He Co-Invented the Transformer. Now: Continuous Thought Machines [Llion Jones / Luke Darlow]

Sakana’s Philosophy: Research Freedom (4:13 - 7:45)

• Sakana AI embraces the philosophy of research freedom, actively protecting its researchers from external pressures.
• As companies grow, pressures for returns on investment and product development often curtail creativity and autonomy in research.
• Llion Jones emphasizes that he encourages Sakana AI employees to work on what they find interesting and important.

The Local Minimum of Large Language Models (7:45 - 18:30)

• The success of large language models (LLMs) has led to a "technology capture," where the industry is overly focused on them, potentially trapping it in a local minimum.
• A parallel is drawn to the era of recurrent neural networks (RNNs), where incremental improvements were made until Transformers provided a "crushingly better" alternative, rendering previous research redundant.
• The current AI landscape sees many papers making slight modifications to the Transformer architecture, which might be a waste of time if a new breakthrough emerges.
• Llion Jones argues that being merely "better" than Transformers is not enough; a new architecture needs to be overwhelmingly superior to shift the industry.
• The concept of "jagged intelligence" is introduced, where LLMs can solve complex problems but also make glaringly obvious errors, suggesting a fundamental flaw in their architecture.

Representation Problems: The Spiral Example (18:30 - 29:12)

• Current neural networks are powerful but may not "want" to represent things in a human-like way, acting as universal approximators that are forced to fit data.
• An obscure paper on "intelligent matrix exponentiation" is cited, showing how a model could represent a spiral dataset as an actual spiral, unlike ReLU MLPs that use piecewise linear approximations.
• This highlights that current models "fake" understanding by fitting data without grasping the underlying concept, leading to poor extrapolation.
• The discussion touches on the "impostor" nature of AI, where models seem intelligent but lack fundamental properties like adaptive computation and constructive learning.
• The speakers express skepticism about video generation models that fix issues like extra fingers through brute-force scaling rather than a deeper, conceptual understanding.

Technical Deep Dive: CTM Architecture (29:12 - 36:00)

• Luke Darlow introduces the Continuous Thought Machine (CTM), an 8-month project at Sakana AI.
• The CTM has three novelties: an internal thought dimension for sequential computation, neurons as individual models (MLPs), and synchronization as a representation of thought over time.
• The maze-solving task is used as a "hello world" problem for CTM, where it solves mazes sequentially, step-by-step, unlike traditional deep learning that processes the entire image at once.
• The CTM's neurons are not simple ReLU activations but rather "neuron level models" that use a finite history of activations to process a single output, building dynamics into the system.
• Synchronization is measured by the dot product between the time series of neuron activations, capturing how neurons fire together over time.

Adaptive Computation & Maze Solving (36:00 - 47:15)

• The CTM uses an autocurriculum system for maze solving, where it learns to predict one more step than it currently can, enabling self-bootstrapping.
• The model observes the full maze image at every step but uses attention to retrieve information and has a fixed number of "thinking steps" (e.g., 100).
• For ImageNet classification, the CTM's loss function encourages adaptive computation by optimizing for both best performance and highest certainty, leading to easy examples being solved quickly and harder ones taking more thought.
• The CTM's stability is attributed to synchronization, which helps with gradient propagation by touching all neurons through all time.
• The use of exponential decay rates in synchronization allows for capturing both short-term and long-term firing patterns between neurons, enriching the representation space.
• The CTM's sequential reasoning and internal thought process are seen as potentially superior to Transformers for tasks requiring planning and discrete reasoning, like the ARC challenge.

Model Calibration & Uncertainty (47:15 - 1:00:43)

• Neural networks are often poorly calibrated, meaning their predicted probabilities don't accurately reflect their confidence.
• The CTM, after training, was found to be nearly perfectly calibrated, which was an emergent property rather than an explicit design goal.
• This adaptive computation time in CTM emerged naturally from the loss function design, unlike previous attempts that required explicit penalties for computation.
• The CTM's ability to build understanding step-by-step and explore trajectories in spaces is seen as crucial for agents that can continuously learn and construct path-dependent knowledge.
• The architecture's approach to carving up the world in different ways could address issues like hallucinations in language models.
• The CTM's design, inspired by biology, allows for different ways of approaching problems to emerge, leading to philosophical questions about intelligence.
• Luke Darlow suggests that language itself can be viewed as a maze, implying the CTM's potential for next-generation language models.
• During training, the CTM was observed to backtrack and explore multiple paths in mazes, a behavior not typically seen in current AI models.
• When constrained for time, the CTM exhibited a "leapfrogging" behavior in mazes, quickly jumping ahead and tracing paths backward, suggesting different algorithmic strategies under constraint.
• The speakers discuss the potential for scaling out CTMs with shared memory structures, creating a "cultural memory" for collective intelligence, which is an active area of research.

Sudoku Bench: Measuring True Reasoning (1:00:43 - 1:13:28)

• Llion Jones introduces the Sudoku Bench, a new benchmark dataset for reasoning, designed to be more challenging than standard Sudoku.
• The benchmark features "variant Sudokus" with highly diverse, handcrafted rules, often requiring complex natural language understanding and even meta-reasoning (e.g., identifying a wrong number in the rule description).
• Current best models achieve only about 15% on the Sudoku Bench, solving only the simplest puzzles, and GPT-5 also struggles with it.
• The dataset was created by scraping thousands of hours of human thought traces from the YouTube channel "Cracking the Cryptic," where experts verbalize their reasoning processes.
• This provides high-quality human reasoning data for imitation learning, addressing Andrej Karpathy's idea of learning from "thought traces."
• The Sudoku Bench is designed to be difficult because each puzzle requires finding a new and unique "break-in" (a specific reasoning step) that current AI models cannot generalize to.
• The "tree of reasoning" analogy is used to describe the vast and complex space of problem-solving, with humans and AI models possessing different parts of this tree.
• Current RL algorithms struggle with the Sudoku Bench because the required breakthroughs are in a "rare space," making sampling inefficient.
• Sakana AI is actively hiring ML engineers and scientists, offering significant research freedom.

"What we're after here is trying to do something biologically plausible... and find a line somewhere where we can take some sprinkling of biological inspiration but still train it with deep learning." — Luke Darlow

"The Sudoku Bench... if anyone was actually be able to beat our benchmark they would necessarily have to have created an extremely powerful reasoning system." — Llion Jones

"What was perfect I realized is they tell you in agonizing detail exactly what reasoning they used to solve those particular puzzles." — Llion Jones

"It's such a rare space, you know, it's it's such an specific kind of reasoning that's required to get to the the specific breakthrough that this kind of technique doesn't work." — Llion Jones