Cass (@x0radi) 's Twitter Profile
Cass

@x0radi

just a cool dog

ID: 1142644956353503232

calendar_today23-06-2019 04:06:01

2,2K Tweet

232 Followers

3,3K Following

Cass (@x0radi) 's Twitter Profile Photo

I think this is a big limitation right now, I haven't seen anything come out of the big labs that controls mouse and keyboard. I'd want to see AI draw something in photoshop with M&K. It's not as efficient but it's the medium into all the software that we use.

Cass (@x0radi) 's Twitter Profile Photo

Brian Beers Never take advice from others on their younger days. I wouldn’t even trust yesterday's me. Who knows what I was thinking

Cass (@x0radi) 's Twitter Profile Photo

Hyde 💨 I just spent six hours trying to build an AI agent last night. Was it fun? Yes. Did I make any progress? No. Did I learn anything? Also no.

Cass (@x0radi) 's Twitter Profile Photo

mfers elite pawning their way through all that superalignment research. Just let the autistic nerds solve it for you

Cass (@x0radi) 's Twitter Profile Photo

No training run this year is doing 10^27 compute from the looks of it. GPT-5 is halfway there with 5.13 x 10^26 FLOPs at a cost of $612M. Not sure what Grok 2 is doing with $6 billion dollars.

No training run this year is doing 10^27 compute from the looks of it. GPT-5 is halfway there with 5.13 x 10^26 FLOPs at a cost of $612M. Not sure what Grok 2 is doing with $6 billion dollars.
Cass (@x0radi) 's Twitter Profile Photo

Some quick napkin math. It's kind of crazy that S3 prices itself in that range where you'd lose money from customers leaving if you had outages.

Cass (@x0radi) 's Twitter Profile Photo

I noticed an interesting limitation with the new O1 and O1-mini models: they're not using the correct route for their API calls, even when provided with the proper documentation. Tested it with the following prompt: "Using the documentation below, create a Python CLI interface

I noticed an interesting limitation with the new O1 and O1-mini models: they're not using the correct route for their API calls, even when provided with the proper documentation.

Tested it with the following prompt:

"Using the documentation below, create a Python CLI interface
Cass (@x0radi) 's Twitter Profile Photo

Was checking out this repo and it's just wild... If he wrote all this himself, than the dude is a one man army. I'm not even using the UI, just sticking to the CLI for now. Now I just need to use his agents to scrape all the docs from his site so my main AI can learn all the

Cass (@x0radi) 's Twitter Profile Photo

Hmm... I keep reading posts lately about features like the new models being agentic and having better tooling, but I think this is a misunderstanding of function calling. Current AI models are already agentic and self-acting! A function call has two parts, a tool_use and a

Hmm... I keep reading posts lately about features like the new models being agentic and having better tooling, but I think this is a misunderstanding of function calling. Current AI models are already agentic and self-acting!

A function call has two parts, a tool_use and a
Cass (@x0radi) 's Twitter Profile Photo

I just don't understand how Google's Gemini 2 works via API... Okay, so there's AIStudio, which uses an API key, nice but if you try to set it up with any type of function calling, you're rate-limited to like 2 API calls per minute, even for something as simple as a greeting.

Cass (@x0radi) 's Twitter Profile Photo

Was curious why the code written by Claude Code's sub agents was trash... Looks like it's using Haiku 3.5 for it's subtasks... It explains why everyone is spinning up so many versions of claude code, none of it works well. Vibe coding going back to the dark ages

Was curious why the code written by Claude Code's sub agents was trash...

Looks like it's using Haiku 3.5 for it's subtasks... It explains why everyone is spinning up so many versions of claude code, none of it works well. 

Vibe coding going back to the dark ages
Cass (@x0radi) 's Twitter Profile Photo

What’s the best diff format for text replacement with gpt-5? Is the model also trained on your apply_patch tool? Should we give it a shell tool and let it focus its own attention with sed/rg, or is giving it a read tool that reads an entire file not productive? If I wanted to

Cass (@x0radi) 's Twitter Profile Photo

Hey Adam.GPT, could you like talk to someone, so that chatgpt can search it's own damn api documentation instead of being restricted by cloudflare and an auth login? Would make building stuff much easier 🙂

Hey <a href="/TheRealAdamG/">Adam.GPT</a>, could you like talk to someone, so that chatgpt can search it's own damn api documentation instead of being restricted by cloudflare and an auth login? Would make building stuff much easier 🙂
Cass (@x0radi) 's Twitter Profile Photo

Why can't they get their models under control? Codex seems to be excellent at following directions and making the right choice.