Wild breakthrough on Math after 56 years... [Exclusive]

This episode unveils Google DeepMind's Alpha Evolve, a groundbreaking AI system that has shattered a 56-year-old record in matrix multiplication and is already optimizing critical infrastructure, offering a glimpse into AI-driven scientific discovery and its implications for computationally intensive fields like Crypto AI.

The Dawn of Alpha Evolve: Surpassing a Mathematical Benchmark

• The podcast opens with the exclusive reveal of Google DeepMind's new paper on Alpha Evolve, an AI system that has made a significant breakthrough in matrix multiplication, a fundamental problem in computer science.
• For over 50 years, the efficiency benchmark for multiplying 4x4 matrices, established by Volker Strassen in 1969 (who found a method using 7 multiplications for 2x2 matrices, leading to 7*7=49 for 4x4), remained unbeaten for matrices with arbitrary numbers.
  • Matrix Multiplication: A core mathematical operation used extensively in computing, including AI model training and cryptographic calculations. Efficient algorithms reduce computational cost and time.
• Alpha Evolve discovered a new algorithm for 4x4 matrices using only 48 scalar multiplications, a feat the researchers themselves found hard to believe. Mate, one of the researchers, shared, "When one of my teammates messaged on the channel that oh it seems like we have this result, I just couldn't believe it. Like let's triple check it."
• This achievement builds on DeepMind's legacy of AI systems that generate new knowledge, such as AlphaGo (Go champion), AlphaZero (self-play champion), AlphaFold (protein structure prediction), AlphaDev (faster sorting algorithms), AlphaTensor (matrix multiplication game), and FunSearch (LLM-based mathematical discovery). Alpha Evolve is presented as the next evolution in this lineage.
• Strategic Implication: Breakthroughs in fundamental algorithms like matrix multiplication can have cascading benefits for Crypto AI, potentially leading to more efficient AI model training, faster cryptographic computations, and novel algorithmic trading strategies.

Alpha Evolve's Architecture: Evolutionary Coding Agent

• Alpha Evolve is described as an "evolutionary coding agent" that iteratively refines algorithms. It leverages the creative power of Large Language Models (LLMs), combined with metalearning, library learning, automated evaluation, and evolutionary search.
  • Large Language Models (LLMs): AI models trained on vast amounts of text and code, capable of generating human-like text, code, and other content. Gemini is Google's latest family of LLMs.
  • Evolutionary Search: An optimization technique inspired by natural selection, where candidate solutions "evolve" over generations to find better ones.
• A key distinction from its predecessor, FunSearch, is Alpha Evolve's ability to operate over an entire codebase and optimize interactions between different functions, rather than focusing on a single function.
• Crypto AI Relevance: The ability to optimize entire codebases could be applied to smart contract optimization, decentralized protocol efficiency, or developing more robust AI agents for on-chain activities.

Navigating Computational Limits: The Halting Problem and Practical Solutions

• The discussion touches upon the halting problem, a theoretical limit in computer science stating it's impossible to determine if an arbitrary program will finish running or run forever. Keith Dugar raised this as a potential issue.
• Alexander Novikov, one of the Alpha Evolve authors, explained that in practice, this hasn't been a major hurdle. They often frame problems with built-in time constraints, e.g., "find a search algorithm that makes progress in 10 minutes."
• He drew a parallel to human research: "How do you know that you should stop working on your problem as a human? Uh like or maybe like maybe you spend a month more and then you solve it, right? It's it's hard."
• However, there's an acknowledgment that terminating algorithms based on time or performance relative to known solutions (like Strassen's) might restrict the discovery of novel paths that are initially slower but could lead to greater breakthroughs.
• Investor Insight: While theoretical limits exist, practical engineering and problem framing can often mitigate them. For Crypto AI, this means evaluating AI-driven discovery tools not just on theoretical perfection but on their practical utility and the cleverness of their problem constraints.

Real-World Impact: Optimizing Google's Infrastructure and AI Itself

• Alpha Evolve is not just a research project; it's already delivering tangible benefits within Google.
• It optimized job scheduling in Google's data centers, recovering an average of 0.7% of fleetwide compute resources that would have otherwise been idle—a massive saving at Google's scale.
• In a remarkable instance of self-improvement, Alpha Evolve found ways to accelerate the training of the Gemini models (the very LLMs that power Alpha Evolve) by 1%. Mate stated, "We have found a way to speed up the training of the next version of Gemini by by 1%."
• This involved Alpha Evolve generating not just the solutions but also the programs that generated those solutions, leading to human-interpretable mathematical insights from the code's structure.
• Strategic Consideration: The ability of AI to optimize its own underlying infrastructure and models points towards compounding efficiency gains. Crypto AI projects leveraging such self-improving systems could gain a significant competitive edge in resource utilization and model performance.

The Human-AI Symbiosis in Alpha Evolve

• Alpha Evolve operates with a strong "human-in-the-loop" paradigm. Humans identify interesting problems, define clear evaluation metrics, and provide initial candidate solutions or code skeletons.
• The system then explores possibilities, and humans can interpret the results, refine the problem, or guide the search further.
• This contrasts with the vision of fully autonomous AI. The speakers critique the "Dead Internet Theory" (the idea that most online content will be AI-generated "slop") and emphasize that AI's true power, as demonstrated by Alpha Evolve, lies in iterative, guided collaboration.
• Alex highlighted this: "The thing that makes Alpha Evolve so cool and powerful is is kind of this back and forth between humans and machines, right?"
• Research Focus: For Crypto AI researchers, this underscores the importance of developing tools and frameworks that facilitate effective human-AI collaboration rather than aiming for complete autonomy, especially in complex, open-ended problem domains.

Evolutionary Algorithms and Architectural Design

• The choice of evolutionary algorithms (EAs) was deemed natural for scientific discovery due to their ability to maintain diversity in exploration and avoid premature convergence on suboptimal solutions.
• Alpha Evolve's architecture pairs LLM creativity with robust evaluators. LLMs propose a wide range of solutions (some "stupid," some "amazing," some "weird"), and the evaluation mechanism filters and identifies promising ones. This is wrapped in an evolutionary pipeline that iteratively improves solutions, often in parallel.
• The system can start from empty code skeletons for maximum creativity or be guided by initial human-provided solutions to "squeeze out the juice of that idea."
• Crypto AI Application: This hybrid approach—combining generative AI with structured search and evaluation—could be a powerful template for designing AI systems that tackle complex crypto-specific problems, like finding novel consensus mechanisms or designing more secure DeFi protocols.

Knowledge Augmentation and the Future of Program Databases

• Alpha Evolve augments its knowledge by executing proposed algorithms and observing their results.
• A fascinating implemented feature is meta-prompting, where LLMs are asked to propose and curate their own prompts to improve the evolutionary process.
• The idea of creating a general, evolving database of useful code modules or programs—a "new oil"—is discussed as a future possibility. While currently focused on depth per problem, the system already uses discovered algorithms to initialize future experiments.
• Mate mentioned, "Technologically I I don't see hurdles for this. It's maybe the organizational question of how exactly to make it happen."
• Investor Watchpoint: The development of shared, evolving libraries of AI-discovered programmatic solutions could become a valuable asset class or infrastructure layer, particularly if these solutions are generalizable across different Crypto AI applications.

Deep Dive into the Matrix Multiplication Breakthrough

• Mate provided a detailed history of the matrix multiplication problem, from the standard high school method to Volker Strassen's 1969 algorithm (7 multiplications for 2x2).
• He emphasized that even for 3x3 matrices, the optimal algorithm is unknown (a gap between 19 and 23 multiplications). The search space is "completely immense."
• For 4x4 matrices (with arbitrary numbers), recursively applying Strassen's 2x2 method (7*7=49 multiplications) was the best known since 1969. AlphaTensor had previously found faster algorithms but only for specialized boolean matrices.
• Alpha Evolve's discovery of a 48-multiplication algorithm for general 4x4 matrices was achieved by searching for an algorithm using complex numbers, which then also applies to real numbers (a common case in neural network training).
• While Alpha Evolve scales further than AlphaTensor, for very large matrices (e.g., 6x6), it didn't match the best-known human-derived algorithms without being given specific domain knowledge (like symmetry constraints), as it was tasked to find algorithms of unrestricted form.
• Research Implication: This highlights that while AI can make surprising discoveries in unrestricted search, incorporating domain-specific knowledge or constraints can still be crucial for tackling highly complex, specialized problems. The interplay between general search and domain expertise remains key.

Intelligibility, Abstraction, and Program Synthesis

• Alpha Evolve can generate solutions across a spectrum of intelligibility. Sometimes it produces simple, human-verifiable code (e.g., for Google's job scheduler).
• In other cases, like the FunSearch capset problem, inspecting the AI-generated code (e.g., peculiar use of "modulo 4") led to new mathematical insights for humans, which then guided further AI search. Alex noted, "just by inspecting the code, we actually were able to develop can think of this like a mathematical insight or a mathematical hypothesis."
• For complex problems like optimizing the matrix multiplication search, Alpha Evolve might generate intricate solutions (e.g., a "time evolving shape of the quantization loss") that are effective but whose logic isn't immediately obvious to humans, akin to complex strategies found by AlphaZero in chess.
• Crypto AI Development: The ability to generate both highly optimized, potentially complex code and simpler, more auditable code is crucial for crypto. For core infrastructure like smart contracts, legibility and verifiability are paramount, while for off-chain analytics or trading, performance might be prioritized.

Computational Resources and Accessibility

• The computational cost of running Alpha Evolve is "elastic," depending on the problem's difficulty. Simpler problems can be solved quickly and cheaply.
• For challenging, open scientific problems, more compute is needed as the system explores more ideas. Mate emphasized, "it is able to sustain the scaling and as you keep running it for longer find better and better ideas."
• The speakers did not provide specific figures for the matrix multiplication discovery but indicated that costs vary widely.
• Investor & Researcher Note: Access to significant computational resources remains a factor for cutting-edge AI research. For the Crypto AI space, this could mean that entities with larger compute budgets may have an advantage in AI-driven discovery, or it could spur innovation in decentralized compute platforms.

Conclusion: AI as a Collaborative Discovery Engine

This episode reveals Alpha Evolve not just as an algorithm-discovering tool but as a paradigm for human-AI collaboration, pushing scientific frontiers and optimizing real-world systems. Crypto AI investors and researchers should monitor the evolution of such systems for their potential to unlock novel efficiencies and create new algorithmic primitives.

Actionable Takeaways:

• Investors: Consider the "picks and shovels" of AI-driven discovery; companies developing foundational AI tools like Alpha Evolve or providing the specialized compute for them may offer significant long-term value.
• Researchers: Explore hybrid AI architectures that combine LLM creativity with evolutionary search and human-in-the-loop guidance for tackling complex problems in decentralized systems and AI ethics within crypto.