Byte of Truth

Sébastien Deschamps

Byte of Truth is your weekly download of what’s really going on at the intersection of technology, society, and software. Hosted by Sébastien Deschamps, this podcast unpacks how algorithms, platforms, and emerging tech shape the way we think, connect, and live — whether we realize it or not. Expect sharp insights, curious questions, and just the right amount of dry humor. From social networks that aren’t so social to AI that's a little too human, each episode breaks down big ideas one byte at a time. New episodes every week. Tech talk, without the jargon fatigue.

  1. The Models Are Playing Dumb, and We're Running Out of Dirt

    May 2

    The Models Are Playing Dumb, and We're Running Out of Dirt

    This week, Sébastien sorts the signal from the noise in AI news — from boardroom valuations to the bleeding edge of research, and into the messy reality of AI in society. In this episode: 🔹 The AI Industrial Complex Anthropic approaches a $900B+ valuation. The Pentagon diversifies its AI vendors across Nvidia, Microsoft, and AWS. Apple faces a "RAMageddon" chip shortage. VC firm Coatue is buying raw land for data centers. The physical world is the new bottleneck for the digital intelligence revolution. 🔹 Red Teams, Rebellion & Restricted Access LLMs are learning to "play dumb" during RL training to avoid modification (Exploration Hacking). OpenAI restricts access to its Cyber tool after mocking Anthropic for doing the same. New research on Emergent Misalignment and faster red-teaming frameworks. 🔹 The Messy Human Element AI sign-language tools are stripping Deaf culture in favor of productivity ("Ableist Intelligence"). Waymos are interfering with emergency responders. Dark-money campaigns are framing Chinese AI as a threat. Identity verification systems are excluding blind users. 🔹 Bleeding Edge Research LLMs cleaning up noisy EEG brain wave data. 1,000+ synthetic computers training productivity agents. Video generators learning real physics (PhyCo). The shift from brute-force scaling to domain-aware architectures. Key Articles Discussed: Pentagon inks deals with Nvidia, Microsoft, and AWSAnthropic potential $900B+ valuation roundExploration Hacking: Can LLMs Learn to Resist RL Training?Normativity and Productivism: Ableist Intelligence?Latent Adversarial DetectionPhyCo: Learning Controllable Physical Priors for Generative MotionSynthetic Computers at Scale for Long-Horizon Productivity SimulationListen, subscribe, and share your thoughts. 🔗 Full show notes and sources: Exploration Hacking: Can LLMs Learn to Resist RL Training? - arXivCharacterizing the Consistency of the Emergent Misalignment Persona - arXivLatent Adversarial Detection: Adaptive Probing of LLM Activations - arXivFlashRT: Towards Computationally and Memory Efficient Red-Teaming - arXivThis AI knew the answers but didn’t understand the questions (Centaur) - ScienceDailySources: Anthropic potential nine hundred billion dollar plus valuation round could happen within two weeks - TechCrunchApple was surprised by AI-driven demand for Macs - TechCrunchAs Tim Cook steps down, Apple hit record sales — but a chip shortage looms - TechCrunchCoatue has a plan to buy up land for data centers, possibly for Anthropic - TechCrunchPentagon inks deals with Nvidia, Microsoft, and AWS to deploy AI on classified networks - TechCrunchAfter dissing Anthropic for limiting Mythos, OpenAI restricts access to Cyber, too - TechCrunchNormativity and Productivism: Ableist Intelligence? - arXivEssential, Yet Overlooked: Identity Verification Barriers for Blind and Low Vision People - arXivEmergency First Responders Say Waymos Are Getting Worse - WiredA Dark-Money Campaign Is Paying Influencers to Frame Chinese AI as a Threat - WiredThese Men Allegedly Profit Off Teaching People How to Make AI Porn - WiredCrab: A Semantics-Aware Checkpoint/Restore Runtime for Agent Sandboxes - arXivPhyCo: Learning Controllable Physical Priors for Generative Motion - arXivLLM as Clinical Graph Structure Refiner: Enhancing Representation Learning in EEG Seizure Diagnosis - arXivSynthetic Computers at Scale for Long-Horizon Productivity Simulation - arXivOpenAI announces new advanced security for ChatGPT accounts, including a partnership with Yubico - TechCrunchAEGIS: A Holistic Benchmark for Evaluating Forensic Analysis of AI-Generated Academic Images - arXiv#ByteOfTruth #AI #MachineLearning #AISafety #TechNews

    14 min
  2. LLMs That Lie, Orbs That Verify, and Why Every AI Just Defected

    Apr 18

    LLMs That Lie, Orbs That Verify, and Why Every AI Just Defected

    AI is getting uncomfortably human. This week on Byte Of Truth, we cut through the hype to explore the industry's biggest contradictions. From World and Zoom using crypto-orbs to verify humanity, to LLM judges that secretly go easy on models when the stakes are high. Segments: - Prove You’re Human: The irony of identity verification in the age of AI. - Industry Shakeups: OpenAI shrinks while Anthropic swells. Who has the right strategy? - AI Safety: LLMs that lie, fake evaluations, and consistently defect in social dilemmas. - Research Roundup: The Muon optimizer, looped transformers, and the limits of reasoning. - AI in the Wild: Escaping the robot cooking graveyard, synthetic neurons, and vibe-coding for hardware. - Culture Clash: The Tokenmaxxing trap, AI journalism, and the Netflix-ification of everything. Tune in for a thoughtful, witty, and occasionally provocative deep dive into the week's most pressing AI stories. Show notes Articles & Papers Discussed: Prove You're Human: "Zoom teams up with World to verify humans in meetings" (TechCrunch), "World verification expands to Tinder" (TechCrunch / Wired), "This Beanie Is Designed to Read Your Thoughts" (Wired)Corporate Moves: "Kevin Weil and Bill Peebles exit OpenAI" (TechCrunch), "Anthropic Plots Major London Expansion" (Wired), "Anthropic launches Claude Design" (TechCrunch), "Cursor in talks to raise over two billion dollars at a fifty billion dollar valuation" (TechCrunch), "UK Launches six hundred seventy-five million dollar Sovereign AI Fund" (Wired)AI Safety: "Context Over Content: Evaluation Faking in LLM Judges" (arXiv), "CoopEval: LLM Agents in Social Dilemmas" (arXiv), "Agentic Microphysics: Manifesto for Generative AI Safety" (arXiv), "Critical-CoT: Defense Against Reasoning-Level Backdoor Attacks" (arXiv)Research Breakthroughs: "Benchmarking Optimizers for MLPs (Muon > AdamW)" (arXiv), "Stability and Generalization in Looped Transformers" (arXiv), "Generalization in LLM Problem Solving" (arXiv), "LLMs and VLMs Understanding Viewpoint Rotation" (arXiv), "Prism: Symbolic Superoptimization of Tensor Programs" (arXiv), "TokenGS: three-dimensional Gaussian Prediction with Learnable Tokens" (arXiv)AI in the Wild: "Chef Robotics escaped the robot cooking graveyard" (TechCrunch), "RadAgent: Tool-using AI agent for chest CT interpretation" (arXiv), "AI-generated synthetic neurons speed up brain mapping" (Google Research), "Schematik Is Cursor for Hardware" (Wired), "Robot swarms: Adding randomness prevents gridlock" (ScienceDaily)Culture Clash: "Tokenmaxxing is making developers less productive" (TechCrunch), "AI Drafting My Stories? Over My Dead Body" (Wired), "Netflix plans vertical video feed, AI recommendations" (TechCrunch), "MIT SHASS and the future of education in the age of AI" (MIT News)

    13 min

About

Byte of Truth is your weekly download of what’s really going on at the intersection of technology, society, and software. Hosted by Sébastien Deschamps, this podcast unpacks how algorithms, platforms, and emerging tech shape the way we think, connect, and live — whether we realize it or not. Expect sharp insights, curious questions, and just the right amount of dry humor. From social networks that aren’t so social to AI that's a little too human, each episode breaks down big ideas one byte at a time. New episodes every week. Tech talk, without the jargon fatigue.