[ExI] Meet AlphaEvolve, the Google AI that writes its own code

[Stuart LaForge]

On 2025-05-14 11:58, Darin Sunley via extropy-chat wrote:
> Traditionally, recursive self-improvement has been understood as AGIs
> that can use the full force of their IQ to modify their own code to
> improve their own IQ. Repeat until apotheosis.
> 
> With LLM-based AGI, that would involve direct manipulation of weights
> that would probably require mechanistic interpretability knowledge we
> don't have yet (although I'm quite sure the major players are
> significantly ahead of anything they've published publicly in that
> field).

That is an interesting idea. If IQ in LLM is based more on weights, i.e. 
training, than it is on the ANN algorithm itself, then AI could be 
limited by its training set. In other words, AGI could be as susceptible 
to GIGO such as certain cognitive biases, prejudice, and superstition as 
people are.

> On Wed, May 14, 2025 at 11:40 AM BillK via extropy-chat
> <extropy-chat at lists.extropy.org> wrote:
> 
>> Meet AlphaEvolve, the Google AI that writes its own code—and just
>> saved millions in computing costs.
>> Michael Nuñez     May 14, 2025
>> 
>> 
> <https://venturebeat.com/ai/meet-alphaevolve-the-google-ai-that-writes-its-own-code-and-just-saved-millions-in-computing-costs/>
>> Quote:
>> AlphaEvolve pairs Google’s Gemini large language models with an
>> evolutionary approach that tests, refines, and improves algorithms
>> automatically. The system has already been deployed across
>> Google’s
>> data centers, chip designs, and AI training systems — boosting
>> efficiency and solving mathematical problems that have stumped
>> researchers for decades.

Yes, developing a new faster algorithm for matrix multiplication sounds 
impressive.
Here is the White Paper on AlphaEvolve (44 pgs):

https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/AlphaEvolve.pdf

>> AlphaEvolve has been quietly at work inside Google for over a year.
>> The results are already significant.
>> -----------------
>> AIs are rewriting their own code now!
>> This sounds like a major step forward towards AGI.

I hope somebody human is looking through that code for backdoors and 
other mischief. This seems more of a recipe for misalignment than it 
does AGI.

Stuart LaForge