Output‑Centered Safety Instead Of Input Filters

Our 219th episode with a summary and discussion of last week's big AI news!
Recorded on 08/08/2025

Check out Andrey's work over at Astrocade , sign up to be an ambassador here

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:

• OpenAI reveals GPT-5, a consolidated model combining all previous versions, marking notable improvements and introducing a new infrastructure and product update.
• Multiple major releases from leading AI labs, including OpenAI, Anthropic, and Google reflect the ongoing competitive landscape with significant business updates and new model capabilities.
• Discussions on geopolitical influences in AI development highlight China’s evolving stance on AI safety and governance, contrasting with U.S. approaches and raising concerns over export bans and international cooperation.
• Papers from leading AI entities such as OpenAI and Anthropic delve into the complexities of AI alignment and safety, proposing new methodologies for auditing and mitigating risks in model behaviors.

Timestamps + Links:

• (00:00:10) Intro / Banter
• (00:02:14) Plug: Astrocade rolls out AI agent-powered game creation experience so anyone can create games
• Tools & Apps

• (00:03:07) OpenAI’s GPT-5 is here
• (00:17:02) Anthropic Releases Claude Opus 4.1 With Agentic, Coding and Reasoning Upgrades
• (00:21:06) Google rolls out Gemini Deep Think AI, a reasoning model that tests multiple ideas in parallel | TechCrunch
• (00:24:04) Grok Imagine, xAI’s new AI image and video generator, lets you make NSFW content | TechCrunch

• Applications & Business

• (00:26:35) Meta, Microsoft stocks rise on strong earnings and AI spending boom
• (00:29:17) OpenAI to Establish Stargate Norway With 230MW Data Center - Bloomberg
• (00:32:12) Anthropic Revenue Pace Nears $5 Billion in Run-Up to Mega Round — The Information
• (00:37:18) OpenAI Hits $12 Billion Annualized Revenue
• (00:40:06) Noma Security raises $100 million to defend against AI agent vulnerabilities | Ctech

• Projects & Open Source

• (00:42:13) OpenAI Just Released Its First Open-Weight Models Since GPT-2
• (00:53:13) Falcon-H1: A Family of Hybrid-Head Language Models Redefining Efficiency and Performance
• (00:57:39) Meta CLIP 2: A Worldwide Scaling Recipe
• (01:01:12) BFL and Krea release FLUX.1 Krea: Open image model designed for realism

• Research & Advancements

• (01:02:33) Google’s Newest AI Model Acts like a Satellite to Track Climate Change | WIRED
• (01:04:50) Google’s new AI model creates video game worlds in real time
• (01:10:55) AlphaGo Moment for Model Architecture Discovery
• (01:17:22) METR evaluates Grok 4

• Policy & Safety

• (01:20:05) Estimating Worst-Case Frontier Risks of Open-Weight LLMs
• (01:23:14) Anthropic's AI 'Vaccine': Train It With Evil to Make It Good - Business Insider
• (01:27:26) Anthropic unveils 'auditing agents' to test for AI misalignment | VentureBeat
• (01:28:31) Optimizing The Final Output Can Obfuscate CoT (Research Note)
• (01:31:23) Why China isn’t about to leap ahead of the West on compute
• (01:33:15) Inside the Summit Where China Pitched Its AI Agenda to the World | WIRED
• (01:38:47) Nvidia H20 GPUs reportedly caught up in U.S. Commerce Department's worst export license backlog in 30 years — billions of dollars worth of GPUs and other products in limbo due to staffing cuts, communication issues | Tom's Hardware

• (01:42:35) Response to listener comments