“It has been 2 years of the Gemini program and GDM. We’ve come a good distance in that point with many efforts we must always really feel very pleased with. On the similar time competitors has accelerated immensely and the ultimate race to AGI is afoot. I believe we have now all of the substances to win this race however we’re going to should turbocharge our efforts.
Code issues most — AGI will occur with takeoff, when the Al improves itself. Most likely initially will probably be with a number of human assist so crucial is our code efficiency. Moreover this must work on our personal 1p code. We’ve to be essentially the most environment friendly coder and Al scientists on the planet through the use of our personal Al.
Productiveness — In my expertise about 60 hours per week is the candy spot of productiveness. Some of us put in much more however can burn out or lose creativity. Quite a lot of of us work lower than 60 hours and a small quantity put within the naked minimal to get by. This final group isn’t solely unproductive but additionally might be extremely demoralizing to everybody else.
Location — You will need to work within the workplace as a result of bodily being collectively is way more practical for communication than gve and many others. And, due to this fact you want to be bodily colocated with others engaged on the identical factor. We have to decrease reporting traces throughout nations, cities, and buildings. I like to recommend being within the workplace not less than each week day.
Group — We have to have clear accountability and group with excessive functioning teams with shared administration and know-how management.
Simplicity — Lets use easy options the place we are able to. Eg if prompting works, simply try this, don’t posttrain a separate mannequin. No pointless technical complexities (akin to lora). Ideally we are going to really have one recipe and one mannequin which may merely be prompted for various makes use of.
Excellence — whether or not it’s an eval or an information supply or a dashboard or a message in an inside Ul, please be sure all of them work and all are good.
Velocity — we’d like our merchandise, fashions, inside instruments to be quick. Can’t wait 20 minutes to run a little bit of python on borg.
Iterate at small scale — we’d like a number of concepts that we are able to check rapidly. One of the simplest ways to do that is small scale experiments till you possibly can ramp up and hopefully see growing benefit at scale. This is a superb validation. Working an excessive amount of at simply massive scale has a behavior of minor tweaking and overfitting to evals, checkpoint sniping, and many others. We want actual wins that scale.
No punting — we are able to’t preserve constructing nanny merchandise. Our merchandise are overrun with filters and punts of varied sorts. We want succesful merchandise and [to] belief our customers.“
