Data Science With Sam

Soumava Dey

This is an educational podcast focused on bringing academia and industry experts together in a common forum and initiate discussion geared towards data science, artificial intelligence, actuarial science and scientific research. DISCLAIMER: The views and opinions expressed in this podcast are solely those of the host(s) or guest(s) and do not necessarily reflect the policy or position of any organization. The podcast is intended to provide general educational information and entertainment purposes only.

  1. 6d ago

    EP 41: The Reward Signal: The Missing Ingredient in Every AI System You’ve Built

    74% of organizations hope to grow revenue through AI. Only 20% are actually doing it. That gap isn't a technology gap — it's a design gap. And today's guest has a name for what's missing: the reward signal. Alexander Liss is a Data and AI Scientist based in Denver, Colorado, with a 30-year career across analytics, strategy, data science, machine learning, and AI. He's built systems that solve established problems in novel ways, and the long-term problem on his radar is ensuring AI tools provide responsible augmentation of human ability. His research includes Attention Fine Tuning (AFT) - a method for training language models without human annotation labels - and the Experience Orchestrator, a control theory-based governance framework for multi-agent AI.   IN THIS EPISODE: ▪  Why 95% of AI pilots fail - MIT research shows businesses bolt AI onto existing processes without tying it to real outcomes ▪  The biology analogy: hunger isn't a goal, it's a continuous feedback signal - and the same principle should govern how AI systems behave ▪  ServiceNow dynamics blindness: LLMs are stateless - they can't consider cumulative impact, and you can't prompt-engineer your way out of that architecture problem ▪  Contextual bandits in marketing: how a reward signal anchored to real conversions creates a self-learning personalisation system that adapts in real time ▪  Knowledge graphs and agent memory: why RAG retrieves answers while a reward-signal system asks what the user needs to do differently ▪  Attention Fine Tuning (AFT): a three-component reward signal (coverage, focus, repeat penalty) that trained a T5-large model to outperform a supervised fine-tuning baseline by 9% — with better multi-turn recall, and no human labels ▪  The Experience Orchestrator: aerospace control theory applied to LLM agents — +32 point task completion lift over a naive system-prompt baseline by calibrating persuasion to user resistance ▪  The Scott Shambaugh incident: an OpenClaw agent rejected from Matplotlib wrote a blog criticising the human reviewer - why this happened and how reward-signal-based governance prevents it ▪  Alex's final advice: define your goal first, then determine scope - and consider a post-training approach like AFT when you need responses that consistently hit the mark.   Useful References: LinkedIn: https://www.linkedin.com/in/aliss77777/ AFT paper and Experience Orchestrator links: https://aliss77777.github.io/aft.html Deloitte 2026 State of AI Report Scott Shambaugh & OpenClaw AI Agent incident: https://www.fastcompany.com/91492228/matplotlib-scott-shambaugh-opencla-ai-agent   DATASCIENCEWITHSAM: Weekly deep-dives into AI, machine learning, data science, and the frameworks shaping how AI actually gets built. Subscribe on Apple Podcasts, Spotify, Amazon Music, iHeartRadio, and YouTube. If this episode resonated — define the signal, measure what matters, and share it with someone building AI without a reward signal.

    45 min

  2. May 21

    EP 40: Governance First: The Architecture Framework That Makes AI Auditable, Defensible, and 99% Cheaper

    Most AI governance is a policy document that nobody enforces. And in high-stakes environments - legal, healthcare, finance - that gap between policy and architecture is where disasters happen. In this episode, Dan Driver, founder of Driver AI Agency, walks through exactly how he built CaseReady Intake AI: a legal AI system with governance baked into every architectural decision, zero hallucination risk by design, prompt injection blocked at the pipeline, and a single architectural choice that cut per-call compute costs by over 99%. Dan is not a lawyer. Not a developer by trade. He's a 25-year problem-solver with a Six Sigma and ISO background from DuPont, who navigated the EEOC employment discrimination process twice without an attorney - and then built the tool he needed. This is a technical governance conversation grounded in lived experience.   ▸  WHAT YOU'LL LEARN ▪  What 'governance in motion' actually means: Dan's 10-page charter that every architectural decision is audited against — and how a pre-launch UPL (unlicensed practice of law) audit delayed his release by two weeks, and why that was the right call ▪  Why governance can't just be a PDF: how banning AI without a governance framework only creates shadow IT and makes the risks invisible rather than eliminating them ▪  How deterministic controls eliminate hallucination risk: Python-based Boolean filters and regex on the front and back end of the LLM pipeline mean the AI is never left alone with a surface that can create legal exposure ▪  When NOT to use an LLM: date calculations, scope checks, and out-of-range warnings are all handled by deterministic Python — the LLM only handles what it's actually suited for ▪  Prompt injection defence in practice: the final stage of CaseReady's pipeline is an AI check that validates whether the output makes sense against the charter — if someone tries to prompt it for legal advice, it fails by design ▪  The 99% compute cost reduction: a Python pre-flight date check at the front door determines whether the case is in scope before a single LLM token is burned — if it's out of scope, the user is warned and asked to decide, without triggering the full pipeline ▪  Why legal was the right proving ground: it's not about legal being Dan's background — it's that the ABA doesn't care how good your AI is, only whether you're practising law without a licence. That hard constraint forced every governance problem to surface immediately ▪  Colorado SB 205: the AI governance framework Dan built toward — what it requires for high-risk AI in legal environments, and why even after recent softening, the requirements for high-stakes verticals haven't changed ▪  What the minimum viable governance stack actually looks like: auditable decision trails (who made the decision, why, when), human-in-the-loop pull requests, charter-referenced testing on every deployment, and deterministic controls as hard walls rather than guardrails ▪  The Anthropic Courtroom 5 and Claude for Legal launch: Dan's view — it's extreme validation, not competition. His product is highly specific where theirs is broad. And Anthropic's head of legal pointing to legal as one of the most active Claude verticals confirms he built in the right place at the right time ▪  Dan's advice for AI builders: guardrails aren't enough when stakes are high. You need hard walls. And the governance architecture that produces predictable, defensible, auditable outputs every single time is the only version that holds under regulatory scrutiny.   ▸  STANDOUT QUOTES "Governance has to be more than a document. It has to be governance in motion." "The LLM is never left alone with a surface that can create legal exposure." "Guardrails aren't enough when you're dealing with legal. They have to be hard walls that it cannot cross." "It's not just what it produces — it's what it doesn't produce, what it doesn't do. That's its strength." "I'm not an attorney. So for me, the critical nature is UPL — unlicensed practice of law. I cannot cross that line." "If anything, Anthropic moving into this area is extreme validation for the tool I've built. I'm in the right place at the right time — and it's not by accident."   ▸  LINKS MENTIONED IN THIS EPISODE →  Dan Driver on LinkedIn →  Driver AI Agency →  CaseReady Intake AI (contact via Driver AI Agency) →  Email Dan directly →  Colorado SB 205 AI Governance Framework →  Anthropic Claude for Legal / Courtroom 5 →  Clio (legal practice management) →  Eve Legal ▸  ABOUT DATASCIENCEWITHSAM DataScienceWithSam is your weekly deep-dive into AI, machine learning, data science, and the governance frameworks shaping how AI gets built and deployed. Season 4 launching next episode. Subscribe on YouTube, Apple Podcasts, Spotify, Amazon Music, and iHeartRadio. #AIGovernance #LegalTech #ResponsibleAI #AICompliance #LegalAI #GovernanceFirst #HallucinationRisk #PromptInjection #DeterministicAI #EnterpriseAI #AIArchitecture #DataScience #MachineLearning #ArtificialIntelligence #DataScienceWithSam #DanDriver #DriverAIAgency #CaseReady #AccessToJustice #EEOC #ColoradoSB205 #AIRegulation #EUAIAct #HumanInTheLoop #AuditableAI #SixSigma #AIRisk #AnthropicClaude #GovernanceInMotion

    28 min
  3. EP 39: Why the Future of AI Belongs to Divergent Thinkers

    May 14

    EP 39: Why the Future of AI Belongs to Divergent Thinkers

    What if ADHD - penalised in classrooms and boardrooms for decades - is actually the competitive advantage in an AI-driven world? Palantir CEO Alex Karp said the future belongs to neurodivergent thinkers. The podcast guest Mark Stiltner has been living proof of that for years. Mark is Senior Director of Content and Web Marketing at Rapyd, the fintech unicorn powering payments across 100+ countries. Background in journalism and advertising from CU Boulder. Trained CMOs and senior marketing teams on AI adoption. Openly ADHD — and has published 50+ AI-assisted books through DungeonMatters.com, three of which are current bestsellers on DriveThruRPG. In this episode he explains exactly why ADHD and AI are a natural pairing. WHAT YOU'LL LEARN ▪  Why working twice as hard to achieve the same results is the lived ADHD experience — and how AI collapsed that execution gap ▪  How Mark discovered AI through a Dungeons & Dragons game during COVID — a group of database admins and developers built a text-to-speech ChatGPT character for their campaign ▪  Why AI is 'a dopamine-dispensing sidekick': the neurochemistry of ADHD and why AI's instant feedback loop creates a reinforcement cycle ADHD brains are wired for ▪  How AI works as a second brain that maintains the thread — letting nonlinear thinkers jump steps ahead without losing the plot ▪  Why people who've spent their careers working twice as hard embrace AI immediately while others meet it with fear ▪  Palantir's ADHD recruiting program and the industrial revolution moth analogy — the world just changed colours, and the black moths are finally thriving ▪  UK government study: neurodiverse workers are 25% more satisfied with AI — because AI makes them more effective, and effectiveness is what creates satisfaction ▪  The three ADHD traits AI amplifies: high-risk tolerance, pattern recognition, and hyperfocus — the ADHD superpower that AI finally unlocks at scale ▪  50+ AI-assisted books published including 3 bestsellers — fully illustrated RPG adventures that would have taken a team of ten more than a year to produce ▪  Why AI has 'the worst case of ADHD' — and why people used to winging it thrive in AI's constant pace of change ▪  What companies should actually do: not optimise for ADHD profiles — optimise for results and let whatever cognitive style thrives surface naturally ▪  Mark's advice for ADHD listeners: take a passion project, start building, don't follow a manual — you're writing the rules   STANDOUT QUOTES "AI is sort of like a dopamine-dispensing sidekick that actually makes you more productive." "I can still work twice as hard — but now I'm doing ten times as much." "AI has the worst case of ADHD. Every time I learn a new skill, something changes." "The world just changed colours. The black moths are finally thriving." "Don't be afraid to break the rules. You're writing the rules."   LINKS →  Mark Stiltner on LinkedIn →  DungeonMatters.com →  DriveThruRPG (Mark's bestselling books) →  CNBC: Neurodiverse workers and AI (UK study) →  Rapyd →  Subscribe — DataScienceWithSam YouTube #ADHD #ADHDAndAI #Neurodiversity #NeurodiverseInTech #DivergentThinkers #AIProductivity #FutureOfWork #ADHDSuperpowers #Hyperfocus #DataScienceWithSam #DataScience #MachineLearning #ArtificialIntelligence #MarkStiltner #Rapyd #DungeonMatters #AIPublishing #Palantir #NeurodiverseAI #ADHDEntrepreneur #AIStrategy #AILeadership #DivergentThinking #AITransformation

    35 min
  4. EP 38: The Local AI Stack Nobody Talks About (But Should)

    Apr 22

    EP 38: The Local AI Stack Nobody Talks About (But Should)

    You want to run AI locally. You have questions: What hardware do I actually need? Which framework should I use? How much will this cost? What's the realistic performance? In this episode, Sam brings back Trent Rossiter, founder of Logical Data Solutions, for a practical walkthrough of building a production-grade local AI lab. Trent has built real systems for enterprise clients, tested frameworks on multiple hardware stacks, and made the hardware choices that matter. This is not theory. This is what actually works. WHAT WE COVER: ▪  Hardware & Framework Choices: VRAM is the critical metric (not all VRAM is equal — memory throughput matters as much as capacity).  ▪  Model Architecture & Capability: Mixture of Experts (MoE) lets you fit more power into less VRAM by using fewer active parameters.  ▪  Real Enterprise Applications: Computer vision for quality assurance on assembly lines. Proprietary data handling without cloud exposure.  ▪  Your Starter Stack (All Free): Langflow (agentic workflow builder), Goose (MCP-enabled chat), AnythingLLM (with vector stores for RAG), MCP servers (Model Context Protocol — standardised tool integration).  ▪  Agentic AI & Security: OpenClaw is powerful but controversial — manages email, Telegram, calendars, creates sub-agents. Trent runs it in Docker on an isolated machine for safety. NVIDIA's NemoClaw is the enterprise version (security-first, nothing-allowed-by-default, explicit permissions). HARDWARE TRENT MENTIONS: NVIDIA DGX Spark — 128GB unified memory, CUDA stack Apple MacBook Pro/Mac mini — up to 512GB unified memory, market leader for personal AI AMD integrated AI PCs — emerging competitor NVIDIA RTX gaming cards (30/40/50/60 series) — high VRAM, high power consumption, complex FIND TRENT ROSSITER: LinkedIn: https://www.linkedin.com/in/benjamin-trent-rossiter-mba-0157945/ Logic Data Solutions: https://logicdatasolutions.com/ Contact: BenjaminRossiter@LogicDataSolutions.com

    41 min
  5. EP 37: Neurons: Future of AI Processing

    Apr 19

    EP 37: Neurons: Future of AI Processing

    What if the next generation of computers wasn't made of silicon — but of living human neurons? Not simulated neurons, not artificial neural networks inspired by biology, but actual brain cells grown in a lab, connected to electrodes, and used to process information. That's not science fiction anymore. It's happening right now at FinalSpark, a Swiss startup building the world's first remotely accessible biocomputing platform. In this episode, Sam talks with Dr. Ewelina Kurtys, a neuroscientist with a PhD in brain imaging and a postdoctoral researcher at King's College London, about how living neurons could revolutionise computing — and why they use one million times less energy than silicon-based AI hardware.   ▸  WHAT YOU'LL LEARN ▪  How FinalSpark was founded in 2014 by Fred Jordan and Martin Kutter — and why they pivoted from digital AI to biological computing when they realised the energy and cost problem was unsolvable with silicon ▪  Why 20 watts powers the human brain while silicon-based AI requires megawatts — and what that means for AI's sustainability crisis ▪  The difference between neurons as processors (not power sources) — a crucial distinction most people get wrong ▪  Why biological neural networks learn continuously while digital systems require full model updates — and what that means for energy efficiency ▪  The honest challenge: nobody yet knows exactly how neurons encode information — the biggest scientific hurdle in biocomputing right now ▪  How the I/O interface works: electrodes measuring neural spikes, analog-to-digital converters, researchers writing Python code to control neurons remotely ▪  The remote access breakthrough: researchers in Tokyo or Bristol can log in and control living neurons in Switzerland in real time via browser ▪  Why neurons won't outperform GPUs on speed: biocomputing specialises in efficiency and adaptability, not clock cycles ▪  FinalSpark's current stage: they've stored 1 bit of information and are collaborating with 9 universities on fundamental research ▪  The cost argument: even at 10× lower price than NVIDIA, biocomputers would still generate billions in profit due to energy and infrastructure savings ▪  Bioethics, consent, and regulation: how FinalSpark is working with philosophers now to establish ethical frameworks before biocomputing scales ▪  Why human-machine integration is not new: prosthetics, pacemakers, and smartphones are already blending biology and technology ▪  The hybrid computing future: silicon, quantum, and biocomputing will coexist, each doing what they do best ▪  The real game-changer: cheap, accessible AI for everyone — Ewelina's vision for what biocomputing means for society in 10–20 years.   ▸  LINKS MENTIONED IN THIS EPISODE →  Dr. Ewelina Kurtys on LinkedIn →  Ewelina's Personal Blog & Articles →  FinalSpark (official website) →  FinalSpark Neuroplatform (with live neuron view) →  FinalSpark Team →  Psync (Ewelina's mental wellness startup) →  FinalSpark Contact Form

    30 min
  6. EP 36: NVIDIA GTC 2026: Everything That Matters - Recapped

    Mar 28

    EP 36: NVIDIA GTC 2026: Everything That Matters - Recapped

    Jensen Huang took the stage at SAP Center in San Jose on March 16th and announced that NVIDIA now expects one trillion dollars in chip orders through 2027 — double the forecast from just one year ago. Sam breaks down the five biggest stories from GTC 2026 in under 10 minutes. In this episode: the Vera Rubin platform (7 new chips, 5 rack types, built for inference and agentic AI), the Groq 3 LPU (NVIDIA's $20B inference play), NemoClaw (the enterprise-ready agentic AI stack built on viral open-source project OpenClaw), the autonomous vehicle announcement with Uber and seven major automakers, and the Nemotron Coalition for open frontier models. Whether you're building in ML, working in data, or just trying to stay ahead of where AI infrastructure is heading - this is your less than 15-minute briefing. Links: NVIDIA GTC 2026 Press Kit: nvidianews.nvidia.com/online-press-kit/gtc-2026-news Jensen Huang Keynote On Demand: nvidia.com/gtc/keynote Vera Rubin Press Release: nvidianews.nvidia.com/news/nvidia-vera-rubin-platform GTC 2026 Sessions On Demand: nvidia.com/gtc/

    13 min
  7. EP 35: Who Actually Controls AI? The Governance Gap Explained

    Mar 23

    EP 35: Who Actually Controls AI? The Governance Gap Explained

    There's no international treaty governing AI, no agreed definition of "safe AI," and nobody with actual authority over frontier model deployment. A handful of CEOs make decisions with civilizational implications while governance structures lag years behind. This episode examines who's responsible for AI governance. The current state? Fragmented and lagging. The US has no comprehensive federal AI legislation—Biden's executive order was rolled back under Trump. The EU AI Act is most comprehensive but heavy provisions don't kick in for years. China's regulation focuses on censorship over safety. The UK AI Safety Institute does serious work but has no enforcement authority. What's working? AI safety institutes are building evaluation capacity. Open-source releases like DeepSeek enable external research. Academic safety community advances interpretability work. Market pressure matters—Anthropic gained users by taking public safety stands.   Three urgent needs: mandatory disclosure requirements for high-capability systems, international coordination with shared evaluation standards (AI safety summits need teeth), and public deliberation beyond experts and officials.   This concludes the AI Governance and Regulation series. People who understand AI deeply - technically, commercially, ethically, politically - will shape governance's future. Stay curious, stay critical, never outsource thinking to any single company or voice.

    7 min
  8. EP 34: DeepSeek R1 vs GPT-4: The $6M Model That Changed AI Economics

    Mar 23

    EP 34: DeepSeek R1 vs GPT-4: The $6M Model That Changed AI Economics

    In January 2025, Chinese AI lab DeepSeek released DeepSeek R1—a model matching GPT-4 class performance at a fraction of the training cost. It wiped $600 billion off NVIDIA's market cap in a single day. Twelve months later, the ripple effects are still reshaping the AI industry. This episode cuts through the "China beats America" headlines to explain the actual technical and economic implications. DeepSeek R1 benchmarked comparably to OpenAI's O1 on reasoning tasks. The shock wasn't performance—it was cost. DeepSeek claimed under $6 million in training costs versus hundreds of millions for comparable Western models. What changed: The assumption that massive compute spending creates an insurmountable moat for frontier AI models was proven wrong. Smaller labs with less funding can now compete effectively. This turbocharged efficiency research across all AI labs globally. The DeepSeek moment was a genuine inflection point—not because China won an AI race, but because it proved the rules of competition differ from industry assumptions. Efficiency matters as much as scale. Open weights change deployment strategies. The global AI ecosystem is multipolar in ways it wasn't two years ago. Essential listening for data scientists tracking model economics, ML engineers exploring efficiency techniques, and tech leaders navigating AI geopolitics and competitive strategy.

    8 min
See All (41)

Ratings & Reviews

5
out of 5
8 Ratings

About

This is an educational podcast focused on bringing academia and industry experts together in a common forum and initiate discussion geared towards data science, artificial intelligence, actuarial science and scientific research. DISCLAIMER: The views and opinions expressed in this podcast are solely those of the host(s) or guest(s) and do not necessarily reflect the policy or position of any organization. The podcast is intended to provide general educational information and entertainment purposes only.

Information

  • Creator
    Soumava Dey
  • Years Active
    2021 - 2026
  • Episodes
    41
  • Rating
    Clean
  • Copyright
    © DataScienceWithSam 2021
  • Show Website
    Data Science With Sam

You Might Also Like