Software 1.0: Write an algorithm that has the right behavior. @karpathy’s Software 2.0: Optimize differentiable blob to have correct behavior. medium.com/@karpathy/softwar… Software 3.0: Figure out the right prompt to make your meta-learning language model have the right behavior? :P
7
62
2
310
Love the idea for Software 3.0 😂. Programming moving from curating datasets to curating prompts to make the meta learner "get" the task it's supposed to be doing. LOL 🤣👌
1
9
1
102
Replying to @karpathy @ch402
😂👍

1:24 AM · Jun 19, 2020

8
56
7
309
Replying to @karpathy @ch402
I actually did spend about an hour playing around with the OpenAI GPT-3 API over the weekend trying to get the davinci model to recommend me the most interesting movies (e.g. asking it for more rare movies, etc). Little did I know I was 3.0 programming.
1
4
1
68
I guess the ultimate Autopilot is giving a big multimodal GPT-10 the DMV handbook in the prompt concatenated with the last 10 seconds of sensor data and asking it very nicely to follow that 🤔
7
9
2
112
Replying to @karpathy @ch402
Andrej I am just wondering how much computing power is used on The FSD computer with everything
0
0
0
1
Replying to @karpathy @ch402
Andrej, can you hire me as a remote dataset labeler for AutoPilot please? 😀
0
0
0
0
Replying to @karpathy @ch402
Thanks for the work you and your team is doing. It is magical using the autopilot system. 👍
1
0
0
2
Replying to @karpathy @ch402
Wasn't ready for the design the prompt 🤣
0
0
0
0
Replying to @karpathy @ch402
When can we expect FSD coast-to-coast run?
0
0
0
0
Software 4.0: sit back and relax
0
0
0
4
Replying to @karpathy @ch402
Interesting to note the pattern in AI research advances: learning algorithms (backprop, sgd), to architectural design (resnet, relu), to initialization strategy (transfer learning, self-supervised), and perhaps ultimately to "prompt/input design" :)
0
0
0
0