Well if any labs want to scrape my knowledge of such things into model weights or faster training runs, I do consulting at movsoft.io :-)
It really is [as simple](https://github.com/ratatui/ratatui/pull/281#issuecomment-160...) as the Molly rocket video about terminals makes it seem; I am sure the models will catch up eventually. They have to overcome a lot of bad training data in most languages, but I find in Rust they usually do ok the first time.
I'm sure their numbers show the tablet in use, as corporate desires. I place my order with the waitress and she picks up the tablet to place it, citing "that's what they want us to do"
"That's not how citations work." (Dude on the interwebz, 2026)
But more seriously, this is my first time seeing that as well, and I'm not sure I like it. Citing an LLM is a little like citing Wikipedia to me, you cite the primary source the LLM is quoting directly, not the secondary source.
Far worse from my perspective is that it's not a fixed citation that it's not falsifiable. It's just a mutable, generalized source, which is stochastic in nature.
It’s the formatting and formality that bother me. Say “I looked it up with Gemini,” don’t give me some weak attempt at “proper citation” to legitimize the bare minimum effort you put in.
So much of present day feels like being back in middle school around the turn of the millenium, with classmates (and even some teachers!) new to the internet insisting that "source: internet" or "source: google" was totally fine.
They’re buying a brand, some employees, and some momentum — not any of their software. HF probably does have propriety goodies to make it all run efficiently, but certainly not a billion dollars worth, much less 13!
We are so used to these numbers being thrown around in the AI era that something one needs a reminder that this is 13 billion, not million. Insane exit by HF.
Was looking for this comment, the tech world has gotten completely insane with ”valuations”. I would love to hear why it was 13 and not 5. It would still be completely insane at 5 billion, but someone though they should add another 8…
Opportunity cost for Nvidia to prevent HF from going public or being sold to someone else. It's an insight into Nvidia's market prediction and what HF told them they'd do if left alone.
No company is going to be acquired for their software alone with multi-billion valuations ever again. It should be clear by now that code alone is not where the value is anymore, if it ever was.
I resent the accusation that no one uses the all caps version :-)
I found support for generating it to be hit or miss with the free online generators, but it really improved scannability for a QR I was pasting onto a vehicle!
what is the state of the art now, that runs on-device?
the article didn't mention the Q$ family, but I haven't found anything better than that plus mixing/co-judging with a dynamic-time-warping based approach
Perhaps the fact that 99+% of today's workloads could be running on the client if it were as easy as shipping Rayforce and the data directly to the client.
Besides that, pure C that you can embed into your app is much easier to deploy for some (and likely 100x more performant) than stuff that comes via Helm chart [cries in JVM 'big'-data solutions]
I've been experimenting in this space, where might I find a guide for what to build that would be useful to you? I suspect most existing approaches are an order of magnitude slower and harder to use than they need to be.
I would reject the premise that the field of molecular/biological simulation is underexplored nor that existing approaches are "slower and harder to use than they need to be". This is a field that has been explored, in fact, by the most brilliant minds and the difficulty arises more in theoretical considerations (that is, devising algorithms to faithfully approximate the developed physics) rather than an obvious no-brainer application of AI.
The field of molecular and biological simulation is far more than simply "Newtonian mechanics". There is indeed a field called molecular dynamics (MD) that relies on "classical mechanics" yet it's defined usually in the Lagrangian formalism. Furthermore, there has been tons of work over the past few decades in developing more accurate numerical approximation algorithms. There is a ton of a theory in this field and if you're interested, the "MD Bible" is "Understanding Molecular Simulation" by Daan Frankel.
Now, MD is just the tip of the iceberg. Almost all chemistry simulations are built entirely from making subtle approximations to quantum mechanics and carefully building up frameworks. For example, Hartree-Fock theory (HF), Density Functional Theory (DFT), Couple Cluster theory (CCSD(T)), etc. Then there is a field known colloquially as semi-empirical methods which are a sort of combination of the above two methods. And that's just on the side of chemical simulations (i.e. I'm excluding physics-specific simulations etc).
And now, more recently there has been effort in building machine-learned interatomic potentials, machine-learned density functionals, equivariant graph neural networks, etc etc.
If you're still interested in these class of problems, consider trying to build a good model for OMol25: https://arxiv.org/abs/2505.08762
reply