@simonw We use fine-tuned models to achieve super low time to first token at https://t.co/Nh0n815OIj
Naturally, the larger models result in higher latency so we squeeze more juice out of smaller models.
@profplum99 I don't get the incentive for @getbasiccapital though. Isn't the average person going to dump the loan into an index fund? If so, why is Basic Capital better off loaning it to me than putting into an index fund?
Presumably they are making more money than an index fund meaning...
@karpathy Is there any research on system prompts in latent space? I've been curious if the KV cache of a system prompt can be treated as a learned parameter tuned with RL
@swaaanson Honest question, what if I just don't do those things instead? I don't mean that I currently don't, but rather are those small things really that valuable? Get rid of them, they weren't that important anyways. That seems like a higher return.
alright I'll take the bait and give a countertake you probably haven't heard before
I dug into your not-tweet-length argument on this on the @freeCodeCamp podcast and if I understood correctly it's along the lines of "copilot and friends make you ineffective at learning". this is true, when coding with copilot you are learning copilot, not the underlying tool. but what's wrong with that?
yes, the calculator argument is trite, but more subtly nobody ever said we had to be good at manual arithmetic. my friend noted when google maps was down once that he was absolutely lost and couldn't navigate, and that navigation has suddenly become an atrophied skill, very similar to your observation about copilot
the problem is that nobody ever said you needed to know how to navigate to survive. in fact, for the vast majority of people it's a net gain that their phone absolves them of doing something reasonably complicated
so yeah, copilot atrophies your ability to learn coding but nobody ever said the goal of the job was to be an effective go or lua coder. the goal of the job is to spit out whatever product the boss wants. it could be in assembly for all they care
so if it makes you more effective at spitting things out, then github copilot makes you a better dev. it has nothing to do with speed
anyways I signed up for twitter premium to post this so hope you enjoy