Last news from LODA@home: AI method replaced Boinc-brutforce crunch race.

Question: Does Number Fields project transfer to AI-search methods?

Last news from LODA@home: AI method replaced Boinc-brutforce crunch race.

Question: Does Number Fields project transfer to AI-search methods?

If I understand what you are asking, AI cannot improve the search method, at least not yet.

If I understand what you are asking, AI cannot improve the search method, at least not yet.

On an AI related topic, wondering if you have considered using AI as a coding tool to optimize the current NumberFields gpu application for speed? Some boinc volunteers have been using AI to optimize the gpu code. For example here in the Einstein@home project: https://einsteinathome.org/content/brp7-and-o4as-nvidia-cuda-special-app-now-windows-too. The custom optimized app is at least 50% faster than stock app. Depending on the gpu card and local settings, could be much higher, with at least 2-5x faster, from what I've seen esp on newer gen cards. Save us a lot of electricity cost and not to mention less carbon footprint.

From the sf6 batch status, do we still have to complete DS13x12, DS14x12, DS15x12, DS16x11 and DS16x12 batches? Hope you can share your insight to this.

P.S. We almost finish sf6 DF16x10.

If I understand what you are asking, AI cannot improve the search method, at least not yet.

On an AI related topic, wondering if you have considered using AI as a coding tool to optimize the current NumberFields gpu application for speed? Some boinc volunteers have been using AI to optimize the gpu code. For example here in the Einstein@home project: https://einsteinathome.org/content/brp7-and-o4as-nvidia-cuda-special-app-now-windows-too. The custom optimized app is at least 50% faster than stock app. Depending on the gpu card and local settings, could be much higher, with at least 2-5x faster, from what I've seen esp on newer gen cards. Save us a lot of electricity cost and not to mention less carbon footprint.

From the sf6 batch status, do we still have to complete DS13x12, DS14x12, DS15x12, DS16x11 and DS16x12 batches? Hope you can share your insight to this.

P.S. We almost finish sf6 DF16x10.

I would be very impressed if AI could make faster code that actually worked correctly. But anybody is welcome to try using AI to do it.

Regarding the batch status, all the cases you listed are needed. I am currently looking into a new method to speed things up - breaking the search into several pieces based on the Galois group of the field. It's tedious work, so will probably take a couple more months before I have something.

The custom optimized app is at least 50% faster than stock app. Depending on the gpu card and local settings, could be much higher, with at least 2-5x faster, from what I've seen esp on newer gen cards. Save us a lot of electricity cost and not to mention less carbon footprint.

Back in the days of Seti, one person's efforts to optimise the GPU code for Nvidia hardware resulted in Tasks that would take 3 hrs on a high end (at the time) CPU, taking 30sec or less on a high end (at the time) GPU.
AI, used appropriately, can certainly help with the optimisation and verification work.

---

Grant
Darwin NT, Australia.

Last news from LODA@home: AI method replaced Boinc-brutforce crunch race.

Question: Does Number Fields project transfer to AI-search methods?

If I understand what you are asking, AI cannot improve the search method, at least not yet.

Did you check this though? Seemingly written by AI. https://github.com/alplix/passiflora
I was wondering if it or similar techniques can be adopted for the official builds

Last news from LODA@home: AI method replaced Boinc-brutforce crunch race.

Question: Does Number Fields project transfer to AI-search methods?

If I understand what you are asking, AI cannot improve the search method, at least not yet.

Did you check this though? Seemingly written by AI. https://github.com/alplix/passiflora
I was wondering if it or similar techniques can be adopted for the official builds

Yes, I was wrong. I knew AI was doing software programming, but had no idea it had come this far to optimize a degree 10 polynomial discriminant and get an improvement as good as we are seeing.

I still need to completely test it before releasing it, and I've been busy this weekend. Soon...

Last news from LODA@home: AI method replaced Boinc-brutforce crunch race.

Question: Does Number Fields project transfer to AI-search methods?

If I understand what you are asking, AI cannot improve the search method, at least not yet.

Did you check this though? Seemingly written by AI. https://github.com/alplix/passiflora
I was wondering if it or similar techniques can be adopted for the official builds

Yes, I was wrong. I knew AI was doing software programming, but had no idea it had come this far to optimize a degree 10 polynomial discriminant and get an improvement as good as we are seeing.

I still need to completely test it before releasing it, and I've been busy this weekend. Soon...

I will also be adding CUDA, MUSA, and Metal versions soon; you can add them as well.

Yes, I was wrong. I knew AI was doing software programming, but had no idea it had come this far to optimize a degree 10 polynomial discriminant and get an improvement as good as we are seeing.

I still need to completely test it before releasing it, and I've been busy this weekend. Soon...

Here is what Terence Tao said recently about AI and math:

“We can have these 100,000-line proofs that we have to verify, but no one understands them,” he said. “Even the humans who entered in the prompts to generate the proofs might not understand them. It is not enough to generate and verify the proofs. The proofs need to be explained well enough that they can be communicated and understood by the community.”

I believe that's what you are trying to do and understand in that AI generated code. We can either prove or disprove that the results with large enough samples (via the boinc community here) but to understand why the new algorithm/method works or not is a different story and require a true mathematician and computer language/programming skill to pour over the code.

Excerpt from Terence Tao: https://www.simonsfoundation.org/2026/08/13/fields-medalist-terence-tao-on-artificial-intelligence-and-why-we-do-math/