Aurel Prosz

CTO at PharosBio @PharosBio. Interested in cancer genomics and applying LLMs in science. Follow me for news and repos I find online about AI and science.

[Aurel Prosz @aurel_pr]

Yesterday I finally set up one of the largest, open-source LLM models in the world, GLM-5.2, on Gefion, the Danish supercomputer, to find new synthetic lethal relationships in cancer. It's going strong at around 5 tokens/second (I have to optimize the inference a bit...), evaluating a new gene pair every ~15 mins.

Qualitatively, the reasoning capability of the model already looks better than that of Qwen 32B-instruct, but requires 2 full nodes, 2x8 H100 GPU-s to run it, and it's not quite trivial to set up the inter-node sharding.

See comments for our manuscript, now in peer review.

[Aurel Prosz @aurel_pr]

@emollick Thanks, we had a blast experiencing the apocalypse with this guy. In the next version I'd like to also be able to share a tea with him.

[Aurel Prosz @aurel_pr]

If you want to understand the current breakthrough in AI about resolving the Navier-Stokes Millenium Prize problem, check out this interactive textbook I made with GPT-6 Astra. You can select a explanation level from kindergartener to math professor.
https://t.co/NP0gYHVuIR

If anybody is interested in the technical details:
- I downloaded the pdf containing the proof from OpenAI website
- Used GPT-6 PRO to create a detailed prompt for Codex
- Used Coded with gpt-6-astra extra-high mode to generate the app
- Deployed on Netlify

Build time ~2 hours end-to-end.

Keep in mind, this is an educational tool, LLMs can hallucinate, be mindful when interpreting the results, and all the usual stuff.

[Aurel Prosz @aurel_pr]

@mark_k I'd like to join

[Aurel Prosz @aurel_pr]

@mark_k Please add me as well :)

[Aurel Prosz @aurel_pr]

What are the "conjectures" for cancer genomics?

[Jake Wintermute 🧬/acc @SynBio1]

https://t.co/39D9ejtsLf

[Aurel Prosz @aurel_pr]

https://t.co/WaEZfYnmY4

[Aurel Prosz @aurel_pr]

One of the reasons I'm optimistic about AI progress is that it unlocks our creativity in new ways to communicate our research: over the weekend I built a working 3D video game for fun that lets you control a miniature spaceship inside a cancer cell, grabbing proteins and letting the original algorithm used in our synthetic lethality project calculate the probability of those being a good candidate for further evaluation. Funnily enough, this is a super inefficient way to run the algorithm, but it's a good way to show what our research is about to people who are not familiar with the field, instead of just showing a graph or a list of numbers. And if you are lucky, you can actually pick the right combination, but good luck, there are around 400 million to choose from.

Some technical details:

- Okay, to be fair, it was codex from OpenAI and gpt 5.6-sol-xhigh that built the game, I just supplied the vision, the general direction and encouraged the AI on the way. I have zero expertise in game or 3D asset development, but this is also kind of the point.

- All assets, animation, textures were generated by codex. The whole build time was around 4 hours, and utilized ~30 subagents

- I used @mattshumer_ 's Gauntlet Loop for sub agent orchestration, link to it is in the comments

- The game currently does not use real physical simulations on how the molecules move in the cell, very limited amount of genes/proteins to pick, etc, and is illustrative at this point. If there is interest I'll keep working on it and open-source it so anybody can try it out.

What do you think about this approach to communicate research? What will you build?

[Aurel Prosz @aurel_pr]

@kimmonismus visualized as a timeline

[Aurel Prosz @aurel_pr]

@PrimeIntellect @omouamoua Thanks for sharing this, I'm having a lot of success with Prime Agent in my research (cancer genomics). If anyone is interested I built an observation tool for Prime Agent and similar frameworks, so you can visualize how your harness evolves in real time:
https://t.co/MDu6Qecj8y

[Aurel Prosz @aurel_pr]

I built a debugger and visualization platform for deep and wide recursive agent swarms, so you can monitor how your harness evolves in real time.
RLM Agent Observatory shows topology, execution history, per-agent output, causal traces, and learned knowledge in one live view.
Focusing on Prime Agent today, but more frameworks coming next. 🧵

[Aurel Prosz @aurel_pr]

Open source: https://t.co/EtDd09tlnW
What would you inspect first? I'm using it on launching agent swarms to find novel synthetic lethal targets in cancer, but comment your ideas below.

[Aurel Prosz @aurel_pr]

I built a debugger and visualization platform for deep and wide recursive agent swarms, so you can monitor how your harness evolves in real time.
RLM Agent Observatory shows topology, execution history, per-agent output, causal traces, and learned knowledge in one live view.
Focusing on Prime Agent today, but more frameworks coming next. 🧵

[Aurel Prosz @aurel_pr]

Knowledge belongs to the run.
Memories, skills, prompt notes, and refinements keep their provenance. Save or load the complete run and knowledge together. It natively matches the Prime Agent's harness state, so the project self-learns.

[Aurel Prosz @aurel_pr]

It's interesting to see that so far our reality follows quite closely the AI 2027 predictions (https://t.co/27dWnRU4Nx), started a year ago. If you haven't read it already, give it a go. I wonder if it eventually diverges somewhere significantly, since the predictions for the next 6 months are quite out of a sci-fi book. By early 2027, AI 2027 expects AI to move from assisting research to materially accelerating the development of better AI.

[Aurel Prosz @aurel_pr]

@arcprize @AnthropicAI Awesome, I’m looking forward to the ARC-AGI-3 results!

[Aurel Prosz @aurel_pr]

I've conducted a series of tests with the newly deployed GLM5.2 model on Gefion, comparing it with smaller models such as Qwen 2.5-32B-Instruct and Ministral-3B-2512-Reasoning to assess their ability to identify known synthetic lethal genes in cancer from experimental data. Interestingly, GLM5.2 performs quantitatively and qualitatively better on both known positive and negative ground-truth cases, and there is a clear correlation between performance and parameter count. The video shows two such cases from thousands.

Despite this performance gain, I'm not entirely convinced that the improvement is not simply the result of the model being able to memorize more biological knowledge from its training data and use it more effectively than smaller models, rather than demonstrating genuinely better reasoning. I'm currently developing a test suite to determine this explicitly.

[Aurel Prosz @aurel_pr]

Fine-tunes a 650M-parameter protein language model to classify enzyme families while controlling homology leakage, calibrating uncertainty, and mapping predictions back to important residues.

https://t.co/C14HXpjClk

[Aurel Prosz @aurel_pr]

Extracts biomarkers from uploaded blood-test reports, checks them against deterministic ranges, then uses an LLM to explain each result and generate a readable summary.

https://t.co/EpEOlQg55a

[Aurel Prosz @aurel_pr]

https://t.co/sGAXsCYoqM
Here is the whole figure with the output tokens included.

[Aurel Prosz @aurel_pr]

What if you could just install one skill for Claude or ChatGPT and reduce ~50% token usage? Since using tokens is starting to look like a utility bill, I set out to find ways to reduce my usage (and also, because nowadays I hit my weekly limit in 24 hours).

For this, I tested Caveman, a skill, that simplifies LLM output to a simpler way, more compact format, while promising to keep the same quality.

First, I made a prompt instructing Codex with and without the Caveman skill, to build me a protein visualization and dynamic simulation app from scratch, and I measured the token usage.

From the figures you can see that it indeed significantly reduced the token usage on this problem, but it's worth exploring in the future if the quality of the output won't degrade on different tasks. More specifically, enabling Caveman reduced total session token usage by about 52%. Although Caveman does not directly compress input per conversation turn, its shorter responses produced a smaller conversation history, and Codex completed the task with fewer model requests and tool calls.

[Aurel Prosz @aurel_pr]

Some observations:
- This is not a comprehensive comparison, take it as an exploratory one. For sure I'd need to rerun this experiment n times to get a good estimate on the reduction.

- Qualitatively the results were the same for both runs, but the end results looked a bit different: buttons placed differently, etc

- This was a relatively straigthforward task for the LLM, I'd be interested how this performs on more complex and/or novel tasks

- The Caveman-OFF run just hosted the app on my computer, Caveman-ON tried to host it using OpenAI sites feature, but I never asked for this. A bit of disalignment in my goals.