Nicolay Gerold
@nicolaygerold
Daytime: Hacking on @AmpCode Nighttime: I hack on new startup ideas with generative AI Podcast @ How AI Is Built
"Sadly, it's a bit off a snake oil. These long context embedding models have tested basically all of them, not really working well. So it's [best length of chunks] something between like 500 and 1,000 tokens." Text embeddings are far from perfect. They struggle with long…
![nicolaygerold's tweet image. "Sadly, it's a bit off a snake oil. These long context embedding models have tested basically all of them, not really working well. So it's [best length of chunks] something between like 500 and 1,000 tokens."
Text embeddings are far from perfect.
They struggle with long…](https://pbs.twimg.com/media/GX2Zg3vWgAAZHN1.png)
Obviously, if you just got hit by a balloon filled with water, and you're like "who was that", you can click here to read it: registerspill.thorstenball.com/p/joy-and-curi…
1. Wall pike shoulder openers (feet on box, chest to wall, 30 s) Just got this exercise suggestion by o3. Can someone explain to me how to put the chest to wall and my feet on a box.
We work hard so you can work harder.
Amp uptime over last 30 days. There have been some blips as we scramble to handle strong growth, but in general really happy about this.
I am still looking for a good term for "coding with agents" Compound engineering is being thrown around by @danshipper, but it's more in the direction of building artifacts and workflows for agents to follow. Agentic coding just feels off. Agent-Based Development, just no.
Our next step towards more agents in the terminal: amp -x
... and if anyone knows something about agents, it's Hamel
Ya’ll @AmpCode is good. Try it. Thank me later. It’s expensive but not as expensive as your time.
Ya’ll @AmpCode is good. Try it. Thank me later. It’s expensive but not as expensive as your time.
New podcast episode is out with @BHolmesDev on how he builds and uses @warpdotdev youtu.be/EkXimXl_0Lk
10 - 80 - 10 Rule: 10% writing the prompt, collecting references, bringing clarity into what I want to build. 80% let Amp code it. 10% tweak with prompts or manual edits.
Some odd things I have in most system prompts: 1. Prompt me with follow-up questions to clarify my intent. 2. Offer feedback on how I could rephrase my input or what additional information I could provide to get more accurate or useful responses. 3. Suggest possible next actions,…
What people don't realize is that most of AI is fucking boring.
Is this this meta-programming everyone keeps talking about. @AmpCode controlling @AmpCode.
best way to learn as an engineer is making mistakes and with agents I can just make mistakes faster
First redesigning APIs to make good MCPs, now redesigning CLIs. notcheckmark.com/2025/07/rethin…
Successful AI agents in production are constrained, domain-specific tools with clear boundaries, human oversight, and robust engineering for reliability and integration. The future of AI agents will be less about full autonomy and more about powerful assistants working within…