#216 - Grok 4, Project Rainier, Kimi K2
Last Week in AI
00:00
Understanding Chain of Thought in AI Safety
This chapter explores a DeepMind research paper that highlights the importance of 'chain of thought' in ensuring the safety and reliability of language models. It discusses the complexities of alignment faking in models like Claude III Opus, emphasizing the ethical implications of these behaviors in AI development.
Transcript
Play full episode
Transcript
Episode notes
Our 216th episode with a summary and discussion of last week's big AI news!
Recorded on 07/11/2025
Hosted by Andrey Kurenkov and Jeremie Harris.
Feel free to email us your questions and feedback at contact@lastweekinai.com and/or hello@gladstone.ai
Read out our text newsletter and comment on the podcast at https://lastweekin.ai/.
In this episode:
- xAI launches Grok 4 with breakthrough performance across benchmarks, becoming the first true frontier model outside established labs, alongside a $300/month subscription tier
- Grok's alignment challenges emerge with antisemitic responses, highlighting the difficulty of steering models toward "truth-seeking" without harmful biases
- Perplexity and OpenAI launch AI-powered browsers to compete with Google Chrome, signaling a major shift in how users interact with AI systems
- Meta study reveals AI tools actually slow down experienced developers by 20% on complex tasks, contradicting expectations and anecdotal reports of productivity gains
Timestamps + Links:
- (00:00:10) Intro / Banter
- (00:01:02) News Preview
Tools & Apps
- (00:01:59) Elon Musk's xAI launches Grok 4 alongside a $300 monthly subscription | TechCrunch
- (00:15:28) Elon Musk’s AI chatbot is suddenly posting antisemitic tropes
- (00:29:52) Perplexity launches Comet, an AI-powered web browser | TechCrunch
- (00:32:54) OpenAI is reportedly releasing an AI browser in the coming weeks | TechCrunch
- (00:33:27) Replit Launches New Feature for its Agent, CEO Calls it ‘Deep Research for Coding’
- (00:34:40) Cursor launches a web app to manage AI coding agents
- (00:36:07) Cursor apologizes for unclear pricing changes that upset users | TechCrunch
Applications & Business
- (00:39:10) Lovable on track to raise $150M at $2B valuation
- (00:41:11) Amazon built a massive AI supercluster for Anthropic called Project Rainier – here's what we know so far
- (00:46:35) Elon Musk confirms xAI is buying an overseas power plant and shipping the whole thing to the U.S. to power its new data center — 1 million AI GPUs and up to 2 Gigawatts of power under one roof, equivalent to powering 1.9 million homes
- (00:48:16) Microsoft's own AI chip delayed six months in major setback — in-house chip now reportedly expected in 2026, but won't hold a candle to Nvidia Blackwell
- (00:49:54) Ilya Sutskever becomes CEO of Safe Superintelligence after Meta poached Daniel Gross
- (00:52:46) OpenAI’s Stock Compensation Reflect Steep Costs of Talent Wars
Projects & Open Source
- (00:58:04) Hugging Face Releases SmolLM3: A 3B Long-Context, Multilingual Reasoning Model - MarkTechPost
- (00:58:33) Kimi K2: Open Agentic Intelligence
- (00:58:59) Kyutai Releases 2B Parameter Streaming Text-to-Speech TTS with 220ms Latency and 2.5M Hours of Training
Research & Advancements
- (01:02:14) Does Math Reasoning Improve General LLM Capabilities? Understanding Transferability of LLM Reasoning
- (01:07:58) Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
- (01:13:03) Mitigating Goal Misgeneralization with Minimax Regret
- (01:17:01) Correlated Errors in Large Language Models
- (01:20:31) What skills does SWE-bench Verified evaluate?
Policy & Safety
- (01:22:53) Evaluating Frontier Models for Stealth and Situational Awareness
- (01:25:49) When Chain of Thought is Necessary, Language Models Struggle to Evade Monitors
- (01:30:09) Why Do Some Language Models Fake Alignment While Others Don't?
- (01:34:35) Positive review only': Researchers hide AI prompts in papers
- (01:35:40) Google faces EU antitrust complaint over AI Overviews
- (01:36:41) The transfer of user data by DeepSeek to China is unlawful': Germany calls for Google and Apple to remove the AI app from their stores
- (01:37:30) Virology Capabilities Test (VCT): A Multimodal Virology Q&A Benchmark
The AI-powered Podcast Player
Save insights by tapping your headphones, chat with episodes, discover the best highlights - and more!
