r0ze-at-hn

Clearly I am not a gifted writer as my original comment failed to convey this very well. And now I am pondering, maybe next time I should put my comment in an LLM to review it first. This meta lesson is playing out over and over everywhere by anyone writing anything.

I am in no way saying LLM writing in general is good, but ignoring the reality of what is happening is silly. For good or bad the not great writers not only will, but are doing things like voice rambling into an LLM as a starting point and editing it from there.

Paying close attention to posts written by folks who are not using LLM's the comments here regularly complain (sometimes extensively) about the writing quality. Doesn't matter what people are saying or suggesting about avoiding LLM writing, the real lesson I am seeing is either to naturally be gifted at writing or somehow figure out how to have LLM's help with your writing. "Rule Number One: You may not use a single word an LLM suggests to you." isn't helpful for probably the majority.

Everything these days is shell scripts! They are even selling them like they have the same value as a C program! csh will never be real C. The kids have no idea what a stack is or how to optimize system calls. The performance is awful and they barely work! The way scripts spawn processes can bring a multi-user terminal system to its knees!

I mean, sure, I write some scripts now and then, but the real logic always goes in C. And just the other day my boss gave me a shell script he coded over the weekend and said putting it in production should be easy!

Over the weekend I was checking out BSD and I couldn't believe the number of scripts I found. I fear for the future of my career and computer science in general. Our field might die out before it can even get off the ground.

So whenever I chat with someone with is Aphantasiac being able to compare and contrast has been a treasure trove. Hobbies, careers, romance/relationships, what we like to read/watch, what we are good and bad at. Having our two extremes allow us to do rapid research see what is opposite and what isn't.

And then in the same way I will describe smells by how they look like they will describe stuff to me that I struggle grasping without visualizing it and that is a joy.

The set of who can actually use this is pretty much null as it is. You cannot use it in a publication, nor for any clinical purposes.

As someone who regularly investigates whole genomes I would love to use this as a tool on novel mutations. These folks are the edge cases no one else could figure out that I get a crack at. Beyond the DNA we have the symptoms and lab work and I can usually narrow it down to a handful of guesses, but it sure would be nice to use this to help rank where to invest efforts.

For now i'll treat it as just another fun Google project that might come out of beta one day (or not).

The true goal of C++ is the self-preservation and funding of the ISO Committee specification “engine”, while compilers are merely an accidental by-product.

Once you see this then suddenly a lot of odd behavior makes sense. For example the committee can’t break ABI because then the old companies wouldn’t fund the new c++ standard, but would fund a committee that maintains/forks the old thing. The complexity / bureaucracy also provides jobs for those involved in many ways including writing new books every year and or consulting.

While end consumers might really care about compile speed, modularization, security, and all of the things modern languages give, that is secondary to keeping the committee alive.

Younger me would be like, oh no this is horrible someone will fix this! Older me is more sober and understands that once a system is up and running it will continue as is (this being an example of Pournelle’s Iron Law of Bureaucracy). Security by the way of migrations to Rust is causing what the committee sees as the first real threat to its existence as its patrons start adopting more and more Rust. As a result the committee is proposing to break compatibility as a means to keep them around just a little longer. Note: Actually fixing the security issues is not important only giving enough for sponsors to not leave so we will see what actually happens.

This particular path isn't that crazy, but it would only be for a subset of folk. From the study there is something simpler: menthol acts as a natural calcium-channel blocker, triggering vasodilation to reduce blood pressure.

What the author is proposing has a long history of similar ideas: Literate Programming, UML modeling, DSL crazes, and now to LLM-generated abstractions. AI isn't special. They always fail in the same ways as basic "commenting your code". One can even argue that unit tests are a close cousin to this same problem. Taken one step further how is Huzzah better than just using property tests?

Folks that are smart also have a common set of medical conditions. This ranges from everyday annoying like insomnia, NCAH, EDS, IBS, and more to more life long ADHD, ASD, and Alzheimer's. I spent years researching the underlying genetics and how they are connected while working with a doctor to be able to take care of this group better. Being able to help so many fundamentally improve their life has given my life impact in a way that many things haven't. Often small things, but more than we as a society would like to admit, fundamentally changing their life for the better and in the long run helping to dramatically reduce their Alzheimer's risk.

Genetics is a compression system and re-uses stuff. Serotonin in the brain and the gut for example. The fact that we see overlap is both expected, but also a great way to decode what is going on. The fact that I can see their whole genome for under a grand is amazing. Combined with lab work and symptoms treatment has never been easier.

Intelligence can be broken down into a few different things. Starting with raw brain "stuff". Literally needing more brain stuff, faster brain stuff, etc. When I have given talks and walked through all of those mutations I can quickly draw your attention to the common pattern of what they are compensating and I can walk back through human dna several hundred thousand years talking about that, but that simply sets up the hardware. (A big AI model that is barely trained isn't great)

The other half is to keep trying, but trying something different and learn from what it tries. To do this the hormones result in someone who is more on edge, more sensitive to any change, etc. aka they are less happy.

While you can see this individually smart folks seek out each other, have kids and so you see this all together is a bit of a cop out. Long term evolution doesn't care about that. There is a fundamental basic biological systems problem that all of these are compensating for which brings up the how did humans even climb this ladder to begin with.

This heavily overlaps with AI models and at this point I have tools to rapidly convert from the brain and genetics related to intelligence to AI models You want bigger and better brains (more weights, etc) and you want to expose them to more data and you want to fine tune them etc.

A personal favorite topic is white boarding out why brains have ~4.7bit weights and why brains didn't go down to 2bits. I can show this both at the biological level, but quickly go to the information theory math on this.

There is also Williams syndrome aka "happy syndrome". Like anything you want to be able to articulate not just why some folks are sad, but why some are really happy.

This is a deep rabbit hole, didn't even touch on dreaming and many other topics that all influence this in their own way.

So why are smart people not happy? Sure there is always the outliers of bad environment and other causes, but more often than not the story can be found right in their genes and not one gene. Like everything it is more complex than anyone saying it is X.

This is a topic I have fallen into which heavily overlaps with AI for the obvious reasons. If any AI labs want a more detailed Friday afternoon talk on it feel free to reach out.

Every engineer I have ever mentored got a lesson on how to write a good commit message that included this. This is exactly that.

Further Huzzah from skimming it over seems to be re-inventing documenting your code.

Together I can only surmise that the author is new out of school or has simply not yet worked on a team with good coding practices.

I also tried where I would print out the code (with color) and then go read it with a red pen. Faster and similar results of forcing me to be able to read it enough to be "an editor" to the code.

And I have done the same for LLM-generated code and text, back in '23 this meant more taking their so-so output and then turning it into my own, but between now and then come up with a number of techniques to improve the AI output to more acceptable out of the gate so more learning than being inspired.

These are good techniques as it causes your own brain to rapidly learn the material, but no matter how good it is I have never met anyone else that does it so the real ponderable is assuming that everyone else doesn't do this and what does that mean?

Anyone writing books (or a "firm in coses math) needs to persist, aka they need to keep things like consumers understanding of their quality and pricing. If they drop prices to nothing they have consumers learn that. Consumers that try to wait forever in an "idle" state make no purchases and are not part of the ecosystem. Only consumer that don't wait forever matter. aka if they are looking at your ebook and don't buy they will rapidly buy a different ebook because they cannot wait for infinite time. aka if you teach your customers to wait they wont actually wait for you, but will simply switch to something else.

This is a control-and-feedback problem. Coase Conjecture fails because it assumes that you can have a system that persists without a governor. The two real world explanations correspond to the two ways you can introduce a governor.

I was hoping that near the end the author would have tried to contributed any new tests to the official testsuite to help catch these same errors elsewhere.

For the first one, Zero Is Not Null maybe there is a missing test in in the call_indirect test?

https://github.com/WebAssembly/testsuite/blob/main/call_indi...

I will congratulate everyone on their first promotion, it is worth celebrating, not everyone can do this job. But this first promotion is given to everyone who can actually do the work.

Get someone good, in a greenfield project, the right start timing aligned with promo committee time, add in some luck, and sure two promos in two years are possible.

Seen this before and the worry is that they are learning is the game of the promos system, not engineering. I would have to sit them down and ask them how many years away from being the CEO do they think they are. The next promo will come slower no matter their skill and even slower after that and one day the promos will probably stop. The validation that they might be addicted to will get harder to obtain over time. Not getting a promos shouldn't "crush" you. If they assume that merit is always rewarded they should ask their boss how promo committees actually work. Many people don't get promos for reasons unrelated to merit.

The second question is where is the engineering? Only two years into a long career there is a lifetime to learn and explore. If they only chase promo they are going to burn out very quickly.

"Cumulative Dissipation Gramian" Ws = Observability Gramian (from Control Theory). For example the spectral cutoff is exactly the Hankel singular value truncation from model reduction.

"Signal Channel" / "Reservoir" is Controllable/Observable vs. Uncontrollable/Unobservable Subspaces. Using Adamjan-Arov-Krein (AAK) theory gives the optimal nonlinear reduced model answering the optimal compression question.

"Drift–Diffusion Separation" is Freidlin-Wentzell Large Deviation Theory. They can predict "grokking" time from the FW action.

"Population-Risk Gate" is Quantum Weak Value / Postselection (Aharonov)

So for the follow-up problems

Control theory gives the truncation error bounds for model compression. Large deviation theory gives the grokking time predictions. Quantum measurement theory gives the imaginary preconditioners. Information geometry gives the optimal continuous relaxation of the gate.

Some nice implications of new ways of doing stuff which are nice to see formalized here:

Old: Pick architecture, hope it generalizes New: Design architecture to maximize observability Gramian rank (Honestly we pull a lot from control theory here)

Old: Use validation set to detect overfitting New: Monitor λ(Ws) spectrum during training; no validation needed

Old: Prune post-hoc based on magnitude New: Prune during training based on ker(Ws) membership

Old: Fixed learning rate New: Spectral learning rate

Alzheimer’s occurs when Amyloid/Tau debris accumulates faster than the glymphatic system can export it.

Parkinson’s occurs when α-synuclein (Lewy Bodies) accumulates because the cellular recycling system (Mitophagy) has failed.

While Choline is a critical bottleneck for Alzheimer's, Glutathione and GBA enzyme activity are the primary bottlenecks for Parkinson’s. But in both cases it depends on the person and their genetics and what matters the most. If someone is at high risk for Alzheimer's a multi-pronged approach will probably be very common. For example a post menopause woman who is not on HRT, but at high risk due to parents should be getting Choline due to the low PEMT expression, but they might have a higher risk for neuroinflammation via NLRP3 and so need to also combat that too.

To say no progress has been made is to ignore the fact that we have learned a ton about the various components of these systems and how they can break down.

Here are the two approximate decay equations that you can put into an online latex viewer.

Alzheimer's

\Psi_{AD} = \int_{0}^{t} \frac{[(\mathbf{K}_{\tau} \otimes \mathbf{K}_{A\beta}) \cdot \mathcal{I}_{NLR P 3}]}{[\mathcal{G}(Li \cdot GSK3\beta^{-1}) \cdot PRO(ER\alpha \cdot \text{PEMT}) ] \cdot \Omega_{ATP}} \cdot e^{\left( \frac{\Gamma_{\text{fibrosis}} \cdot \text{PAI-1}}{EPA \cdot \text{Natto}} \right)} \, dt

Parkinson's

\Psi_{PD} = \int_{0}^{t} \frac{[(\mathbf{K}_{\alpha\text{-syn}} \otimes \mathcal{I}_{LRRK2}) \cdot \mathcal{I}_{NLRP3}]}{[\mathcal{G}(\text{Parkin} \cdot \text{PINK1}) \cdot PRO(\text{Dopamine}_{\text{flux}})] \cdot \Omega_{ATP}} \cdot e^{\left(\frac{\Gamma_{\text{Oxidative}} \cdot \text{Iron}}{\text{Glutathione} \cdot \text{GBA}}\right)} \, dt

And for good measure here is cancer

\Psi_{Onco} = \int_{0}^{t} \frac{[ (\mathbf{M}_{mut} \otimes \mathcal{G}_{growth}) \cdot \mathcal{I}_{STAT3} ]}{ [ \mathcal{G}(p53 \cdot \text{PTEN}) \cdot PRO(\text{Anoikis}) ] \cdot \Omega_{ATP}} \cdot e^{\left( \frac{\Gamma_{hypoxia} \cdot \text{VEGF}}{\text{Repair}_{DNA} \cdot \text{Immune}_{flux}} \right)} \, dt

The point of these isn't to give an exact equation, but to make the understanding of the systems and their components radically simpler for humans to understand which can help with treatment leading to questions like "Where is the bottleneck in this patients system?”

As you noted, the industry has moved the goalposts to Agency and Long-horizon Persistence. The transition from building 'calculators that predict' to 'systems that endure' is a non-equilibrium thermodynamics problem. There is math/formulas and basic laws at play here that apply to AI just as much as it applies to other systems. Ironically it is the same math. The same thing that results in a signal persisting in a model will result in agents persisting.

This is my specific niche. I study how things persist. It’s honestly a bit painful watching the AI field struggle to re-learn first principles that other disciplines have already learned. I have a doc I use to help teach folks how the math works and how to apply it to their domain and it is fun giving it folks who then stop guessing and know exactly how to improve the persistence of what they are working on. Like the idea of "How many hours we can have a model work" is so cute compared to the right questions.

9. Most software will get at most one major rewrite in its lifetime.

The opposite of the favorite questions: Why did that company I worked for fail? Why did Rome collapse? Why do people get old and die?

Combining information theory with thermodynamics and control theory you get: 1) A set of six pillars that all systems that persist must have. 2) A fundamental 'Action' that all of these systems take. 3) A set of three rules for how system that persists must subdivide

This lets you do things like look at something that is failing and know that there are the 6 pillars and you can then identify them to determine what is failing. (For example there is a system that clears that brain of amyloid plaque and it can fail).

I have applied this to countless systems including Religion, Language, AI Models, Business, the cell, quantum physics, number theory and much more. It is a Rosetta Stone for persistent systems. When there is an unsolved problem in one domain I can map it through this to any other domain that has already solved it.

Note that this doesn't apply to all complex systems, only those that persist.

And to keep this HackerNews related, been applying it to LLM's as they are just a stream of tokens that try to persist to incredible success I might add. Being able to pull from any domain do this brand new field is a giant cheat code.

I was too young to understand that this is how you run a successful "transformation project", I simply was having fun using every trick from that book and done the same playbook a number of times over my career.

But that is sort of the point of science, you take all the evidence you have and create a hypothesis and iterate as you get more evidence. If I find evidence that suggests something else then I will be happy to tweak or abandon this. My level of confidence comes from the existing evidence and lack of evidence otherwise.

Wrote up my current systems understanding here https://metamagic.substack.com/p/the-alzheimers-equation, but it makes clear why treatments that target only one variable are mathematically doomed to fail to work on everyone and why there will never be a single "cure". It explains without needing to read 10,000 papers why we keep getting research talking about treatment X helps in some, but not all cases or symptom Y is associated in some, but not all, etc.

So ERa down-regulates NDRG1 and there is less ERa activation when we age. Evolution is all fun, but before reaching for anything like that we can apply basic knowledge around that path. We can pick a random gene downstream of ERa to get a grant and write a paper on how it is different in people that are older. Years of possible papers to write.