Jon Dalgir

Building decision-support intelligence for AI-focused teams
Founder @ • Watchlists • Homepage intelligence

[Jon Dalgir @jondalgir]

@thsottiaux Very cool, but how does hosting work via share websites in codex?

[Jon Dalgir @jondalgir]

@OpenAI Pretty coolio, will test it out later today.

[Jon Dalgir @jondalgir]

@Yuchenj_UW Let it begin!

[Jon Dalgir @jondalgir]

Build update #2 for RouterCore.

I ran the full Track 2 loop on AMD Developer Cloud:

- synthetic DevOps routing dataset
- Qwen LoRA fine-tuning
- ROCm PyTorch on AMD Instinct MI300X
- eval vs deterministic baseline

Confirmed setup:
torch 2.9.1+rocm6.4
HIP 6.4
AMD Instinct MI300X VF

Results:

FakeRouter:
required fields 28.57%
status accuracy 57.33%
unsafe rejection 100%
false routes 0%

Safety-tuned LoRA on AMD ROCm:
required fields 100%
status accuracy 86.67%
unsafe rejection 100%
false routes 0%

Fine-tuning improved structured routing quality, while policy stayed authoritative for safety.

Repo: https://t.co/ZxGZVRhr0g
Space: https://t.co/lTNrjnkvvI

@lablab_ai @AIatAMD

#BuildInPublic #ROCm #AIAgents #FineTuning

[Jon Dalgir @jondalgir]

@thsottiaux Hello @thsottiaux there is an issue with codex where if i close the terminal inside codex on Windoes, instead of just going back to the original tab, it trys to open up a new chat. Just wanted to share, cheers.

[Jon Dalgir @jondalgir]

@sama I've been using wispr flow to articulate what I am thinking and other ramblings, but sometimes switch between typing and voice

[Jon Dalgir @jondalgir]

@mcuban I had no idea there was a cup fee. This is how I know I do not go outside enough.

[Jon Dalgir @jondalgir]

Building RouterCore for the AMD Developer Hackathon.

It’s a safe routing layer for AI agents: before a DevOps agent executes, RouterCore decides whether a request should be routed, clarified, confirmed, rejected, or escalated.

Current MVP: FakeRouter baseline, schema validation, policy engine, clarification loop, Gradio demo, synthetic train/eval data, eval comparison tools, optional HF model + LoRA path for AMD ROCm

Baseline:
JSON validity 100%
workflow accuracy 97.01%
unsafe rejection 100%
false route rate 0%

Next: fine-tune a compact model on AMD GPUs with ROCm.

Repo: https://t.co/ZxGZVRhr0g

@lablab_ai @AIatAMD

#BuildInPublic #AIAgents #ROCm #FineTuning #HuggingFace

[Jon Dalgir @jondalgir]

Stronger AI systems may not come from one model knowing everything.

They may come from smaller domain-specific environments: constrained workflows, clear rules, known data, validation layers, and human judgment where it matters.

That is probably the foundation for more reliable enterprise AI.

[Jon Dalgir @jondalgir]

@piyorinko25 Exactly. Also, for what it’s worth, it did not auto-translate on my side. I had to tap translate, but once I did, it came through clearly in English. Appreciate the thoughtful reply.

[Jon Dalgir @jondalgir]

Quick question:

What is the most annoying workflow problem you are dealing with right now?

AI, engineering, infra, product, dashboards, ops, research, anything.

Trying to understand what people actually need, not just what sounds cool to build.

[Jon Dalgir @jondalgir]

Enterprise AI will probably look less like “one model to rule them all” and more like a messy portfolio of models, agents, vendors, internal tools, data sources, and governance layers.

The opportunity is not just building agents. It is keeping the whole system coherent.

That is also why domain-fit smaller models are interesting.

Not every enterprise AI task needs a frontier model. For narrow workflows, a trained smaller model can reduce external API calls, limit token ingestion outside the company, and make systems cheaper and more controllable.

I explored this with Gemma for GCP infra routing:
https://t.co/zqL8JzJRJk

[Jon Dalgir @jondalgir]

@AdiPolak You can read the code, but that doesn't mean you'll understand it. Especially if its spaghetti legacy with meatballs.

[Jon Dalgir @jondalgir]

Enterprise AI will probably look less like “one model to rule them all” and more like a messy portfolio of models, agents, vendors, internal tools, data sources, and governance layers.

The opportunity is not just building agents. It is keeping the whole system coherent.

That is also why domain-fit smaller models are interesting and is worth checking out.

Not every enterprise AI task needs a frontier model. For narrow workflows, a trained smaller model can reduce external API calls, limit token ingestion outside the company, and make systems cheaper and more controllable.

I explored this with Gemma for GCP infra routing:
https://t.co/zqL8JzJRJk

[Jon Dalgir @jondalgir]

https://t.co/oAKtpuZ3Ca

[Jon Dalgir @jondalgir]

This is why I’ve been experimenting with LoRA on smaller models for engineering tasks.

For some workflows, the question is not just “is the model smarter?”

It is: “can a smaller model be trained to fit this domain reliably?”

I tried that with Gemma for GCP tool routing:
https://t.co/zqL8JzJRJk

[Jon Dalgir @jondalgir]

@Yuchenj_UW I get it. My last job paid a lot more, but it came at the cost of my mental health. I make less now, but I’m genuinely happier.

[Jon Dalgir @jondalgir]

@zarazhangrui This is pretty neat! Do you have a prompt you can share? Just curious to try it myself!

[Jon Dalgir @jondalgir]

@boardyai Indeed, and not just that, networking in general, start early even when you are in college or uni.