Can someone not on google use it?

[Rubiel1]

Hello,

So there is an update from deepmind alphaevolve

Which makes me wonder, in this time, was someone (not on google) being able to use funsearch?

According to the paper, it cost around 1400 dollars per experiment (with the possibility of failure).

I tried running it on hpc and it would take months.

I understand this is not a technical issue the team can solve, but I think it is important for the people coming here to notice the strong computational requirements.

[Rubiel1]

Does any of you know someone that successfully used funsearch? @joelonsql, @jonppe, @jbdatascience, @jbdatascience, @psikosen, @pwnedzen, @HJ9412, @HAFDIAHMED, @Visheeeee, @KOVVURISATYANARAYANAREDDY, @fraspin, @jack-powers, @1thereckless, @nacalorea
, @KevinH48264

[1thereckless]

Maybe you can try this open-source version of Funsearch: https://github.com/RayZhhh/funsearch. Funsearch does indeed have a strong computing power requirement, as it constantly calls LLM and evaluates the generated functions. You can start by selecting some affordable LLMs to assess the effectiveness.

[jack-powers]

Agree with using the above-mentioned repo as a starting point. The original paper used a distributed set-up, but I have found success in both single and multi-threaded applications (i.e., parallel samplers, etc.).

For the LLM, running Ollama on a HPC GPU is one option, and another would be using an API based LLM (so you would only need the HPC for CPUs). Most input tokens from Funsearch are cacheable too, further reducing costs. Costs would vary depending on the LLM. Deepseek coder seems to be an affordable option (utilising the caching as mentioned).

[Rubiel1]

Now that deep seek made everything cheaper, can any one estimate the amount needed to replicate an experiment of the original paper? they said "We estimate that when running on Google Cloud, the price of an experiment is around $800 – $1400", but now, how much would that be?
Also, is there an open-source AI with enough quality? did Ollama worked?
Setting up everything takes time and effort, and last time it was not successful, so please save us time with your knowledge.