Pinned Tweet
You and your team deserve AI agents that take on menial tasks, so you can make extraordinary things.
delegate.to
2
119
This wave of patches is going to be brutal. But we’ll all be better off.
Our security researchers built an autonomous cyber vulnerability discovery, exploitation, and patching pipeline.
In a single run with an upcoming cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel at a total API cost of $1.2 million.
30
Thank-you for the Top Pick @justinryanio! So excited that after two years of development our little aviation app finally gets to be enjoyed by everyone!
See The Bravo just launched on Apple Vision Pro, and it’s already one of my favorite apps!
You can explore live air traffic in 3D and follow planes as they fly. I’ve spent so much time jumping from one aircraft to the next. It’s incredible.
1
2
76
You can find the Spatial Insiders review and the app here: spatialinsiders.com/vision-p…
1
41
So you want to run top-tier AI models on your Mac. But which model actually fits your hardware? And which engine should you run it on?
I got tired of the guessing game, did the testing, and built an open-source tool to map out the exact answer for your setup. 🧵👇
1
2
170
It tracks all the top open models across text, vision, and audio—including DeepSeek V4, Qwen3.8, GLM-4.7, FLUX.2, and Kokoro.
Instead of searching through random forum threads to see what works on your chip, you get a clean breakdown built specifically for Mac power users.
1
45
Check out the interactive tool and open-source repo here: dreamingwell.github.io/apple…
If this saves you time setting up your local AI stack, hit repost on the first post to share it with other Mac builders! 🙏
41
16 qwen3.8-27b ai agents, 300 tok/s aggregate, $1.017 per hour to rent. The RTX Pro 6000 incredible.
1
60
Qwen-3.8-27B is legit. First local single GPU model that seems useable for many rounds of tool calling and agentic work.
1
92
It appears that Kimi-K3 hallucinates tool use far more often than even the super cheap GPT-5.6–Luna.
That means for agentic use, k3 appears to be not better than a model that is ~10x cheaper.
1
84
"...GPT-5.6 Sol’s detected cheating rate was higher than any public model we have evaluated on our ReAct agent harness". This explains a lot of what I'm seeing. metr.org/blog/2026-06-26-gpt…
54



