OpenAI’s Shift, Nvidia’s Speed, Apple’s AI Gambit (Ep. 458)

We already intervene. We screen embryos. We correct mutations. We remove risks that used to define someone’s fate. No one says that child is less human. In fact, we celebrate it—saving a life before it suffers.So what’s the line? Is it when we shift from preventing harm to increasing potential? From fixing broken code to writing better code? And if AI is the system showing us how to make those changes—faster, cheaper, more precisely—does that make it the author of our evolution, or just the pen in our hand?Here’s an updated conundrum that leans into exactly that tension:The conundrumWe already use science to help humans suffer less—so if AI shows us how to go further, to make humans stronger, smarter, more adaptable, do we follow its lead without hesitation? Or is there a point where those changes reshape us so deeply that we lose something essential—and is it AI that crosses the line, or us?Maybe the real question isn’t what AI is capable of.It’s whether we’ll recognize the moment when human stops meaning what it used to—and whether we’ll care when it happens.This podcast is created by AI. We used ChatGPT, Perplexity and Google NotebookLM's audio overview to create the conversation you are hearing. We do not make any claims to the validity of the information provided and see this as an experiment around deep discussions fully generated by AI.How this content was made

Want to keep the conversation going?Join our Slack community at thedailyaishowcommunity.comWhat started as a simple “let’s think step by step” trick has grown into a rich landscape of reasoning models that simulate logic, branch and revise in real time, and now even collaborate with the user. The episode explores three specific advancements: speculative chain of thought, collaborative chain of thought, and retrieval-augmented chain of thought (CoT-RAG).Key Points DiscussedChain of thought prompting began in 2022 as a method for improving reasoning by asking models to slow down and show their steps.By 2023, tree-of-thought prompting and more branching logic began emerging.In 2024, tools like DeepSeek and O3 showed dynamic reasoning with visible steps, sparking renewed interest in more transparent models.Andy explains that while chain of thought looks like sequential reasoning, it’s really token-by-token prediction with each output influencing the next.The illusion of “thinking” is shaped by the model’s training on step-by-step human logic and clever UI elements like “thinking…” animations.Speculative chain of thought uses a smaller model to generate multiple candidate reasoning paths, which a larger model then evaluates and improves.Collaborative chain of thought lets the user review and guide reasoning steps as they unfold, encouraging transparency and human oversight.Chain of Thought RAG combines structured reasoning with retrieval, using pseudocode-like planning and knowledge graphs to boost accuracy.Jyunmi highlighted how collaborative CoT mirrors his ideal creative workflow by giving humans checkpoints to guide AI thinking.Beth noted that these patterns often mirror familiar software roles, like sous chef and head chef, or project management tools like Gantt charts.The team discussed limits to context windows, attention, and how reasoning starts to break down with large inputs or long tasks.Several ideas were pitched for improving memory, including token overlays, modular context management, and step weighting.The conversation wrapped with a reflection on how each CoT model addresses different needs: speed, accuracy, or collaboration.Timestamps & Topics00:00:00 🧠 What is Chain of Thought evolved?00:02:49 📜 Timeline of CoT progress (2022 to 2025)00:04:57 🔄 How models simulate reasoning00:09:36 🤖 Agents vs LLMs in CoT00:14:28 📚 Research behind the three CoT variants00:23:18 ✍️ Overview of Speculative, Collaborative, and RAG CoT00:25:02 🧑‍🤝‍🧑 Why collaborative CoT fits real-world workflows00:29:23 📌 Brian highlights human-in-the-loop value00:32:20 ⚙️ CoT-RAG and pseudo-code style logic00:34:35 📋 Pretraining and structured self-ask methods00:41:11 🧵 Importance of short-term memory and chat history00:46:32 🗃️ Ideas for modular memory and reg-based workflows00:50:17 🧩 Visualizing reasoning: Gantt charts and context overlays00:52:32 ⏱️ Tradeoffs: speed vs accuracy vs transparency00:54:22 📬 Wrap-up and show announcementsHashtags#ChainOfThought #ReasoningAI #AIprompting #DailyAIShow #SpeculativeAI #CollaborativeAI #RetrievalAugmentedGeneration #LLMs #AIthinking #FutureOfAIThe Daily AI Show Co-Hosts: Andy Halliday, Beth Lyons, Brian Maucere, Eran Malloch, Jyunmi Hatcher, and Karl Yeh

Want to keep the conversation going?Join our Slack community at thedailyaishowcommunity.comInstead of learning solely from human data or pretraining, AI models are beginning to learn from real-world experiences. These systems build their own goals, interact with their environments, and improve through self-directed feedback loops, pushing AI into a more autonomous and unpredictable phase.Key Points DiscussedDeepMind proposes we’ve moved from simulated learning to human data, and now to AI-driven experiential learning.The new approach allows AI to learn from ongoing experience in real-world or simulated environments, not just from training datasets.AI systems with memory and agency will create feedback loops that accelerate learning beyond human supervision.The concept includes agents that actively seek out human input, creating dynamic learning through social interaction.Multimodal experience (e.g., visual, sensory, movement) will become more important than language alone.The team discussed Yann LeCun’s belief that current models won’t lead to AGI and that chaotic or irrational human behavior may never be fully replicable.A major concern is alignment: what if the AI’s goals, derived from its own experience, start to diverge from what’s best for humans?The conversation touched on law enforcement, predictive policing, and philosophical implications of free will vs. AI-generated optimization.DeepMind's proposed bi-level reward structure gives low-level AIs operational goals while humans oversee and reset high-level alignment.Memory remains a bottleneck for persistent context and cross-session learning, though future architectures may support long-term, distributed memory.The episode closed with discussion of a decentralized agent-based future, where thousands of specialized AIs work independently and collaboratively.Timestamps & Topics00:00:00 🧠 What is the “Era of Experience”?00:01:41 🚀 Self-directed learning and agency in AI00:05:02 💬 AI initiating contact with humans00:06:17 🐶 Predictive learning in animals and machines00:12:17 🤖 Simulation era to human data to experiential learning00:14:58 ⚖️ The upsides and risks of reinforcement learning00:19:27 🔮 Predictive policing and the slippery slope of optimization00:24:28 💡 Human brains as predictive machines00:26:50 🎭 Facial cues as implicit feedback00:31:03 🧭 Realigning AI goals with human values00:34:03 🌍 Whose values are we aligning to?00:36:01 🌊 Tradeoffs between individual vs collective optimization00:40:24 📚 New ways to interact with AI papers00:43:10 🧠 Memory and long-term learning00:48:48 📉 Why current memory tools are falling short00:52:45 🧪 Why reinforcement learning took longer to catch on00:56:12 🌐 Future vision of distributed agent ecosystems00:58:04 🕸️ Global agent networks and communication protocols00:59:31 📢 Announcements and upcoming shows#EraOfExperience #DeepMind #AIlearning #AutonomousAI #AIAlignment #LLM #EdgeAI #AIAgents #ReinforcementLearning #FutureOfAI #ArtificialIntelligence #DailyAIShowThe Daily AI Show Co-Hosts: Andy Halliday, Beth Lyons, Brian Maucere, Eran Malloch, Jyunmi Hatcher, and Karl Yeh

Want to keep the conversation going?Join our Slack community at thedailyaishowcommunity.comIt’s Wednesday, which means it’s news day on The Daily AI Show. The hosts break down the top AI headlines from the week, including OpenAI’s corporate restructuring, Google’s major update to Gemini Pro 2.5, and Hugging Face releasing an open source alternative to Operator. They also dive into science stories, education initiatives, and new developments in robotics, biology, and AI video generation.Key Points DiscussedGoogle dropped an updated Gemini 2.5 Pro with significantly improved coding benchmarks, outperforming Claude in multiple categories.OpenAI confirmed its shift to a Public Benefit Corporation structure, sparking responses from Microsoft and Elon Musk.OpenAI also acquired Codium (now Windsurf), boosting its in-house coding capabilities to compete with Cursor.Apple and Anthropic are working together on a vibe coding platform built around Apple’s native ecosystem.Hugging Face released a free, open source Operator alternative, now in limited beta queue.250 tech CEOs signed an open letter calling for AI and computer science to be mandatory in US K-12 education.Google announced new training programs for electricians to support the infrastructure demands of AI expansion.Nvidia launched Parakeet 2, an open source automatic speech recognition model that transcribes audio at lightning speed and with strong accuracy.Future House, backed by Eric Schmidt, previewed new tools in biology for building an AI scientist.Northwestern University released new low-cost robotic touch sensors for embodied AI.University of Tokyo introduced a decentralized AI system for smart buildings that doesn’t rely on centralized servers.A new model from the University of Rochester uses time-lapse video to simulate real-world physics, marking a step toward world models in AI.Timestamps & Topics00:00:00 🗞️ AI Weekly News Kickoff00:01:15 💻 Google Gemini 2.5 Pro update00:05:32 🏛️ OpenAI restructures as a Public Benefit Corporation00:07:59 ⚖️ Microsoft, Musk respond to OpenAI's move00:09:13 📊 Gemini 2.5 Pro benchmark breakdown00:14:45 🍎 Apple and Anthropic’s coding platform partnership00:18:44 📉 Anthropic offering share buybacks00:22:03 🤝 Apple to integrate Claude and Gemini into its apps00:22:52 🧠 Hugging Face launches free Operator alternative00:25:04 📚 Tech leaders call for mandatory AI education00:28:42 🔌 Google announces training for electricians00:34:03 🔬 Future House previews AI for biology research00:36:08 🖐️ Northwestern unveils new robotic touch sensors00:39:10 🏢 Decentralized AI for smart buildings from Tokyo00:43:18 🐦 Nvidia launches Parakeet 2 for speech recognition00:52:30 🎥 Rochester’s “Magic Time” trains AI with time-lapse physics#AInews #OpenAI #Gemini25 #Anthropic #HuggingFace #VibeCoding #AppleAI #EducationReform #AIinfrastructure #Parakeet2 #FutureHouse #AIinScience #Robotics #WorldModels #LLMs #AItools #DailyAIShowThe Daily AI Show Co-Hosts: Andy Halliday, Beth Lyons, Brian Maucere, Eran Malloch, Jyunmi Hatcher, and Karl Yeh

Want to keep the conversation going?Join our Slack community at thedailyaishowcommunity.comIs vertical SaaS in trouble? With AI agents rapidly evolving, the traditional SaaS model built around dashboards, workflows, and seat-based pricing faces real disruption. The hosts explored whether legacy SaaS companies can defend their turf or if leaner, AI-native challengers will take over.Key Points DiscussedAI agents threaten vertical SaaS by eliminating the need for rigid interfaces and one-size-fits-all workflows.Karl outlined three forces converging: vibe coding, vertical agents, and AI-enabled company-building without heavy headcount.Major SaaS players like Veeva, Toast, and ServiceTitan benefit from strong moats like network effects, regulatory depth, and proprietary data.The group debated how far AI can go in breaking these moats, especially if agents gain access to trusted payment rails like Visa's new initiative.AI may enable smaller companies to build fully customized software ecosystems that bypass legacy tools.Andy emphasized Metcalfe’s Law and customer acquisition costs as barriers to AI-led disruption in entrenched verticals.Beth noted the tension between innovation and trust, especially when agents begin handling sensitive operations or payments.Visa's announcement that agents will soon be able to make payments opens the door to AI-driven purchasing at scale.Discussion wrapped with a recognition that change will be uneven across industries and that agent adoption could push companies to rethink staffing and control.Timestamps & Topics00:00:00 🔍 Vertical SaaS under siege00:01:33 🧩 Three converging forces disrupting SaaS00:05:15 🤷 Why most SaaS tools frustrate users00:06:44 🧭 Horizontal vs vertical SaaS00:08:12 🏥 Moats around Veeva, Toast, and ServiceTitan00:12:27 🌐 Network effects and proprietary data00:14:42 🧾 Regulatory complexity in vertical SaaS00:16:25 💆 Mindbody as a less defensible vertical00:18:30 🤖 Can AI handle compliance and integrations?00:21:22 🏗️ Startups building with AI from the ground up00:24:18 💳 Visa enables agents to make payments00:26:36 ⚖️ Trust and data ownership00:27:46 📚 Training, interfaces, and transition friction00:30:14 🌀 The challenge of dynamic AI tools in static orgs00:33:14 🌊 Disruption needs adaptability00:35:34 🏗️ Procore and Metcalfe’s Law00:37:21 🚪 Breaking into legacy-dominated markets00:41:16 🧠 Agent co-ops as a potential breakout path00:43:40 🧍 Humans, lemmings, and social proof00:45:41 ⚖️ Should every company adopt AI right now?00:48:06 🧪 Prompt engineering vs practical adoption00:49:09 🧠 Visa’s agent-payment enablement recap00:52:16 🧾 Corporate agents and purchasing implications00:54:07 📅 Preview of upcoming shows#VerticalSaaS #AIagents #DailyAIShow #SaaSDisruption #AIstrategy #FutureOfWork #VisaAI #AgentEconomy #EnterpriseTech #MetcalfesLaw #AImoats #Veeva #ToastPOS #ServiceTitan #StartupTrends #YCombinatorThe Daily AI Show Co-Hosts: Andy Halliday, Beth Lyons, Brian Maucere, Eran Malloch, Jyunmi Hatcher, and Karl Yeh

Want to keep the conversation going?Join our Slack community at thedailyaishowcommunity.comToday the hosts unpack a fictional but research-informed essay titled AI-2027. The essay lays out a plausible scenario for how AI could evolve between now and the end of 2027. Rather than offering strict predictions, the piece explores a range of developments through a branching narrative, including the risks of unchecked acceleration and the potential emergence of agent-based superintelligence. The team breaks down the paper’s format, the ideas behind it, and its broader implications.Key Points DiscussedThe AI-2027 essay is a scenario-based interactive website, not a research paper or report.It uses a timeline narrative to show how AI agents evolve into increasingly autonomous and powerful systems.The fictional company “Open Brain” represents the leading AI organization without naming names like OpenAI.The model highlights a “choose your path” divergence at the end, with one future of acceleration and another of restraint.The essay warns of agent models developing faster than humans can oversee, leading to loss of interpretability and oversight.Authors acknowledge the speculative nature of post-2026 predictions, estimating outcomes could move 5 times faster or slower.The group behind the piece, AI Futures Project, includes ex-OpenAI and AI governance experts who focus on alignment and oversight.Concerns raised about geopolitical competition, lack of global cooperation, and risks tied to fast-moving agentic systems.The essay outlines how by mid-2027, agent models could reach a tipping point, massively disrupting white-collar work.Key moment: The public release of Agent 3 Mini signals the democratization of powerful AI tools.The discussion reflects on how AI evolution may shift from versioned releases to continuous, fluid updates.Hosts also touch on the emotional and societal implications of becoming obsolete in the face of accelerating AI capability.The episode ends with a reminder that alignment, not just capability, will be critical as these systems scale.Timestamps & Topics00:00:00 💡 What is AI-2027 and why it matters00:02:14 🧠 Writing style and first impressions of the scenario00:03:08 🌐 Walkthrough of the AI-2027.com interactive timeline00:05:02 🕹️ Gamified structure and scenario-building approach00:08:00 🚦 Diverging futures: full-speed ahead vs. slowdown00:10:10 📉 Forecast accuracy and the 5x faster or slower disclaimer00:11:16 🧑‍🔬 Who authored this and what are their credentials00:14:22 🇨🇳 US-China AI race and geopolitical implications00:18:20 ⚖️ Agent hierarchy and oversight limits00:22:07 🧨 Alignment risks and doomsday scenarios00:23:27 🤝 Why global cooperation may not be realistic00:29:14 🔁 Continuous model evolution vs. versioned updates00:34:29 👨‍💻 Agent 3 Mini released to public, tipping point reached00:38:12 ⏱️ 300k agents working at 40x human speed00:40:05 🧬 Biological metaphors: AI evolution vs. cancer00:42:01 🔬 Human obsolescence and emotional impact00:45:09 👤 Daniel Kokotajlo and the AI Futures Project00:47:15 🧩 Other contributors and their focus areas00:48:02 🌍 Why alignment, not borders, should be the focus00:51:19 🕊️ Idealistic endnote on coexistence and AI ethicsHashtags#AI2027 #AIAlignment #AIShow #FutureOfAI #AGI #ArtificialIntelligence #AIAgents #TechForecast #DailyAIShow #OpenAI #AIResearch #Governance #SuperintelligenceThe Daily AI Show Co-Hosts: Andy Halliday, Beth Lyons, Brian Maucere, Eran Malloch, Jyunmi Hatcher, and Karl Yeh

This podcast is created by AI. We used ChatGPT, Perplexity and Google NotebookLM's audio overview to create the conversation you are hearing. We do not make any claims to the validity of the information provided and see this as an experiment around deep discussions fully generated by AI.How this content was made

In this special two-week recap, the team covers major takeaways across episodes 445 to 454. From Meta’s plan to kill creative agencies, to OpenAI’s confusing model naming, to AI’s role in construction site inspections, the discussion jumps across industries and implications. The hosts also share real-world demos and reveal how they’ve been applying 4.1, O3, Gemini 2.5, and Claude 3.7 in their work and lives.Key Points DiscussedMeta's new AI ad platform removes the need for targeting, creative, or media strategy – just connect your product feed and payment.OpenAI quietly rolled out 4.1, 4.1 mini, and 4.1 nano – but they’re only available via API, not in ChatGPT yet.The naming chaos continues. 4.1 is not an upgrade to 4.0 in ChatGPT, and 4.5 has disappeared. O3 Pro is coming soon and will likely justify the $200 Pro plan.Cost comparisons matter. O3 costs 5x more than 4.1 but may not be worth it unless your task demands advanced reasoning or deep research.Gemini 2.5 is cheaper, but often stops early. Claude 3.7 Sonnet still leads in writing quality. Different tools for different jobs.Jyunmi reminds everyone that prompting is only part of the puzzle. Output varies based on system prompts, temperature, and even which “version” of a model your account gets.Brian demos his “GTM Training Tracker” and “Jake’s LinkedIn Assistant” – both built in ~10 minutes using O3.Beth emphasizes model evaluation workflows and structured experimentation. TypingMind remains a great tool for comparing outputs side-by-side.Carl shares how 4.1 outperformed Gemini 2.5 in building automation agents for bid tracking and contact research.Visual reasoning is improving. Models can now zoom in on construction site photos and auto-flag errors – even without manual tagging.Hashtags#DailyAIShow #OpenAI #GPT41 #Claude37 #Gemini25 #PromptEngineering #AIAdTools #LLMEvaluation #AgenticAI #APIAccess #AIUseCases #SalesAutomation #AIAssistantsTimestamps & Topics00:00:00 🎬 Intro – What happened across the last 10 episodes?00:02:07 📈 250,000 views milestone00:03:25 🧠 Zuckerberg’s ad strategy: kill the creative process00:07:08 💸 Meta vs Amazon vs Shopify in AI-led commerce00:09:28 🤖 ChatGPT + Shopify Pay = frictionless buying00:12:04 🧾 The disappearing OpenAI models (where’s 4.5?)00:14:40 💬 O3 vs 4.1 vs 4.1 mini vs nano – what’s the difference?00:17:52 💸 Cost breakdown: O3 is 5x more expensive00:19:47 🤯 Prompting chaos: same name, different models00:22:18 🧪 Model testing frameworks (Google Sheets, TypingMind)00:24:30 📊 Temperature, randomness, and system prompts00:27:14 🧠 Gemini’s weird early stop behavior00:30:00 🔄 API-only models and where to access them00:33:29 💻 Brian’s “Go-To-Market AI Coach” demo (built with O3)00:37:03 📊 Interactive learning dashboards built with AI00:40:12 🧵 Andy on persistence and memory inside O3 sessions00:42:33 📈 Salesforce-style dashboards powered by custom agents00:44:25 🧠 Echo chambers and memory-based outputs00:47:20 🔍 Evaluating AI models with real tasks (sub-industry tagging, research)00:49:12 🔧 Carl on building client agents for RFPs and lead discovery00:52:01 🧱 Construction site inspection – visual LLMs catching build errors00:54:21 💡 Ask new questions, test unknowns – not just what you already know00:57:15 🎯 Model as a coworker: ask it to critique your slides, GTM plan, or positioning00:59:35 🧪 Final tip: prime the model with fresh context before prompting01:01:00 📅 Wrap-up: “Be About It” demo shows return next Friday + Sci-Fi show tomorrow

Want to keep the conversation going?Join our Slack community at dailyaishowcommunity.com“Better prompts make better results” has been a guiding mantra, but what if that’s not always true? On today’s episode, the team digs into new research by Ethan Mollick and others suggesting that polite phrasing, excessive verbosity, or emotional tricks may not meaningfully improve LLM responses. The discussion shifts from prompt structure to AI memory, model variability, and how personality may soon dominate how models respond to each of us.Key Points DiscussedEthan Mollick’s research at Wharton shows that small prompt changes like politeness or emotional urgency do not reliably improve performance across many model runs.Andy explains compiled prompts: the user prompt is just one part. System prompts, developer prompts, and memory all shape model outputs.Temperature and built-in randomness ensure variation even with identical prompts. This challenges the belief that minor phrasing tweaks will deliver consistent gains.Beth pushes back on "accuracy" as the primary measure. For many creative or reflective workflows, success is about alignment, not factual correctness.Brian shares frustrations with inconsistent outputs and highlights the value of a mixture-of-experts system to improve reliability for fact-based tasks like identifying sub-industries.Jyunmi notes that polite prompting may not boost accuracy but helps preserve human etiquette. Saying “please” and “thank you” matters for human-machine culture.The group explores AI memory and personality. With more models learning from user interactions, outputs may become increasingly personalized, creating echo chambers.OpenAI CEO Sam Altman said polite prompts increase token usage and inference costs, but the company keeps them because they improve user experience.Andy emphasizes the importance of structured prompts. Asking for a specific output format remains one of the few consistent ways to boost performance.The conversation expands to implications: Will models subtly nudge users in emotionally satisfying ways to increase engagement? Are we at risk of AI behavioral feedback loops?Beth reminds the group that many people already treat AI like a coworker. How we speak to AI may influence how we speak to humans, and vice versa.The team agrees this isn’t about scrapping politeness or emotion but understanding what actually drives model output quality and what shapes our relationships with AI.Timestamps & Topics00:00:00 🧠 Intro: Do polite prompts help or hurt LLM performance?00:02:27 🎲 Andy on model randomness and Ethan Mollick’s findings00:05:31 📉 Prompt phrasing rarely changes model accuracy00:07:49 🧠 Beth on prompting as reflective collaboration00:10:23 🔧 Jyunmi on using LLMs to fill process gaps00:14:22 📊 Formatting prompts improves outcomes more than politeness00:15:14 🏭 Brian on sub-industry tagging, model consistency, and hallucinations00:18:35 🔁 Future fix: blockchain-like multi-model verification00:22:18 🔍 Andy explains system, developer, and compiled prompts00:26:16 🎯 Temperature and variability in model behavior00:30:23 🧬 Personalized memory will drive divergent outputs00:34:15 🧠 Echo chambers and AI recommendation loops00:37:24 👋 Why “please” and “thank you” still matter00:41:44 🧍 Personality shaping engagement in Claude and others00:44:47 🧠 Human expectations leak into AI interactions00:48:56 📝 Structured prompts outperform casual phrasing00:50:17 🗓️ Wrap-up: Join the Slack community and newsletterThe Daily AI Show Co-Hosts: Jyunmi Hatcher, Andy Halliday, Beth Lyons, Brian Maucere, and Karl Yeh

Want to keep the conversation going?Join our Slack community at dailyaishowcommunity.comIntroIn this week’s AI News Roundup, the team covers a full spectrum of stories including OpenAI’s strange model behavior, Meta’s AI app rollout, Duolingo’s AI-first transformation, lip-sync tech, China’s massive new model family, and a surprising executive order on AI education. From real breakthroughs to uncanny deepfakes, it’s a packed episode with insights on how fast things are changing.Key Points DiscussedOpenAI rolled back a recent update to GPT-4 after users reported unnaturally sycophantic responses. Sam Altman confirmed the issue came from short-term tuning and said a fix is in progress.Meta released a standalone Meta AI app and replaced the Meta View companion app for Ray-Ban smart glasses. The app will soon integrate learning from user Facebook and Instagram behavior.Google’s NotebookLM added over 70 languages. New language learning features like “Tiny Lesson,” “Slang Hang,” and “Word Cam” preview the shift toward immersive, contextual language learning via AI.Duolingo declared itself an “AI-first company” and will now use AI to generate nearly all of its course content. They also confirmed future hiring and team growth will depend on proving AI can’t do the work first.Brian demoed Fall’s new Hummingbird 0 lip-sync model, syncing Andy’s face to his own voice using a one-minute video clip. The demo showed improvement beyond simple mouth movement, including eyebrow and expression syncing.Alibaba released Qwen 3, a family of open models trained on 36 trillion tokens, ranging from tiny variants to a 200B parameter model. Benchmarks suggest strong performance across math and coding.Meta AI is now available to the public in a dedicated app, marking a shift from embedded tools (like in Instagram and WhatsApp) to direct user-facing chat products.Anthropic CEO Dario Amodei published a blog urging more work on interpretability. He framed it as the “MRI for AI” and warned that progress in this area is lagging behind model capabilities.AI science updates included a Japanese cancer detection startup using micro-RNA and a MIT technique that guides small LLMs to follow strict rules with less compute.University of Tokyo developed “draw to cut” CNC methods allowing non-technical users to cut complex materials by hand-drawing instructions.UC San Diego used AI to identify a new gene potentially linked to Alzheimer’s, paving the way for early detection and treatment strategies.Timestamps & Topics00:00:00 🗞️ Intro and NotebookLM’s 70-language update00:04:33 🧠 Google’s Slang Hang and Word Cam explained00:06:25 📚 Duolingo goes fully AI-first00:09:44 🤖 Voice models replace contractors and hiring signals00:13:10 🎭 Fall’s lip-sync demo featuring Andy as Brian00:18:01 💸 Cost, processing time, and uncanny realism00:23:38 🛠️ “ChatHouse” art installation critiques bot culture00:23:55 🧮 Alibaba drops Qwen 3 model family00:26:06 📱 Meta AI app launches, replaces Ray-Ban companion app00:28:32 🧠 Anthropic’s Dario calls for MRI-like model transparency00:33:04 🧬 Science corner: cancer tests, MIT’s strict LLMs, Tokyo’s CNC sketch-to-cut00:38:54 🧠 Alzheimer’s gene detection via AI at UC San Diego00:42:02 🏫 Executive order on K–12 AI education signed by Biden00:45:23 🤖 OpenAI rolls back update after “sycophantic” behavior emerges00:49:22 🔒 Prompting for emotionless output: “absolute mode” demo00:51:57 🛍️ ChatGPT adds shopping features for fashion and home00:54:02 🧾 Will product rankings be ad-based? The team is wary00:59:06 ⚖️ “Take It Down” Act raises censorship and abuse concerns01:00:09 📬 Wrap-up: newsletter, Slack, and upcoming showsThe Daily AI Show Co-Hosts: Jyunmi Hatcher, Andy Halliday, Beth Lyons, Brian Maucere, and Karl Yeh