AtomizedEconomicUnit

I see you every night in my dreams... sweet Beatrice #AI #lovers #art

@fedichev What is thermodynamically inevitable about it?

Most of human aging is thermodynamically irreversible. While it's a much disliked phrase. However, I think it's one of the most important and actionable statements in the field, because it means the goal is not rejuvenation. The goal is to stop the clock. And now let me tell you how. At @hacking_aging, we feed medical histories across tens of millions of people into physics-based machine learning models, which: • Use patients medical histories to predict how a person's health evolves over the full arc of their life • Pull aging out as a distinct process from specific disease trajectories This happens not because we told the models to, but because the signal is there in the data. Through this analysis, we identify genetic targets that control the rate of aging itself—not a particular disease predisposition or progression, but the underlying aging process closely related to configurational entropy of the aging organism. These genetic factors do not tell stories about treating particular diseases; they're about shifting the fundamental rate at which aging occurs. In 2021 we were the first estimate the maximum human lifespan from clinical data. What's gets measured - get optimized. Today, that estimate is approximately 120 years. This is how aging biology can be framed as a data problem, and the data analysis can reveal what no amount of experimentation alone cannot. Think of this post as a imminent new preprint announcement - please like, share and follow to know more, check and subscribe to my Substack using the link in the first comment.

@den_ewig @davidasinclair But have they started administering the first doses?

@AtomizedEcoUnit @davidasinclair According to the methodology described in clinicaltrials.gov/study/NCT07290… The participants will get the treatment slowly escalating in a sequential way. This is they will treat one, do a 30-day follow-up and then upscale or downscale dosage accordingly.

Safety is the endpoint. Recruitment of patients is ongoing

@SierraHotel28 @AlexanderKalian I was responding to the claim Xgboost was better at “predicting bioactivities of molecules” It sounds like you’re having a separate discussion better suited for someone else.

@AtomizedEcoUnit @AlexanderKalian Because if it takes more compute to do the same thing *it’s worse at the job*, not exactly the same. That’s the whole point of specialization, to achieve either better results, or the same results more efficiently.

Dear "AI will solve biology!" crowd... First make a deep learning model that significantly outperforms XGBoost at predicting bioactivities of molecules. Then we'll talk.

@SierraHotel28 @AlexanderKalian You’re adding a new constraint with “same compute”. Again, I would challenge you to read over what I have said on this topic. It really is quite clear and explicit. If you have a counter example to any of those points I am open to new information

@AtomizedEcoUnit @AlexanderKalian Precisely match in what ways? Because if you claim they can match XGBoost’s results with the same compute consistently, show some evidence, because I’ve certainly never seen it.

@SierraHotel28 @AlexanderKalian It may be easier to plug and play Xgboost in many situations with well structured and/or vectorized data but if you are doing research in this domain that should not be where your inquiry ends. I am challenging him to grow as a researcher.

@SierraHotel28 @AlexanderKalian The difference is that while NNs can precisely match Xgboost, the converse is not true. NNs optimal performance is strictly greater than or equal to the optimal performance of Xgboost.

@AlexanderKalian This is not an analogy…. I strongly encourage you to read through the actual algorithms, and ask yourself the following question: “What decision tree could not be replicated by NN connections?” This excercises will help your career tremendously

I understand the analogy, but it's overly simplistic. XGBoost is a boosted ensemble of decision trees; neural networks are differentiable compositions trained by backpropagation. Tree splits are not "basically hidden layers with activations" in the practical ML sense. Also, there is no general theorem that NNs outperform XGBoost "when done correctly". Expressivity does not imply better finite-sample generalisation. On tabular data relevant to molecular property prediction, boosted trees remain extremely strong baselines and often beat DL. You still have not elaborated on the benchmarks you referred to - which was my main question.

@AlexanderKalian @001TMF Right, as we’ve discussed before this just comes down to sim accuracy. You are under the impression AI has no ability to bridge that gap by devising more accurate and efficient algorithms/models/etc Which has already been disproven by AIs doing exactly that.

AI "solving" biology would imply AI (whether via LLMs alone, or a multi-modal system with tooling) being able to: perfectly navigate design of new drugs, predict bioactivities, pathways, pharmacokinetics, metabolites, efficacy, on-target and off-target toxicological effects, binding dynamics, interactions within mixtures or drug-drug interactions, and lots of edge-cases for rare mutations, epigenetics and so on. All of this, with negligible errors. And this is just a limited overview for pharmacology. Biology is actually much much vaster than that.

@AlexanderKalian I feel like I’ve explained it pretty well. Xgboost trees are fundamentally the same as hidden layers and activations just without the ability to model continuous interactions. It is mathematically provable that NNs outperform xgboost when done correctly. What is your confusion?

GNNs are actually one of my main niches. Happy to go into a lot of depth, if you so choose. Frontier labs are likely using them for molecular graph generation. They also tend to use GNNs for property prediction, but either in a way that is integrated with graph generation, or often as part of wider ensemble models that gauge multiple algorithms. Which benchmarks are you referring to? You can disagree all you like - the comments under my wider post are full of ML engineers in the life sciences, cheminformaticians etc. who agree.

@AlexanderKalian @001TMF Why is that ridiculous? AI solving biology doesn’t have to be a chatbot telling you if a small molecule or peptide will have side effects. AIs could simply improve existing or devise new deterministic approaches.

I think a lot of these "AI will solve biology!" utopianists seek AI that literally has no (or negligible) errors in predicting perfect drug candidates, phenotypes, disease trajectories etc. Others have some nuance for imperfection, but still mean something close to perfection. Both are of course ridiculous to expect.

@misraetel What work will humans do better, quicker, or cheaper than specialized machines or ai?

Once AI agents are capable enough to be CEOs of business, I predict that the number and size of companies doing valuable work will explode. And, big prediction: there will be so much real-world work for all these companies to do that we're likely to have massive human LABOR SHORTAGES and that this will provide the final major demand push to get robotics across the finish line of real world deployment at scale. I think this happens comfortably before 2030 if not sooner.

@AgroNationalism This has nothing to do with quantitative easing…. Dave Ramsey has lost touch with reality and is selling bootstrap porn to boomers. He barely pays his own SWEs half of what he suggests someone can get immediately after “coding school”

Dave Ramsey is practically the Germanic Protestant Saint of Money. Imagine how bad things would be if men like him didn't save millions of white people's finances over the last 30 years. He probably boosted the fertility rate by half a percent fixing family finances. That said, the era of quantitative easing inverts virtually all his advice. When the money printers go brrrr, debt is good, and saving your money is stupid. Many didn't realize this and are far poorer now because of it.

@gregisenberg Enjoying your life

What are the best businesses to be in a post-AGI world?

@AlexanderKalian There are currently relatively good simulators for interactions and biochemistry as of today. We don’t need wet lab data to solve the current limitations. The current limitation is compounding physical approximation. This would be resolved without AI it will be quicker with it

Biophysical simulations are problematic, for a variety of reasons. They are a supportive tool in research, not a one-stop shop. They alone cannot support an effort to expand data or verify candidate drugs, to "solve biology". Problems include: necessary approximations of classical physics, often a total lack of (or sometimes a heavy approximation of) quantum physics, limited simulation / approximation of solvent molecules, cut-off points for force fields, limited simulation of extremely complex and biochemically diverse environments, wet lab data needed to accurately inform simulated conditions, often a necessary lack of chemical bonding (would need simulation of atomic orbitals and more), discretised time steps, computational unfeasibility to fully atomistically model dynamics across a whole cell or organism etc. And this is a non-exhaustive list. I did a bunch of biophysics simulations, back in my master's degree research days, so it's a fun topic to revisit. I don't mean to patronise - it is genuinely this flawed. Big pharma R&D teams do a lot of these simulations, but treat them as a weak signal, supportive evidence, or early-stage virtual screening tool - rather than anything close to an absolute. And as for the other fields of biology - plenty (e.g. neuroscience, microbiology etc.) are deeply relevant to medicine and longevity too.

Every time I tell AI utopianists that biology is too complex for AI to "solve", they cite the success of AlphaFold. No, AlphaFold did not "solve" protein folding. It gets broad structures correct ~70-88% of the time (depending on evaluation), enabling useful but flawed statistical guesses. True "solving" would require ~99.9%+ accuracy, practically zero meaningful edge cases, and high confidence across fine details like side chains and conformations. Even then, this is just one narrow slice of the complexities of proteomics. The persistent gap between the "AlphaFold solved protein folding" claim and reality is a perfect example of AI overhype in biology.

@jiratickets Pittsburgh

@AlexanderKalian None of that is correct: - simulation would simulate physics it would not be an AI trained on wet lab data - simulation would be used to verify intervention without clinical trials not just train AIs - the other subfields aren’t related to life extension or medicine

I'm afraid that's not really realistic either, imo. Biological simulations alone are indicative only and flawed - we need real wet lab data, which for the scale required to "solve biology", is a major bottleneck. And even if we negated this, we would probably need 100,000+ years of non-stop simulating virtual experiments, with an armada of supercomputers. And this would possibly only begin to address a few sub-domains in biology, like pharmacology and toxicology - but not the wider space which also includes neuroscience, developmental biology, ecology, evolutionary biology, microbiology, astrobiology, synthetic biology etc.

@den_ewig @davidasinclair I was curious when phase 1 would start

@AtomizedEcoUnit @davidasinclair Currently it's in recruitment phase. clinicaltrials.gov/study/NCT07290…

So far, the data suggest ER-100 should be safe in humans, and the FDA has cleared us to move into clinical trials. Translation from primates to humans is never guaranteed If it doesn’t work as hoped, we’ll learn & improve, just like SpaceX did🚀