So any local model that doesn't hit that Opus4.5 level will be very limited. It can help doing some basic tasks, or help a bit with your coding, but you cannot even ask it to summarize morning news without doing most steps (such as getting the articles) manually.
That's only partially true. Some of that hunger for frontier comes from the fact that under a certain threshold, the LLM is mostly a useless toy, and above that threshold it suddenly becomes very useful.
Before Opus4.5, LLMs were good for search, small scripts, autocomplete, but not much else.
I just learnt about approval voting and decided that it would solve all of the problems of modern societies.
Also these models will be served close to inference price by a variety of providers. As opposed to closed labs serving models with ever changing conditions and 3-5x margins.
> New paper: LLMs are useless, biased and will make your little kid cry
> Looks inside
> LLMs used in the study: 4o, sonnet 3.7
Every. Fucking. Time.
The spa placements is oddly specific, strange, and it's only message is that "no doctors would trust this, so we don't even try" and that's not a good message for such a machine.
I'm not saying this cannot work.
I'm saying that without explaining how it will improve outcomes of certain conditions it's nothing.
It's like saying "I am building a vehicle that runs on alcohol and the seats will be covered by red silk" "what kind of vehicle?" "one with red silk seats."
The second one is debated. If a test with low sensitivity indicates other tests that are invasive or causes considerable radiation exposure than depending on many other factors it can be even negative to public health. See Theranos with its false positive cancer diagnoses.
It can be legitimate, I don't know the details, but this blogpost is written like a low tier grift or a joke that doesn't even try to look legitimate.