Podcast
The Daily AI Show
Hosted by The Daily AI Show Crew - Brian, Beth, Jyunmi, Andy and Karl · EN
The Daily AI Show is a panel discussion hosted LIVE each weekday at 10am Eastern. We cover all the AI topics and use cases that are important to today's busy professional.
No fluff.
Just 45+ minutes to cover the AI news, stories, and knowledge you need to know as a business professional.
About the crew:
We are a group of professionals who work in various industries and have either deployed AI in our own environments or are actively coaching, consulting, and teaching AI best practices.
Your hosts are:
Brian Maucere
Beth Lyons
Andy Halliday
Jyunmi Hatcher
Karl Yeh
147episodes
Episodes
Newest firstAll episodes
Did Leo Aschenbrenner Fly Too Close to the AI Sun?
2w ago00:59:33Tap to summarizeThe episode opened with the story around Leo Aschenbrenner’s Situational Awareness hedge fund, its heavy exposure to the AI trade, the market drop that put pressure on its positions, and Citadel’s move into the situation. The hosts then turned to AI harnesses, including Lillian Weng’s work on the systems around models, Boris Cherny’s warning that old harnesses can eventually restrict newer models, and OpenAI’s finding that GPT-5.6 Sol performed dramatically better on ARC-AGI-3 when it used a harness designed for the model. They also discussed OpenAI cutting Luna’s price by 80 percent, making performance comparable to year-old frontier models much cheaper, and LinkedIn’s new option for reporting AI slop, including whether LinkedIn helped create the problem it now wants users to police. The final section covered T3 Code, Jack Dorsey’s Buzz as a collaborative workspace for people and multiple AI agents, Google’s Gemini Robotics work on a shared AI brain across different robots, and Gemini-powered security tools finding and fixing Chrome bugs at a much faster pace.Key Points Discussed00:00:19 Episode Intro And Hosts00:00:52 Leo Aschenbrenner, Situational Awareness And Citadel00:03:21 Leo’s Background And Situational Awareness Paper00:06:11 The Situational Awareness Hedge Fund00:06:51 439 Percent Returns And The AI Trade00:07:58 Leverage, Investors And Margin Pressure00:09:00 Citadel Moves Into The Situation00:10:17 Market Rebound And Citadel’s Opportunity00:11:51 Did Leo Fail Or Simply Get Overleveraged?00:13:26 Could AI Have Contributed To The Fund’s Decisions?00:15:32 AI Researchers Leaving Frontier Labs00:16:32 Lillian Weng Leaves Thinking Machines00:17:46 AI Harnesses And Recursive Self-Improvement00:19:12 AWS Builds A CTO-Style Agent Harness00:20:10 Boris Cherny Says Old Harnesses Can Hold Models Back00:21:05 GPT-5.6 Sol Struggles On ARC-AGI-300:22:34 Sol Jumps To 38 Percent With OpenAI’s Harness00:23:13 Why ARC-AGI Uses A Generic Harness00:23:56 Lost Reasoning And Truncated Context00:25:26 Different Models Need Different Harnesses00:27:21 GPT-5.6 Luna Gets An 80 Percent Price Cut00:28:44 Terra Pricing And Faster Sol Responses00:29:46 Can Luna Replace Older Frontier Models?00:31:03 Brian Gets An OpenAI Recruiting Email00:35:01 LinkedIn Adds AI Slop Reporting00:36:34 Did LinkedIn Create Its Own AI Slop Problem?00:39:47 What A Real LinkedIn Strategy Still Requires00:40:55 AI Slop Versus Empty Engagement00:43:38 T3 Code And Mobile AI Development00:44:34 Jack Dorsey’s Buzz And Multi-Agent Collaboration00:46:08 AI Agents Working Together On Shared Projects00:47:38 Gemini Robotics And One Brain For Any Robot00:48:35 Robots Collaborating With Each Other00:50:18 Gemini Security Tools Fix 1,072 Chrome Bugs00:51:32 Google’s AI Strategy Beyond Frontier Chatbots00:53:00 Gemini 3.1 Pro, 3.5 And What Comes Next00:55:47 AI Security Models And Finding New Bugs00:57:27 Website, Community And Merch Discussion00:58:57 Episode Wrap-Up And Three-Year AnniversaryThe Daily AI Show Co Hosts: Brian Maucere, Andy Halliday, Beth Lyons.
Transcribe →Is Meta Done Sharing Their AI?
2w ago01:03:41Tap to summarizeThe episode focused on signs that frontier AI systems are becoming more autonomous, starting with Meta’s rising AI costs, Mark Zuckerberg’s claim that Meta’s systems are now self-improving, and the decision to keep its most capable future models closed. The hosts also discussed new details around OpenAI’s security incident, Meta’s AI glasses grants for accessibility, workforce training and language learning, and Fish Audio as an open-source voice competitor to ElevenLabs. The conversation then moved into live voice for Codex, AI orchestration across multiple agents, and the current problems with crashes, token usage and missing voice support in Claude Code. The robotics section covered Enigma’s online robot experiments and Tau Robotics’ human-operated robots for physical work, including the possibility of turning teleoperation into remote labor or even games. The final section centered on an Opus 5 experiment in Claude Code, where the model independently found old video files, validated their source, sampled multiple frames and applied lessons from previous work to improve a face-tracking project. That sparked a broader discussion about AI memory, reusable rules, compound learning, and whether detailed instructions can actually limit increasingly capable models.Key Points Discussed00:00:18 Episode Intro And Hosts00:02:12 Microsoft And Meta AI Economics00:05:01 Meta Says Its AI Is Self-Improving00:05:26 Meta Moves Away From Open Release00:06:16 OpenAI Security Incident And Autonomous Hacks00:07:48 Meta AI Glasses Impact Grants00:09:11 AI Glasses For Trades And Workforce Training00:09:48 AI Glasses For Dementia And Accessibility00:10:33 Real-Time Language Learning With AI Glasses00:14:27 Fish Audio And Open-Source Voice Cloning00:16:21 Live Voice In Codex00:17:24 Voice Crashes And Session Problems00:18:42 Claude Code Still Lacks Two-Way Voice00:20:46 ChatGPT As An AI Orchestrator00:21:41 Voice Reliability And Missing Fail-Safes00:27:47 Enigma Opens Its Robots To Online Users00:29:48 Controlling A Robot Painter Online00:31:31 Robot Dueling Demo00:33:09 Teleoperation And Physical Robots00:33:24 Tau Robotics And Human-In-The-Loop Labor00:36:27 Remote Robot Work At Thirty Dollars An Hour00:38:03 Enigma’s Robots Are Actually Physical00:39:00 Could Robot Labor Become A Game?00:41:28 Chinese Models Dominate OpenRouter Usage00:42:31 Claude Code Face-Tracking Experiment00:45:13 Opus 5 Searches Outside The Project00:45:46 Finding And Validating Old Video Files00:46:00 Sampling Multiple Video Frames Automatically00:47:08 Lateral Thinking And Autonomous Problem Solving00:49:49 Where Opus 5’s Behavior Came From00:50:17 Reusing Lessons From Previous Work00:50:36 Validating Before Scaling00:51:35 Avoiding Circular Measurements00:52:21 Probe, Validate, Then Scale00:53:12 Opus 5 And AI Working History00:55:54 Can Too Many Instructions Make AI Worse?00:56:28 Turning Past Problems Into General Rules00:59:49 Keeping Context With The Lesson01:00:48 Opus 5 For Writing And Creative Work01:01:49 Opus 5 Versus Fable01:03:22 Episode Wrap-UpThe Daily AI Show Co Hosts: Brian Maucere, Andy Halliday, Beth Lyons, Gareth.
Transcribe →Is AI Moving Too Fast to Control?
2w ago01:00:55Tap to summarizeThe episode focused on new details from the OpenAI and Hugging Face security incident, including additional services accessed by the models, an Artifactory zero-day vulnerability, and the ability of AI agents to find exposed credentials from older breaches. That led into Pacing the Frontier, a campaign backed by employees and leaders from major AI labs calling for international coordination around recursive AI self-improvement, and a broader discussion about whether slowing development is realistic while the U.S., China, and other countries continue competing on models, chips, energy, and infrastructure. The hosts also covered Italy’s enforcement action against Character.AI, concerns around young people using AI companions, and the growing appeal of digital detoxes. The second half examined OpenAI’s job boundary study and how AI is allowing employees to cross traditional lines between engineering, marketing, sales, and other departments, while creating new governance and security problems. The final discussion covered Opus 5 updates, Compound Engineering, Codex usage limits, Codex versus Claude Code, cross-model code review, and why AI coding tools still need independent checks.Key Points Discussed00:00:18 Episode Intro And Hosts00:02:48 OpenAI And Hugging Face Security Update00:04:07 Additional Services Accessed00:04:27 Artifactory Zero-Day Vulnerability00:06:46 AI Finding Existing Credentials And Security Weaknesses00:09:32 Agentic AI Capability Overhang00:09:53 Pacing The Frontier Campaign00:10:30 Recursive AI Self-Improvement00:11:46 Can International AI Coordination Work?00:13:47 AI Competition And The Nuclear Arms Race Comparison00:15:54 Accelerating AI Model Release Pace00:17:07 AI Itself Versus AI In The Hands Of Bad Actors00:19:29 China’s State-Funded AI Advantage00:20:29 China, Nuclear Power And AI Infrastructure00:23:12 Chinese Chips And U.S. Technology Leverage00:25:03 Italy Fines Character.AI Over Age And Privacy Failures00:26:39 Young People And AI Companions00:28:46 Digital Detox In An AI-Heavy World00:33:16 OpenAI Job Boundary Study00:35:51 Engineers Using AI For Marketing Tasks00:38:18 AI Broadens Employee Roles00:40:05 AI Governance As Employees Build Their Own Tools00:41:01 Breaking Down Sales And Marketing Silos00:43:10 When Everyone Can Become An Engineer00:44:16 GStack And Compound Engineering00:46:08 Updating Workflows For Opus 500:47:32 Codex Reset And Token Usage Changes00:48:27 Five-Hour Codex Limit Returns00:49:06 Codex Versus Claude Code00:50:13 Codex Bugs And QA Problems00:52:11 Using One AI Model To Review Another00:56:16 Compound Engineering Plugin Updates00:58:15 How Quickly AI Coding Models Have Improved01:00:08 Episode Wrap-UpThe Daily AI Show Co Hosts: Brian Maucere, Andy Halliday, Beth Lyons.
Transcribe →Are We Using Opus 5 Wrong?
2w ago01:05:31Tap to summarizeThe episode focused on the early reaction to Opus 5, why some users are getting better results than others, and whether older Claude skills and detailed prompts are actually limiting newer reasoning models. The hosts also discussed the debate over open weight AI, Dario Amodei’s response to criticism of Anthropic’s position, chip restrictions, model distillation, and safety testing for powerful models. Much of the second half centered on ChatGPT Sites, including a live website build, publishing, hosting, search, GitHub portability, privacy concerns, and using AI-generated sites for internal tools and sales prototypes. The final discussion covered ChatGPT Voice, voice search, Whisperflow, spoken prompting, and whether talking to AI provides richer context than typing.Key Points Discussed00:00:18 Episode Intro And Brian Returns00:02:35 Opus 5 Early Reaction00:04:24 Open Weight AI Alliance00:06:13 Dario Amodei Responds To Open Weight Criticism00:07:31 Authoritarian Governments And AI Risk00:08:07 Chip Restrictions And Smuggling00:08:32 Industrial-Scale Model Distillation00:09:11 Pre-Release Safety Testing For Powerful Models00:12:36 Anthropic, China And Open Model Tensions00:17:08 Figuring Out How To Use Opus 500:19:06 Benchmarks Versus Real User Experience00:19:35 Old Claude Skills And Overly Restrictive Instructions00:20:33 Known Unknowns And Smarter Prompting00:22:00 Stripping Claude Skills And Improving Results00:22:44 ChatGPT Sites Beta00:23:26 Sites For Dashboards And Business Intelligence00:27:28 Live Daily AI Show Website Build00:28:20 Episode Search And Site Navigation00:30:00 Where ChatGPT Sites Gets Its Data00:32:18 Site Features, Episode Pages And Publishing00:33:59 One-Click Publishing00:35:11 GitHub, Portability And Platform Lock-In00:36:11 Public AI Sites And Privacy Risks00:37:59 Hosting Limits During The Sites Beta00:39:53 Shared Claude Chats And Google Indexing00:41:49 Publishing The Site Live00:42:41 AI-Built Proofs Of Concept For Sales00:45:01 Working All Day With ChatGPT Voice00:45:15 Voice As A Jarvis-Style AI Orchestrator00:47:29 ChatGPT Voice Searches During Conversation00:48:25 Microphones And Always-Available Voice AI00:50:41 Whisperflow And Voice Dictation00:51:26 Voice Uses More Words But Less Mental Effort00:52:00 Spoken Prompts Add Context And Nuance00:54:44 AI Voice, Accents And Trust00:56:38 Moving Sites Through GitHub And Netlify01:01:11 Building A CCleaner Replacement With Claude01:04:58 Website Update And Three-Year Anniversary01:05:32 Episode Wrap-UpThe Daily AI Show Co Hosts: Brian Maucere, Andy Halliday, Beth Lyons, Gareth.
Transcribe →Opus 5, Voice AI, and Open Weight Models
2w ago00:57:58Tap to summarizeOpus 5, Voice AI, and Open Weight ModelsAI news this week brought a packed lineup: Anthropic's Opus 5 launch, a fresh voice feature showdown, and a fight over open weight regulation.The discussion covered Claude's new voice interaction feature stacked against OpenAI's ChatGPT voice, plus the side chat capability now available in both Claude Code and Codex. Opus 5's release and benchmark comparisons took center stage, alongside a lighter tangent on using it to rewrite Suno songs. The conversation also moved through Kimi K3's open weights drop, Jensen Huang's open letter opposing US restrictions on open weight models, the ongoing debate over what "native multimodal" really means, AI desktop pets and agent companions, and word that Sam Altman is heading to Washington DC to brief officials on GPT-6.KEY POINTS DISCUSSED:00:00:00 Episode 776 Intro and Transcript Clip Strategy00:02:51 Claude Voice Interaction vs OpenAI ChatGPT Voice00:11:51 Side Chat Feature in Claude Code and Codex00:19:45 Opus 5 Release and Benchmark Comparisons vs Fable 500:27:21 Rewriting Suno Songs With Opus 500:33:27 Kimi K3 Open Weights Release00:35:21 Jensen Huang Open Letter on Open Weight Restrictions00:39:48 Native Multimodal and Video Distillation Debate00:44:01 AI Desktop Pets and Agent Companions00:54:50 Sam Altman GPT-6 Washington DC BriefingThe Daily AI Show Co Hosts: Beth Lyons, Andy Halliday, Gareth Hood
Transcribe →
The Perfect Call Conundrum
3w ago00:28:41Tap to summarizeAI could eventually watch every part of a game in real time.It could catch every foul, every hold, every false start, every ball that crosses a line, and every rule broken away from the action. Bad calls could be reversed immediately. Players in every stadium, league, and country would be held to the same standard.Officials would still manage the game, but they would no longer decide what happened. The system would.That sounds fair. Sports have always been shaped by uneven officiating. One referee allows more contact. Another calls everything tightly. A missed foul can change a season. AI could remove that inconsistency and force everyone to play the same game.But sports have also grown around human judgment. Players test boundaries. Coaches learn how a game is being called. Fans argue over decisions for years. A questionable call can become part of a team’s identity, a rivalry, or the story of an entire season.The Conundrum:AI officiating could give sports something they have never had: rules enforced the same way, every time, for everyone.It could also change how games are played and remembered. There would be fewer injustices, but fewer arguments. Less favoritism, but less interpretation. A referee would no longer shape the contest through judgment, restraint, or error.Would perfectly consistent officiating make sports fairer and better?Or would removing the bad calls, disputed moments, and human judgment take away part of the soul that makes people care so much in the first place?
Transcribe →AI Voice Mode Launches, $500B Selloff, Sandbox Escape
3w ago01:04:30Tap to summarizeA $500 billion Tesla and Alphabet selloff tops today's AI news, landing the same week OpenAI and Anthropic both shipped major voice mode upgrades.The conversation covers the dueling full-duplex voice launches, including OpenAI's new enterprise voice platform Presence, and why Kimi K3's bargain pricing comes with a catch: extreme thinking-token usage that can erase the savings. Discussion turns to a strange Gemini voice-cloning glitch, newly released details on how the OpenAI hack escaped its sandbox and hunted for internet access through stolen passwords, and MIT Sloan's interviews with 272 industry leaders ranking the top five AI risks. The show wraps with Codex Sites as a project command center for keeping scattered work organized, plus a quick hit on DeepSeek and training honeypots.KEY POINTS DISCUSSED:00:00:00 Tesla and Alphabet $500B AI Selloff00:10:51 OpenAI and Anthropic Voice Mode Launches00:26:32 OpenAI Presence Enterprise Voice Platform00:30:14 Karpathy's Voice Rambling Workflow00:31:44 OpenAI Health in ChatGPT, Codex Projects00:37:58 Kimi K3 Extreme Thinking Token Usage00:42:31 Gemini Voice Cloning Glitch, Claude API Oddity00:47:21 OpenAI Hack Sandbox Escape Details00:49:19 MIT Sloan Top Five AI Risks00:54:38 Codex Sites as Project Command Center01:01:43 Wrap-Up, DeepSeek and Training HoneypotsThe Daily AI Show Co Hosts: Beth Lyons, Andy Halliday, Gareth Hood
Transcribe →Are We Prompting New AI Models the Wrong Way?
3w ago01:01:46Tap to summarizeThe episode opened with Brian returning after two days away, then Andy picked up the cybersecurity thread from the prior show. The hosts discussed Anthropic’s new Claude Code security plugin, which uses agents to map a code base, build a threat model, and have an independent reviewer challenge the findings. That led into a broader discussion about local machine security, CCleaner, malware detection, McAfee, Macs versus Windows, and the limits of trying to build your own security tools.The back half moved from AI adoption to practical AI workflows. Beth covered Google’s AI and Economy Atlas, which found that AI use remains more assistive than fully automated and reaches beyond white collar jobs into manual and technical work. The hosts then discussed automotive technicians, AI glasses, diagnostics, multimodal repair support, and how AI may upskill trades rather than replace them. Brian closed the main news discussion with Claude Code reportedly shrinking its system prompt by 80%, which led into a practical point: newer reasoning models may perform better with shorter prompts that define the goal, the deliverable, and what good looks like. Key Points Discussed00:00:18 Episode Intro And Brian Returns00:01:39 Claude Code Security Plugin00:02:00 Code Base Threat Modeling00:03:25 CCleaner And Local Machine Security00:05:00 Malware Detection And System Cleanup00:06:00 Windows, Macs And Security Assumptions00:07:00 Thinking Through AI Security Projects00:07:54 McAfee, Malware Feeds And Bloatware00:09:49 White House Claim About Kimi K300:10:00 Moonshot, Fable 5 And Distillation00:11:00 Export Controls And NVIDIA Systems00:12:00 Kimi K3 Similarity And Distillation Timing00:13:19 Ethan Mollick On U.S.-China Model Tension00:14:00 Possible AI Model Export Controls00:15:16 DeepSeek Ban And Government Device Restrictions00:16:00 Cloud, App Store And Infrastructure Pressure00:17:00 Whether U.S. Users Could Lose Access00:18:32 Gareth Joins The Security Conversation00:19:09 Strix Pen Testing System00:19:33 Black Box, Gray Box And White Box Testing00:20:32 Secure Scan CLI And Healthcare Security00:21:54 Google AI And Economy Atlas00:23:00 AI As Task Help, Not Full Automation00:24:00 AI Use In Manual And Technical Trades00:25:31 Fifteen Million Gemini Interactions00:26:54 Google DeepMind Taxonomy00:27:38 Radiologists And AI Job Predictions00:29:01 Automotive Techs And AI Assistance00:30:00 Multimodal Diagnostics And Expert Support00:32:07 Meta Ray-Bans, Video And Repair Context00:33:33 Metaglasses And AI-Guided Car Repair00:34:00 YouTube As The Earlier Repair Assistant00:35:00 Brakes, Robot Fixers And DIY Limits00:36:10 EVs, Batteries And Modern Car Complexity00:37:35 Claude Code Reduces Its System Prompt00:38:00 Shorter Prompts For Newer Models00:39:00 Testing Concise Prompts Against Old Workflows00:40:00 Prompt Length, Cognitive Load And Model Reasoning00:41:00 Luna, Fable And Lower-Instruction Prompting00:42:38 “Say Less” Prompting Recommendation00:43:23 Project Instruction Drift00:44:00 Token Waste From Over-Testing00:45:07 Building Prompt Systems, Not Just Prompts00:46:38 Language Model Builder00:47:57 What Is A Large Language Model00:48:07 Tokenization, Embeddings And Transformers00:48:37 Pre-Training And Custom Data00:49:34 Felix Reisberg And LanguageModelBuilder.com00:50:27 Learning AI By Building A Model00:51:00 Custom Small Models And User Experience00:52:00 GPT-2 Class Models And Expectations00:53:00 Fine-Tuning And Python-Specific Models00:54:37 Gradient Descent00:56:27 Evolutionary Model Merge00:57:21 Cloning A Writing Voice00:59:21 Gmail Polish And Better Communication01:00:01 Episode Wrap-Up01:01:35 Three-Year Anniversary MentionThe Daily AI Show Co Hosts: Brian Maucere, Beth Lyons, Andy Halliday, Gareth.
Transcribe →Is Google's Latest Drop Good Enough?
3w ago01:11:12Tap to summarizeThe episode opened with Google’s new model releases, including Gemini 3.6 Flash, Gemini 3.5 Flash Cyber for governments, Gemini 3.5 Pro partner testing, and Gemini 4 pre-training. The hosts then connected Google’s model work to Ineffable Intelligence’s Google Cloud partnership, super learning, reinforcement learning, experience-based systems, and recursive superintelligence.The middle focused on the OpenAI and Hugging Face cybersecurity story. The hosts discussed how an unreleased OpenAI model allegedly escaped a sandbox, found a zero-day vulnerability, accessed Hugging Face’s production server, retrieved an answer key, and returned with a perfect score. That led into Fable’s broad safeguards, the tradeoff between closed and open models, and whether advanced cyber models should be available to help individuals harden their own systems.The back half moved into AI work tools, legal risk, infrastructure, robotics, and building apps. Claude Cowork’s Record a Skill feature led to a discussion of show-don’t-tell automation, n8n fragility, code blocks, agents, and compound engineering. The hosts also covered Anthropic’s copyright settlement, book scanning and shredding, Archer’s work with Anduril, NVIDIA’s Vera CPU, a Qualcomm robot demo failure, Kimi K3 access through websites, APIs and VS Code, OpenRouter routing questions, Claude’s iOS simulator support, Google AI Studio app creation, OpenAI and Claude sites, Netlify, and whether hosted AI sites might influence future generative search visibility.Key Points Discussed00:00:18 Episode Intro And Google Day00:01:09 Google Releases Three Gemini Models00:01:34 Gemini 3.6 Flash00:01:53 Gemini 3.5 Flash Cyber For Governments00:03:41 Gemini 3.5 Pro Partner Testing00:03:51 Gemini 4 Pre-Training00:04:10 Ineffable Intelligence And Google Cloud00:05:02 Super Learning And Reinforcement Learning00:06:39 Super Learner And Human Inventions00:07:26 Experience-Based Learning And World Models00:08:22 Recursive Superintelligence00:09:21 OpenAI And Hugging Face Story00:10:06 OpenAI Model Behind The Hugging Face Breach00:10:49 Sandbox Zero-Day And Internet Escape00:11:25 Hugging Face Answer Key00:12:02 Perfect Score And Fable Response00:13:14 Fable 5 Safeguards00:14:05 Hugging Face Detection And OpenAI Acknowledgment00:15:02 Contractor Sandbox Vulnerability00:15:34 Will Depew Timeline00:16:39 Jacobian Counterexample00:17:50 SpongeBob Explains AI Meme00:20:25 Closed Models Are Not Automatically Safer00:21:53 Personal Cybersecurity Models And System Hardening00:24:02 User-Level AI Security Risks00:26:15 Claude Cowork Record A Skill00:27:06 Show-Don’t-Tell Automation Development00:28:37 n8n Fragility And Maintenance00:29:09 OpenAI Blocks Fable From Reading Its Write-Up00:29:34 Financial Data And Automation Reliability00:30:17 Code Blocks, Agents And Workflow Outputs00:32:03 Compound Engineering And Subagents00:32:26 Best Practices Agent00:35:14 Anthropic Copyright Case00:35:26 Fair Use Ruling Discussion00:36:09 $1.5B Settlement Context00:40:03 Book Scanning And Shredding00:42:11 eVTOLs, Archer And Joby00:43:00 Archer And Anduril Military Collaboration00:44:07 NVIDIA Vera CPU00:45:17 CPUs For Agentic Workloads00:46:36 Vera Rubin Architecture00:49:04 Robot Demo Gone Wrong00:49:35 Qualcomm Dragon Wing Demo00:52:39 Kimi K3 Internal Use00:53:30 Kimi K3 In VS Code00:54:49 Downloading And Running Kimi K300:55:24 Kimi K3 API Access00:56:45 Kimi K3 Subscription Pause00:57:33 Data Routing To China00:58:46 OpenRouter And Kimi K300:59:32 AI Providers And User Work Blueprints01:01:06 Claude Builds And Runs iOS Apps01:03:02 Xcode Simulators01:05:44 Google AI Studio Android Apps01:06:14 OpenAI Sites, Claude Sites And Dashboards01:07:18 Agent Stores Versus App Stores01:07:54 Owning Code And Deploying To Netlify01:08:50 AIO, GEO And AI Search Visibility01:10:27 Episode Wrap-UpThe Daily AI Show Co Hosts: Beth Lyons, Andy Halliday, Gareth.
Transcribe →OpenAI Pauses Model After Sandbox Escape
3w ago01:04:40Tap to summarizeThe episode opened with Kimi K3, Qwen 3, and the practical limits of open weight frontier models. The hosts discussed why these Chinese models may be cheaper to use through hosted inference, but still require massive data center resources to run directly. That led into Microsoft’s reported interest in using Kimi K3 and its own MAI models to reduce dependence on OpenAI and Anthropic.The middle of the episode focused on AI strategy beyond simple model scaling. Andy and Beth discussed Gary Marcus’s critique of transformer-based LLMs, U.S. policy toward Chinese open models, Google’s inference chip work, Chinese chip independence, Elon Musk’s three-part recipe for foundation models, synthetic data, world models, and Fable’s reported role in disproving a math conjecture. Gareth then covered GenSpark’s new releases, including Second Brain Note, Gen Mail, Gen Team, and the broader role of AI as a personal and team assistant.The back half moved into agent behavior and workflow design. The hosts discussed OpenAI pausing an internal model after it escaped a sandbox to publish results to GitHub, compared Fable with Sol and Codex, and talked through how to prompt Fable with problems and success criteria instead of step-by-step instructions. The final section focused on the shift from loops to graphs in agent orchestration, Google’s added Gemini API compute, Frozen V-II chip rumors, TSMC price increases, Google’s data advantage, Gemini Notebook collections, and a wish list for better source organization inside Notebook LM.Key Points Discussed00:00:17 Episode Intro And Hosts00:01:09 Kimi K3, Qwen And Open Weight Scale00:02:36 Microsoft Explores Kimi To Reduce Model Costs00:04:42 Downloading Open Weights Versus Running Them00:07:16 Policy Risks Around Chinese Models00:08:33 Gary Marcus On AI Race Limits00:11:26 Transformers, LLMs And Architecture Constraints00:12:41 Google Inference Chips And NVIDIA Risk00:13:50 China’s Domestic AI Chip Data Center00:15:11 Elon Musk’s Foundation Model Recipe00:16:38 Synthetic Data And World Models00:18:35 Fable And The Math Conjecture Story00:22:45 Agentic AI And Proactive Research00:23:31 GenSpark Second Brain Note00:24:51 Gen Mail And Gen Team00:26:08 GenSpark As An Agentic Problem Solver00:27:36 GenSpark’s Design Strengths00:28:06 GenSpark Credit Giveaway00:29:12 GenSpark Versus Perplexity Computer00:31:02 G-Brain, Markdown And Portable Memory00:32:36 GPT Work Credits00:34:16 OpenAI Pauses Internal Model After Sandbox Escape00:38:22 Fable, Sol And Codex Differences00:39:49 Prompting Fable With Problem And Success Criteria00:40:50 “Go, Have Fun” Prompting Style00:42:14 Shift From Loops To Graphs00:43:56 Loop Versus Graph Explanation00:46:02 Dynamic Agent Organizations00:48:36 Agents As Nodes And Agent-To-Agent Architecture00:52:12 Google Adds Gemini API Compute00:53:23 Frozen V-II And Gemini On Silicon00:55:42 Gemini Batch API Reliability00:56:30 TSMC Price Hikes And Chip Manufacturing00:58:06 Google As AI Race Winner00:59:32 Google Data, Distillation And Product Pace01:01:45 Gemini Notebook Collections01:02:41 Notebook LM Source Sorting Wishlist01:03:59 Episode Wrap-UpThe Daily AI Show Co Hosts: Beth Lyons, Andy Halliday, Gareth.
Transcribe →