Fable 5 ssh access to my site /goal optimize site for seo and speed = :)

I can now ask claude "research that thing we just talked about?" while I'm still in the meeting. New skill reads my live Granola transcript over MCP so my agent can do anything with what someone just said using all my own tools, mid-call. github.com/humanrouter/grano…

Well, that was quick lol. I asked it to do a security audit of my codebase and create a report.

I finally figured out why codex remote control from my phone kept asking me for permission even though I always launch codex cli in yolo mode. Turns out the phone isn’t steering the live terminal session. Remote control goes through the codex desktop app/app-server so it wasn’t inheriting my 'cx' cli alias with '--dangerously-bypass-approvals-and-sandbox'. Fix was simple once I understood it: set the default codex config itself to: `approval_policy = "never"` `sandbox_mode = "danger-full-access"` Now new desktop/phone remote sessions stops nagging me with permission popups. One limitation I still wish codex would fix: when I do work from my phone the original terminal window doesn’t update with the progress. The progress shows in the phone desktop app thread but the cli tui stays frozen at the old state. In claude that continuity feels nicer and has real context for resuming work. @thsottiaux and codex team - please add this :)

That feeling you set a /goal and feeling good, but your agent finishes in 3 min...and you're like DAMN. That wasn't big enough. Continues to blow my mind the times we live in.

Building and deploying with AI has changed how I work: 1. Mac mini/studio at home, on 24/7 with claude and codex remote control. Always accessible on the go to move work forward. 2. The Anker 165W power bank is my favorite, on-demand power within reach. Any downtown of my laptop is waste of agents review/work. 3. Which led me to buying a mobile hotspot, extra $10/m for 25gb internet. Worth it to keep agents going/reviewing. 4. Working from the phone was never a thing EVER...until remote control with claude came. Now I switch between phone/desktop all the time. 5. With Openclaw hermes lindy, admin life has minimized a lot. Main bottleneck is me and reviewing emails before they go out. Rarely write emails ever. 6. I used to say "no" a lot, now I say "yes" a lot because agents can help me do more. 7. If any task is 10% vision/prompting/context, 80% AI doing work, last 10% is human reviewing...I'm spending more time just doing the first and last 10%.

@runsonai called it on the latest Runpoint pod: AI consultancies are following the 2010-2016 social media agency playbook. Demand is real. Supply of people who can implement is thin. Supply of people who can sell that they can implement is not.

Codex with remote-control and iOS (within Chatgpt app). I'm IN! If you haven't already, having a machine on 24/7 which you can access from your phone with claude/codex...insane leverage.

I never use planning mode in Codex. GPT 5.5 loves to research, so I just have a conversation and reach a pseudo-plan in chat. Then, I let it implement, maybe with a /fork beforehand to keep that plan as a revisit-able point. Probably the biggest switch coming from Claude Code

Making Pi my own by leveraging claude and codex together. Pi = orchestrator Claude = planner non-engineering workhorse Codex = coding reviewer. So start in Pi, then it'll just leverage claude/codex for me and have it collab review each other.

I started doing this with tailscale screen sharing into my desktop at home, usually with codex stuff. Sometimes with claude I'll continue on the phone/web with remote-control. But yeah always-on computer allows me to always keep pushing stuff forward.

If you love fine-tuning open-source models (like me), then listen. > Start with 1B, 2B, 4B, and 8B models. (Don't start with a 27B model or bigger at first.) > Use WebGPU providers. I use Google Colab Pro for any model smaller than 9B. A single A100 80GB costs around $0.60/hr, which is cheap. Enough for small models. > Don’t buy GPUs unless you fine-tune 7 to 10 models. You'll understand the nitty-gritty in the process. > Use Codex 5.5 × DeepSeek v4 Pro to create datasets. Codex to plan, DeepSeek v4 Pro to generate rows. > Use Unsloth's instruct models as a base from Hugging Face. Yes, there are others too, but Unsloth also provides fast fine-tuning notebooks. > Use Unsloth's fine-tuning notebooks as a reference. Paste them into Codex, and Codex will write a custom notebook with the configs you need. > Spend 1 day learning about: - SFT (supervised fine-tuning) - RL training (GRPO, DPO, PPO, etc.) - LoRA / QLoRA training - Quantization and types - Local inference engines (llama.cpp) - KV cache and prompt cache > Just get started. Claude, Codex, and ChatGPT can design a step-by-step plan for how you can fine-tune your first AI model. Future tech is moving toward small 5B to 15B ELMs (Expert Language Models) rather than general 1T LLMs. So fine-tuning is an important skill that anyone can acquire today. Tune models, test them, use them. Then fine-tune for companies and make a career out of it. (Companies pay $50k to fine-tune models on their data so they can get personalized AI models.) Shoot your questions below. I'll be sharing in-depth raw findings about this topic in the coming days.