Algorithms that evolve to be better versions of themselves.
Deepmind's AlphaEvolve herald the next era of agents, the illusion of conscious AI.
Happy Friday! Let’s dive in.
-Tara
DeepMind’s recently launched coding agent AlphaEvolve marks the beginnings of a new era in how software is created. AlphaEvolve goes a step beyond generating code. It can invent, test, and improve the algorithms it creates.
What sets AlphaEvolve apart is its ability to continuously optimize itself through “evolutionary” processes. What this means is that it doesn’t just output a solution and call it a day. Instead, it evaluates, iterates, and refines its own code, much like how nature selects the most efficient adaptations over generations.
So far, it has:
Optimized Google’s data centers by freeing up 0.7% of global compute resources, saving millions in operational costs.
Sped up AI model training: cut the training time for DeepMind’s own Gemini models by 1%
Enhanced chip performance: it suggested improvements to TPUs that sped up matrix operations by 23%.
Surpassed a 50-year-old math milestone: It discovered a faster way to multiply 4x4 matrices, reducing the number of required multiplications from 49 to 48.
This video demos a list of changes proposed by AlphaEvolve to discover faster matrix multiplication algorithms. In this example, AlphaEvolve proposes extensive changes across several components, including the optimizer and weight initialization, the loss function, and hyperparameter sweep.
A quick primer on how it self-improves (more in the AlphaEvolve paper):
Initial Code Generation: AlphaEvolve uses large language models (LLMs), such as Gemini, to generate initial algorithmic code based on a given problem.
Automated Evaluation: Each generated algorithm is automatically evaluated for performance metrics relevant to the task, such as efficiency or accuracy.
Evolutionary Optimization: The system employs evolutionary strategies—selecting the best-performing algorithms, introducing variations (mutations), and iteratively refining them over successive generations.
Iterative Refinement: This cycle of generation, evaluation, and selection continues, allowing AlphaEvolve to progressively enhance algorithm performance beyond initial versions.
If algorithms can self-improve, we’re looking at machines that can continuously outthink themselves. This could reshape everything from drug discovery to algorithmic trading, where efficiency gains compound exponentially.
Beyond just LLMs, self-improving models look to be emerging in the application layer, too. The belief is that evolution hinges on the user feedback loop.
Popular coding IDE Windsurf, which was recently acquired by OpenAI for $3B, just announced their first family of frontier models that has been tuned for agentic software engineering.
Windsurf’s Head of Research Nicholas Moy says: “The model today is good, but we see a path to making it exceptional. How? The data flywheel.
We have a best-in-class AI product where we constantly get signals on where the models are failing – and where we need to make them better.
Our belief is that the path to full automation of real work requires a tight feedback loop with the work being done, and that is the App layer.”
Likewise, Google DeepMind is developing a user interface with its People + AI Research team with Alpha Evolve and plans to launch an Early Access Program for selected academic researchers.
This is a glimpse of what evolutionary software can become: machines that learn not just from data but from their own improvements. It’s early, but it’s significant. We’re teaching algorithms to write and rewrite themselves.
Self-improving algorithms could ultimately lead to entire systems that evolve on their own, pushing the boundaries of efficiency, creativity, and problem-solving far beyond what we can currently engineer. AlphaEvolve might be the first of its kind, but it certainly won’t be the last. I have a feeling this is the beginning of a new way to think about software.
Nvidia and AMD partner with Humain to develop AI data centres: Nvidia and Advanced Micro Devices (AMD) have partnered with Saudi Arabian AI company Humain to supply semiconductors for a large-scale data centre project. The two parties have committed up to $10bn for the deployment of 500MW of AI computing power over the next five years.
Stability AI releases an audio-generating model that can run on smartphones: Stability AI released Stable Audio Open Small, a “stereo” audio-generating AI model that the company claims is the fastest on the market – and efficient enough to run on smartphones. Designed for quickly generating short audio samples and sound effects (e.g., drum and instrument riffs), Stable Audio Open Small can produce up to 11 seconds of audio on a smartphone in less than 8 seconds, claims Stability AI.
TikTok launches TikTok AI Alive, a new image-to-video tool: TikTok is launching “TikTok AI Alive”, its first image-to-video AI feature that allows users to turn static photos into videos within TikTok Stories.
Alibaba’s ‘ZeroSearch’ lets AI learn to google itself – slashing training costs by 88 percent: Alibaba’s Researchers at Alibaba Group have developed a novel approach that could dramatically reduce the cost and complexity of training AI systems to search for information. The technique, called “ZeroSearch,” allows LLMs to develop advanced search capabilities through a simulation approach rather than interacting with real search engines during the training process.
Anthropic released web search capabilities in the API. The feature allows web developers to build applications that can search the web for up-to-date information and provide grounded answers with relevant citations.
Perplexity's agentic browser, Comet, will launch in 3-5 weeks. CEO Aravind Srinivas said the browser will have native virtual meeting recording, transcription, and searches. However, it won't be a part of the first release, but a follow-up update.
AI power rankings upended: OpenAI, Google rise as Anthropic falls, Poe report finds: Poe’s latest usage report shows OpenAI and Google strengthening their positions in key AI categories while Anthropic loses ground and specialized reasoning capabilities emerge as a crucial competitive battleground.
The illusion of conscious AI. Neuroscientist Anil Seth lays out three reasons why people tend to overestimate the odds of AI becoming conscious.
Tobias Rees interviews CTO and EVP of Al Microsoft, Kevin Scott on craft and his approach to living life with AI
