Codex limits are non-existent now it seems - and GPT 6.1 Sol worse than Opus 5.5 - what's a guy to do?! How can I have used over 60% of my weekly limit in 15 hours, and barely done anything?

Deepseek just replied to my request with "Jiaotong University" in Chinese, then went back to English to answer fully - anybody had this before? 🤔

📹 ✅

Smart ring delivered at noon, new firmware created by 3pm, bricked to never recover by 4pm 😅 I knew I should have stayed away from hardware

72% of Just Eat restaurants are shown with a perfect 5/5 rating - but it should probably be 18% 📈 I analysed more than 70,000 UK listings, and 10m reviews, to see why Just Eat say so many are 5/5, and I stumbled into a few oddities. The investigation, evidence and a way to check your own area are here 📍 ashleyhindle.com/writing/can…

See you there? @LaraconAU

I gently hacked this microphone to remotely control my Mac 🎙️

Are OpenAI speed running a developer migration from Codex to Claude? Opus 5.5 is better than Sol 6.1. And now my $200/mo plan has been cut in half? "Starting October 30, 2026, your included usage in Codex will decrease from 20 times to 10 times the ChatGPT Plus allowance"

Turns out gpt-6-sol is really lazy I wish updating to the latest models was unproblematic and straight forward 😅 I've not experienced this for a while while using 5.6-sol, I thought they'd fixed it, but I think there's a lot of "GO ON THEN, BLOODY DO IT!" in my future 🔮

Aahhh yes, thank you for all the detail Codex 😀 "Ran commands" -> "Ran command" 🤫

Jev was 3x faster, 65% cheaper, and safer on this benchmark 🤯 Fuel classifies the first message in a new chat to see what level of intelligence will be needed for that chat. If we get it right we reduce the user's token cost, improve latency, and user satisfaction scores increase. If we get it wrong we piss the user off and increase their costs, usually because we've given a less-intelligent model a difficult problem. Our benchmark for this has 75 test cases, and we found GPT-OSS-120B does great, but Jev from @typesafeai does it even better 😍 There are so many ways to use Jev to improve coding harnesses & AI software factories, lots more to do here 🙌

The very first conference talk I did was for @LaravelLiveUK and I had the pleasure of practicing it on Zoom for the @TightenCo team and in the room they were awesome, and afterwards incredible with the ridiculously useful feedback they gave 🫶 Really helped with my talk structure and confidence - great team, but also wonderful humans 😍

These are the subagents I'm spending roughly a billion tokens on daily 🙏 I'm optimizing for cost & quality over speed, so my choices here might be different to yours, and they're surprisingly expected I'd say My main agents: - explorer -> return a short map of how things work <- deepseek-v4-flash-0731 - observer -> vision for non-vision models <- kimi-k2.7 - oracle -> give me advice on this task <- gpt-5.6-sol high - apex -> give me guidance on a broader scope <- gpt-6-astra medium Oracle & apex are the trickiest to get the primary agent to use at the appropriate times. I'm experimenting with apex right now to see if it's worth its inclusion, and where it's better than oracle. I'll share more as I learn more 😁

I knew doing this would be good, but my brain wouldn't let me because of a weird mental block 😅 Early access users are pleased I got out of my own way 😆

Changing my agent harness almost doubled subagent usage 📈 and reduced costs by 57% 📉 Fuel used to spawn the Codex app directly for tasks, and our prompts would nudge Codex to use subagents to reduce cost, but subagents would only get used 35.8% of the time. Now all work in Fuel goes to our own agent, with similar subagent tools to Codex & no prompt changes, and subagents are suddenly used 69.9% of the time with the same models 🤯 Really surprised me to see subagents get used so well, after fighting with Codex for so long to get it to use them better 😅