Alex Carmichael

Here is a consolidated list of articles on my persistent, LLM agnostic and telegram enabled personal agent setup: alexcarmichael.com/halsey-setup

@AndrewYNg Makes 0 sense

The new White House policy requiring green card applicants to apply from outside the US is a capricious attack on legal immigration. It will hurt families, leave us with fewer doctors, teachers and scientists, and hurt American competitiveness in AI.

@dhh Do you run persistent harnesses with each model (toggle between them)?

For complicated agent work, it's amazing how much GPT5.5 has improved. I found 5.2 to be very far behind Opus. Now using Opus 4.7 after 5.5 feels like a big step backwards. Gotta love this level of competion! Strong comeback for OpenAI.

@ericzakariasson Curious how "blocks files over 1k lines" is enforced. hard reject at PR time, or does the agent propose the split and you approve the decomposition?

the most used skill internally at cursor right now /thermo-nuclear-code-quality-review - deletes complexity instead of moving it - blocks files over 1k lines - flags thin wrappers and leaked logic - rejects PRs that work but make code messier

@jxnlco Does the persona extraction step uses clustering on message embeddings or just free-form identify patterns? the latter may hallucinate consistency that isn't there.

codex prompt tip 13 "read my past 400 slack messages, identify my personas and make a skill on how to message people one each one, then do the same thing for emails and twitter so you jnow how to write in my voice"

@dair_ai at diversity threshold does the compiled model start degrading? how narrow does the workflow scope need to be to hold that quality bar?

NEW paper worth reading. A full agentic workflow can be distilled into model weights and run at roughly 100x lower inference cost while preserving near-frontier task quality. The workflow includes multi-step LLM calls, tool invocations, intermediate scratchpads, and decision structure. Instead of expressing all of that at runtime through a framework, the paper amortizes the behavior into a compiled model through targeted distillation. This is the strongest economic argument for agent compilation so far. Runtime loops are flexible, but expensive. Compiled workflows trade some flexibility for a massive inference-cost reduction. Paper: arxiv.org/abs/2605.22502 Learn to build effective AI agents in our academy: academy.dair.ai

@thejustinwelsh I have no opinion on this

An easy way to be at peace is to not have an opinion on everything in the world.

@gregisenberg Makes sense in theory, but what if each customer has minor variations in the workflow? Your process is exactly what I’ve followed building custom agents in my businesss, but it’s not obvious that they could be lifted and replaced to another company with a different tech stack

How to build a vertical AI agent cash-flowing startup: find painful workflow in a boring industry → talk to 10 people who do that workflow every day → map every step, every tool, every spreadsheet, every phone call → do the workflow manually first → be the agent before you build the agent → find the edge cases that break everything → document them in obsidian as structured markdown → set up your agent stack → hermes for the harness → obsidian vault as the knowledge base → composio for authentication across apps → build your first 1-3 skills that solve the core pain → use claude code or codex to build the product → use agents to set up other agents → use perplexity MCP and context7 for up-to-date docs → let the agent handle the scaffolding while you focus on the workflow logic → ship the agent to your first 5 customers for free → watch what they actually use it for → they will surprise you → the thing you built for isn't always the thing they need most → build content around the niche → not "building in public" content → useful content → the tips, the shortcuts, the pain points that only someone who does this workflow would know → become the person for that niche → charge per outcome not per seat → per lease renewed, per claim processed, per candidate sourced → the ROI conversation takes 10 seconds when it's tied to a result → set up watchdogs and alerts → your agent emails you when a cron job breaks or a skill fails → the customer should never have to tell you something is broken → connect to open router → see exact costs per model per task → use GPT 5.5 for tool calls → use open source for lightweight tasks → route the right model to the right job → watch your margins double → let hermes write to its own memory after every task → the agent compounds → the longer it runs the better it gets → that accumulated memory becomes your moat → a competitor can clone your product but they can't clone 6 months of context → expand the workflow → you started with one step → add the next → then the next → now you own the entire workflow end to end → you went from a tool to the operating system for that vertical → stack the agents → one agent is a side project → five agents across five customers is a business → each one runs in its own environment → you check in once a day → raise only if you need capital not credibility → most agent businesses should never raise → the margins are too good to give away equity → stay lean → stay profitable → repeat i'm rooting for you

@jxnlco Persistent task state across sessions and a structured decision log — right now every context reset throws away implicit reasoning that's hard to reconstruct.

If you're using codex desktop app today, what features do you feel like are still missing? Let me know and I’ll summarize all the feedback and share internally.

@SahilBloom In my experience, people miss a lot of chances to keep their mouths shut

I’m convinced that knowing when to keep your mouth shut is the ultimate life hack.

No, after you, I insist.

@jxnlco Does Codex pull from that repo at inference time, or is it baked into the system prompt context on session start?

If you're using Codex, you might not know about this repo `openai/skills` if codex is open right now just ask codex `take a look at skills installer, what are some skills I should install` and it'll know how to scan this repo for skills and install the ones you need

europeans after enabling their out of office auto reply until september

@dair_ai Does the memory model itself require retraining when the domain shifts, or is the update mechanism truly online?

// Memory as a Model // The paper augments any LLM with a separate trained memory model that stores, retrieves, and integrates facts on its behalf. It decouples memory updates from base-model weight updates. It achieves continual-learning robustness without catastrophic forgetting, which is a property that RAG fails to deliver. A vector store is a database with a learned encoder bolted on. MeMo is a learned subsystem with explicit interfaces. That distinction matters, as agents need to be able to ingest fresh knowledge weekly without retraining or vector-DB churn. At its core, the position here is that memory in agents should be modular, learned, and gated, not a context-window hack. Paper: arxiv.org/abs/2605.15156 Learn to build effective AI agents in our academy: academy.dair.ai

@SahilBloom My problem is that the excitement "collides" with the exhaustion and sleep quality lacks as I lay awake thinking about how excited I am for the next day. Curious how others handle this.

Build a life where you're energized in the morning and exhausted at night. Energized in the morning means you're excited about the things you get to work on and the people you get to work on them with. Exhausted at night means you gave your all to those things and people. Your best life is built in that collision.

@trq212 I think you meant to say "let's get this bread"

the future is bright, lets get to work

@thejustinwelsh I know many who've built a business who would disagree with you. It's all tradeoffs

Once you've built your own business, the idea of being told what to do by another adult seems ridiculous.

Pure gold

This is the greatest video I’ve ever seen. No notes. The lifeless clanker carcass just laying there. No crowd reaction, anything. Just Billie Jean. Until its lifeless shell is shamefully dragged off. Purely amazing.

I've completely stopped using Notion UI. It still exists as a backend database for my Agents to use but my workflow is 95% Telegram messaging my agents to do things. It happened gradually then all at once.

@danshipper Something good

what did karpathy see