I’m not an expert either. I mostly use Opus at work and I don’t pay for it - the company does. I’ve noticed Fable almost seems like a waste for a lot of tasks, so mostly I don’t use it. I’ve not tried dialing it down to Sonnet or even Haiku to see how things fare, but keep meaning to try that out.
You’ll see Fable 5.2 is at the very top with 61.1. Even Opus is 58.7.
Here, you can see the very top is 54.9 from Kimi and I’m not sure which of these you can host yourself. My understanding is that just because something is open weight doesn’t mean you can download it and in some instances, even when you can, the hardware to run it is not something exactly consumer-class (but a company might choose to build or rent that hardware themselves).
When you remember how people went gaga over this in 2022…imagine using a model that “lags” some of the top-tier stuff but applied to more mundane things - it’s not like the capabilities from 2022 or whatever evaporated. Suppose you assume you can run something at decent enough speeds, but it “lags” the others by 6 months to a year?
At some point, it just seems to me there will be diminishing returns to keep paying for the most expensive option. My company offers me Fable now and I don’t find that I need it for day to day things. For all I know, Haiku might be enough for some things.
I’m not an expert either. I mostly use Opus at work and I don’t pay for it - the company does. I’ve noticed Fable almost seems like a waste for a lot of tasks, so mostly I don’t use it. I’ve not tried dialing it down to Sonnet or even Haiku to see how things fare, but keep meaning to try that out.
If you look at leaderboards like:
https://llm-stats.com/leaderboards/llm-leaderboard
You’ll see Fable 5.2 is at the very top with 61.1. Even Opus is 58.7.
Here, you can see the very top is 54.9 from Kimi and I’m not sure which of these you can host yourself. My understanding is that just because something is open weight doesn’t mean you can download it and in some instances, even when you can, the hardware to run it is not something exactly consumer-class (but a company might choose to build or rent that hardware themselves).
https://llm-stats.com/leaderboards/open-llm-leaderboard
When you remember how people went gaga over this in 2022…imagine using a model that “lags” some of the top-tier stuff but applied to more mundane things - it’s not like the capabilities from 2022 or whatever evaporated. Suppose you assume you can run something at decent enough speeds, but it “lags” the others by 6 months to a year?
At some point, it just seems to me there will be diminishing returns to keep paying for the most expensive option. My company offers me Fable now and I don’t find that I need it for day to day things. For all I know, Haiku might be enough for some things.