What if LLMs could find fundamentally new solutions to hard problems? AlphaEvolve is an evolutionary coding agent built on top of Gemini for scientific discoveries. 🧵👇 🔵 Matrix Multiplication Re-invented (0:00) 🔵 AlphaEvolve's Architecture (3:08) 🔵 Math to Google's Core (3:37) 🔵 Spectrum of Discovery (4:21) 🔵 Creativity (5:14,... 7:11) 🔵 Representational Power (6:26) 🔵 Self-Improving Systems (8:18) 🔵 Value of Programmatic Discovery (8:46) 🔵 Synergy: AlphaEvolve & LLMs (11:58) 🔵 Managing Expensive Validations (14:53) 🔵 Future: "Library of Algorithms" (17:17) Clips from our interview with Google DeepMind researchers Matej Balog and Alexander Novikov on AlphaEvolve.show more

(0:00-3:08): Matrix Multiplication Reimagined by AlphaEvolve For 56 years, Strassen's 49-op recursive method for 4x4 general matrix multiplication was unbeaten. AlphaEvolve found a 48-op solution for complex-valued 4x4 matrices. It evolved a search algorithm for tensor decomposition (matrix multiplication's twin). Starting with a basic Adam optimizer & reconstruction loss, it ingeniously used complex numbers and developed advanced components like specialized initializers _get_init_fn [SEE PAPER FIG 3, TOP RIGHT] and custom loss functions with annealing schedules [SEE PAPER FIG 3, MIDDLE RIGHT].

(3:08-3:37): AlphaEvolve's Architecture: Evolutionary Coding AlphaEvolve uses an LLM ensemble (Gemini 2.0 Flash & Pro) in an evolutionary loop. 🔴 Core: LLMs propose code changes (diffs) to an initial program. 🟡 Evaluation: User-defined functions (h) score new program variants. 🟢 Progression: Successful variants populate an evolutionary database (inspired by MAP-Elites & island models) to guide future improvements. The system (detailed in [SEE PAPER FIG 1]) is asynchronous, maximizing discovery throughput.

(3:37-4:21): Surprising Versatility: From Math to Google's Core AlphaEvolve's impact is broad: 🔴 Fundamental Math: Improved >20% of 50+ open math problems (e.g., Minimum Overlap, Kissing Numbers). 🟢 Google Infrastructure: 🔹 Data Center Scheduling: Evolved a more efficient heuristic for Borg, recovering 0.7% fleet compute [SEE PAPER FIG 5 for the simple heuristic]. 🔹 TPU Hardware: Optimized Verilog for a key TPU arithmetic unit. 🔹 Gemini Kernels: Sped up Gemini LLM training by 23% (kernel-wise) via better tiling heuristics.

(4:21-5:14): The Spectrum of Discovery: Interpretable to Intricate AlphaEvolve produces: 🔴 Simple & Interpretable Code: E.g., concise Python for data center scheduling [SEE PAPER FIG 5] or Verilog for TPU optimizations. Easy to debug & deploy. 🔵 Complex & Novel Algorithms: For hard science like matrix multiplication, it evolved sophisticated search routines with non-obvious techniques (e.g., time-evolving quantization losses [SEE PAPER FIG 3, MIDDLE RIGHT]).

(5:14-6:26 & 7:11-8:18): Creativity: Starting Points & Evolution AlphaEvolve's creativity adapts: 🔴 From Basic Skeletons: With minimal initial code (e.g., Adam optimizer for matrix multiplication), it builds complex solutions, leveraging base LLM knowledge (18 mutations led to the advanced tensor algorithm [SEE PAPER FIG 3]). 🟢 Refining Seeded Ideas: Given more developed initial code (e.g., human domain knowledge), it focuses on maximizing that concept's potential. Evolution maintains diversity for broad exploration.

Quantum computers can break today’s encryption in seconds.🔑 Quantum tech will reshape our digital lives. Governments & hackers are preparing for the quantum era. How about you?🛡️ Don’t get left behind!🧠 Cybersecurity Dictionary for Everyone can help:

This has the potential to be revolutionary. Imagine algorithms evolving themselves to solve problems we haven't even conceived of yet. That's a whole new level of innovation.

An intriguing prospects for computational breakthroughs.

If AlphaEvolve can propose novel solutions efficiently, it may revolutionize computational research methods significantly.

AlphaEvolve could significantly impact current problem-solving strategies.

This approach could significantly advance scientific problem-solving methods.