Broken Clock
And for coding tasks that are not in the top 5% of difficulty you can ask V4 0731 to create the code cheaply, have SOL 5.6/Kimi/ Fable review it and have again V4 implement modifs + last review from SOTA model. You use maybe 75% less token that way for result…
In practice when things get complex it is not that easy. You better have a good handling of model capabilities and more importantly clearly define their roles and avoid all forms of bloating. You need to rely more on the SOTA models than I thought, so while you do save $, it's less than half of what I expected
I also think Claude Code & Codex provide a lot more scaffholding by default for everyone not already a good or great SWE. Which all things considered might be worth the extra cost
Broken Clock
I longed some YEN, will add if we go lower
Added more, still leaving room to add
It's kind of funny that you can put the transcript of a podcast you're not a master of into chatgpt and have it destroys all the thesis enumerated by the guy in 10min.
This guy sometimes spent anywhere between 2y to 40y getting where he is and he worse than a 4-6 months years old AI system at his domain
This guy sometimes spent anywhere between 2y to 40y getting where he is and he worse than a 4-6 months years old AI system at his domain
Broken Clock
In practice when things get complex it is not that easy. You better have a good handling of model capabilities and more importantly clearly define their roles and avoid all forms of bloating. You need to rely more on the SOTA models than I thought, so while…
Yeah don't bother touching anything else than Sol 5.6 high, Fable, Kimi K3 and Opus 5 for dev work if you're not a SWE. It's just not worth it. It doesn't matter if it's Deepseek or Luna it's just not good enough.
They can't even increase sales despite the biggest wealth effect in history (china boom excluded) they're so fucked.
Worth reading that you agree or not https://x.com/goodalexander/status/2089906200817066108?s=46
X (formerly Twitter)
goodalexander (@goodalexander) on X
My take on society and crypto is really simple. If you have societal level investment, you are accountable for societal level returns. Things are not getting better.
We said we have "AI warfare"…
We said we have "AI warfare"…
One counter thesis I have to his when he says Uber didn't get better, I agree, but we have moved to a world so complex that almost nothing is fixable.
The value of AI will be does it deliver Uber like service better, but it won't be Uber that does
The value of AI will be does it deliver Uber like service better, but it won't be Uber that does
There is also a ceiling in terms of how much better you can make smt / costs involved
Uber app is fine for what it does, I don't want a 10x better Uber app if it increase the cost of my rides even 10%
Physic reality that could do with an upgrade, planes tickets have increased outrageously while the quality of the offering has clearly not followed. But I'm not sure AI will improve my lie flat seat. Maybe having a VR immersive headset will make it more palatable