"I Co-Invented the Transformer. Now I'm Replacing It." & Continuous Thought Machines - Llion Jones and Luke Darlow [Sakana AI]

Jones argues the AI field is trapped in a Transformer local minimum similar to the RNN era. Before Transformers, endless RNN variants (LSTMs, GRUs) achieved incremental improvements (1.26→1.25 bits/char), but Transformers immediately jumped to 1.1, making all that research obsolete. Current Transformer tweaks may be similarly wasted effort when the next breakthrough arrives.

• •RNN research produced endless architectural variants with tiny improvements before Transformers made them all obsolete
• •Transformers are 'too good' - they work well enough that incremental improvements prevent exploration of fundamentally better architectures
• •Current research freedom is constrained compared to the Transformer era due to commercialization pressure and reduced autonomy
• •Moving the industry requires being 'crushingly better' not just marginally better - Transformers set this high bar

Sakana AI is built on Kenneth Stanley's 'Why Greatness Cannot Be Planned' philosophy - researchers should follow gradients of interestingness rather than fixed objectives. Jones emphasizes protecting researcher freedom from commercialization pressures, publication demands, and committee-driven agendas that narrow exploration and prevent breakthrough discoveries.

• •Company philosophy based on Kenneth Stanley's ideas about epistemic foraging and avoiding objective-driven research
• •Main challenge: protecting research freedom as company grows and pressure for returns/products increases
• •Researchers told to 'work on what you think is interesting and important' - CTM paper is proof this works
• •Reduced freedom in current AI landscape compared to Transformer era due to competitive pressure and commercialization

Jones uses the 'matrix exponentiation' paper's spiral classification task to illustrate fundamental limitations. ReLU networks solve spirals with piecewise linear boundaries - technically correct but lacking true understanding. Better architectures would represent spirals as spirals, enabling proper extrapolation and generalization rather than brute-force fitting.

• •ReLU networks solve spiral classification with piecewise linear boundaries - correct but not understanding the pattern
• •Matrix exponentiation layers naturally represent spirals as spirals, enabling proper extrapolation
• •Current models are 'too powerful' - with enough compute/data they can brute-force any problem without true understanding
• •Video generation fixing finger counts may be brute force rather than fixing underlying representation problems

Luke Darlow explains CTM's three core novelties: (1) internal sequential thought dimension for multi-step reasoning, (2) neuron-level models where each neuron is a small model processing temporal history, and (3) synchronization-based representations measuring how neuron pairs fire together over time, creating a richer d²/2 dimensional representation space.

• •Internal thought dimension enables sequential reasoning - maze solving becomes 'go up, go right' rather than single-shot image prediction
• •Each neuron is a small model (MLP) processing temporal activation history, inspired by biological neurons
• •Synchronization representation: dot product of neuron activation time series creates d²/2 dimensional space
• •Exponential decay rates allow different neuron pairs to synchronize at different timescales (fast vs slow)
• •Architecture is biologically inspired but not biologically plausible - designed for deep learning training

CTM naturally exhibits adaptive computation without requiring carefully tuned penalty losses (unlike Alex Graves' ACT paper). The loss function trains on both lowest-loss and highest-certainty points, causing easy examples to solve in 1-2 steps while hard examples use full thinking time. Remarkably, the model is nearly perfectly calibrated without post-hoc tricks.

• •Adaptive computation emerges naturally from loss design - no hyperparameter sweeps needed unlike prior ACT work
• •Easy ImageNet examples solved in 1-2 steps, hard examples use full 50 steps automatically
• •Model is nearly perfectly calibrated (90% confidence = 90% accuracy) without calibration tricks
• •Synchronization representation helps gradient propagation through long sequences, avoiding RNN training issues

During maze training, CTM spontaneously developed sophisticated strategies: mid-training it would explore one path, realize it's wrong, backtrack and try another. Under time constraints, it invented a 'leapfrogging' algorithm - jumping ahead to approximate positions and filling in paths backwards, a faster algorithm humans don't typically use.

• •Mid-training behavior: model explores paths, realizes mistakes, backtracks and tries alternatives
• •Under time constraints, invented leapfrogging: jump ahead, trace backwards, repeat - faster than sequential solving
• •Auto-curriculum training: model learns step N+5 once it masters step N, enabling self-bootstrapping
• •These emergent algorithms suggest the architecture naturally discovers efficient problem-solving strategies

CTM's sequential reasoning enables path-dependent understanding - how you arrived at a conclusion matters, not just the conclusion itself. This allows agents to explore trajectories, construct understanding that 'carves the world up by the joints,' and potentially avoid the shortcut learning and hallucination problems of feed-forward models.

• •Path-dependent understanding: the trajectory to a solution matters, not just the final answer
• •Enables exploration of different 'carve-ups' of the world - different valid reasoning paths
• •Addresses shortcut learning: cost function alone isn't enough, the reasoning process must be mechanistically sound
• •Potential solution to hallucination: autoregressive token generation may carve up the world incorrectly

Darlow is actively exploring applying CTM to language modeling, viewing language as a type of maze with ambiguity. Future directions include multi-agent systems with shared memory structures - agents solving mazes with limited visibility but shared cultural memory, enabling collective intelligence and long-term memory capabilities.

• •Language can be viewed as a maze task with ambiguity - natural fit for CTM architecture
• •Exploring multi-agent systems with shared memory: agents with limited local view but global memory access
• •Memory systems key to future AI - agents need to save/retrieve memories to avoid repeating mistakes
• •Parallel agents with weight sharing could enable collective intelligence and cultural memory

Jones introduces Sudoku Bench, a reasoning benchmark using variant Sudokus with handcrafted unique constraints requiring strong NLU and meta-reasoning. Includes thousands of hours of expert reasoning traces from 'Cracking the Cryptic' YouTube channel. Current best models achieve only ~15% on simplest puzzles; GPT-5 shows improvement but still fails on human-solvable puzzles.

• •Variant Sudokus have unique handcrafted constraints requiring creative breakthroughs ('break-ins'), not just search
• •Includes expert thought traces from 4-hour solving videos - true human reasoning data Karpathy wanted
• •Current models fall back to boring trial-and-error (try 5, try 6, try 7) rather than finding break-ins
• •RL doesn't work - breakthrough reasoning is too rare to sample, showing limits of current techniques
• •Meta-reasoning required: professionals recognize problem types and select appropriate reasoning 'LEGO blocks'
• •Benchmark tests ability to reason about reasoning itself, backtrack failed approaches, and discover novel solution paths