https://www.youtube.com/watch?v=3pdlTMdo7pY

About the Video

Title - John Carmack (Keen Technologies): Research Directions @ Upper Bound 2025

Event: Upper Bound 2025

Speaker: John Carmack

• Description
• About: John Carmack
• About: Keen Technologies
• About: Upper Bound 2025 (Event)

<aside> ⚠️

AI Warning

These section overviews, key points and quotes are comprehensively processed with AI.

This is explained further here in the Video Database.

As always, AI-generated content makes mistakes.

These are sanity-checked, formatted and lightly edited by a human, but not as thorough as the content itself, and said human also makes mistakes.

</aside>

Summary

Based on the talk, John Carmack and his company, Keen Technologies, are pursuing a focused and contrarian research direction toward achieving Artificial General Intelligence (AGI). Instead of following the mainstream hype around Large Language Models (LLMs), they are tackling what Carmack sees as the fundamental, unsolved scientific problems of learning and intelligence.

In general, they are:

1. Rejecting the LLM-Only Path to AGI: Carmack argues that while LLMs are magical tools, their architecture is fundamentally flawed as a model for general intelligence. He points out that they are trained by "putting all of human knowledge in a giant blender" and cannot learn sequentially from new experiences without catastrophic forgetting. He believes true intelligence, like that seen in "cats and dogs, let alone small children," requires a different approach.
2. Using Reinforcement Learning on Atari as a "Crucible": Keen has deliberately chosen the classic Atari Learning Environment as its primary research testbed. Carmack defends this choice by arguing that Atari games are an unbiased, diverse, and sufficiently complex environment to isolate and solve core AI problems. Unlike modern games, they prevent researchers from "cheating" by accessing internal game data, forcing the AI to learn from pixels alone, just as a human would.
3. Grounding Research in Physical Reality: To test the robustness of their algorithms, they built a physical robot that plays an actual Atari console. This "stunt" is a research tool designed to force their AI to confront real-world challenges that are abstracted away in simulators:
   • Latency: The robot must cope with ~180ms of end-to-end latency, which breaks many state-of-the-art algorithms that assume instantaneous feedback.
   • Imperfect Perception: The system has to learn to read the score off a real screen, a task that proved surprisingly difficult and brittle.
   • Physical Actuation: The robot controller has physical limitations, like taking time to move the joystick, which the AI must learn to account for.
4. Focusing on Core Unsolved Problems: The company's primary work is aimed at making breakthroughs in areas where current AI consistently fails:
   • Sequential Multi-task Learning: Teaching a single agent to learn multiple games in a row without it forgetting how to play the previous ones.
   • Transfer Learning: Enabling an agent to use knowledge from past games to learn new, unseen games faster, rather than starting from scratch like an "infant opening its eyes for the first time."
   • Sparse Rewards & Intrinsic Curiosity: Developing agents that can learn and explore effectively without the constant, dense reward signals that video games provide but the real world does not.
5. Building a New Benchmark for the Community: Carmack is not just working in a silo. He is actively developing and advocating for a new, open-source benchmark for the entire AI research community. His goal is to create a standardized "harness" that sequences agents through multiple Atari games, rigorously measuring their ability to learn continuously and transfer knowledge. He believes that having a common, difficult, and "uncheatable" benchmark is essential for driving real, verifiable progress in the field.

Talking Points