daemonk

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?

It's a different skillset and way of thinking. Engineers tend to think vertically deep on technical problems. With AI, you have to think horizontally broad and vertically up on the architectural problem. The trick is to be comfortable relegating the details to AI.

One concrete example of this realization was when I was researching how to optimize my claude code environment with agents, skills, etc. I read a lot of technical documents on how these supplemental plugins work and how to create them. After an hour of reading through all this, I realized I could just ask Claude to optimize the environment for me given the project context. So I did, and it was able to point out plugins, skills, agents that I can install or create. I gave it permission to create them and it all worked out.

This was a case of where I should not think more technically deeper, but at a more "meta" level to define the project enough for Claude to figure out how to optimize the environment. Whether that gave real gains is another question of course. But I have anecdotally observed faster results and less token usage due context caching and slightly more tools-directed prompts.

Pretty cool technique using complementary overhangs and toehold sequences to generate a 3-way heteroduplex, ligate knick, and then remove barcode duplex.

They don't give much details on how the barcode duplex is removed though. I guess ultimately the barcode duplex strands can just be melted off and the ligated strand can be used to template off of.

If this can be made into an easy to use kit, can really make vector generation much easier and hopefully not locked into proprietary systems.

I can imagine a company that bioinformatically generates libraries of common long oligos with corresponding barcode and allow end-users to select oligos to modularly ligate together in a one pot reaction. Cool stuff.

I am unsure of the impact of this to the regular consumer as this seems like a pretty niche area. But it's kinda shitty that interpersonal relationships of two companies are impacting their customers negatively.

The onus is not up to public opinion or customer politics to resolve your schoolyard differences. We just want to buy your products and not get loaded with baggage. We don't owe you loyalty on top of the price we paid you for the product.

It's a bad look for the parties involved.

It reeks a bit of tech self importance. From an overall business context, you balance time/effort vs short/long term reward and design or fix things when you need to.

There are many dials to manipulate in a business to make it more efficient. And not all businesses are tech startups. Software, fortunately or unfortunately, has a lot of room to be inefficient due to cheap hardware as long as it gets the job done in a lot of industries.

I have a technical background as a data scientist and developer. If I look back 5-10 years ago, I can definitely recognize my bias towards over-engineering and premature perfectionism. Identifying that sweet spot between over-designing and unplanned failure is key.

So the queen can lay 3 types: hybrid female, Ibiricus male, structor male. Did they do karyotyping? Is it actually that the queen somehow removed its own genetic material from the nuclei or does it somehow get silenced when the M. structor genetic material is present in the nuclei (which is interesting by itself). Perhaps some kind of complex imprinting is happening.

The haplotype phasing strategy is the key method here. The phasing method is described in a previous paper: https://www.nature.com/articles/s10038-022-01049-6

I think this sentiment has probably been echoed through the ages.

It feels like there's an assumption that we've reached some kind of a complexity ceiling and compressing complexity below us will just make us dumb? What if we've black-boxed complexity below us so we can explore more complexity above us?

Maybe the argument is that the rate of compressing complexity below us is faster than expanding the complexity space above us? And the result is that it makes us run out of knowledge of digest and explore? Perhaps the answer to that is to make people more curious to go out and explore the complexity above us so we can generate that knowledge.

It might not be the mainstream on HN, but most popular polls I've seen show similar trends of a lesser proportion of people wanting to live longer, citing the same societal collapse concerns. In any case, whether something is espoused by the majority or the minority doesn't really add much weight.

I don't think there is an "anti-aging agenda". Not everything needs to be seen through the lens of an ideological movement. But I do think that there is an unhealthy persistent cynicism underneath the current popular culture. This cynicism makes people not want to be optimistic/idealistic in fear of being wrong or looking naive. I am not suggesting we should all tint our lenses rose colored, but I do think allowing people to expand their optimistic ceiling is warranted; especially when it is so easy to imagine a dystopic future currently.

Nonetheless, I thoroughly enjoyed your sardonic reply.

This isn't an unpopular opinion. I would argue this is the mainstream argument.

I think all medical advances benefit the wealthy first and then becomes more affordable over time.

The term "aging" seems to trigger a lot of people and lead to philosophizing over the importance and morality of death. They are important topics to discuss, but I also think it is worthwhile to also hear out the optimist perspectives rather than the endless dystopic cynicism we hear on the daily basis.

At a higher level, MCP seems wants to enforce a standard where no standard exists. I get that the low level technical implementation allows AI to utilize these tools.

But there doesn't seem to be any standardization or method in how to describe the tool to the AI so that it can utilize it well. And I guess part of the power of AI is that you shouldn't need to standardize that? But shouldn't there at least be some way to describe the tool's functionality in natural language or give some context to the tool?

Learning biology is a great insight to how humans think about complex systems. We tend to utilize a reductionist and engineering approach to figure out how things work. This is a perfectly valid approach when building up a complex system from knowable individual parts.

But when analyzing a complex biological system, we tend to make analogies to our own engineered components (motherboards, power sources, circuits). There are definitely a lot of similarities and it is a great way to understand one facet of the system. But it can also sometimes make us lose sight of the intertwining relationships among all of these parts through evolution.

The analogy of our genome to an informational blueprint is one of the best examples of the multi-faceted nature of biology. While the sequence of bases contained within DNA (primary structure) is informational, the complex structures the molecule itself (chromatin structure) also have mechanistic purposes.

We build engineered components to be controllable and independent so we can better assess how the system is working. However, that is not an explicit goal with biology. Biological "components" settle into the best form for the given environment over time even if it creates a potential "mess" of connections and relationships.

I think this is also why biology takes a long time to "sink in" while learning compared to other technical fields. It's very easy to over-train your mental biological model on one facet of the system and lose sight of the others.

It is not uncommon to see the same terminology used in very different ways in various sub-fields of biology. Gene Ontology is also a great example of this as people have found there are biases associated with what the originator lab is studying at the time. Genes with pleiotropic function will tend to get assigned function that's more relevant to what the lab is interested in.

Sometimes I wonder if we are really equipped to navigate it and understand it. Maybe AI/computation is really the only way to try to have a holistic view of the complexities. Perhaps trying to understand biology with our biological brain has inherent limitations like a piece of software trying to understand the hardware it resides in.

I don’t think we are going to agree on this.

At the end of the day we are really just disagreeing on how seriously we should take AI-ification of software engineering.

I take a lighter stance that it’s mostly harmless and has a lot of educational positives. I guess you foresee a lot of potential harm.

I do appreciate people who are willing to be the canary in the coal mine. I just find that a lot of the criticism appears to be more gatekeeping than legitimate worries.

Sure. But it's not really about you. I wouldn't break into someone's house and tell them they painted their house an uninformed color.

If something sucks or won't scale, it will sort itself out in the market.

[flagged]

I think this is just subject to normal market forces and very product dependent?

If you have engineering skills that are not easily found, then you can ask for more. If the founders don't give it to you, then they might not be able to build their idea.

If the product do not require niche or sophisticated engineering, then the founder will just move on to the hordes of candidates that can do the job.

I don't think "fairness" can be intrinsic (nor objective) in a market-driven economy. It's always going to be a push-pull, negotiation, calculating business. And you disenfranchise yourself out of this economy if you don't have the stomach for it.

TXTL systems already works. And there are antibodies (not many) that can only be produced cell free.

The problem is really scale.

Cell-free synthesis is the next step. Theres slow progress, but very interesting operational models if it works.

Nanopore spits out data as dna goes through the pores. And depending on the flowcell you are using, you can load or top up the flow cell in smaller amounts. So you can potentially have quick turn around time without having to wait for enough samples to pile up before you batch sequence on a larger illumina machine.

Illumina machines are cost efficient in terms of cost per basepair, but only at large batches. They are trying to rectify this after seeing other benchtop sequencing machines (Element Biosci) moving into the mid-throughout niche and doing well. Their solution is the miseq i100 that they just announced.

But at the end of the day, these are all still constrained by having to think in terms of multiplexed batches which has a lot of operational complexity involved (equi-molar pooling, barcoding, etc).

Ultimately, for commercial sequencing labs, one of the more difficult problem to solve is the operational complexity of how to optimally load the sequencer for lowest cost while balancing failure rates/low coverage rates rather than the technicalities of dna prep/lib prep. Given unlimited and consistent intake samples, the problem gets easier. But most labs have some kind of seasonality or project cycle built in which means it’s not about maximizing a yearly capacity, it’s more about how many samples max you can pump through within a few days.

Productionizing and scaling nanopore sequencing is definitely an achievement. With that operational obstacle out of the way, you get to think about sequencing in terms of “streams” of data rather than “batches” of huge amounts of data (illumina). That confers a huge operational and commercial benefit.

For most instruments, it's an interplay between resolution and consistency. There isn't a "true" reading in any sense of the word. There is just how much resolution you are measuring and how consistent the measurement is between independent readings of the same event.

The overwhelming amount of pre-existing clinical data gathered by cuff-monitors means that any novel measurement methods benefit from tying/correlating their output to the existing data for compatibility. It's a bit unfortunate in some sense as cuff-monitor data lacks resolution and sometimes can be a bit of a drag on novelty.