I took a quick look at agentic coding IDE. I think they are taking this in an interesting direction. Sonnet only as far as I could tell for now but it formalizes a workflow that a lot of strong AI coders use re: acceptance planning first.

if you put the timeline start at Babbage or before, the rise of AI has to be the canonical form of "slowly then all at once".

Amazons agentic ide. Looks like it makes plan building a first class citizen. Will try it out today!.

Architecture = grok 4 heavy or o3 pro via playground (system prompt doesn't interfere).Quick question = grok 4 or o3.Bug help = grok 4 then g 2.5 then o3.Hard bug help = grok 4 heavy / o3 pro.Really hard bug help = o3 pro deep research + git work tree and Claude code opus in yolo.

AI is making me rip out tons of SaaS. Build vs buy calc changed. Insidious how hard it is. Many systems in place to ratchet costs back up. Killing the card and then having a frank convo is one of the only solutions I'm having success with.

Sometimes I get frustrated with AI because it introduced a subtle bug (my fault) that cost me time, but then I remind myself that I literally wouldn't have gotten started without AI and then I think. no I am still mad. Stupid AI.

Groq + Kimi 2 in aider is incredibly fast.

Kimi-2 is really compelling but too slow to be bearable for coding tasks through open router atm.

I hate this but. I think it won't matter for a while that apple is missing AI. They have the best devices by a mile. Huge lock-in, including for social reasons which is massive. They can figure it out after the dust settles at the model level, probably.

Built with grok - AI supply chain extraction of unstructured data into graph which can be used to answer a variety of high value questions for retailers. No prior knowledge, 10 minutes to process data and infer relationships. 200+ spreadsheets. @elonmusk

Grok 4 is too good, they are going to take it from us. A frontier AI that talks to me like an adult, listens to my custom instructions, defers to my moral judgement and grants me agency? Too good.

Welp the fact that I can is probably more good than bad net net. Maybe not for my social life.

Ever wish you could automatically find all the code related to a failing test and compose a prompt you can paste in to grok, chatgpt, etc?. This uses LLMs to figure out which source code is relevant + includes the failure stdout/stderr, copy pasteable output:

Claude code is awesome at finding where something broke in the git history. I use it a lot for this.

In general the MCPs I've used leave the model "looking through a key-hole." Making them only narrowly useful for "do this one specific task" and getting broader context takes a ton of repeated calls that frequently lead to divergence from original task.

Side project: full agent coding. Go nuts. Learning code base : Claude code, auto ops & env.Debugging tests: Claude code with supervision.Real world simple: Claude + plan mode.Real world hard stuff: Aider + grok 4. Precise edits + git.Very hard: aider + cursor side by side. Grok.

Quiet Grok is

Peak grok maxing. This is mostly AI writing AI, btw.

RT @__YuWang__: Introducing The Most Advanced Memory System for LLM Agents. MIRIX is by far the most advanced memory system in the world, d….