Subject: Re: end of SaaS debate
some of them have tons of credibility.
https://www.fabricatedknowledg......

but so do the generalists !

I'm really neither a generalist nor a technocrat, and I think the distinction is a little reductive.

For context (hah) I use Claude Code from the terminal via a cloud provider's foundational model service. I am on 4.5 of the models, and have switched to 4.5 Opus. I have used an IDE based AI assistant, and I have used the other providers (Gemini, OpenAI) for investigation and research tools. I haven't used Grok because I don't consider Elon Musk trustworthy or ethical. I use these tools for troubleshooting, development, and architecture and I use it in a semi-sophisticated way with subagents and multistage planning.

There are people a lot better than me, but I can smell bullshit on the topic and O'Laughlin has a whiff of bullshit.

Anyway, I know stock and finance halfway well, also. The amount of capital spent on AI is very likely to result in a dot-com like retrenchment. The foundational model providers are the ISPs, Nvidia is Cisco, and the cloud providers are the "critical players in the new economy" (2nd time around for ORCL and MSFT).

Like the dot-com, I expect a great deal of value from these. I expect it to be uneven and with growing pains. I fully expect at least one very large scale software infrastructure failure in the next 0-5 years, because these tools and automated monitoring and resolution tooling behind are maturing faster than our understanding of their systemic interaction, which is combinatorially greater than their complexity in isolation.

That's where UIs, API, and such become important. Every single part of what I typed and what you and I are viewing now is just artfully arranged high and low energy electric states interpreted and evaluated as boolean algebra. It works because the 1s and 0s are evaluated in hardware specific ways, the recognized in software specific ways, then exchanged between systems in platform agnostic represenations, over stateful and stateless protocols built upon each other many layers down.

Then most modern SaaSes are built on Kubernetes, descended from Google's internal compute resource management system, Borg. A humble container in a kubernetes pod is a way for multiple systems to share a kernel (lxc containers, based on cgroups and namespaces). Those kernels they are sharing are typically virtual machine kernels, sharing physical resources (network, storage, compute) highly customized hardware running - usually - cloud provider specific hypervisors, which use processor level instruction sets to enforce hardware level resource isolation.

This is all ungodly complex.

As a result, you can train LLMs on code, on various levels, very well and it offers a lot of value in assistance in their ability to simultaneously "pay attention" to all the layers of complexity in ways the human brain is not as good at doing. Code is highly structured textual information that has very clear signals for correctness / rapid gradient descent. It is essentially a best-case target to apply LLMs.

It can embed human knowledge rather well, too, and "Please implement the scaffolding for a web service in Rust such that it implements based healthz/ and livez/ endpoints, supports oauth 2.0, and implements [some technical description]. Then add unit tests and static validation." is an entirely valid request. It will do this and it will save you time.

There is an ungodly amount of information outside the context of that request. Literally and figuratively. Then you see things like MCP architectures (model-context-protocol) and system prompts to enrich requests in ways so you don't have to be repetitive or mind-numbingly nitpicky. Like "please ensure you use a specific base image", "please use a specific set of vetted libraries", "please add checks for various regulatory checks", and so on.

You can build all that, too. But people still have to work with it at some points, and layers of abstraction are necessary in both human and computer systems, and very particularly along the interface of those boundaries. They may look different but those boundaries will still exist.

I expect this will result in a diminished or stalling job market for software developers. But calling for the end of UIs and such sounds very much like 1950s-like fantastical, detached thinking that we'd all be taking nuclear powered flying cars to our jobs in floating cities on the moon. Human beings will be in the loop, just like there are still highly skilled humans building cars alongside sophisticated robots.

This is all a very substantial over-extrapolation of jump discontinuities of technological improvement. Not the first, unlikely to be the last time. There is no consequence to being wrong, and lots of reward for making outrageous predictions, and it results in some real dopey behavior.

Anyway, please accept they are interesting or powerful and may change the way that all or some people work, but that it's not a binary outcome. Pun intended.