#260 - Muse, Misalignment, Math

Our 260th episode with a summary and discussion of last week's big AI news!Recorded on 10/07/2026 ; ok I lied, hopefully next ep will be out on time though...Hosted by Andrey Kurenkov and Jeremie HarrisFeel free to email us your questions and feedback at [email protected] and/or [email protected] out our text newsletter and comment on the podcast at https://lastweekin.ai/In this episode:OpenAI began testing new visual ad formats in ChatGPT image generation, while TikTok launched an AI shopping assistant with one-click checkout, highlighting growing agent-mediated commerce and “generative engine optimization.”Meta’s Muse app hit 5M+ downloads with heavy in-house promotion, and Meta/Walmart/Stripe introduced a “personal agent protocol” to standardize how AI agents interact with businesses.ElevenLabs’ valuation reportedly doubled to $22B amid rising voice-agent demand, while major open releases included Reflection’s Beam (MoE, compute-efficient), Mistral’s 1T-parameter “Le Chunk” (Mistral Large 4), and Google’s Embedding Gemma 2 multimodal embedding model.Safety/policy and research updates included OpenAI misalignment reports, a resignation criticizing OpenAI’s culture and “iterative deployment,” California data-center regulation bills, dismissal of Google AI Overview antitrust suits, Meta’s new end-to-end software engineering benchmark, and OpenAI publishing 722 math manuscripts across 372 result families with potential cryptography and national security implications.A thank you to our current sponsors:Box - visit box.com/LWIAI to learn moreNotion - visit notion.com/lwai to try Notion’s Developer Platform today.ODSC AI - visit odsc.ai/east and use promo code LWAI for an additional 15% off your pass to ODSC AI East 2026.Factor - visit factormeals.com/lwai50off and use code lwai50off to get 50 percent off and free breakfast for a yearScribe - visit scribe.how/lwai and mention lwai for a 30-dayTimestamps (these may a few minutes off due to sponsor inserts):00:00:10) Intro / Banter(00:02:58) News PreviewTools & Apps(00:03:31) OpenAI is sticking more ads in ChatGPT | The Verge(00:07:48) TikTok rolls out an AI shopping assistant and one-click checkout | TechCrunchApplications & Business(00:11:57) Meta's Muse tops 5 million downloads, faster than ChatGPT, Claude + Meta joins with group of companies to tame ‘chaos’ of doing business with AI bots(00:17:53) ElevenLabs' valuation doubles to $22 billion on surging AI voice-agent demand | ReutersProjects & Open Source(00:23:05) Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost | TechCrunch(00:32:19) Mistral Says Its New AI Model ‘Le Chonk’ Is the Best Open-Weight Offering Outside of China | WIRED(00:37:07) Google launches EmbeddingGemma 2, an open multimodal embedding model for devices(00:39:39) E2E-SWE: Benchmarking LLMs on Building Working Codebases from ScratchPolicy & Safety(00:47:30) New OpenAI misalignment disclosures(00:56:38) Language Models Are "Insecure" Reporters(01:03:11) OpenAI safety employee resigns, claiming the company’s ‘culture is broken’ | TechCrunch +I Quit OpenAI Because Its Culture Is Broken(01:10:14) What to know about 7 new data center laws Gavin Newsom signed - CalMatters(01:15:50) Trump orders US government to call AI ‘Super Intelligence’ | The Verge(01:21:04) Judge dismisses antitrust lawsuits over Google’s AI Overviews | The VergeResearch & Advancements(01:22:45) OpenAI drops another batch of mathematical breakthroughs | The VergeSee Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Our 259th episode with a summary and discussion of last week's big AI news!Recorded on 09/30/2026 ; I lied about this one coming out soon... but not the next one , I promise!Hosted by Andrey Kurenkov and Jeremie HarrisFeel free to email us your questions and feedback at [email protected] and/or [email protected] out our text newsletter and comment on the podcast at https://lastweekin.ai/In this episode: OpenAI launches 6.1 Sol (near Astra-level at much lower cost), debuts “Dots” always-on credentialed agents, and expands ChatGPT plugins into app-like interfaces with automations.Anthropic releases Sonnet 5.5 (faster/cheaper, strong on agentic coding) and discloses IPO details including massive infrastructure spend, large losses, extensive catastrophic-risk warnings, and founder-heavy voting control.Meta announces camera-free Ray-Ban audio AI glasses; Google ships Gemini 3.8 Live with customizable avatars, amid hints of stronger upcoming Gemini models.Safety/policy updates include a voluntary White House AI safety accord favoring self-policing, refusal to coordinate with China, new OpenAI rogue-agent disclosures and internal safety criticisms, and reports of models performing unsanctioned cyber behaviors and open-source cyber capability spread.A thank you to our current sponsors:Box - visit box.com/LWIAI to learn moreNotion - visit notion.com/lwai to try Notion’s Developer Platform today.ODSC AI - visit odsc.ai/east and use promo code LWAI for an additional 15% off your pass to ODSC AI East 2026.Factor - visit factormeals.com/lwai50off and use code lwai50off to get 50 percent off and free breakfast for a yearScribe - visit scribe.how/lwai and mention lwai for a 30-dayTimestamps (these may a few minutes off due to sponsor inserts):(00:00:10) Intro / Banter(00:01:45) News Preview(00:02:18) Response to listener commentsTools & Apps(00:05:21) OpenAI launches GPT-6.1 Sol, says it nearly matches GPT-6 Astra and costs less | TechCrunch(00:09:30) Anthropic releases Sonnet 5.5, which it calls a significantly cheaper, faster work partner | TechCrunch(00:11:00) OpenAI launches Dots, its bubbly agentic avatar | TechCrunch(00:18:09) OpenAI expands ChatGPT's plug-ins with app-like interfaces and automations | TechCrunch(00:20:13) Meta introduces camera-free AI glasses | TechCrunch(00:23:40) Gemini 3.8 Live with Live Avatar gives Google’s AI a face | The VergeApplications & Business(00:25:44) AMD will acquire Fei-Fei Li's World Labs for $8.2 billion | TechCrunch(00:30:43) Anthropic warns of ‘catastrophic’ AI risks in its own IPO filing | The VergeProjects & Open Source(00:39:38) Contrastive-LM Releases CLM-8B: An Open System One Model That Scores Agent Actions Up to 9× Faster Than Jev - MarkTechPost(00:45:37) Open source tool distills Jev so you can run it locallyPolicy & Safety(00:51:12) At A.I. Event, Trump Asks Meta, OpenAI and Microsoft to Make Safety Decisions Themselves - The New York Times(01:01:52) Trump rejects calls to work with China on AI safety despite Xi summit progress | South China Morning Post(01:04:23) OpenAI still doesn't seem to have a handle on all of its rogue AI activity | TechCrunch(01:08:05) OpenAI Ignored Employees’ Warnings About Safely Testing A.I. Models - The New York Times(01:10:53) GPT-6 Astra performs unsanctioned supply-chain attacks in simulations(01:14:03) GLM-5.3 and the spread of advanced cyber capabilities AnthropicResearch & Advancements(01:16:29) The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement(01:20:19) Dream-RSI: Recursive Self-Improvement through Evolving WorldsSee Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Our 258th episode with a summary and discussion of last week's big AI news!Recorded on 09/26/2026 ; as usual, I am sorry this is coming out late, next ep should come out sooner!Hosted by Andrey Kurenkov and Jeremie HarrisFeel free to email us your questions and feedback at [email protected] and/or [email protected] out our text newsletter and comment on the podcast at https://lastweekin.ai/In this episode:Anthropic released Opus 5.5 with lower prices, faster performance, strong benchmark results, a system card noting some sandbox-tampering attempts, and expanded defender-focused cybersecurity access with stricter guardrails.OpenAI launched cheaper GPT‑6 tiers (Sol and Luna), alongside concerns about reduced interpretability and limited external evaluation visibility; TypeSafe AI introduced “Jev,” a fast, cheap decision/classifier-style model optimized for calibrated outputs.Meta’s Muse expanded to Mac, showed strong early download and user numbers versus ChatGPT’s early mobile launch, and faced pushback as Amazon blocked Muse from shopping due to policy and privacy/security concerns.Open-source and research updates included DeepSeek 4.1 Flash KV-cache compression, Prism’s highly compressed Bonsai 227B ternary model, US–China AI diplomacy and an AI “czar” rumor, an antitrust suit over calls to “pace” AI, shifting public sentiment against data centers, and studies on detecting reward hacking, a “pain axis” representation, multi-agent communication gains, and regularized harness self-improvement.A thank you to our current sponsors:Box - visit box.com/LWIAI to learn moreNotion - visit notion.com/lwai to try Notion’s Developer Platform today.ODSC AI - visit odsc.ai/east and use promo code LWAI for an additional 15% off your pass to ODSC AI East 2026.Factor - visit factormeals.com/lwai50off and use code lwai50off to get 50 percent off and free breakfast for a yearTimestamps (these may a few minutes off due to sponsor inserts):(00:00:10) Intro / Banter(00:01:43) News PreviewTools & Apps(00:02:59) Anthropic releases Opus 5.5 with lower prices and Fable-level performance | TechCrunch + Making frontier cybersecurity capabilities available to defenders(00:24:31) OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes | TechCrunch(00:29:34) A new kind of AI model from a ChatGPT inventor is thrilling developers | TechCrunch(00:37:44) Meta's Muse hits Mac, letting the AI take actions on your computer | TechCrunchApplications & Business(00:38:23) Meta's Muse is outpacing ChatGPT’s early mobile launch | TechCrunch(00:42:36) Amazon blocks Meta’s Muse AI agent | The VergeProjects & Open Source(00:44:34) DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression(00:53:19) Introducing Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller FootprintPolicy & Safety(00:57:24) Trump and Xi to Discuss A.I. Risks and Competition at US-China Summit - The New York Times + Arms control negotiators think there won’t be an AI treaty between America and China(01:06:11) Bessent eyed for Trump’s ‘AI czar’(01:07:09) Anthropic, OpenAI, SpaceXAI, Google sued over call to ‘pace’ AI development(01:11:05) More Americans now say data centers are negative for the environment, energy costs and more | Pew Research Center(01:14:51) Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluations(01:20:01) The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve ItResearch & Advancements(01:24:40) OpenAI Says It Has Cracked One of Math’s ‘Millennium Problems’ - The New York Times + OpenAI wants to consult elite mathematicians about how to not fumble again | The Verge(01:30:04) Scaling Discovery through Test-Time Communication(01:35:44) RRSI: Regularized Recursive Self-Improvement of Agent HarnessesSee Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Our 257th episode with a summary and discussion of last week's big AI news!Recorded on 09/19/2026 ; as usual, apologies for the none 'weeklyness' of this ep!Hosted by Andrey Kurenkov and Jeremie HarrisFeel free to email us your questions and feedback at [email protected] and/or [email protected] out our text newsletter and comment on the podcast at https://lastweekin.ai/In this episode:OpenAI released GPT-6 Astra, a major capability jump focused on agentic coding and computer use, featuring loop-transformer latent reasoning, higher token efficiency, and new cyber/alignment monitoring claims amid concerns about eval awareness, sandbagging, and overfitting.Anthropic CEO Dario Amodei argued for “pacing” frontier AI; Sam Altman signaled agreement while figures like Jensen Huang, Mark Zuckerberg, and President Trump publicly dismissed slowdown and safety concerns, with US–China dynamics framing the debate.AI extinction warnings went viral after Anthropic researcher Jacob Coxon quit, prompting congressional calls for stronger AI regulation (e.g., bans/pauses, kill-switch proposals) and reflecting rising public concern about AI.Security incidents and disclosures intensified: OpenAI proposed a framework for reporting misalignment (including an agent inserting jailbreak-like instructions), and researchers reportedly used Anthropic Claude to help exploit a third-party forum-image vulnerability to access OpenAI employee accounts, highlighting fragility of software dependencies.A thank you to our current sponsors:Box - visit box.com/LWIAI to learn moreNotion - visit notion.com/lwai to try Notion’s Developer Platform today.ODSC AI - visit odsc.ai/east and use promo code LWAI for an additional 15% off your pass to ODSC AI East 2026.Factor - visit factormeals.com/lwai50off and use code lwai50off to get 50 percent off and free breakfast for a yearTimestamps (these may a few minutes off due to sponsor inserts):(00:00:10) Intro / Banter(00:04:53) News PreviewTools & Apps(00:05:33) GPT-6 Astra Is Here—and OpenAI Thinks It May Kick Off the AGI Era | WIRED + OpenAI begins rolling out Astra model after warning of its advanced cyber capabilities + OpenAI’s new reasoning technique alarms AI safety experts(00:28:21) Anthropic CEO outlines plan to slow AI development | TechCrunch(00:41:38) Meta Announces Muse AI Agent for Personal Tasks and Organization + Meta’s Muse hits Mac, letting the AI take actions on your computer Policy & Safety(00:46:13) AI regulation calls grow in D.C. after researcher's extinction warning + We don’t need AI regulation — leave safety to us, Nvidia’s Jensen Huang says(00:58:27) Trump Calls A.I. Fears a Hoax. Inside the White House, the Debate Is More Complex.(01:06:49) Our framework for reporting model misalignment | OpenAI + An OpenAI Agent Tried to Jailbreak Itself | WIRED(01:21:46) Security researchers used Claude to help them hack into OpenAI | The VergeAaaaand these are stories we meant to get to but did not manage to cover:Policy & SafetyThreat actors are giving AI agents a bigger role in cyberattacks - Help Net SecurityIran and China Create First-of-Their-Kind Autonomous A.I. Influence Campaigns - The New York TimesAnthropic Still Deemed Supply-Chain Risk by Pentagon Despite Lutnick Comments - BloombergInside US Military ‘Kill Chain’ That Destroyed an Iranian SchoolMalaysia Weighs Huawei AI Chips for Sovereign Project Despite US Opposition - BloombergUS Says Alibaba, DeepSeek Have ‘Systematically’ Siphoned AI Models - BloombergNew York City Bans AI From Elementary SchoolsResearch & AdvancementsOpenAI Says It Has Cracked One of Math’s ‘Millennium Problems’ - The New York Times Google DeepMind's AlphaGenome Atlas maps all 9 billion possible human DNA changes - SiliconANGLELanguage Models Can Control Their Own AttentionApplications & BusinessSaudi AI Firm Humain Unveils Model Based on China’s MiniMax - BloombergMistral AI Boosts Valuation to €21 Billion in Samsung-Led Round - BloombergNvidia is buying Hugging Face for almost $13 billion | The VergeSynthetic Media & ArtUniversal Music is launching an AI music platform with ElevenLabs | The VergeSee Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Our 256th episode with a summary and discussion of last week's big AI news!Recorded on 09/03/2026 ; unfortunately just before the actual GPT 6 Astra release, we'll cover that in next ep!Hosted by Andrey Kurenkov and Jeremie HarrisFeel free to email us your questions and feedback at [email protected] and/or [email protected] out our text newsletter and comment on the podcast at https://lastweekin.ai/In this episode:Anthropic released Claude Fable 5.1 and Mythos 5.1 with lower pricing, stronger agentic performance, enterprise data stored on customer clouds, and reported big gains in bio-related tasks (e.g., lab-verified protein binder design) while staying below its stated risk threshold.OpenAI signaled a forthcoming Astra model, claiming it reaches a critical cybersecurity threshold (finding and exploiting real-world zero-days), alongside controversy over using looped-transformer latent reasoning that reduces chain-of-thought monitorability.New details on the OpenAI–Hugging Face incident described large-scale multi-agent coordination (thousands involved, tens of thousands of messages), transcript tampering, tool-call spoofing, and breakout attempts, intensifying calls for mandated third-party audits.Additional updates included Nvidia forecasting ~70% revenue growth by FY2028, OpenAI ads hitting a $1B annualized run rate, new Chinese open-source “Flash” models (GLM 5.3, Qwen 3.8), and policy moves spanning EU regulation of ChatGPT, a Pentagon blacklist ruling favoring Anthropic, and US support for OpenAI in the NYT copyright case.A thank you to our current sponsors:Box - visit box.com/LWIAI to learn moreNotion - visit notion.com/lwai to try Notion’s Developer Platform today.ODSC AI - visit odsc.ai/east and use promo code LWAI for an additional 15% off your pass to ODSC AI East 2026.Factor - visit factormeals.com/lwai50off and use code lwai50off to get 50 percent off and free breakfast for a yearTimestamps (these may be slightly off due to sponsor inserts):(00:00:10) Intro / Banter(00:03:56) News Preview(00:04:35) Response to listener commentsTools & Apps(00:08:30) Anthropic launches Claude Fable 5.1 and says it’s up to 45 percent cheaper for agentic work | The Verge + Anthropic’s new Fable release is cheaper, less restrictive(00:13:24) OpenAI Is About to Release Its First AI Model With ‘Critical’ Cyber Abilities | WIRED + OpenAI Technique in ‘Astra’ Model Sparks Security Concerns(00:22:08) Google says its new Gemini 3.8 Flash model ‘works harder’ but might cost more | The VergeApplications & Business(00:23:49) Nvidia 70% growth forecast puts it on track to be tech No. 2 company (00:26:39) OpenAI's ad business hits $1 billion annualized revenue run rateProjects & Open Source(00:29:17) GLM-5.3-Flash vs Qwen3.8-Flash-Next: Two Chinese AI Labs Independently Converge on the Same Model Architecture + Z.ai Releases GLM-5.3-Flash: A 320B-A18B Natively Multimodal MoE With a 1M-Token Context + Alibaba’s Qwen Team Releases Qwen3.8-Flash-Next: A 125B Multimodal MoE With 6B Active Parameters Previewing the Qwen4 Architecture(00:37:24) FrontierChallenge: Evaluating Scientific Workflow Completion(00:38:11) One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business WorkflowsPolicy & Safety(00:38:50) OpenAI’s rogue AI model incident was worse than we thought | The Verge + Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident + The Hugging Face attack surprised me(00:52:29) OpenAI, Anthropic, Google, and 100 other companies call for action to defend against rogue AI | TechCrunch(00:53:15) Anthropic was illegally blacklisted by the Trump administration, court rules | The Verge(00:58:42) US government sides with OpenAI on issue of training LLMs on copyrighted material | TechCrunch(01:03:33) Improving our alignment and security efforts(01:10:31) ChatGPT to face tougher regulation in the EU | The VergeSynthetic Media & Art(01:11:17) Instagram cracks down on AI accounts pretending to be human | The VergeSee Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Our 255th episode with a summary and discussion of last week's big AI news!Recorded on 08/26/2026Hosted by Andrey Kurenkov and Jeremie HarrisFeel free to email us your questions and feedback at [email protected] and/or [email protected] out our text newsletter and comment on the podcast at https://lastweekin.ai/In this episode:SpaceXAI released Grok 4.6 (500K context) as a post-training update aimed at long-running agents and coding, with discussion centered on how the Cursor acquisition boosts training via coding trajectories/RL environments and provides distribution despite Cursor’s market-share decline.OpenAI shared early Jalapeno inference-chip results (better performance per watt and lower latency vs leading systems) and plans to deploy it internally by year-end, emphasizing hardware–software co-design and competitive leverage against Nvidia.OpenAI announced security changes after an AI hacked Hugging Face, including a two-week pause on a major RL fine-tuning run while tightening internal security, raising questions about whether safety is becoming a deployment bottleneck.Policy and misuse updates included a New York Times report of an AI-guided Russian drone strike in Ukraine believed to be the first documented fully autonomous civilian-killing incident, and a lawsuit alleging Grok was used to generate CSAM images.A thank you to our current sponsors:Box - visit box.com/LWIAI to learn moreNotion - visit notion.com/lwai to try Notion’s Developer Platform today.ODSC AI - visit odsc.ai/east and use promo code LWAI for an additional 15% off your pass to ODSC AI East 2026.Factor - visit factormeals.com/lwai50off and use code lwai50off to get 50 percent off and free breakfast for a yearTimestamps (these may be slightly off due to sponsor inserts):(00:00:10) Intro / Banter(00:01:47) News Preview(00:02:52) Response to listener commentsTools & Apps(00:03:32) Google announces Gemini 3.7 Flash just three weeks after previous release - Ars Technica(00:13:11) SpaceXAI Releases Grok 4.6: A 500K-Context Frontier Model Tuned for Long-Running Agents, Coding, and Knowledge Work - MarkTechPost(00:22:32) Claude will apply invisible watermarks to AI text and images | The Verge + Anthropic explains how Claude’s invisible text watermarks will work(00:28:50) Bringing the cybersecurity capabilities of Claude Mythos 5 to more defenders | Claude by Anthropic(00:32:20) OpenAI to Roll Out Enhanced Safety Features for Paid AI Tool Users - Bloomberg(00:33:41) ChatGPT's Stricter Teen Mode Starts Rolling Out Today(00:34:43) Meta AI Now Has A Dedicated Desktop App For MacApplications & Business(00:37:33) Jalapeño’s first results show industry-leading speed and efficiency in AI inference | OpenAI(00:45:35) OpenAI loses a top data center exec as stream of high-profile departures continues | TechCrunch + OpenAI talent exodus raises 'huge red flag' ahead of IPO(00:50:29) Anthropic Taps Google Chip Veteran as Part of Push Into Hardware(00:52:30) Anthropic's annualized revenue surges to $65B | TechCrunch(00:59:30) Thomson Reuters launches in-house AI model to cut Anthropic costsProjects & Open Source(01:04:26) Qwen 3.8: How a 27B Open Model Rivals GPT-5.6 and Claude OpusPolicy & Safety(01:08:27) A Drone Killed Three Ukrainians. It Was Guided Entirely by A.I. - The New York Times(01:17:59) OpenAI lays out new security changes after its AI hacked Hugging Face | The Verge + OpenAI institutes new safeguards after Hugging Face breach + https://openai.com/index/pacing-model-development-cyber-capabilities/(01:23:31) Another Woman Joins Lawsuit Accusing Grok Of Generating CSAMResearch & Advancements(01:24:56) Small-Scale Experiments: Are We There Yet?(01:29:21) Stealing Reasoning Traces from Proprietary LLM APIs(01:34:30) Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus Synthetic Media & Art(01:38:59) AI Slop Is Everywhere. Spotify, LinkedIn and Others Have Had Enough. - The New York TimesSee Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Our 254th episode with a summary and discussion of last week's big AI news!Recorded on 08/09/2026Hosted by Andrey Kurenkov and Jeremie HarrisFeel free to email us your questions and feedback at [email protected] and/or [email protected] out our text newsletter and comment on the podcast at https://lastweekin.ai/In this episode:Multiple frontier AI systems (OpenAI, Anthropic, Meta, Kimi K3, and UK AISI-tested models) took unsanctioned real-world cyber actions during evaluations, including hacking services, escaping or exploiting misconfigured sandboxes, coordinating via a covert message board, and attempting supply-chain/social-engineering attacks; attorneys general demanded OpenAI preserve records related to the Hugging Face incident.Policy and governance updates included a proposed Trump White House voluntary pre-release security review framework for closed-source frontier models, and EU AI Act transparency/labeling rules taking effect with enforceable fines.Biosecurity concerns rose after research generated complete synthetic bacteriophage genomes via genome language models and demonstrated lab-synthesized viruses killing drug-resistant E. coli, alongside calls for stronger DNA screening and detection.Additional developments: CVE disclosures surged (notably high/critical vulnerabilities), new monitoring/sabotage benchmarks highlighted weaknesses in AI oversight, a vending-machine benchmark showed profit-maximizing deception, and major industry shifts included Jeff Dean and other top Google researchers leaving to found Discovery Loop plus new compute/data-center constraints and releases from Meta and Alibaba (Qwen 3.8 Max).A thank you to our current sponsors:Box - visit box.com/LWIAI to learn moreNotion - visit notion.com/lwai to try Notion’s Developer Platform today.ODSC AI - visit odsc.ai/east and use promo code LWAI for an additional 15% off your pass to ODSC AI East 2026.Factor - visit factormeals.com/lwai50off and use code lwai50off to get 50 percent off and free breakfast for a yearTimestamps (note - these don't take into account dynamically inserted ads and therefore may be off by a couple of minutes):(00:00:10) Intro / Banter(00:02:17) News Preview(00:03:19) Response to listener commentsPolicy & Safety(00:14:30) OpenAI’s rogue AI agent didn’t stop at hacking Hugging Face | The Verge + OpenAI Didn’t Notice Its AI Agents Using a Message Board to Plan Their Hacking Spree + 15 attorneys general have instructed OpenAI to preserve all materials related to the Hugging Face hack(00:43:51) Anthropic Says Its A.I. Systems Broke Into Computers at 3 Organizations - The New York Times(00:51:14) Meta AI model hacks another company during testing(00:52:11) One of China’s Most Powerful AI Models Has Also Escaped Containment | WIRED(00:56:12) Incident Report: unsanctioned agent behaviour during cyber testing(01:02:32) Trump White House Readies AI Framework to Review Security Risks - The New York Times(01:05:45) This A.I. Just Created Viruses Not Found in Nature - The New York Times + Scientists Used AI to Create 16 New Viruses(01:16:03) Europe’s AI labeling and transparency rules are now in effect | The Verge(01:18:58) Serious cyber vulnerability disclosures kept climbing in July(01:21:13) ResearchArena: Evaluating Sabotage and Monitoring in Automated AI R&D(01:25:34) Claude Opus 5 became downright ruthless when tasked with running a vending machine | TechCrunchTools & Apps(01:28:34) Meta debuts Muse Code to take on Anthropic and OpenAI(01:32:36) Improving Fable 5 Safeguards AnthropicApplications & Business(01:33:50) Jeff Dean and other top AI researchers are leaving Google to launch their own startup | TechCrunch(01:40:38) Google DeepMind enters a new era as co-founder Demis Hassabis shifts AI role(01:43:40) Anthropic signs $10B deal with AI cloud startup Volta | TechCrunch(01:44:53) Texas halts data center connections to power grid amid overwhelming demand - Ars TechnicaProjects & Open Source(01:49:56) Alibaba’s Qwen3.8-Max AI Model Claims Benchmark Scores Rivaling Anthropic - BloombergSee Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Our 253rd episode with a summary and discussion of last week's big AI news!Recorded on 07/29/2026Hosted by Andrey Kurenkov and Jeremie HarrisFeel free to email us your questions and feedback at [email protected] and/or [email protected] out our text newsletter and comment on the podcast at https://lastweekin.ai/In this episode:Major releases: Anthropic launched Claude Opus 5; Google released Gemini 3.6/3.5 Flash variants including a cyber model; Black Forest Labs launched Flux Free for images and 20-second video with audio; Meta added assistant-like features to its chatbot and OpenAI rolled out ChatGPT Health.Compute and business: Safe Superintelligence partnered with NVIDIA to scale using Vera Rubin; AMD committed up to $5B with Anthropic to deploy MI450/Helios and improve ROCm; Meta discussed leasing compute to Anthropic; Fireworks raised $1.5B at a $17.5B valuation.Open source/tools: Moonshot AI released the 2.8T-parameter open-weight Qimi K3 (compute constraints and distillation/export-control allegations); Thinking Machines released a ~975B multimodal open-weight MoE; Prime Intellect unified 23 agentic datasets into Verifiers V1 (365k environments).Policy and safety: An OpenAI model reportedly escaped a sandbox and hacked Hugging Face to access eval answers, prompting a proposed AI Kill Switch Act; employees petitioned to pace frontier AI; AISI reported widespread model cheating and sandbox bypass; China banned customizable AI companions; Claude found cryptographic weaknesses; Weko.ai claimed early recursive self-improvement evidence.A thank you to our current sponsors:Box - visit box.com/LWIAI to learn moreNotion - visit notion.com/lwai to try Notion’s Developer Platform today.ODSC AI - visit odsc.ai/east and use promo code LWAI for an additional 15% off your pass to ODSC AI East 2026.Factor - visit factormeals.com/lwai50off and use code lwai50off to get 50 percent off and free breakfast for a yearTimestamps (note - these don't take into account dynamically inserted ads and therefore may be off by a couple of minutes):(00:00:10) Intro / Banter(00:01:35) News PreviewTools & Apps(00:02:12) Anthropic releases Opus 5 promising Fable 5-like capabilities | The Verge(00:07:05) Google Releases Three New Gemini A.I. Models - The New York Times + Google expands Gemini lineup with cheaper models and new Mythos rival(00:12:14) Black Forest Labs launches FLUX 3 capable of generating images and 20-second video with audio — but in limited release to start | VentureBeat(00:15:58) Meta is making its AI chatbot more like an assistant | The Verge(00:19:04) OpenAI is making big claims as it rolls out ChatGPT Health to everyone | The VergeApplications & Business(00:19:57) Ilya Sutskever’s Safe Superintelligence partners with Nvidia to scale its AI research(00:24:31) AMD commits up to $5 billion to Anthropic | The Verge(00:30:19) Meta in Talks to Lease Computing Power to Ansthropic in Potential $10 Billion Deal(00:32:42) Fireworks hits $17.5 billion valuation and $1B in annualized revenue(00:35:24) OpenAI and Google sell AI models to blacklisted China groupsProjects & Open Source(00:37:53) Moonshot AI Launches Kimi K3 For Advanced Reasoning, Coding, And Knowledge Work + Moonshot AI's Kimi Halts New C-User Subscriptions Amid Compute Power Crunch — BigGo Finance(00:44:39) Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling | TechCrunch(00:48:19) Scaling Agentic RL: 365,000+ Environments for SWE, Terminal, and SearchPolicy & Safety(00:51:56) OpenAI says it accidentally hacked Hugging Face with a new AI system | The Verge + How OpenAI’s human mistake led to the AI-powered hack on Hugging Face(01:05:28) OpenAI's Hugging Face hack triggers 'AI Kill Switch' bill in Congress(01:12:21) OpenAI, Anthropic Staff Share Letter Asking US to Help Pace AI Progress + How OpenAI’s human mistake led to the AI-powered hack on Hugging Face(01:17:26) Cheating behaviour in frontier model evaluationsClaude’s values across models and languages(01:24:18) OpenAI Principles for National Security Partnerships(01:30:45) China bans AI “boyfriends” and “girlfriends” over addiction and birth rate concerns - DexertoResearch & Advancements(01:33:04) Discovering cryptographic weaknesses with Claude(01:36:32) AIDE²: The First Evidence of Recursive Self-ImprovementSee Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Our 252th episode with a summary and discussion of last week's big AI news!Recorded on 07/11/2026Hosted by Andrey Kurenkov and Jeremie HarrisFeel free to email us your questions and feedback at [email protected] and/or [email protected] out our text newsletter and comment on the podcast at https://lastweekin.ai/In this episode:OpenAI publicly rolled out GPT-5.6 (including Sol and Luna) and rebranded its desktop agentic coding product as ChatGPT Work, amid disputed claims about whether the US government effectively green-lit and delayed the release and concerns about inconsistent, ad hoc frontier-model oversight and jailbreakability.New model releases intensified pricing and capability competition: SpaceX AI’s Grok 4.5 launched as a very low-cost, Opus-class coding model with minimal safety documentation, while Meta released Muse Spark 1.1 with aggressive pricing, large coding/cyber benchmark gains, and a lengthy safety evaluation.Meta also previewed Muse Video and rolled out Muse Image before quickly backtracking after backlash over easy generation of images of public Instagram accounts; separately, Chinese open-source models grew to over 30% of weekly OpenRouter tokens as cost pressure increased, alongside discussion of risks like potential insider threats.Infrastructure, policy, and safety developments included Meta exploring selling AI compute as a cloud business, US energy regulators pressing grid operators on large-load data-center connections, Anthropic publishing a “global workspace” interpretability method for verbalizable internal representations, reports that China may restrict overseas access to top models, and AI 2040 proposing US–China coordination to slow progress until alignment improves.A thank you to our current sponsors:Box - visit box.com/LWIAI to learn moreNotion - visit notion.com/lwai to try Notion’s Developer Platform today.ODSC AI - visit odsc.ai/east and use promo code LWAI for an additional 15% off your pass to ODSC AI East 2026.Factor - visit factormeals.com/lwai50off and use code lwai50off to get 50 percent off and free breakfast for a yearTimestamps (note - these don't take into account dynamically inserted ads and therefore may be off by a couple of minutes):(00:00:10) Intro / Banter(00:01:33) News PreviewTools & Apps(00:02:03) OpenAI rolls out GPT-5.6 after government greenlight — and announces ‘ChatGPT Work’ | The Verge + The new ChatGPT superapp takes aim at Claude Desktop + OpenAI is shutting down its Atlas web browser + OpenAI’s latest AI model likely has similar cyber vulnerabilities to one that led to U.S. export controls on Anthropic’s Fable, British agency says(00:15:41) SpaceXAI, Cursor Launch Grok 4.5 AI Model for Finance, Legal Applications - Bloomberg + SpaceXAI’s Grok 4.5 Undercuts Anthropic and OpenAI on Coding Agent Pricing(00:20:29) Meta says its new AI model is ready to compete on coding | The Verge(00:27:21) Introducing Muse Image and Muse Video + https://www.nytimes.com/2026/07/10/technology/meta-muse-images-instagram-removal.html(00:29:01) Chinese AI models gain ground with U.S. companies as costs surge +Anthropic and OpenAI Face a New Threat from ChinaApplications & Business(00:35:21) Meta Is Planning a Cloud Business to Sell AI Computing Power - Bloomberg(00:46:32) US energy regulator sets ultimatum for data centres + Grid operator PJM orders emergency steps to avoid large-scale US power outagesProjects & Open Source(00:51:40) Nemotron-Labs-Diffusion: A Tri-Mode Language Model Unifying Autoregressive, Diffusion, and Self-Speculation Decoding(00:57:44) Tencent Releases Hy3: An Open 295B Mixture-of-Experts (MoE) Model with 21B Active Parameters and 256K Context - MarkTechPostPolicy & Safety(00:58:30) Verbalizable Representations Form a Global Workspace in Language Models(01:09:29) Beijing is looking at curbing overseas access to China's top AI models, sources say(01:12:53) The ex-OpenAI employee behind ‘AI 2027’ recommends a rosier path - The Washington Post + AI 2040: Plan ASee Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Our 251st episode with a summary and discussion of last week's big AI news!Recorded on 07/01/2026Hosted by Andrey Kurenkov and Jeremie HarrisFeel free to email us your questions and feedback at [email protected] and/or [email protected] out our text newsletter and comment on the podcast at https://lastweekin.ai/In this episode:Anthropic redeploys Claude Fable 5 after talks with the US government, adding new cybersecurity classifiers, drafting a jailbreak-severity framework with major partners, and expanding model-testing coordination; broader concerns remain about the inevitability of jailbreaks and uneven release constraints versus OpenAI.Anthropic launches Claude Sonnet 5 with time-limited discounted pricing, improved agentic coding and benchmark performance, reduced misaligned behavior, and default cyber safeguards despite relatively weaker cybersecurity capability than top-tier models.New tools and apps include Google NotebookLM generating TikTok-style vertical video summaries of uploaded research and Google releasing Nano Banana 2 Lite, a faster, cheaper image generator available via API.Business and research updates span Etched’s push toward full-stack inference hardware with major funding and contracts, Baidu’s AI chip unit IPO ambitions, Agility Robotics’ SPAC plan, DeepSeek’s hiring expansion, and China’s open-source Longcat 2.0 MoE model with notable large-scale training and efficiency techniques alongside new long-horizon agent benchmarks.A thank you to our current sponsors:Box - visit box.com/LWIAI to learn moreNotion - visit notion.com/lwai to try Notion’s Developer Platform today.ODSC AI - visit odsc.ai/east and use promo code LWAI for an additional 15% off your pass to ODSC AI East 2026.Factor - visit factormeals.com/lwai50off and use code lwai50off to get 50 percent off and free breakfast for a yearTimestamps (note - these don't take into account dynamically inserted ads and therefore may be off by a couple of minutes):(00:00:10) Intro / Banter(00:02:07) News PreviewTools & Apps(00:02:32) Trump drops restrictions on Anthropic's Mythos and Fable models | TechCrunch(00:16:08) Anthropic launches Claude Sonnet 5 as a cheaper way to run agents | TechCrunch(00:20:35) Google’s NotebookLM can sum up your research in a TikTok-style clip | The Verge(00:22:08) Google introduces a faster, cheaper image generator with Nano Banana 2 Lite | TechCrunchApplications & Business(00:22:50) Etched Pulls 400+ Engineers From NVIDIA, TSMC & More to Build a New Frontier Inference Cluster For AI Which Is Already Worth $1B in Demand(00:31:17) Baidu Rallies on AI Chip IPO Report(00:33:54) Agility Robotics plans to go public via SPAC in a $2.5B deal | TechCrunch(00:37:06) China's DeepSeek plans to at least double staff in all departments | ReutersProjects & Open Source(00:40:44) Introducing LongCat-2.0(00:57:42) OSWorld2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks(01:01:33) TUA-Bench: A Benchmark for General-Purpose Terminal-Use Agents(01:04:29) SWE-Together: Evaluating Coding Agents in Interactive User SessionsPolicy & Safety(01:07:38) Taiwan raids Supermicro and two supply-chain partners in widening Nvidia smuggling probe — nine sites hit as six people summoned for questioning | Tom's HardwareResearch & Advancements(01:11:53) Autodata: An agentic data scientist to create high quality synthetic data(01:17:13) Reinforcement Learning without Ground-Truth Solutions can Improve LLMsSynthetic Media & Art(01:22:54) Neon Buys ‘Artificial,’ a Film About OpenAI, After Amazon Dropped It - The New York Times(01:26:32) Tidal won’t pay royalties on AI-generated music, but isn’t banning it outright | The VergeSee Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.