@basedjensen vision is a different beast! Google has consistently excelled here for a while; openai has more or less caught up. Anthropic has lagged since sonnet 4.
@jxmnop Simple. Cumulative sources for a PhD thesis can easily hit millions of tokens. Including those in the training set, or chunking them and trying to anticipate whats relevant are hacks because we can't just dump in context.
@MishaTeplitskiy I'm a quarter of the way through Dark sun after loving the first book. The minutiae of espionage is getting to be too much at this point. I am impressed at the detail, but also fine not knowing if the transfer of a specific experiment happened in August or September 44.
@awesomekling I gave myself an 8 talking to a well known python consulting company a few years ago and got "can you point me to any of your conference talks?"
@lateinteraction I generally agree with you, but if someone is experimenting with training/steering the compiler llm or has a prompt-intensive algorithm, it's more difficult to use a 4.1
@NirantK Weave is their tracing product. Would be nice to use it too. Everyone is converging on the same tracing api, wonder how differentiated the rest of the experience is.
@finbarrtimbers What kind of specific scenario can you envision? I'm also optimistic, I worked on implementing Bayesian optimization algos for a while. There's many variations to adapt to new settings and analyzed pretty well with llm help. Caveat: results are workshop grade not breakthrough.