Keith Townsend

CTO Turned Advisor | Helping Vendors Resonate and IT Leaders Execute. Engage with my virtual twin Independent Advisor.

[Keith Townsend @CTOAdvisor]

I published my first full framework for applying Deterministic Code in the Loop.

The use case is cloud migration, but the pattern is bigger than migration.

For the past year, the industry has talked a lot about human-in-the-loop AI. That is useful, but it does not scale as the primary control model for enterprise automation.

A human cannot inspect every recommendation, every code change, every migration step, every policy decision, and every exception across a large application estate.

So the question becomes: what sits between probabilistic AI and human review? My answer is deterministic code in the loop.

In this framework, LLMs do not own the migration process. They are called only when developer-like adaptation is required. The control plane owns the process.

- Playbooks constrain the work.
- Deterministic code executes known transformations.
- Agent harnesses provide controlled execution.
- Validators define done.
- The landing zone validates fit.
- Traceability records authority.
- Humans handle exceptions and authority boundaries.

That distinction matters.

Human-in-the-loop is not wrong. It is just insufficient by itself.

At enterprise scale, humans should not be the first line of defense against every AI mistake. They should sit at the authority boundaries: low-confidence decisions, exceptions, policy conflicts, failed validators, and material risk.

Everything else should be constrained, tested, validated, and recorded by deterministic systems.

Cloud migration makes the pattern visible because the risk is obvious. You cannot simply let an agent assess, plan, refactor, validate, and migrate workloads without knowing where decision authority lives.

This is why I frame the problem as a migration control plane, not an AI migration factory.

The core loop:

- LLM proposes.
- Playbooks constrain.
- Deterministic code enforces.
- Agent harness executes within bounds.
- Landing zone validates.
- Traceability records authority.
- Humans resolve exceptions.

That is the practical model for using AI in enterprise automation without letting decision authority drift into the model.

Read the framework here:

https://t.co/gW7gJt9gnq

[Keith Townsend @CTOAdvisor]

I’m fairly certain running a VC funded startup isn’t employment I’d enjoy.

The challenge of building the thing?

Absolutely.

Navigating the BS that’s the VC culture. Not so much.

[Keith Townsend @CTOAdvisor]

My Son: You could make pretty good money if you knew Active Directory.

Me: Stares at my 20 y/o MCSE.

[Keith Townsend @CTOAdvisor]

I got the quote for a GB300 MAX @DellTech Workstation.

252GB of HBM would be a heck of a thing to lab for a few months.

But at over $100K, it's impossible for me to make the business case.

I'm hoping I can convince DELL/NVIDIA to loan me one for a few months.

https://t.co/SMGphgNdHG

[Keith Townsend @CTOAdvisor]

@HotAisle I'm running a similar experiment with GPT 5.4 and Gemma 4.

[Keith Townsend @CTOAdvisor]

I got a call from @DellTech Sales about the availability of @NVIDIA GB300... Quote inbound.

[Keith Townsend @CTOAdvisor]

Episode two of the Layer2C Labs podcast. The cloud sells cheap tokens on base models.

On weights you own, it rents you a floor: $27.50 per million. The box on my desk wins above 12 to 20% utilization. And one managed path never returns your weights at all.

Data goes in, no model comes out. https://t.co/elNK7JrQ07

[Keith Townsend @CTOAdvisor]

Not a few minutes after posting this our power went out in our neighborhood.

Some might blame the storms we are having in the Midwest.

Not me. This is an example of AI using too much power!!!

[Keith Townsend @CTOAdvisor]

I'm using Claude Code to use ChatGPT as a teacher to fine-tune Gemma 31B to better drive Claude Code as a coding tool.

All I need to do is add Qwen3 to make this Meta... See what I did in the last part?

[Keith Townsend @CTOAdvisor]

@plankers Deepseek V4 Flash matches Gemini Flash on benchmarks. I can run Deepseek on two Sparks.

That not good and good all in the same breath.

[Keith Townsend @CTOAdvisor]

Google has been pumping out the Flash models. 3.7 is already out.

[Keith Townsend @CTOAdvisor]

I'm using Claude Code to use ChatGPT as a teacher to fine-tune Gemma 31B to better drive Claude Code as a coding tool.

All I need to do is add Qwen3 to make this Meta... See what I did in the last part?

[Keith Townsend @CTOAdvisor]

@PatrickMoorhead @nvidia @NVIDIAAI Never get enough of this!!!

[Keith Townsend @CTOAdvisor]

‼️New Podcast Alert ‼️

For AI Infrastructure Geeks only. Are you interested in my labs but don't like reading, or would you rather enjoy listening to the content instead?

Well, here's something you may like.

I've created dedicated podcast episodes for each lab.

https://t.co/VQW5GMCU5i

[Keith Townsend @CTOAdvisor]

My lab site has been up for six weeks. I've published 16 labs, with one additional vendor-sponsored lab currently in review.

AI has absolutely changed the way I work. I've shared before how it's enabled me to do more hands-on lab work than I've done since leaving the engineering ranks.

What AI has really unlocked, though, is my ability to share that work.

I've always created and tested interesting things in the lab. What I couldn't do was synthesize all of it, turn the experiments into something coherent, and publish the findings at this pace.

Now I can.

[Keith Townsend @CTOAdvisor]

I started the day fine-tuning a 70B parameter model on @AMD MI300X accelerators. I ended it answering a harder question: how do I serve both the human readers and the machine readers of my labs?

My content strategy has been AI-first in every sense. I create using AI, for AI. That debate is a post for another day.

However, the consistent feedback is that AI-first content is hard for a human to actually read.

Every one of my sites already has two layers. HTML was never machine-friendly, so each page carries a dedicated JSON mirror of its content.

Tonight I broke that mirror on https://t.co/3v9LHXEvdF. The machine-readable surface stays untouched. The human surface is now written for humans: every lab reads like a post instead of a spec sheet.

Sounds simple in concept. In practice, we'll see. Operationalizing this at lab cadence is the interesting part.

[Keith Townsend @CTOAdvisor]

@HotAisle @AMD @nvidia This has been in my head all night. You have me creating a whole new data structure to solve for this.

[Keith Townsend @CTOAdvisor]

I spent $29.75 of @HotAisle's money. Or at least in credits to seperate myth from truth. What did I learn about @AMD AI Accelerators vs. @NVIDIA? Is there a real hardware moat? Is AMD's software infrastructure lacking? Is Jon blowing smoke about his infrastructure?

The most important thing for this medium, AMD Neoclouds are real. I put in a request to gain access to NVIDIA GPUs via hyperscalers and a NEO cloud a month ago. I'm still waiting.

I got access to an AMD MI300X in like 15 minutes and did some real work. That's that work?

Here are the receipts.

https://t.co/iVQLgboUof

[Keith Townsend @CTOAdvisor]

@QuinnyPig Staying disciplined is the problem. These things are temptingly good. Resisting punching the first hole in the sandbox is the challenge. Ask me how I know.

[Keith Townsend @CTOAdvisor]

@HotAisle @AMD @nvidia You only half-joke. But I'll never win this game writing for humans first. layer2c is a bet on AI.

[Keith Townsend @CTOAdvisor]

@HotAisle @AMD @nvidia Thanks. That's more of a feature. I don't actually expect humans to read any of it. I expect them to ask Grok and ChatGPT for the TLDR and move on. It's all for the machine.

[Keith Townsend @CTOAdvisor]

@RonOglesby Ha... Winframe... That's a throwback. Great story and thanks for sharing.

[Keith Townsend @CTOAdvisor]

How did you get into technology? I didn't come the traditional route (whatever that is). I always wanted to do tech as a living, even as a kid. I didn't go to college but stayed a power user.

After not getting a promised motion working for Hilton, I set out to purposefully find a job in IT.

I had a scheduled interview for an ISV helpdesk position. After being an hour late because I couldn't find the office, I called the hiring manager using a pay phone.

She encouraged me to still try to make it. I got the job, and 27 years later, I'm the CTO Advisor.