Subvolts take Understand LLM limits and the real path to AGI beyond pattern matching. Mar 17, 2026 - Developers
- Understand the mathematical underpinnings of current models and their limitations.
- Business
- Grasp the fundamental differences between current AI and future AGI capabilities.
This episode features a researcher explaining core AI concepts and future directions, offering valuable technical context.
What's Missing Between LLMs and AGI - Vishal Misra & Martin Casado
Audio streams from a16z. Your place is remembered on this device.
From Subvolts
Summary and analysis by Subvolts editorial, AI-assisted.
Vishal Misra joins Martin Casado on AI + a16z to discuss the inner workings of large language models. Misra explains research showing transformers update predictions predictably. He argues current LLMs are not conscious and contrasts pattern matching with the cause-and-effect understanding needed for AGI, which requires continued learning post-training.
This episode explains how AI models like ChatGPT process information predictably, like a calculator. It clarifies why this is not consciousness and what's truly needed for AI to learn and understand the world like humans do, moving beyond just recognizing patterns.
This episode offers insight into the fundamental limitations of current LLMs and the path toward true artificial general intelligence. It is for researchers, AI engineers, and tech leaders tracking the evolution of AI beyond today's models. The discussion on learning and understanding cause and effect is highly relevant to frontier labs like Google Gemini and OpenAI as they push the boundaries of AI capabilities.
- Hosta16z
- GuestVishal Misra
- Guest: Vishal Misra
- Show: AI + a16z
- Length: 48 minutes
- Transformer prediction update: Precise, mathematically predictable
- AGI requirements: Keep learning after training, understand cause and effect
Vishal Misra returns to explain his latest research on how LLMs actually work under the hood. He walks through experiments showing that transformers update their predictions in a precise, mathematically predictable way as they process new information, explains why this still doesn't mean they're conscious, and describes what's actually required for AGI: the ability to keep learning after training and the move from pattern matching to understanding cause and effect.
Resources:
Follow Vishal Misra on X: https://x.com/vishalmisra
Follow Martin Casado on X: https://x.com/martin_casado
Check out everything a16z is doing with artificial intelligence here, including articles, projects, and more podcasts.
Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.
- LLMs
- Vishal Misra
- Large Language Models, AI systems that understand and generate text
- AGI
- Vishal Misra
- Artificial General Intelligence, AI with human-like cognitive abilities
- transformers
- Vishal Misra
- A type of AI model architecture common in LLMs
24 minHow AI Is Rewriting the Rules of Cybersecurity | Truffle Security & Socket
WithDylan Ayrey, Feross Aboukhadijeh
47 minInside vLLM: The Engine Powering Open-Source AI
WithSimon Mo
41 minBuilding Self-Accelerating AI with Mirendil
WithBehnam Neyshabur, Harsh Mehta
43 minIdeogram’s Open-Weights Image Model and the Future of AI Design
WithMohammad Norouzi
1 h 16Confronting the CEO of the AI company that impersonated me
WithShishir Mehrotra
2 h 18NVIDIA's $1 Trillion Prediction, Anthropic Beats OpenAI, Tesla vs. TSMC & The CS Job Collapse | 240
WithSalim Ismail, Dave Blundin, Alexander Wissner-Gross
1 h 01
1 h 07Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI
WithAndrej Karpathy
Questions people ask
- How do transformers update predictions?
- Experiments show transformers update their predictions in a precise, mathematically predictable way as they process new information.
- What is required for AGI?
- AGI requires the ability to keep learning after training and a move from pattern matching to understanding cause and effect.
Page generated Oct 19, 2026. Summaries are Subvolts' own; the audio, artwork and notes belong to a16z.