We really need to add some kind of risk to people making these claims to make it more interesting. I listened to the type of advice you're giving here on more occasions than I can remember, at least once for every major revision of every major LLM and always walked away frustrated because it hindered me more than it helped.
> This is actually amazing now, just use [insert ChatGPT, GPT-4, 4.5, 5, o1, o3, Deepseek, Claude 3.5, 3.9, Gemini 1, 1.5, 2, ...] it's completely different from Model(n-1) you've tried.
I'm not some mythical 140 IQ 10x developer and my work isn't exceptional so this shouldn't happen.
The dark secret no one from the big providers wants to admit is that Claude is the only viable coding model. Everything else descends into a mess of verbose spaghetti full of hallucinations pretty quickly. Claude is head and shoulders above the rest and it isn't even remotely close, regardless of what any benchmark says.
I've been working on macOS and Windows drivers. Can't help but disagree.
Because of the absolute dearth of high-quality open-source driver code and the huge proliferation of absolutely bottom-barrel general-purpose C and C++, the result is... Not good.
On the other hand, I asked Claude to convert an existing, short-ish Bash script to idiomatic PowerShell with proper cmdlet-style argument parsing, and it returned a decent result that I barely had to modify or iterate on. I was quite impressed.
Garbage in, garbage out. I'm not altogether dismissive of AI and LLMs but it is really necessary to know where and what their limits are.
dns_snek|6 months ago
> This is actually amazing now, just use [insert ChatGPT, GPT-4, 4.5, 5, o1, o3, Deepseek, Claude 3.5, 3.9, Gemini 1, 1.5, 2, ...] it's completely different from Model(n-1) you've tried.
I'm not some mythical 140 IQ 10x developer and my work isn't exceptional so this shouldn't happen.
ramesh31|6 months ago
delta_p_delta_x|6 months ago
Because of the absolute dearth of high-quality open-source driver code and the huge proliferation of absolutely bottom-barrel general-purpose C and C++, the result is... Not good.
On the other hand, I asked Claude to convert an existing, short-ish Bash script to idiomatic PowerShell with proper cmdlet-style argument parsing, and it returned a decent result that I barely had to modify or iterate on. I was quite impressed.
Garbage in, garbage out. I'm not altogether dismissive of AI and LLMs but it is really necessary to know where and what their limits are.
Sharlin|6 months ago