hosh

LLMs tends towards closed-worlds and bias towards decidability. Risk assessment is terrible, as are creating opportunities — no animal instincts to draw from.

Despite having this in the training data, I am constantly having to remind my coding agents to reason and design with grammars. On the other hand, once all of that is settled, and the task is no longer open-ended, then they are very good at driving towards completion.

Crafting the automation itself can be fun. That is one of my favorite part of designing systems running on Kubernetes.

I have also come to the conclusion that, right now, AIs lack imagination. So designing human-AI collaboration that lets human imagination in the loop would be interesting.

If we’re going by water, land, etc as an argument against wasteful activities … Coming from the permaculture world, I can tell you just how wasteful the pre-AI world is. The ideas in the mainstream environmental-friendly practices are themselves very wasteful.

I have sometimes had to work in bed while running a low fever.

I do that because I believe in the mission and work is assessed by outcome.

The flip side too — I learned well before COVID that the Kanban principle of “make work visible” applies to WFH

I wonder if the employer is in violation of ADA laws, and whether exemption from monitoring is “reasonable accommodations”.

Not that I think the broader practice for monitoring remote workers is something I think is a good idea.

Code as documentation works better when the code is declarative or a DSL. These capture intent and promises (as in promise theory) better.

When it is not, it has to be reasoned out and does not work well for documentation.

Other things that code and tests alone do not capture well:

- promises (as in Promise Theory) made to other parties. Claude already has PT in its training data.

- Constraints-inducing-properites, as in Roy Fielding / Christopher Alexander. While tests, and property testing can capture properties, there is no formal connection to the constraints that induces fhem. By constraints, I am not talking about business requirements and business value — those are better understood through Promise Theory. I am talking about things like at-least-once delivery or total ordering (from append-only constraint). Claude already has Fielding’s dissertation and Alexander’s works and ideas in its training data.

- grammars, as in pattern panguages (not just patterns) a la Alexander / Fielding are also not captured in code alone. These tell both humans ans AI how to extend a pattern, and how to identify anti-patterns (when they violate a constraint-inducing-property)

- LLMs are trained with many different worldviews and bounded contexts at the same time, and is very capable of translating across it. However, these need to be soelled out, otherwise it would talk in whatever it infers

Specifications written for the exact way components are wited together run into that stale doc problem. Although it takes much more human attention and token burn to describe things in terms of pattern language and promise theory, it becomes easier over time. The actual implementation plan tends to fall out more cleanly when all those other stuff are at least considered. This is where I have been spending most of my time.

If people are happy with Actions API, that sounds great.

I’m not happy with the Actions API. I think Gitlab’s cicd design is much better, and I’m not fighting it all the time when I use it.

Smalltalk also had extensibility for one. But it largely missed the explosion of collaboration when open source code became shareable and people can work on it together.

People also shared Hypercard apps.

I think it is less about client/server and more about how ideas can be shared. Sometimes, data might be shared.

I think as we see more AI participation in code, github stars are going to matter less.

Maybe they got sucked into the cool tech?

I remember my first pass reading of Stephenson’s Snow Crash and Diamond Age, I was bedazzled by some of the tech … it was later readings that the comments like “biomass” (the recent echo in AI “meat proxy”), or how the Primer was ultimately a failure sank in later readings.

Not to mention language communities with constant supply chain attacks because its standard library story is poor, and everyone keeps reinventing new, often half-baked solutions?

Or even that, the very same ecosystem congratulates themselves on the typing system but still relies on linters because the language and runtime themselves allow whole categories of dumb ideas to be written?

This looks a lot less annoying than Typescript, particularly how dynamic() is a lot more useful than any()

I also wonder if this works well with Ruby’s duck-typing and monkeypatching.

In my twenties, I came across an idea from one of Ken Wilbur’s books that helped settle a conundrum for me. What happens when one wants to honor all life? Are they all equal?

He made a distinction between intrinsic value and extrinsic value. Plankton is not as complex of a lifeform as whales, yet whales cannot live without plankton. One has more intrinsic value and the other has more extrinsic value. There is an interrelationship that does not have to flatten value for everything and everyone.

LLMs are trained from the language corpus of our collective consciousness. It reflects our collective, all the wonderful, beautiful, and horrific things we can dream of and put into words.

What people are finding now when building agenic AI orchestrators is that they are reinventing some of the ideas behind BEAM and OTP. The big one being queues (mailboxes), but also fault isolation.

Async runtimes can’t really do preemptive scheduling so those end up using other methods that, while getting low latency, may not improve reliability or reduce variance in latency.

Orchestrating tasks across a number of different services that are not in your control becomes a distributed system where not everything is reliable.

Is it fair to say that Mesos is more top-down while Kubernetes is more bottom-up in scheduling approach?

That is not very insightful. Your thesis started with the idea of elastic horizontal scaling.

Mesos was designed from ideas with HPC and catered to that. Large hardware capex, which is not elastic. Containers did not make sense in that world when it was architected, and it was retrofited into Mesos's architectural foundation.

Kubernetes was designed for a wide variety of workloads, and designed to be composable, versatile, and extensible. It has a far more decentralized approach, an antithesis to Mesos's Data Center as an OS approach. It isn't that Kubernetes did not try to do too much. It's that they laid a much more flexible foundation.

It turns out, Kubernetes was a better fit for a many more use-cases, even beyond large and medium sized enterprises. Kubernetes works pretty well at the edge and locally as well, and it runs many ML workloads on the other end of scale.

This is not theoretical.

I build production Kubernetes and cloud infra for work. When I run Kubernetes locally, it is because I am developing operators or manifests for application workloads. Kubernetes is not the “value add” for my dev workflow, it is literally what I am developing.

I have run Linux laptops before. After running it for five years, I came to the conclusion that it did not make a good laptop for my use-case. Poor suspend-resume support, poor wireless networking support means I can not just pick and go. (And no one has yet to replicate Apple’s trackpad experience). So yes, I run Apple laptop with MacOS and use my TUI tools, sometimes with Linux running in an VM, sometime remotely to a full headless VM with my full dev suite via mosh because I use cli and TUI for dev.

Your turn. You still have not defined “poor man’s Linux”.

They could have chosen Mesos instead. Kubernetes had other characteristics that allowed it to be adopted far and wide besides the ability to scale horizontally.

How are resource management distinct from scheduling in Mesos?

I think people are putting together pi clusters for their homelab these days.

Not all clusters are elastic. Cloud infrastructure can be, but HPC setups before the cloud were not.

Scale isn’t the only reason. Sometimes you want resource isolation and self-healing, something that is useful if you want a personal swarm of AI agents.

I am not sure I understand this argument. Kubernetes typically runs on Linux. I use an Apple laptop, work mostly with headless Linux VMs and Kubernetes. What is a “poor man’s Linux”?

A local model that can do better than Siri or Alexa as a personal or home assistant is, in my eyes, very useful. Being able to run on a phone or watch or glasses translates to me, low-powered AI, and not necessarily that I want my phone, or watch, or glasses to run things for me.

My Siri use has narrowed down to just setting timers. And even then, I still have my phone call people in the middle of the night. Siri is pretty dumb and does not do what I want it. I’d rather be able to customize an assistant to myself.

I am also thinking of automation in my day to day workflow for work.

I don’t know at what threshold a complicated system becomes complex.

For example, at a level of scale, Kubernetes start having emergent behavior.

On the other hand, it doesn’t take much to produce a complex system. The Boids simulation is a complex adaptive system in the form of a flock, yet each member of the flock concurrently follows only three basic rules.

I was not trying to be exhaustive. I am sure you can come up with more characteristics.

They all influence each other to one extent or another.

And, the Cynefine Framework defines “complexity” a bit differently than the intuitive way it’s often used.

The simple domain is a single dimension. The complicated domain is a system of factors. I think when most people say “complex”, they are really talking about what Cynefine labels as “complicated”.

The Cynefine complex domain is not so easily solved or reduced. It has emergent behaviors. The act of measuring tends to perturb the system. No single solution will ever solve something in the Cynefine complex domain, because the complex system will shift behavior, making solutions that worked before start working against it.

Examples are ecosystems and economies. Software systems tend not to be complicated, not complex, until you start getting into distributed systems.

One of the key insights of Cynefine is understanding that each of the domains has its own way of solving things and that often times, people use solutions and methods from one domain to solve problems characterized by a different domain.

You don’t solve problems in the complicated domain with methods from the simple domain. And you don’t solve problems in the complex domain with methods that work for complicated domains.

Complexity, if it can be reduced to a single measurable dimension, is only one of several factors in a solution space.

There are other properties such as, maintainability, scalability, reliability, resilience, anti-fragility, extensibility, versatility, durability, composability. Not all apply.

Being able to talk about tradeoffs in terms of solution spaces, not just along a single dimension, is one of what I consider the differentiator between a senior and staff+ developer.

Azure’s core hypervisor orchestrator was half-baked at launch and it has never been fixed. This long read blog series explains a lot for me — for example, why the FedRamp certification program was never able to get a straight answer from Azure about how they handled secrets.

https://isolveproblems.substack.com/p/how-microsoft-vaporize...

https://www.kunalganglani.com/blog/microsoft-fedramp-failure...

It is not necessarily monetary or even transactional form of support. Reciprocity builds relationships.

Not necessarily even code contributions. It could be professional networking. It is a bit different if the person is not a stranger.

Something I learned about being a part of an ecosystem: if you want it, you need to support it and help it stay alive.

That applies to local shops as it does open source projects.