> I'm in this process of running something, get logs, tune, get logs, tune... and I told the thing multiple tmes... hey create scripts if you need to run something multiple times... the thing burned like crazy in fable 5.1 effort medium... until you put in caps in expletives what's going on... then writes the script and 16 tokens to check the results...
Anthropic already says Fable is token hungry and should be used for 'ambitious projects'; IME, i use it only if I need to do some open ended research where the answers are candidates for further research. You are using it to do the work of Opus or Sonnet, and of course it will be terrible.
Like the other commenter says, it's like hiring Ken Thompson to write tps reports
Great metaphor, and I tend to agree. At some point we have to become fine with not using the newest shiniest model, because we will or already have, hit a limit to the necessity of the newest shiniest model. When you visit a foreign city and need a translator, do you hire a PhD in linguistics, or do you hire someone who can speak both your language and the local language?
Anthropic feels free to do whatever they want for better or worse in their pursuit of building the machine god.
They can A/B test on users, "pace the frontier" but then actually concentrate upon recursive self improvement. Just like Adverts, you are the product, not the customer -- $100/mo individual plan is nothing compared to the enterprise and government sales.
Building AI models requires massive compute power—it's not as easy as it sounds
> I'm in this process of running something, get logs, tune, get logs, tune... and I told the thing multiple tmes... hey create scripts if you need to run something multiple times... the thing burned like crazy in fable 5.1 effort medium... until you put in caps in expletives what's going on... then writes the script and 16 tokens to check the results...
Anthropic already says Fable is token hungry and should be used for 'ambitious projects'; IME, i use it only if I need to do some open ended research where the answers are candidates for further research. You are using it to do the work of Opus or Sonnet, and of course it will be terrible.
Like the other commenter says, it's like hiring Ken Thompson to write tps reports
What kind of work? That sort of makes a big diff. If you hire Einstein to run Starbucks you are going to get quantum field theories on making coffee.
Great metaphor, and I tend to agree. At some point we have to become fine with not using the newest shiniest model, because we will or already have, hit a limit to the necessity of the newest shiniest model. When you visit a foreign city and need a translator, do you hire a PhD in linguistics, or do you hire someone who can speak both your language and the local language?
Fable 5.2 is basically useless on a 20$ pro subscription.
Anything even slightly interesting worth trying gets 5h limit hit in 10 minutes. (Code audit on Fable 5.2 on high with "use subagents")
So the "sneak peek" one time reset courtesy gives me only real 20 minutes of Epstein brain to run my Starbucks - meager meat on the bone.
So why do I need Fable 5.2 at all?
Anthropic feels free to do whatever they want for better or worse in their pursuit of building the machine god.
They can A/B test on users, "pace the frontier" but then actually concentrate upon recursive self improvement. Just like Adverts, you are the product, not the customer -- $100/mo individual plan is nothing compared to the enterprise and government sales.
I believe this is the sound of the music ending.