AlphaEvolve

When you have them in a harness, a harness is likely more sample efficient than having them in the weights, and so it depends what trade-off you want.

So if you were doing something like AlphaEvolve, where you have a very specific problem, how do we do four-by-four, uh, uh, complex matrix multiplication more efficiently, uh, then building a harness may well be the best thing you can do, a neurosymbolic AI to solve specific problems.

Um, and this is what you get in all of these, um, wonderful papers, things like AlphaEvolve, uh, also things like, uh, the Darwin girdle machine or hyper-agents from Jenny Zhang.

Um, they're all doing harness engineering, and they're great at solving specific problems.

But what we, what we've observed is that this doesn't tend to generalize, and what we wanted to do is build a system which you can then apply to a completely different problem, a quite a difficult long horizon problem, which is to replicate a problem in a-- replicate a paper in a completely different area of ML research.

They might even be able to use recursive progress to make themselves better and better.

To see how that might work, let's take a look at one of today's best AI coders, Alpha Evolve.

Basically, Google trained a large language model on tons of functions, pieces of computer code that performed specific operations, and let it start spitting out its own code, and paired it with an automated evaluator to check whether its functions actually worked.

All that meant FundSearch went through the whole process of attempting a function, learning from its mistakes, refining approaches and inputting those new functions to get even more successful outputs all by itself.

In other words, FundSearch could engage in aspects of recursive self-improvement.

And in early 2025, Google expanded on FundSearch to create Alpha Evolve, an evolutionary coding agent that trains on whole code bases, not just single operation functions.

Evolutionary coding agents like Alpha Evolve mimic natural evolution like the kind you see in nature by generating potential solutions, mutating them at random, selecting the ones that perform the best, and repeating that whole process until it gets something that really works.

And that means Alpha Evolve can learn to tackle all kinds of problems, from building a website to open mathematical research problems to, yeah, coding new models of AI.

Wow.

Otto is saying, sounds a bit like Alpha Evolve.

Yeah, it does.

And like, that sounds good.

The Irish airline said it would use Gemini Enterprise to develop custom AI agents to automate some decisions, improve crew scheduling, and reduce disruption.

The carrier said it would also use Google DeepMind models, including AlphaEvolve and WeatherNext, to support its fleet operations and mani- uh, maintenance scheduling.

Uh, financial terms of the agreement were not disclosed.

"The agreement demonstrates how deploying generative AI at scale can help industry leaders scale securely, reduce operational costs, and redefine the travel experience," said Maureen Costello, Cloud Google's vice president for United Kingdom, Ireland, and Sub-Saharan Africa.

En cuanto al uso, Gemini superó los mil millones de usuarios activos mensuales el miércoles, lo que coloca a dos asistentes de IA por encima de ese umbral en el mismo mes, después de que ChatGPT lo cruzara a finales de julio.

Y AlphaEvolve, el agente de codificación de Google DeepMind impulsado por Gemini, ya está disponible comercialmente a través de Google Cloud para clientes empresariales como Klarna, Schrödinger y WPP un año después de su presentación.

El sistema está generando una reducción del 30 % en las tasas de error en secuenciación de ADN y un error diez veces menor en circuitos cuánticos, según las cifras publicadas por DeepMind.

De vuelta con Amber para el panorama de capital.

Para lo último en la frontera de la investigación, le pasamos la palabra a Amber.

Un artículo firmado por diez autores, publicado el lunes en Archive, con AlphaEvolve de Google DeepMind como herramienta central y los mismos matemáticos que establecieron los últimos dos récords, ha llevado el exponente de multiplicación de matrices omega por debajo de dos coma trescientos setenta y uno mil ciento setenta y siete, frente al valor anterior de dos coma trescientos setenta y uno mil trescientos treinta y nueve.

La diferencia parece mínima, pero omega gobierna el costo asintótico de prácticamente todo cómputo reducible a productos de matrices, desde el álgebra lineal hasta los algoritmos de grafos, y cada avance en su reducción ha requerido históricamente técnicas matemáticas genuinamente nuevas.

De vuelta contigo, Edo.

That's a genuinely new mathematical object published in Nature.

And AlphaEvolve found a way to multiply four-by-four matrices in forty-eight scalar multiplications, improving on Strassen's forty-nine, which had stood since nineteen sixty-nine.

And I have to be precise here, that's in the non-commutative setting.

There's real prior art dispute in weaker settings, and we shouldn't overstate it.

Google's decision to open its first AI laboratory in Puebla, scheduled for early 2026, reflects the country's rising capacity to support advanced R&D ecosystems.

Furthermore, next generation capable of multi-step discovery and self-evolving systems like Alpha Evolve highlight the direction global innovation is taking and signal the types of capabilities Mexican organizations exceeded twenty-four billion dollars, projections ranging from twenty-five billion dollars.

With 2025, infrastructure identified as priority sectors due to nearshoring and the demand for high resilience digital platforms.

As global enterprises expand in Mexico, the strategic question becomes how local companies can align with the architectural standards expected by multinational organizations, including governance models to prevent agent sprawl verifiable agent identities using standards like SPIF least privilege access deterministic guardrails.