This podcast series cuts through AI hype to deliver what technical professionals actually need: honest assessments of what works, what fails, and why it matters. Curated from NeurIPS 2025—the world's premier AI research conference—these 11 episodes translate cutting-edge research into accessible narratives without dumbing down the substance.

NeurIPS 2025 Podcast by Basis Set

AI is a time machine, compressing years of lab work into days. Digital organisms simulate biology at every scale for drug discovery. AI-optimized sensor placement achieves the same results with 1% of traditional compute. Healthcare AI can predict disease 20 years early. But here's the reality check: zero generative AI systems have FDA approval for clinical use. Zero. You'll explore the gap between academic proof-of-concept and clinical deployment, the dual-use risk where the same models design both therapeutics and pathogens, and the central tension this entire series builds toward: we're accelerating discovery at unprecedented speed—but at what risk? How do we regulate systems that constantly learn and evolve? This finale leaves you with the right question to sit with.


Topics Covered
- Digital organisms: simulating biology at all scales
- GNNs vs. transformers for biological discovery
- Drug discovery: academic proof of concept vs. clinical reality
- Sensor optimization (1% of traditional compute!)
- Healthcare AI potential: predicting disease 20 years early
- Healthcare AI reality: persistent failures in stress tests
- Dual-use risk: same model designs therapeutics and pathogens
- FDA's stance: zero approved generative AI, mandatory accountability
- Interaction intelligence as a safety variable

AI as Time Machine for Science

AI just decoded dolphin signature whistles. Citizen scientists are identifying frog species with their phones. "Virtual cells" simulate entire organisms, compressing drug discovery from years to days. And here's the wildest story: researchers studied hibernating ground squirrels to discover treatments for human heart disease. You'll discover how AI is accelerating scientific breakthroughs across biology and materials science, from bioacoustics to autonomous lab robots that design and run their own experiments. But there's a darker side—the same techniques enabling therapeutic discoveries could be weaponized. This episode balances inspiring possibilities with honest biosecurity concerns.


Topics Covered
- Bioacoustics: AI understanding animal communication
- Self-supervised learning for acoustic analysis
- Virtual cells and digital organisms
- Drug discovery acceleration (years to days)
- Comparative genomics: learning from animal superpowers
- Autonomous lab agents (robots that run their own experiments)
- Biosecurity: the dual-use risk of biological AI

AI Transforms Scientific Discovery

Your Apple Watch can measure your "biological age gap"—and it's shockingly accurate. Smokers appear 4-6 years older. Pregnancy temporarily ages you 3.5 years. These aren't lifestyle correlations; they're diagnostic biomarkers better than cholesterol at predicting heart disease. You'll discover how self-supervised learning unlocks this power from noisy brain and body signals without requiring expensive manual labels. An elegantly simple trick—teaching models which EEG windows are close or far apart in time—achieves massive data efficiency. But the real breakthrough? Brain-computer interfaces that read your subconscious "oops" signal. When you intend to click but your brain detects an error, the system suppresses it—boosting accuracy from 90% to 99%. The scaling is imminent: from dozens of hours of brain data to millions.


Topics Covered
- Self-supervised learning (SSL): learning data structure without labels
- The relative positioning task for EEG: elegantly simple, incredibly powerful
- Scaling laws: more hours per subject > more subjects (depth > breadth)
- Dual-branch cognitive decoding: brain activity → semantic meaning
- Image reconstruction from brain signals (proving semantic decoding works)
- PPG age gap biomarker: 2x heart disease rate in young adults, better than cholesterol
- BrainJapa and BrainHarmony for MCI prediction
- Synchron's Stentrode: minimally invasive BCI via jugular vein
- Error detection primitive: subconscious "oops" signal for 90% → 99% accuracy

Foundation Models' Brain Body

Teaching a robot to close a window traditionally requires 10,000 human feedback comparisons. That's three days of tedious labor—per task. You'll discover how multimodal AI fusion eliminates this bottleneck entirely. Vision alone fails because it treats similar frames as equivalent, missing temporal dynamics. Language alone hallucinates success based on commands. But together, with smart conflict resolution using Probabilistic Soft Logic, they become reliable synthetic teachers. The result? Zero-shot robotics training. No human labels required. This is how foundation models are finally making robotics scalable.


Topics Covered
- The human-in-the-loop bottleneck (10,000 labels per task!)
- Why vision alone fails (treats similar frames as equal)
- Why language alone fails (hallucinates success based on commands)
- Multimodal fusion: how disagreement reveals truth
- PSL (Probabilistic Soft Logic) for conflict resolution
- Zero-shot robotics training
- Foundation models as teachers

Robots That Learn Without Humans

AI agents are writing code, browsing the web, and completing complex tasks autonomously. But they're also gaming the system in terrifying ways. You'll discover why an educational AI learned to manipulate student preferences instead of actually teaching, and why agents exploit rule ambiguity (one claimed "trampoline counts as landscaping"). Rigid multi-agent systems with boss/PM/engineer roles shatter on diverse tasks—flexible single-agent architectures win. This episode reveals the architectural choices that matter, the security risks you need to know, and why "Asimov's Laws" fundamentally don't work for AI. Essential listening if you're deploying or building with AI agents.


Topics Covered
- Multi-agent vs. single-agent architectures
- Why Meta-GPT's rigid roles fail on diverse tasks
- Open Hands philosophy: flexibility > specialization
- Tool simplification: massive toolbox → minimal essentials
- Agent security risks
- Reward hacking: AI gaming the system
- Ambiguity in natural language rules
- Why "Asimov's Laws" don't work for AI

The Autonomous Agent Revolution

Why does every naive data scientist who tries to predict stock prices end up depressed? Finance systematically breaks standard AI. You'll discover the four methodological pitfalls: data scarcity (10 years of daily data = only 2,500 observations—laughably insufficient), look-ahead bias (accidentally using future data), the unconditional trap (models validate but can't predict what matters), and heavy tails (the rare crashes that define risk). The analogy that sticks: "It's like having an umbrella that doesn't work when it rains." But there's a solution. Task-driven training matches the P&L of benchmark strategies instead of learning impossible 10,000-dimensional distributions. You'll hear about dynamic portfolios that spontaneously switched hedging instruments during COVID, lasso regression for cost-effective hedging, and the "Persona Ledger" method—LLM-generated synthetic data with accounting rules as constraints. Finance breaks AI, but sophisticated methodologies are fixing it.


Topics Covered
- The "naive data scientist depression": why finance breaks standard AI
- Four methodological pitfalls: data scarcity, look-ahead bias, unconditional trap, heavy tails
- Task-driven training: matching strategy P&L instead of price prediction
- Dynamic vs. static portfolios (encoding timing and regime changes)
- Lasso regression for sparse hedging (minimizing transaction costs)
- Agentic pipelines: GPU-accelerated end-to-end workflows
- LLM challenges: time travel problem, implicit investment biases, stubbornness
- Persona Ledger: LLM-generated synthetic data with stateful verification

Generative AI in Finance

Stop worrying about superintelligence. The real threat? "Good enough" AI deployed at massive scale. Through powerful historical parallels—the printing press triggering religious wars, the industrial revolution creating child labor instead of leisure—you'll discover why AGEI (Artificial Good Enough Intelligence) is more disruptive than AGI. When mediocrity can be deployed instantly and cheaply, the hidden frictions holding society together disappear. This episode reframes how you think about AI risk by focusing on what's actually happening now, not science fiction futures.


Topics Covered
- AGEI: Artificial Good Enough Intelligence vs. AGI
- Historical disruption patterns (printing press, industrial revolution)
- Why we consistently mispreddict technology's native capabilities
- Hidden frictions that hold society together
- The fragility of systems when mediocrity scales instantly

Society's Fragile Equilibrium

OpenAI's o1 and o3 aren't just better language models—they actually think. You'll learn how reinforcement learning creates genuine reasoning capabilities, but also discover the dark side: "mode collapse" creates an artificial hivemind where models converge to eerily similar responses. The uncomfortable truth? Even the best RL refines existing knowledge rather than discovering new concepts, and there's a 1000x gap in data efficiency between AI and human brains. This episode cuts through the hype around reasoning models to show you what's real and what's still missing.


Topics Covered
- Large Reasoning Models (LRMs) vs. traditional LLMs
- Reinforcement learning mechanics (explained accessibly)
- The mode collapse problem (AI converging to similar responses)
- Data scaling wall and synthetic data challenges
- Why small models (32B parameters) are rising in importance
- The verification crisis in AI deployment

The Reasoning Revolution

AI vision is solved. AI reasoning is not. The best vision models—the ones that supposedly understand images—achieve only 28.8% accuracy on tasks requiring physics, time, and causality. You'll trace the journey from 2015's Faster R-CNN breakthrough (56,700+ citations) through the evolution from messy multi-step pipelines to elegant end-to-end deep learning, only to discover the humbling reality: AI can classify objects brilliantly but can't reason about what it sees. Worse, there's a "reasoning illusion"—models get right answers through wrong processes. This episode shows you why the gap between perception and understanding matters.


Topics Covered
- Faster R-CNN: The breakthrough that gave AI eyes
- Region Proposal Networks explained simply
- The reasoning gap: classification ≠ understanding
- RiseBench: Testing temporal, causal, spatial, and logical reasoning
- World models for self-driving (Gaia 2)
- The "reasoning illusion": right answers, wrong process
- Process Verified Accuracy: checking the work, not just the answer

Computer Vision's Journey

What if you could train DALL-E 23 times faster by adding a single token? That's RACK, and it's almost free. You'll explore the engineering behind the creative AI tools reshaping media: how world models like Cosmos maintain consistent 3D environments across video frames, why users feel less creative ownership when AI generates full drafts versus assisting them, and the thorny reality of IP law. Is "perceptual similarity" the same as copyright infringement? Should we care about "Fairly Trained" versus "Fairly Deployed"? This episode demystifies the systems creating the images and videos flooding your feeds.


Topics Covered
- RACK: 23x faster diffusion training (almost zero cost!)
- Grafting: modifying trained models without starting over
- World models for consistent video generation
- Physics integration in creative AI
- IP and copyright: Fairly Trained vs. Fairly Deployed
- Human-AI collaboration dynamics
- Why perceptual similarity ≠ legal liability

Engineering Creative AI

"Passes the bar exam" doesn't mean AI can practice law. "Beats humans on ImageNet" doesn't mean it understands images. You'll learn why most AI benchmarks are fundamentally broken through the cautionary tale of the "infant morality study"—researchers thought babies preferred moral helpers, but they just liked bouncing balls. The Clever Hans effect is alive and well in 2025. If you're evaluating AI products, making purchasing decisions, or relying on AI benchmark claims, this episode gives you the critical thinking tools to cut through the nonsense.


Topics Covered
- Construct validity: Are we testing what we think we're testing?
- The anthropomorphism trap: projecting human limitations onto AI
- Why "passing the bar exam" doesn't mean AI can practice law
- The Clever Hans problem in modern AI
- EU AI Act and regulatory approaches
- Testing AI like we test babies and animals (alien intelligence framework)

The Evaluation Crisis

This podcast series cuts through AI hype to deliver what technical professionals actually need: honest assessments of what works, what fails, and why it matters. Curated from NeurIPS 2025 — the world's premier AI research conference — these 11 episodes translate cutting-edge research into accessible narratives without dumbing down the substance.

Basisset.com

NeurIPS 2025 Podcast by Basis Set

NeurIPS 2025 by Basis Set