You need to come up with structures and principles to make this process work for you and your team - but a regular reassessment of whether what you are executing is worth it is essential.

I also recommend using X% of your time for evolving a pool of ideas until they become ready for execution.

Importantly this is not a call to execute on fewer ideas - you need to test as many ideas as possible - such as 200-1000 possible variations on modelling choices (inductive biases, priors, data, tests) - but you need to take the time to come up with good ideas.

I tried voice mode a few times and found it quite unusable for anything complex.

I mean time writing, reading, responding.

Very exciting future to come.

9/n

Now my journey transitioning to mapping out how TF-CRE-GRN dynamics underpins how cells know what to do (the cell plasticity landscape), their collective roles/functions/goals within the organism and what models+data are needed to program cells and tissues to design the new therapies.

8/n

In addition to fascination with multicellular life and how brains become you, a big part of what motivates me to drive these challenging tasks forward is gaining understanding needed to accelerate medicine, last few months at Relation helped make it more practical.

7/n

- working towards decoding DNA syntax code of human organogenesis: how transcription factors interact with each other & with regulatory DNA elements (cell2state, ongoing work, @steglelab.bsky.social @bayraktarlab.bsky.social @teichlab.bsky.social @mhaniffa.bsky.social @leopoldparts.bsky.social)
6/n

This journey started with learning how proteins use peptide motifs to interact (@epetsalaki.bsky.social), building modeling foundation with @leopoldparts.bsky.social, then comprehensively mapping spatial tissue architecture (cell2location @bayraktarlab.bsky.social @steglelab.bsky.social )
5/n

- as well as how none of that is possible without collaborative and open science.

4/n

- learning and embracing how to understand cells with large scale multimodal cell atlases (towards whole organism), big biophysical AI/ML models (theory+data guide modelling and questions), and data-driven approach to discovering organising principles (catalogue then formalise)

3/

My my past 8 years at Wellcome Sanger Institute & EMBL-EBI (Bayraktar, Stegle, Teichmann, Parts, Haniffa, Petsalaki labs) have been both productive and formative

2/n

That internal use can be covered by internal regulations and RCA/MTA as well as GDPR - and could already be quite useful especially for large already collaborative organisations such as EMBL, Sanger, Crick.

Re purpose - I would suggest reading the full thread as it communicates examples.

Do you mean data protection/GDPR or do you mean legal regulation of how scientists interact (eg RCA/MTA)?

This full thread is more about the vision rather than implementation - but I think staged implementation with parameter sharing in specific institutions/labs could help test this.

There is also a risk that by not building an AI system that helps researcher community and mandating its use - non ethical use of generative AI will break the current system leading to AI use bans, people only trusting papers from those you know and halting the discovery and medical progress.

So if you use someone’s ideas suggested by the system - the system would show funders who contributed. You also still need to be the best person for the job and capable of delivering the grant.

It is possible to design the fine tuning steps in a way that allows attributing credit.

Staged rollout could help understand how original ideas from one chat are used for subsequent suggestions.

bsky.app/profile/vita...

Funders will be aware that the grant was written with help from this co-pilot system and the system will link related grants.

This system will have ethical use constraints by design - because it will be designed as a co-pilot for research community. Someone writing grants by getting ChatGPT to generate most of the proposal will do worse than someone who uses our community co-pilot to help shape the project.

But it doesn’t understand the full context of what I am doing, to get good results I need to write very detailed prompts that give what context I think is necessary. It also doesn’t have useful templates for a more structured discussions. I also don’t want OpenAI to train a public model on my chats.