Ask HN: What will the state of AI be in 10 years?
5 comments
I can offer a perspective as someone whose graduate studies in machine learning were during the "AI Winter", the time period after the first hype wave of NN and before the second wave we're currently in today.
The math progress has not been substantial, frankly. The very same ideas were there before. Sure, there are some new ideas, but they too are the result of the real fundamental change, which was the scale-up in compute and in memory.
So where I see a good analogy is were computing was in the age of the mainframe and these huge computers that Bill Gates and his chums were fiddling with in assembly, compared with personal computing.
ML is going to be everywhere, the way databases are everywhere. To an extent, it'll be an extension of SQL, in the sense that if SQL is about "counting stuff on data" then ML is a more sophisticated take on that idea. Where you'll have data, you'll have an engine that extracts stuff from it the ML way. I am guessing transfer learning will play a substantial part in this, but I'm less certain on that statement.
Just like with personal computers, our expectations of what you can "just do" will adjust. But all of that will not stem from progress in ML, just that current technology becomes more widespread. As far as the cutting edge of ML, judging by the sort of papers I come across, I think we're currently at the stage of doing more-of-the-same only bigger.
Again, I find history a really useful teacher in that sense. Take a trip to a science museum to see that people have always first taken any technology they've had to its utmost boundary, alongside the diminishing gains, before something new came along and the curve restarts. We're already in diminishing-gains territory, but it's hard to say how long that kind of momentum lasts.
The math progress has not been substantial, frankly. The very same ideas were there before. Sure, there are some new ideas, but they too are the result of the real fundamental change, which was the scale-up in compute and in memory.
So where I see a good analogy is were computing was in the age of the mainframe and these huge computers that Bill Gates and his chums were fiddling with in assembly, compared with personal computing.
ML is going to be everywhere, the way databases are everywhere. To an extent, it'll be an extension of SQL, in the sense that if SQL is about "counting stuff on data" then ML is a more sophisticated take on that idea. Where you'll have data, you'll have an engine that extracts stuff from it the ML way. I am guessing transfer learning will play a substantial part in this, but I'm less certain on that statement.
Just like with personal computers, our expectations of what you can "just do" will adjust. But all of that will not stem from progress in ML, just that current technology becomes more widespread. As far as the cutting edge of ML, judging by the sort of papers I come across, I think we're currently at the stage of doing more-of-the-same only bigger.
Again, I find history a really useful teacher in that sense. Take a trip to a science museum to see that people have always first taken any technology they've had to its utmost boundary, alongside the diminishing gains, before something new came along and the curve restarts. We're already in diminishing-gains territory, but it's hard to say how long that kind of momentum lasts.
I think we’ll have started integrating the three laws of robotics into all ai systems by that point. I know the Asimov books show lots of ways that can go wrong but it still makes me feel safer.
The only roadblock is I can’t think of any way to build in the three laws in such a way that they can’t easily be removed.
The only roadblock is I can’t think of any way to build in the three laws in such a way that they can’t easily be removed.
Asimov-level AI is still fiction and most likely remain that way for a long time.
I personally think that there is still enormous room for improvement in applying even modern cutting edge techniques to real world usecases and industry.
Also, I think interpretability and the introduction of better understood priors will lead to extremely large, sparse networks being used across multiple modalities in applications and software.
Also, I think interpretability and the introduction of better understood priors will lead to extremely large, sparse networks being used across multiple modalities in applications and software.
- AutoML will perform most of the model development [ This will occur within the next 5 years or even less].
- Custom chips will perform the inference and probably the training. For example, Tesla managed to create a chip, which is 10x cheaper than a GPU for training.
- There will be some breakthrough that will combine logic and stat. This would eventually lead to AGI.
- Custom chips will perform the inference and probably the training. For example, Tesla managed to create a chip, which is 10x cheaper than a GPU for training.
- There will be some breakthrough that will combine logic and stat. This would eventually lead to AGI.
> AutoML will perform most of the model development [ This will occur within the next 5 years or even less].
Similar to how languages that are 'designed for non-programmers' fared?
Similar to how languages that are 'designed for non-programmers' fared?
No. The current "no-code" tools which really have their origin in visual programming in the '80s are trying to convert icons to logic, which cannot work well. I.e. you can generate CRUD, but you cannot convert logic.
The ML dev process is an optimization process and not a decomposition process (I.e. it is stateless).
Instead, it is much more experimental (and hence declarative). I.e. you can define a search space, and substitute the data scientist time, with computation time.
This is the idea of "software 2.0"
https://www.youtube.com/watch?v=zywIvINSlaI
This is how Tesla operate today for its autopilot project.
The ML dev process is an optimization process and not a decomposition process (I.e. it is stateless).
Instead, it is much more experimental (and hence declarative). I.e. you can define a search space, and substitute the data scientist time, with computation time.
This is the idea of "software 2.0"
https://www.youtube.com/watch?v=zywIvINSlaI
This is how Tesla operate today for its autopilot project.
Tesla managed to create a chip, which is 10x cheaper than a GPU for training
Really? Where can I buy that chip?
Really? Where can I buy that chip?
Whoever leads in artificial intelligence in 2030 will rule the world until 2100
https://www.brookings.edu/blog/future-development/2020/01/17...
https://www.brookings.edu/blog/future-development/2020/01/17...
> It is useful to think of technical change as having come in four waves since the 1800s
> steam engine, electric power, information technology (IT), and artificial intelligence (AI).
Just no. AI has not come, it hasn't even started. What people call AI today is literally faster IT infrastructure in implementing math. The proper term for what we see today is machine learning and it is based purely on statistics.
> steam engine, electric power, information technology (IT), and artificial intelligence (AI).
Just no. AI has not come, it hasn't even started. What people call AI today is literally faster IT infrastructure in implementing math. The proper term for what we see today is machine learning and it is based purely on statistics.
I'm a graduate student studying ML and Theoretical Neuroscience, and I've been thinking about what the state of ML will be a decade from now.
What are your thoughts?