Marcin J. Suskiewicz

So does it exist in the end?

Beautiful work!

Banner image with screenshot of scientific article from nature Medicine, as well as two panels from the study method and results

⚠️ Despite all the hype, chatbots still make terrible doctors. Out today is the largest user study of language models for medical self-diagnosis. We found that chatbots provide inaccurate and inconsistent answers, and that people are better off using online searches or their own judgment.

On the balance of knowledge - Nature Reviews Immunology Ruslan Medzhitov shares his thoughts on the balance between generating data and developing theories in immunology, with a focus on exploring the rules that govern complex systems.

Too much data, too little thinking.
A important essay from Ruslan Medzhitov on the importance of understanding data, not just generating it. A must read.
@Yale
www.nature.com/articles/s41...

Lowering the temperature allows transitions from the ‘sol-sol’ to the ‘sol-gel’ region, which manifest with a jump in the total volume fraction of the protein-rich phase.

The interplay between biomolecular assembly and phase separation.

🔗 buff.ly/lvTQhIM

🧪🧬New preprint We present cryo-EM structures of reconstituted CTCF–nucleosome complexes, showing CTCF dimerization drives nucleosome oligomerization into defined higher-order assemblies. Disrupting CTCF–CTCF interfaces in mESCs reduces looping and impairs differentiation. tinyurl.com/CTCF-nucleos...

Finally, a detailed review about C-mannosylation: mechanism of C-mannosyltransferases and the influence of the C-mannosylation on the function of several proteins. Highly recommended to everyone in this type of post-translational modification!
tinyurl.com/3ywjmfk3

#glycotime

I also wondered if any of the dimerisation contacts are the same as in the crystal structures 8SSS, 8SSQ? It doesn't seem so, though.

This is stunning! Super work. Would it look the same regardless of linker DNA length?

Look at this 🙂 #CryoEM

Structural basis for CTCF-mediated chromatin organization by @lucas.farnunglab.com @voslab.org

www.biorxiv.org/content/10.6...

In memory of Professor Israel Silman (1935–2025) Click on the article title to read more.

I was saddened to hear (with delay) of Israel Silman ('Sili')'s passing. A kind and very intelligent protein biochemist whom I was lucky to briefly work with.

A near-complete map of human cytosolic degrons and their relevance for disease

We measured degron potency of >200,000 30-residue tiles from >5,000 human proteins, and trained a model to predict degrons from sequence

Led by @vvouts.bsky.social in @rhp-lab.bsky.social

doi.org/10.1126/scia...

I am excited to share a new review @cp-molcell.bsky.social written in collaboration with @tanjamittag.bsky.social , Mikayla Eppert, and Ambuja Navalkar where we review the current evidence for and against the role of density transitions in regulating transcription www.cell.com/molecular-ce...

New preprint from the lab! We identify the ZnF protein Mulberry as a condensation-dependent structural regulator of genome topology that organizes “multi-way regulatory hubs” in early Drosophila embryos.
www.biorxiv.org/content/10.6...

Absolutely! - and all the insights into the usefulness of evolutionary couplings for structure determination, as well.

When certain patterns are implicitly present in large high-quality human data, one should not give too much credit to the machine learning algorithm that mines these data for patterns.

And AlphaFold was only possible because a huge number of experimental structures from which one could generalise was produced and catalogued by scientists over the years (and similarly for protein sequences used for MSAs).

(Because keeping it ontological can also prevent you from excluding any human who doesn't [yet] manifest a given quality in a way we want it to be manifestable, as I tried to argue under the previous post).

A part of it is whether the term is going to be demeaning/insulting. Even then, I can see the force of your objection, but I would still stand by the fact that keeping categories ontological and thus ammenable to being protected by the mimicry objection is a double-edged sword with a positive side.

I, personally, think that something that is more technical and focuses on operation would be better than any already widely used, 'casual' term. People do not object to 'efficient' or 'useful'. Perhaps one can say more than that, but in a similar vein.

This might be interesting:

Me later, 2011-2018. It would be easier to make everyone agree if we (or Turing) took a new term with less ontological and emotional baggage. I think you cannot just decide a term that is widely (and diversely) used to be easily definable.

One last thing I would add. We cannot escape the fact that our knowledge of humans (and perhaps other animals by extension) is largely through introspection. So it is not only the 'opaque' knowledge of how brain works. And we can have reasonable doubts this is extendable to LLMs.

(By the way, it seems we were both at the IMP at some point @giorgio.gilest.ro . Thanks for the discussion!)

My answer is in later posts, where I argue for the need to define this as an ontological capacity that might not be (yet) manifested - as in trainees, which is why they are trainees. Human beings developing in time is very different from a machine that you might want to benchmark at a specific point

I think freezing 'intelligence' as operational/measurable in any other way could be dangerous too, as is, in my view, extending it to machines that, as we know, were actually deliberately designed to mimic manifestation of intelligence without its ontological underpinnings.

... as there in principle rather than only when manifested, knowing that manifestation is contingent on lacks of obstacles and presence of supporting factors etc. The ellusive nature of the definition is a guardrail - just like any attempts to freeze it before ('IQ' etc.) had an effect of excluding.

And the Phillis Wheatley example is, I think, dangerous. Making certain commitments ontological might make them more difficult to extend - but it also prevents using them to exclude. I do think there is a certain robustness (and truth) in acknowledging a shared 'ontological' human capacity...

I think anyone trying using LLMs knows that a little 'push' in the right direction from a human operator at each step makes all the difference in almost any slightly more complex process - I don't think that's an illusion.

My counterargument would be that, perhaps, the moving goal post phenomenon is not dishonesty, but rather trying to name a certain reality that is difficult to put into words but is genuine, and is neither sentience nor consciousness per se, but is indeed about navigating intellectual questions.