codekilla

Thanks for the reference. Have you ever worked with Maude? Curious what the advantages of one over the other might be. Maude seem like it might be more focused on being a meta logic, and I'm guessing it is probably easier to write programs in Prolog.

I guess Turing couldn’t see the trillions of base pairs of DNA, complex methylation states, dendritic spines of the neurons, etc., just for starters.

The author seems to be of the opinion that the creators of DeepSeek will either be unable to, or will not see the value of optimizing the 'second stage' RL component of the 'new' (post pre-training RL) way of training frontier foundation models. Every competent programmer in China is now looking for low level ptx optimizations for EVERY SINGLE STAGE of the pipeline. They will now, likely not publish any of it.

I don’t disagree with any of these points, it’s just I’ve been involved in open science circles, where these things are always mentioned, and I just don’t see any material progress (maybe I’m not looking that closely though). I think the reason for the lack of progress is mainly funding—so until someone gets serious about funding (billionaires or taxpayers), it just seems like the same merry go around. It’s very expensive to replicate biomedical studies—but it’s the only thing that works. Maybe the tide is turning though and simply incentivizing/protecting grad students to become whistleblowers will do more good, but I fear this case was more the exception than the rule.

> To expect a national-level organization to essentially take over the duties of peer-review journals is a very big ask

Fair, but what is the alternative that would actually work? What is the budget of all of the journals compared to the NSF+NIH? Is medical research that is true, and certainly actionable worth as much as an F whatever fighter jet? People will have to decide.

Sure, and that's on the private funders to ensure they are getting what they pay for. Google pays for plenty of research--since they are the payee, it's their responsibility to ensure it's accuracy to whatever degree satisfies them. Institutions like the FDA are supposed to regulate private research when it comes to market (.i.e pharmaceuticals and the like). Whether or not the FDA and related agencies are effective is a different, but just as important question. Taxpayers desperately need a formal, funded system to verify the science they are paying for--particularly for biomedical research where the incentives for fraud are so high.

Peer (or any type of institutional) review needs to be implemented at the national level--same as the funding for the original research. Why would you pay for research and not check that it is correct? Congress needs to fund a new science agency that explicitly does this. I have suggested before that part of graduate training should be replicating select studies that are published (a national review board could select those that seem the most high impact). State-funded schools could take this on, and students would probably learn at least as much doing this as they do in their other studies.

For dependent types, I would look at Idris [1]. Adding Univalence in a satisfying way is I think still somewhat of a research question (I could be wrong, and if anyone has any additional insight would be interested to hear), i.e. see this thread about Univalence in Coq [2]. There are some implementations in Cubical Type Theory, but I am not sure what the state of the art is there [3]

[1]https://www.idris-lang.org [2]https://homotopytypetheory.org/2012/01/22/univalence-versus-... [3]https://redprl.org

Bob Harper wrote a really good blog entry that expounds on this as Computational Trinitarianism [1].

Michael Shulman also wrote about the extension to Homotopical Trinitarianism [2]

For a good summary with links there is [3]

[1] Computational Trinitarinism, https://existentialtype.wordpress.com/2011/03/27/the-holy-tr...

[2] Homotopical Trinitarinism, http://home.sandiego.edu/~shulman/papers/trinity.pdf]

[3] nCatLab, https://ncatlab.org/nlab/show/computational+trilogy

Thanks, yes, I was thinking along the lines of HPC type applications in industry.

Thanks, this is helpful. It seems like (based on your reply) there are people successfully using Regent for scientific computing (I'm assuming); do you think the language is a viable choice for industry, or are there particular milestones you're looking reach?

How would you characterize/contrast Regent vs. Chapel? What do you see are the main drawbacks/benefits of each?

Bought current house without agent (very desirable part of Los Angeles). Selling now without agent. If you buy with an agent you put yourself at a disadvantage because the selling agent will need to split the commission (typically 2.5% a piece or so). When you make an offer on a home without a buying agent, suddenly your offer looks a lot more attractive to the selling agent, who is the only point of contact the seller has into what is happening with their property in terms of offers. People wonder how we got our house so cheap—bank on the real estate agents being greedy. They are the worst, period. I have not met a single one who will not double end a deal in 10 years in the LA market. Not sure how the current sale will go, but I will not work with an agent, I’ve dealt with too many to make that mistake.

The Quantenkoffer by qutools is an interesting kit for learning about quantum stuff [1]

Also, Qiskit pulse was fun to play around with (haven't in awhile, don't know what the current capabilities are) [2]

[1] https://qutools.com/quantenkoffer_science-kit/

[2] https://qiskit.org/documentation/apidoc/pulse.html

Mathematical Biology/Bioinformatics. We have to think very carefully about every step in the process of extracting information from large, diverse datasets--often writing things from scratch, combining/transforming things in novel ways, and implementing new mathematical ideas efficiently enough to be computable on large datasets.

Absolutely spot on. I actually do algorithm design, usually over a period of weeks (at least), and leet code is a joke for the serious algorist (I’m sure I’d fail an interview based on it). Nothing has so clearly illustrated the robotic nature of the leet code expert quite like this result has.

Try Eugenia Chang’s new book, The Joy of Abstraction. Riehl’s book is very technical.

Something like this: https://en.wikipedia.org/wiki/Theater_of_All_Possibilities, where they were guided by the idea we could be artists, explorers, and scientists.

It’s nice to finally have a name for what I suppose is an incurable addiction (maybe 1K books in my library). The most fun for me is buying almost exclusively used books, and then finding interesting inscriptions in them, or when the books come from an interesting place (many have ex-university library marks/stamps). For instance, I have a first printing of Weiner’s The Human Use of Human beings that came from the Redford Arsenal Library.

Well....for starters there is the monkey tunnel that leads from Broad to the Chen building because....you guessed it....the vivarium was in the basement (and still is) of Broad. Many neuroscience labs were in Broad (a VERY nice building in and of itself).

I believe at least 170 million was slated initially for construction costs, which may have run over. I don't know how much was for new equipment, but I can't imagine it was a significant fraction (and likely it was an additional expense over those numbers).

I worked in the Broad building at Caltech (Eli Broad), and I watched for 2 years as they demoed the parking lot outside my office to build the 200 million dollar Chen neuroscience building. It is a complete and total waste of money, and I say that cognizant of arguments like 'attracting talent'...etc. At one point we calculated the number of neuroscience postdocs we could hire for that money and easily came to the conclusion that we could have pulled off a Manhattan project of neuroscience with that kind of money. It's pathetic that donors delude themselves into thinking that projects like this do ANY good.

Edit: link: https://www.henselphelps.com/project/chen-neuroscience-resea...

Is the pitch directly to taxpayers? There have been lawsuits (by faculty even..including one in which the DOE was deposed) to rein in misappropriation of funds etc., but it seems like the agencies themselves don't care about conflicts of interest at these universities (and basically encourage them). Is it a research organization that says: 'Hey...we only do open source patents'?

Thank you!

Yes, it's admittedly an incredibly tough problem (how do you get those competing to cooperate?), and how to compute on data that is trustworthy. Ultimately, in biotech/pharma you have to ultimately know what an underlying gene or say pathway actually is, it can't be totally obfuscated. Still not sure how to set this up (if it's really even possible) that solves for this use case.

> but a system where pieces of a research puzzle are stored on chain and each user can claim ownership of those findings, a resultant drug's profits could be proportionally split by every entity which contributed to the research.

I think ultimately this is how a research cooperative could work. If distributing and re-allocating fractional ownership is efficient enough, it seems like something like this might be feasible. The idea with multiparty communication (MPC) is that in this setup a research entity would contribute their data in an encrypted fashion, and any parties would be granted access to compute on it based on some set of rules/buy in etc.

This is a really difficult technical approach, as MPC is really only in it's infancy, made only to seem easy by the far more difficult and distant prospect of socializing medicine, which would seem to be of the greatest benefit.

Yes, essentially, though you may create models of interactions etc., but the main idea is to extract information from various aspects of the cell.

As far as in silico, I think absolutely there are probably opportunities here. Generative models might be useful for some type of counterfactual (automated) reasoning with respect to disease course/treatment. I think we're in the relatively early days of collecting high resolution cellular data, so I think in silico approaches like this will be more and more relevant.

I've thought about things like the patent/ip problem, the structure of biomedical research, Pharma research, etc. This is an area where I don't actually see competition as a net benefit, however....it's the reality. The only thing I can come up with is a version of 'data rental'. Rather than Pharma companies locking this data away from others indefinitely, is there a way they could profit from it somehow, while still retaining ownership and not divulging trade secrets? Maybe not.

I've thought that a type of cryptographic data commons based on multi-party communication [1] could possibly be deployed with some effect. Basically you need algorithms that can compute on encrypted data, and a way to securely communicate encrypted data. There might not be huge incentive to use something like this, but maybe a version of this idea could work.

[1] https://en.wikipedia.org/wiki/Secure_multi-party_computation

Interesting....I'll have a close look at this, thank you. I didn't mean to imply drug repurposing was straightforward, certainly as you say this is very challenging. I guess my thinking was that there might be relatively lower hanging fruit here (if Pharma companies have very little incentive to exhaustively search for repurposing targets for off-patent meds, maybe only a non-profit would be willing to do this) than say de-novo development.

There are a lot of modalities being integrated, things like spatial/temporal, ADT/protein, etc. Integrating all of this data is a computational challenge, and of course there are lots of methods for analyzing it that vary in computational demands. It's not simulation, but still a lot of processing.

I have been reading a book recently: The Story of Taxol: Nature and Politics in the Pursuit of an Anti-Cancer Drug, and one of the most fascinating parts was the way they discovered this molecule. Long story short, Taxol is a molecule they isolated from the bark of the Pacific Yew. The interesting part for me was learning about the Cancer Chemotherapy National Service Center [1]. They went around collecting samples of random plants, then tested them for anti-cancer properties very systematically. So in the U.S., at one point, we had a publicly funded drug discovery program targeted at a specific disease, and this is what jump started Pharma research in anti-cancer drugs. I would say we need to restart a program like this, and of course we should also focus on rare diseases--we stand to learn a tremendous amount, and it's difficult to convince industry to do it.

Personally, I'm a computational/mathematical biologist and I work on single cell data targeting multiple myeloma, I'd really like to see serious non-profit Pharma. Drug repurposing seems like the most feasible avenue. What I know of right now is open Pharma [2].

[1] https://dtp.cancer.gov/timeline/flash/milestones/M3_CCNSC.ht... [2] https://www.ospfound.org