Wesley Schlemmer — Notes

GLP-1s and the sovereignty question

Tue, 28 Apr 2026 00:00:00 GMT

I spent a chunk of April reading research briefs on the GLP-1 class of drugs — semaglutide, tirzepatide, retatrutide, the new oral orforglipron that just got FDA approval. The numbers from the trials are striking. Tirzepatide produces ~20% weight loss. Retatrutide hit 28.7% in Phase 3, the highest figure ever published in an obesity trial. Wegovy HD, approved last month, gets 20.7% in plateau patients. These are not modest interventions. They are doing something to metabolic regulation that diet-and-exercise advice has never reproduced at scale.

The instinct, in the freedom-tech / biohacking / sovereignty world I run in, is to be skeptical of any pharmaceutical with that kind of efficacy curve. The skepticism is partly well-earned — there are real discontinuation rates (12-18% at the higher doses), real safety signals being monitored, real questions about long-term effects on muscle mass and metabolic adaptation. Those are worth tracking.

But the skepticism that isn't well-earned is the kind that conflates "I want to be sovereign over my health" with "the answer must be a supplement stack from a podcast sponsor." Sovereignty over your own metabolism doesn't mean refusing the most effective tool because it came out of a pharmaceutical pipeline. It means understanding the tool — the mechanism, the trade-offs, the trial data, the discontinuation curves, what it does and doesn't address — and making a decision that's actually yours.

Health sovereignty, the way I'm starting to think about it, isn't an ideological allergy to medicine. It's the discipline of being the informed principal on your own body. That can mean GLP-1s. It can also mean rejecting them. What it can't mean is outsourcing the question to the loudest voice in your tribe.

The other thing I keep coming back to: metabolic dysregulation at population scale is infrastructure rot. It's the substrate everything else sits on. If a drug class can move that needle 20% — and the pipeline coming behind it (retatrutide, CagriSema, monthly injectables in the late 2020s) might move it 30% — that is a structural intervention, whether the sovereignty crowd wants to admit it or not.

I haven't decided what I'd do for myself. I'm watching the data and the discontinuation rates and the long-term cohorts. But I'm not going to pretend the question is settled by ideology either way.

This is a personal observation, not medical advice. Talk to your doctor. Read the trial data. Make your own call.

Open Source Agents are Catching Up

Tue, 28 Apr 2026 00:00:00 GMT

Hermes Agent, the open source agent framework from Nous Research, hit 100,000 stars on GitHub in April '26. 53,000 of those stars landed in a single week. The repo is the number one trending repository on the platform.

Two years ago, an "AI agent" meant calling OpenAI's API in a loop with some glue code. The infrastructure was closed. The models were closed. The only open piece was the loop itself, and even that was usually sitting inside someone's proprietary product. If you wanted to build with agents, you were building on rented land.

What I'm seeing now is different. Hermes Agent is model agnostic by design. That's their explicit philosophy "everyone should have access to AI. We don't dictate the rules of use for your agent, YOU do." The framework supports Ollama for local models, AWS Bedrock for enterprise, every major closed API as a swappable backend, and a growing skill ecosystem. 300+ models are now reachable through one open subscription via the Tool Gateway.

The pattern is open source agent frameworks pulling closed APIs into their orbit instead of being held inside them.

For people building on freedom-tech stacks, this matters. It means the agent layer is not going to be a permanently rented concession. It means a selfhostable orchestration layer can sit on top of whatever models you trust, local Llamas, Mistrals, Qwens, whatever lands next quarter, without you having to rewrite your stack when a vendor changes terms. It means the moat for closed source agent platforms is shrinking from "we have the orchestration AND the models" to "we have the models, for now."

I don't know whether the 100k-stars compounds or plateaus. Open source momentum is famously lumpy. But the structural piece - agents going open, model layer going pluggable, infrastructure becoming self hostable - feels durable. That's the trend I'm tracking.

If you're building anything that depends on long running agentic capability, building on the open layer is starting to look less like a values choice and more like a survival choice.

Why I chose Astro over Next.js

Tue, 28 Apr 2026 00:00:00 GMT

This site runs on Astro. Not Next.js, not a hand-rolled HTML stack, not a headless CMS feeding React. The decision wasn't about benchmarks. It was about defaults.

Next.js is the answer when you have a SaaS app — auth, dashboards, per-user state, server-rendered React tied tightly to dynamic data. It's a good answer for that shape of problem. It is not the answer for a personal site that hosts essays, project pages, and the occasional note. For that shape, Next.js ships a React runtime to every visitor whether they need it or not. You spend the rest of your time fighting the framework's defaults to keep that runtime small.

Astro inverts the contract. Zero JavaScript by default. A page is HTML and CSS unless you explicitly mark a component as interactive. The framework's posture matches the posture I want the site to have toward visitors: respect their bandwidth, their batteries, their attention, and — most importantly — don't ship code that watches them.

That last part is the values-aligned read. A framework that defaults to no scripts is a framework that defaults to no fingerprinting, no tag managers, no third-party tracking pixels you forgot you wired in three years ago. You can still add those things in Astro if you want them. You just have to do it on purpose. The defaults are on the side of the visitor.

There's a tooling read too. Astro's content collections give me typed Markdown — frontmatter validated by Zod schemas at build time — and that means essays and notes are just files in a folder I can edit in any tool I want. The same .md files that live in my Obsidian vault flow into the repo with no transformation. No CMS lock-in, no proprietary export format, no SaaS that holds my writing hostage.

The third read is sovereignty-by-default. Static output deploys anywhere — Vercel today, a VPS or a Cloudflare worker or a Pi at home tomorrow. The dist/ directory is just files. There's no runtime dependency I have to keep paying for to keep the site online.

I picked the framework whose defaults match my values. The fact that the site is faster and cheaper to host is a bonus, not the point.

Why there are no comments on this site

Tue, 28 Apr 2026 00:00:00 GMT

This site doesn't have a comment section. No Disqus, no Hyvor, no embedded discussion widget. The contact page has my email and the booking link for consulting calls. That's it. That's the whole engagement surface.

The reasoning is short.

The people I actually want to hear from — peers in the Bitcoin and freedom-tech orbit, podcast bookers, partnership and speaking-invite people, aligned consulting clients — already know how to send an email or DM. The friction of writing an email is useful friction: it filters for the people who actually had something to say. A comment box, by contrast, optimizes for the lowest-effort response. That's exactly the wrong selection.

A comment widget is also a third-party JavaScript surface. Disqus is a tracker with a comments feature attached. Even the privacy-respecting options — Hyvor Talk, Cactus Comments, Webmention receivers — add a script, a database, a moderation queue, and an attack surface. For a site whose entire posture is less third-party code, less surveillance, less SaaS dependency, adding a comment system would contradict what the site is supposed to mean.

There's a deeper design point too. Comment sections optimize for engagement-time. The people running them are incentivized to provoke responses, lean into hot takes, and reward reactivity. That's a mode of writing I don't want to build into the structure of the site. Notes here are seedlings — half-formed thoughts I'm working through. If they had a comment box under them, the gravitational pull would be toward defending positions instead of refining them.

The IndieWeb path back is open if I change my mind. Webmentions are the federated answer — someone writes a response on their own site, pings mine, and I display it (or don't) at my discretion. That's the inverse of a comment box: the responder commits to publishing on their own surface first, the friction works the right direction, and the conversation lives in a graph instead of a thread. If a piece here ever earns enough inbound mentions to justify wiring up a Webmention receiver, I will. Until then, the current setup is correct.

If you have something to say to a piece on this site: my email is on the contact page. I read everything that comes in. That's the bar.

Why I self-host Umami

Tue, 28 Apr 2026 00:00:00 GMT

Most sites in 2026 still run Google Analytics. The wiring is automatic — your CMS template ships with the snippet, your marketing person assumes you want it, your hosting provider may even add it for you "to help you grow." It's the surveillance default. Visitors get tagged, fingerprinted, and joined against the rest of Google's ad-graph. Nobody asked them.

Plausible is the popular escape hatch. It's good. It's open-source under AGPL. The hosted version sits in the EU, no cookies, no consent banner. For most people, paying $9 a month to Plausible is a strict upgrade over GA4.

I run Umami instead, and I run it on my own VPS — umami.crossthebridge.io, behind nginx, in a Docker container I can pull the plug on whenever I want. Three reasons.

One: data location matters. When I'm thinking about a visitor — what page they read, what referrer brought them in, how long they stayed — that data is mine to look at. It's not Plausible's, not Cloudflare's, not Vercel's. The data lives on a VPS I rent, behind a domain I own, in a database I can dump and inspect. If the analytics vendor disappears tomorrow, I still have history. If they pivot, raise prices, or get acquired, I'm not migrating.

Two: the difference between analytics and surveillance is whether the visitor is the product. Umami doesn't fingerprint. It doesn't set cookies. It doesn't hash IPs and sell aggregates back. It counts pageviews and referrers and gives me a chart. That's information. Surveillance is what happens when "analytics" tries to identify a person across sessions, sites, and devices to optimize an ad against them. Umami isn't trying to do that. GA4 is.

Three: the bar for what I run on my site should match the bar I write about. I write about freedom tech, sovereignty, opting out of legacy systems. If I were piping visitor IPs into Google's ad graph in the same breath, the writing would be theater. Self-hosted analytics is a small thing. But the small things stack up to whether the site is legibly aligned with what it claims to value.

Plausible Cloud would've been fine. Self-hosted Umami is more aligned. I'd rather pay the maintenance tax than the alignment tax.