AI That Can Change Its Mind? (New Architecture) [w/ Sakana CTO]

Llion Jones, one of the original authors of the “Attention Is All You Need” paper, expresses a sense of “inventor’s remorse,” arguing that the AI industry has been captured by the success of the Transformer architecture. This “success capture” has led to an oversaturated research space focused on incremental tweaks, trapping the field in a local minimum and potentially hindering the discovery of the next major breakthrough. He draws a parallel to the pre-Transformer era of Recurrent Neural Networks (RNNs), where countless small modifications to LSTMs and GRUs were rendered obsolete overnight by the Transformer’s superior performance. Jones suggests we may be in a similar situation now, wasting time on minor architectural adjustments when a fundamentally new approach is needed.

The Problem: Faking Intelligence and The Spiral Analogy

A core issue with current architectures is their tendency to find shortcuts rather than developing a true understanding. Jones uses the “spiral problem” as a potent analogy. A standard ReLU-based neural network can learn to classify the points in a spiral dataset perfectly, but its decision boundary is composed of many tiny, piecewise linear segments. It “fakes” the spiral without understanding the underlying generative concept.

“Yes, if you train these things enough and you push these little piecewise linear boundaries around enough, it can fit the spiral and get a high accuracy. But there’s no feeling when I look at that image that the ReLU version actually understands that it is a spiral.”

This reflects the “jagged intelligence” of modern Large Language Models (LLMs), which can solve PhD-level problems one moment and make glaringly obvious errors the next. They are incredibly powerful function approximators, but they don’t represent concepts in a way that aligns with human cognition, leading to a lack of robust generalization and true reasoning.

Sakana AI’s Philosophy: Fostering Research Freedom

The culture at Sakana AI is modeled on the early, bottom-up innovation environment of places like Google Brain, where the Transformer itself originated from informal discussions rather than a top-down corporate mandate. The company’s philosophy is heavily inspired by Kenneth Stanley’s book, “Why Greatness Cannot Be Planned,” which advocates for allowing researchers to follow their own “gradient of interest.” This approach encourages epistemic foraging and the discovery of novelty, countering the trend where commercial pressures and the “pressure to publish” curtail creative freedom and push researchers toward safer, more incremental work.

A New Direction: The Continuous Thought Machine (CTM)

Luke Darlow introduces the Continuous Thought Machine (CTM), a new architecture inspired by biology, designed to address the shortcomings of current models. The CTM is built on three core novelties:

1. Internal Thought Dimension: The model performs computation sequentially over an internal time dimension. This allows it to “think” step-by-step. Using a maze-solving analogy, Darlow explains that while a standard CNN might try to output the entire path in one shot, the CTM “walks” through the maze, breaking the problem down into a sequence of steps.
2. Rethinking the Neuron: Instead of simple activation functions like ReLU, each neuron in the CTM is a small model itself (an MLP). This allows for more complex and dynamic interactions within the system.
3. Synchronization as Representation: A “thought” is not merely the state of the neurons at a single point in time, but a process that unfolds over time. The CTM captures this by using the synchronization patterns between pairs of neurons as its core representation. This is calculated by measuring the dot product of their activation time series, creating a much richer representational space.

Emergent Properties: Adaptive Computation and Calibration

A key benefit of the CTM is that it exhibits adaptive computation naturally. The model’s loss function encourages it to find the best-performing and most-certain points in its internal thought process.

• For simple tasks, the model quickly arrives at a confident, correct answer in just a few steps.
• For more complex problems, it naturally uses more of its available “thinking time” to ponder the solution.

This behavior emerges from the architecture itself, unlike previous attempts like Alex Graves’s “Adaptive Computation Time” which required carefully tuned penalty terms to prevent the model from always using its full computational budget.

Furthermore, the CTM was found to be nearly perfectly calibrated. This means its confidence scores accurately reflect its probability of being correct, a property most deep learning models lose after extensive training. This emergent quality is another strong indicator that the architecture is a more principled approach to reasoning and uncertainty.

Sudoku Bench: A Benchmark for True Reasoning

To push the community beyond simple pattern matching, Sakana AI developed Sudoku Bench. This is not a test of solving standard Sudokus, but rather “variant Sudokus” with unique, complex, and often ambiguous rules described in natural language. For example, a puzzle might state that one of the numbers in its own rule description is incorrect, forcing the model to perform meta-reasoning.

The benchmark includes a unique dataset of “thought traces” scraped from the “Cracking the Cryptic” YouTube channel, where experts verbalize their complex, multi-step reasoning process in detail. Current state-of-the-art models, including GPT-4, perform poorly, scoring around 15% and only solving the simplest puzzles. They fail to discover the novel “break-ins” or logical leaps required and instead fall back to brute-force guessing. The benchmark is designed to measure and encourage the development of systems capable of flexible, grounded, and human-like reasoning.

• Tags:
• Transformer
• Reasoning
• Continuous Thought Machine
• CTM
• Sakana AI
• Adaptive Computation
• Neural Architecture