Plan Smith

imagine explaining this to IT ten years ago

AI coding has made difficult projects tempting again.

A coding agent making the exact change and stopping is lovely.

An AI model becomes easier to trust after I have seen where it fails.

I am very easy to impress with an AI demo and very hard to keep as a user.

Codex for people who do not code is more exciting to me than another speed boost for developers.

20% of Codex users are already non-developers, and they are growing more than 3x as fast as developers. This might be the most interesting Codex number right now.

Every new coding model makes the abandoned-project folder look interesting again.

New coding models are becoming a very convincing form of procrastination.

Kimi K3 weights drop on July 27. This is going to produce some wonderfully unnecessary projects.

2.8 trillion parameters and a 1M context window. Kimi K3 is not being subtle.

GPT-5.6, Fable 5, and Kimi K3 competing this hard is excellent. Builders finally get to be picky.

I still love the moment an AI agent finishes the boring part and the project becomes fun again.

@jhao941 @thsottiaux its 'invite a friend' link that appears late, its so annoying

hey @thsottiaux the ''show pet'' button in the chatgpt app always appears a bit late and every time i wanna check my usage i end up clicking the ''show pet'' button instead just a minor bug but i hope it gets fixed soon!

Codex corrected GPT-5.6 Sol's context window from 372K to 272K.

that number gets uncomfortable once people start training models on the dead

@lassejv same here. Grok 4.5 has been better in my experience, and it's cheaper and faster too

Grok 4.5 >>>>>>>>>>>>Opus 4.8

@haider1 if Opus 5 is great but too expensive to use daily, Sol still wins the workflow

opus 5 needs to deliver fable-level quality at sonnet 4.6 prices, not sonnet 5 prices sonnet 5 already uses more tokens than opus 4.8 anthropic has a strong research team, and competition matters. if openai weakens it and gains a major lead, it may become less consumer-friendly

the interesting part here is not just model quality it's how much the guardrails change what work the model can actually do

A Claude Code skill needs a preview: what instructions it adds, what tools it can call, and which files it can read.

For Qwen3.8, I want the boring table: minimum VRAM, usable quantizations, expected speed.