Frontier models require frontier speculators. Our team moved mountains to achieve 460 tok/s with our custom DFlash model.
Try it day 0 at https://t.co/8vwHM10ox5.
Today we're announcing our Series C funding: $355M at a $4.65B valuation, led by some great investors @generalcatalyst and @Redpoint.
We've had insane growth in the last year, but we're still very early. So proud of the team and what we have built so far!
@eatonphil this is also a problem with using software written by australian tech companies at an australian tech company ... eventually you hire some of the people who worked on that software ๐
@ulfjack would that require storing ActionResults twice? once keyed by the "true" key, second by the "shortcut" key? I also just found this: https://t.co/uL0b2KJACC via https://t.co/L20LcjVQOg.
just realised something about bazel that seems a bit obvious in hindsight. if you have a chain of targets A..Z, all cached remotely, you need to fetch A-Y before fetching Z. this can take a while if latency is bad. why don't we have enough info to fetch Z immediately? ...
theory: continuous delivery (vs. versioned releases) is better for users so everyone does it, but it's also MUCH HARDER. CD blends planning, building, fixing which feels chaotic. for me to be "left alone" I would have to have no operational responsibilities, which means no CD
@dwarkesh_sp@tylercowen if they are centrally planned and they are inefficient, how many more high quality large businesses could exist if we had a better mechanism than central planning for allocating resources within firms?
... to 100% of users instantly. Even if you have great test coverage, no sane person would want to do that. Why do we accept the same for developer tools? The root of the problem is that you don't invest in any release infra/process for tools in a monorepo ...
... while still keeping the merge experience fast and reliable. Ultimately you need to get things into the hands of your users, and have ways to quickly roll back bad releases. Most of the time the release process is "merge to master". This is akin to deploying new releases ...
I am a bit of a monorepo zealot, but a slightly niche challenge has dawned on me over the past few weeks: at scale, it's really challenging to safely make changes to developer tools that are checked into a monorepo. There are limits to what you can reasonably test in CI ...