daemonk

I wish there was something like LEAN for biology. But there just isn't nearly enough data to build anything close to a good drug discovery simulator. To have a LEAN for drug discovery, we would need to simulate human physiology down to the molecular interactions at a reasonable accuracy.

We currently have disparate datasets that allows us to make some predictions on how drugs will interact with molecules. And we have studies on how these molecules might affect our organs/bodies. But all of this data are at different resolutions, sparsity, and have different experimental context. It is extremely hard to reconcile for a holistic model.

DNA is the great unifier in molecular biology as it is the digital anchor/foundation that everything above relies on. But there is just so much complexity above it. An analogy in the AI space would be model distillation where you are trying to discover all the weights in a LLM by giving it a prompt and seeing what the results are. With biology, we know the prompt (DNA sequence, genotype) and we can kind of quantify the result (phenotype). And we have to figure out all the weights in between (how DNA becomes a fully functioning organism).

It helps with bureaucracy and pipeline generation. That's valuable and frees up time and effort for scientists. But it's not generating new ideas or solving problems like in math/engineering.

The problem is that biology is like trying to reverse engineer alien technology to understand how it works. Math/engineering is more like building up complexity through well defined parts. Math/engineering is a heck of a lot more deterministic than biology. The ultimate bottleneck in biology in context of AI is well curated and breadth of data, not necessarily intelligence.

Claude will just choose similar toolings as everyone. Orchestration is a really nice tool to have and this can definitely empower non-technical bench biologists. But the bottleneck isn't using the tools or even intelligence in this field. It's data generation and ontology. There needs to be more hardware automation and good data/record keeping in biology. With the exception of high-throughput sequencing and some other more niche quantitative methods, I feel like molecular biology is still at similar data throughput of Mendel's pea experiments.

It orchestrated available open source tools to de novo generate and assess the designs. It is still really impressive that it has enough theory to use the tools correctly, but it isn't generating the residues from scratch within the model itself.

I got Opus5 to write me a enzyme design pipeline on AWS last month using available tools and databases. It works fine. It's not as handsoff as you might think when you start looking at more complex proteins with multiple binding pockets and sequential catalytic process within multimeric structures. The bottleneck is still actually testing the candidates out in the lab.

Seems like the TLDR is that the neural network is not randomly placing things together due to constraint, it is actually finding some commonality from the input dataset. We just need to spend more effort trying to trace it.

Perhaps our anthropomorphic understanding of data is biasing what objective features there actually are in the dataset.

I wonder if we can study the graph similarities of various models, do we see a convergence of structures that encapsulate certain concepts. And would that convergence be due to just how humans have narrowly defined the concept in our language, or is there something inherent to the concept that allows it to converge. If we compare the same concept in different languages (ie. chinese vs english), does this convergence disappear?

Yeah the calibration is really what makes it useful in practice for quick, small decisions. Asking a LLM to give scores to a problem will yield inconsistently scaled/anchored results that changes at a whim.

The blog is pretty heavy on statistics. I'll have to study it more when I have time. Is it essentially bootstrapping results to statistically normalize the answers?

They are going to run out of real world data very soon. AI labs need to hire non-cynical domain experts to build physical hardware/automation that can generate experimental data. They are hard to find though. Most domain experts are so tunneled-visioned within their own domain that they can't think outside of it.

I am getting the exact opposite experience. But it is probably because I am in a domain where code/software is a tool, not the product.

I find myself learning exponentially faster and more. For example, I am working with spectroscopy hardware currently (raman, nmr) where I got Claude to write code that interfaces with equipment on a hardware level. Instead of me going through data sheets and writing out a bunch of wrapper code, Claude did it for me.

I am able to progress much faster by using Claude to discuss various techniques, implement them, and test it out. This loop would have probably taken me 5-10x more time previously.

And I am learning so much more about these machines/techniques/data than I would have if I had to expend the mental effort to write menial code just to see a result.

I have more than a decade of experience as a developer. I am glad that we are finally moving towards a world where we can utilize code as a tool rather than constantly trying to think how to make it into a product.

I don't disagree with Bambu from an operational standpoint, but disagree with their handling of this.

They are offering a cloud infrastructure that allows users to remote control the printer via their software. If they don't want users to use a non-approved software to access their cloud, they should just build auth around it and explicitly tell people that. The accessibility for users to utilize the printer without going through official software and cloud is a whole other can of worms of course.

This whole fiasco could have been avoided by not being so confrontational, giving their user base ideological ammo.

I have a data science/engineering background. From my perspective, using AI is like mining the solution space for optimality. The solution space is the combinatorics of the billions of parameters and their cardinalities. You try to narrow down the search space with your prompt and hopefully guide your mining with more semantic-based heuristics towards your optimal solution.

You might hit a local maxima or go down a blind path. I tend to completely start my code base from scratch every week. I would make things more generic, remove unnecessary complexity, or add new features. And hope that can move me past the local maxima.

I AI coded an entire platform for my work. It works great for me. I also recognize that this is not something I want to make into a commercial product because it was so easy that there's just no value.

I think this might be more of an comment on software as a business than AI not coding good apps.

Yeah this is the simpler and also effective strategy. A lot of people are building sophisticated AST RAG models. But you really just need to ask Claude to generally build a semantic index for each large-ish piece of code and re-use it when getting context.

You have to make sure the semantic summary takes up significantly less tokens than just reading the code or its just a waste of token/time.

Then have a skill that uses git version logs to perform lazy summary cache when needed.

Just write a Claude OS already.

I did this in the beginning and realized I never went back to it. I think we have to learn to embrace the chaos. We can try to place a couple of anchors in the search space by having Claude summarize the code base every once in a while, but I am not sure if even that is necessary. The code it writes is git versioned and is probably enough to go on.

Were there any discussion from either company about giving government access to consumer data from the the consumer product?