AI in practice: agents, harnesses, models, tools. The feed is noisy. Most of it is amateur hour. I write the parts that survive contact with real ugly work.
Joined August 2025
- Tweets109
- Following128
- Followers4
- Likes126
I have to ask - how many people have tasks that can scale to that many (100?) agents? So many things have serial development stages and dependencies, bugs and mismatched planning, that scaling over more than a handful of scenarios such as toy problems become untenable.
kewltools retweeted
David Sacks says the Wall Street Journal quoted the Anthropic whistleblower's tweet storm 18 minutes before it was posted
[ But do you think it was like a premeditated coordinated effort? ]
"We know that this was a coordinated media campaign. It wasn't just a spontaneous tweet. How do we know that? Because the Wall Street Journal ran a story quoting his tweet storm 18 minutes before that tweet storm was printed. So, how could the Wall Street Journal have done that? They had the tweet storm in their hands before Jacob Coxon ever posted it. So, obviously that was coordinated. He worked with them on it. That was set up."
[ Well, you could also argue that that is a so called whistleblower trying to get their message out. ]
"Sure. And that's what I'm saying... he tried to get his message out. It got out. It was amplified by certain groups. And I think it's pretty clear now. This definitely plays into a larger political agenda."
For anything complicated, Every model i've seen in Cursor needs a "review your recent work, fix any bugs" - you can just follow that up in your queue.
kewltools retweeted
I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.
The future will be current frontier capability models running on silicon at thousands of tokens per seconds.
Cerebras, Taalas, etc People will have a box next to their machine, or added in as another specific card.
New Frontier models will still exist - as co-ordinators.
A good harness needs to be adaptive - adaptive to the model and to the problem.
This is more than just a pretty system prompt.
It is more than just simple guardrails or digital electric fence to stop models making petty mistakes.
What works: More like a specialist life coach.
I love small models. They are the best to test your harness around, you'll find every instability you could hope for.
Larger models sometimes get there later too, so if you fix for small - your larger tier will benefit.
Waiting for something more reliable for my agentic coding than Deepseek 0713 flash.
Cheap is great, if it works reliably.....
I think we need to be a bit more cautious with OxAlpha. Some of my files became corrupted in a suspicious way.
Anyone else notice Ox Alpha tends to have a lot of sequential thinking rounds? Does that look familiar to anyone re other known models?
First look at Ox Alpha - wow!
I had tasks that were taking ages with Grok 4.6 - not understanding what how a UI is meant to look, user preferences etc
Put the problem into Ox Alpha and it immediately called out all the debt! Clear ideas on refactoring for usability etc WOW
Hi All! Just created this account if anyone has questions, suggestions etc for my site, KewlTools.com