially?

1:08:52Algorithmic Efficiency, Human Labor, and the METR Plot

Algorithmic Efficiency, Human Labor, and the METR Plot

1:11:18METR Eval for Opus 4.5: High Results and Confusing Success Rates

METR Eval for Opus 4.5: High Results and Confusing Success Rates

1:14:02METR's Task List, Noisy Data, and VIBE Reports

METR's Task List, Noisy Data, and VIBE Reports

Part 8: Safety, Regulation, and Technical Monitoring

1:17:22New York Governor Signs RAISE Act to Regulate AI Safety

New York Governor Signs RAISE Act to Regulate AI Safety

1:19:01OpenAI's Support and the Complexity of Lobbying

OpenAI's Support and the Complexity of Lobbying

1:20:08Interoperability: Activation Oracles, Training, and Evaluating LLMs

Interoperability: Activation Oracles, Training, and Evaluating LLMs

1:22:11Activation Oracles: A General-Purpose Interpretability Strategy

Activation Oracles: A General-Purpose Interpretability Strategy

1:24:00Training the Oracle: Data Sets and Compute Efficiency

Training the Oracle: Data Sets and Compute Efficiency