Conrad Ludgate

Replacing the HN app on my phone with my weight loss app is probably the most healthy thing I've ever done, both physically and mentally.

Why did I spend my vacation thinking about congestion control and client side load balancing? In other news, I got bitten by a fish in the Mediterranean Sea the other day

I'm mostly using K3 in Codex right now. It's pretty good. It's not cheap but it's comparable with the duopoly. I use it the same way I used Sol or Opus: 1. Sketch out the plan for the feature implementation - several prompts 2. Implement it - one prompt (+skills) 3. Refinement - many prompts

During implementation phase, it's committing by itself and writing/running tests. While it's running I'm off doing something else. After refinement I usually use another session with another LLM and I get it to review it. I also make sure I provide my understanding of the feature to compare against

Would someone security conscious enough to host an MCP/AI Gateway in their own home then be willing to let the agent run on a VM in the cloud? What if the machine the VM was running on had no direct internal or external Internet access, all access goes through a tunnel service via PrivateLink?

I'm working on an interesting product idea for personal AI that involves lightweight VMs for reproducible LLM evals, but I don't want you to give your API keys to me. If your assistant goes rogue deleting emails you can turn off your gateway and everything stops. All gateway access is logged

I think platforms like github should invest in exposing audit information. An agent should be able to see a log of events that occurred and know whether it was me, another agent, or itself that made some change.

The LLM should know that it was an intentional decision and not a mistake it made. I'm working on an experiment that will give each session their own access tokens, but fundamentally it'll require effort from the services to store some identifiers.

I think platforms like github should invest in exposing audit information. An agent should be able to see a log of events that occurred and know whether it was me, another agent, or itself that made some change.

So anyway, I'm currently thinking that package sources live in an OCI registry of your choosing, and the package metadata (dependencies, version, etc) lives in your PDS. It will support off-platform dependencies via packageurl.org.

Oops I accidentally designed a package manager

Oops I accidentally designed a package manager

I need to stop looking for problems for the atproto hammer...

Codex spent the day setting up my new PDS server 👀 interesting to see the process!

I'm pretty deep in the LLM psychosis right now. It started from ADHD induced burnout for a project I really wanted to work on but I just couldn't get past the mental block. I was able to throw cursor at it and make very solid progress. Now it's also my notes and research assistant.

I haven't written a single line of code in the past two weeks. It's definitely helped me so far. But it's also incredibly addictive. I fire off the prompt, wait, review the plan, wait, build the plan, wait, review the code/test, wait for fixes. Brain will probably explode soon.

Working on an absolutely cracked project idea involving zero-trust federation, MLS, paxos, @iroh.computer and CRDTs 👀 I hope this works out because the idea is really really cool

I'm pretty against using AI in my own software but this latest wave of discourse is very 🍿

Don't get me wrong. Plenty of LLM produced code is slop. And I hate the companies behind the current wave. I'm definitely staying away from depending on them until we get more access to local hosted models. The incestuous nature of current training is also unlikely to end well.

I'm pretty against using AI in my own software but this latest wave of discourse is very 🍿

I'm so ready to set up my own git forge when I move. I'm very much looking forward to ditching github and owning my own code backups. Current plan is to set up a tangled.sh knot, hosted in my home (with backups in s3). I want to keep the social aspect of github, I hope tangled does this well

I respect github charging for github actions self-hosted runners because their orchestrater isn't free. However, github actions sucks ass and it's not a compelling product. Not sure what I plan to replace it with. I've been somewhat bullish on CI=build system. Definitely not using nix though.