Sam Charrington
500 episodes listed below
PRESS PLAY
Machine learning and artificial intelligence are dramatically changing the way businesses operate and people live. The TWIML AI Podcast brings the top minds and ideas from the world of ML and AI to a broad and influential community of ML/AI researchers, data scientists, engineers and tech-savvy business and IT leaders. Hosted by Sam Charrington, a sought after industry analyst, speaker, commentator and thought leader. Technologies covered include machine learning, artificial intelligence, deep learning, natural language processing, neural networks, analytics, computer science, data science and more.
Based on 6 episodes we listened to · September 2026
Sam Charrington turns advanced AI research and production systems into precise conversations about methods, tradeoffs, and unresolved engineering problems.
The TWIML AI Podcast follows artificial intelligence from research questions into systems people are trying to train and deploy. Sam Charrington conducts focused interviews that give guests room to explain their work, then presses on architecture, tradeoffs, and what remains…
A robot that almost does the job still owes you an explanation.
Press play if you want
Short reviews from the PlayNext desk, based on the episodes we processed.
A practical look at multi-agent AI in car shopping and the observability needed to run it in production.
Rashmi Shetty sketches a clear picture of Capital One already using multi-agentic AI to simplify car shopping, with Chat Concierge helping buyers find a car, schedule a test drive, get approved for financing, and estimate trade-in value.
A focused look at why diffusion language models start to matter once serving costs become the constraint.
The episode looks at how diffusion language models might move into production, with Stefano making the case that inference-time economics now carry the most weight: lower serving costs, faster responses, and more tokens per GPU.
Raschka shows why post-training and reasoning now draw more attention than pre-training.
The episode treats current LLM work as a move toward post-training and reasoning. Pre-training still matters because it remains costly and data-hungry, but the sharper questions now sit downstream, where there is still room for practical gains.
A sober look at why humanoid robots still struggle to generate value today.
The episode takes a blunt view of humanoid robotics: no robot earns its keep today if it is judged against the job it is meant to do. Sam Charrington and Nikita Rudin, co-founder and CEO of Flexion Robotics, map the gap between the present state of the field and the larger…
A measured look at model homogeneity, spectrum tuning, and AI's mixed human effects.
The piece takes a clear technical concern and keeps it in focus: models can stay less diverse than expected even across repeated runs, and Yejin Choi uses that to explore both intra-model and inter-model homogeneity.
Autonomous coding is framed as a disciplined, graph-grounded workflow, with a bit of scale gloss around it.
Sid Pardeshi presents a workable system for autonomous software development: multiple agent swarms, sandboxed environments, GitHub commits used as checkpoints, and a graph database that gives the codebase shared structure.
Markets, macro, AI, and crypto in motion.
Daily AI news through the lens of trust, infrastructure, and real-world deployment.
Yejin Choi tells Sam Charrington that her current focus is making small language models reason better, not just scaling large ones.
TWIML AI Podcast host Sam Charrington interviews Siddhant Pardeshi, co-founder of Blitzy, who built an autonomous software development platform that recruits dynamic swarms of agents to handle enterprise codebases — generating hundreds of thousands to millions of lines of validated, tested, pixel-perfect code in a single run, with more than 80% of the work completed autonomously.
["Capital One\u2019s Chat Concierge already uses multi-agentic AI to help car shoppers find inventory, schedule test drives, estimate trade-ins, and pursue financing.", "Rashmi Shetty, Senior Director of Enterprise Generative AI Platform at Capital One, traces her path from a pervasive-computing thesis to enterprise AI and says the team began its GenAI mission in early 2023 with an agent-assist pilot.", "She argues that multi-agent systems fit complex goals because one agent can disambiguate intent, other agents can execute tools, and platform guardrails can enforce policy, cyber, and model-risk constraints.", "The episode also covers observability, closed-loop learning, eval gates, human-in-the-loop checkpoints, and Capital One\u2019s preference for specialized models and student distillation to control personalization and latency."]
["Sam Sherrington interviews Stefano Ehrman on diffusion language models, and Ehrman opens by saying production LLMs are judged by price per token and tokens per GPU.", "At Stanford, he says diffusion work began in 2019 when GANs were unstable and image generation was still blurry.", "He explains why text is harder than images because tokens are discrete, then describes token masking as the way his lab made diffusion work on words.", "He says Mercury 2 at Inception now matches speed-optimized frontier models in quality while running about 5x to 10x faster, which matters for latency-sensitive coding and real-time AI apps."]
Sam Charrington and Sebastian Raschka frame the episode around 2026 LLM trends, with Raschka saying post-training and reasoning now matter more than architecture changes.
["Nikita Rudin says there is not a single humanoid robot today that actually generates value, because many systems still miss the exact task and need more handlers than before.", "He traces that view to ETH Zurich work on a quadruped robot, where GPUs and massively parallel simulation cut reinforcement-learning training from weeks to minutes and produced a stage demo that updated the policy every 15 seconds.", "Rudin says perception, sim-to-real transfer, and humanoid motion standards are still hard, especially once a robot must decide whether to climb stairs, avoid plants, or move through a warehouse without looking unnatural.", "He argues that the near-term path is modular industrial robotics, with locomotion, planning, and VLM-based tool use combining walking, picking, and dropping skills rather than fully autonomous home robots."]
Sam Charrington interviews Akansha Chowdhury about her experience training large language models like Palm and Gemini, and discusses the need to rethink pre-training for agentic tasks.
Sam Charrington interviews Munawar Hayat about Qualcomm's research on multimodal AI and visual AI presented at NeurIPS.
Sam Charrington interviews Scott Stevenson, CEO of DeepGram, about the evolution of audio AI and the shift from separate speech models to unified perception and interaction models.
Sam Charrington interviews Tim Rocktäschel about self-improving AI systems and progress in AI since their last conversation.
An interview with Lucas Garcia from MathWorks about incorporating AI models into safety-critical systems, contrasting with traditional methods like Kalman filters.
Sam Charrington interviews Princeton professor Arvind Narayan about AI agents, their capabilities, failures, and how to benchmark them, as well as his book AI Snake Oil.