Since the training data is basically the “source code” via this reasoning, does that mean there are trade secret claims (source code theft) to be made against GenAI companies? The argument would be that sites like SO, Reddit, etc implicitly grant permission to read the “code”, but not to “run in production”.
Many attempts in the past to make the making of software “simpler” but none panned out … would be interesting to see what happens this time around with this new tool … :-)
Spot on. But I think there is also going to be a massive impact on hiring and teams.
This ladies and gents, is where we won't need to hire people. Software 2.0 could use the same infra, rely on 100 different datasets and produce 100 neural nets instead of a 100 teams. Unfortunately they wouldn't have the domain knowledge to understand most of the why or the domain.
More examples of 1.0 that will be things of the past and could transition to 2.0:
- people, entire teams or even companies writing branching logic for IVR (dialers)
- almost all manually curated nocode drag and drop decision trees can be replaced with a NN.
- ranking & sorting algorithms
- prediction
- ux (what to show first)
- product recommendations
- a/b testing
- workflows with a bunch of ifs and branches
2.0 when you think about it is truly nocode. That said we are having a tough time visualising large neural networks. Radar charts work for < 10 weights. Tables are boring. Its time for innovating in visualising 2.0 as well.
Like always this ignores that the space programs live in is not only high dimensional but hyperbolic, meaning the search space for even trivial programs explodes much faster than you'd expect.
Neural networks are best thought of as oracles, simple single purpose tools which can do amazing things impossible with regular software in human time frames, but ultimately limited by the fact there isn't enough bits in the universe to encode enough training data even for medium difficulty tasks.
Put another way, all the LLMs have been trained on every book ever written. If we put a human being on an island with every book ever written and effective immortality we'd have someone a lot more interesting than gpt4 when we came back in a few million years or so.
9 comments
[ 5.7 ms ] story [ 33.6 ms ] threadTime to write an IDE for developers 2.0 (dataset labelers) and then lay down and die.
This ladies and gents, is where we won't need to hire people. Software 2.0 could use the same infra, rely on 100 different datasets and produce 100 neural nets instead of a 100 teams. Unfortunately they wouldn't have the domain knowledge to understand most of the why or the domain.
More examples of 1.0 that will be things of the past and could transition to 2.0:
2.0 when you think about it is truly nocode. That said we are having a tough time visualising large neural networks. Radar charts work for < 10 weights. Tables are boring. Its time for innovating in visualising 2.0 as well.Neural networks are best thought of as oracles, simple single purpose tools which can do amazing things impossible with regular software in human time frames, but ultimately limited by the fact there isn't enough bits in the universe to encode enough training data even for medium difficulty tasks.
Put another way, all the LLMs have been trained on every book ever written. If we put a human being on an island with every book ever written and effective immortality we'd have someone a lot more interesting than gpt4 when we came back in a few million years or so.