AI Waking Up is Closer Than It Appears
Artificial Intelligence (AI) could “wake up”: it could achieve human level intelligence. The implications are profound.
This could happen sooner than many expect, because of:
- Money going into chips which drive AI,
- Money going into AI directly,
- Decentralized computing is coming of age, and
- There are no fundamental roadblocks.
Let’s explore each.
1. Massive Money Into Chips Drives AI
Money drives better compute substrates; and better substrates means better AI.
Money drives better substrates. This is about Moore’s Law. The $300B semiconductor industry’s core thesis is: spend more money, which improves performance of the silicon compute substrate, which opens new applications and grows the market. The cycle repeats. It’s worked for 50 years, from the space program to enterprise mainframes to PCs to smartphones. If someone tells you that Moore’s Law is dying, be skeptical.
Better substrates means better AI. In AI, researchers have often found that to solve harder problems, we often don’t need fancier algorithms, we just need more compute resources (processing, bandwidth, storage / volume of data). There are many examples of this. In his first book on Genetic Programming (GP), John Koza used a single algorithm to solve a broad variety of AI problems that had been attacked piecemeal by various AI algorithms in the preceding decades. It was ostentatiously wasteful of resources, but Koza didn’t care: GP solved the problems without custom algorithm intervention.
Here’s another example. Today’s deep nets, which have been used to crack some previously-unsolvable problems, are not much different than the multilayer perceptrons of the late 80s, which aren’t much different than the nonlinear perceptrons of the early 60s. Schmidhuber’s team used 80s tech to get competitive results with recent deep nets. They simply had more compute power than was available in the 80s. The algorithms are better, yes, but the hardware that much more so.
One final example: Norvig and others at Google got better results on AI problems by simply using more data. And in fact the more the data, the simpler the algorithms. While AI researchers including myself love to invent cool new algorithms, often all it really takes is compute power.
2. Massive Money Into AI Directly
There’s a lot of money going into AI directly, which catalyzes AI progress. The main driving force is to understand you better, to sell you more ads. This is precisely Google’s business model: the better Google can model how you think, the better the response to your queries, the more ads they sell. Larry Page has said as much, though it’s certainly not something Google likes to emphasize!
“Artificial intelligence would be the ultimate version of Google. The ultimate search engine that would understand everything on the Web. It would understand exactly what you wanted, and it would give you the right thing. We’re nowhere near doing that now. However, we can get incrementally closer to that, and that is basically what we work on.” — Larry Page, 2000 [ref]
Google could easily justify its purchase of AI startup DeepMind for $500M: DeepMind tech improved Google’s search by a few percentages that translated into money that probably paid for itself in a year or two at most.
More money into AI is also the motivation behind everyone wanting to sell more ads: Facebook, Baidu, Microsoft, Twitter, and so on.
The money into AI means much faster progress into state-of-the-art techniques. The current fashion is deep nets, which have improved the UX for AI, making it easier for non-pros to get good results, and pros to get great results, for certain classes of problems. $ goes into other AI problems too; for example there’s a big push right now in unsupervised learning (learning “X” rather than a mapping from “X →y”).
There’s even interplay between money into chips, and money into AI: AI has become a sufficiently important market that companies are now building specialty chips for AI. Roughly speaking, you get a 10x in performance for every step from microprocessor → GPU → FPGA → digital ASIC → analog ASIC. (ASIC = application specific integrated circuit.) Hardware neural network experiments go back decades, though most of those were research efforts that were not commercially justifiable. But more recently, that’s changed. For several years, Nvidia has been optimizing their GPU hardware and software for deep learning. Even though shipping an ASIC on a modern process costs more than $50M, it’s finally worth it: several startups are doing it, in addition to internal efforts at large enterprises. Even Google recently announced a dedicated chip for their deep learning library (a Tensor Processing Unit, for TensorFlow).
3. Decentralized Computing is Coming of Age
Sparked by Bitcoin, we are in a new age of decentralized computing. We’re right at the beginning of it. We’re seeing waves of tech and hype starting with Bitcoin, on to blockchain, to smart contracts, to DAOs (Decentralized Autonomous Organizations), and finally to AI DAOs.
There’s real money going into this. Billions of venture capital dollars have been invested in Bitcoin, blockchain, and smart contracts startups. Even DAOs are getting money, most notably the recent $150M into The DAO (well, almost into).
DAOs catalyze AI waking up, because DAOs are an easy path for AIs to get real resources. AI DAOs are way more powerful — and scary — than AIs their own or DAOs on their own.
4. There Are No Fundamental Roadblocks
I’ve discussed the large financial incentives to make AI more powerful, via chips and directly on AI technology itself. So, the money is flowing. These are the drivers towards a day when AI might wake up. Of course, many say that’s not possible, citing many reasons. Two of the most common are that humans are uniquely creative, and that our brain/body is special. Let’s discuss each.
“Humans are uniquely creative.” This is dead wrong. There are plenty of examples of machine creativity.
“Our brains and bodies are unique.” This is dead wrong too. A human brain + body is just a dynamical system that can be instantiated in one of many substrates. Fortunately, because our brains & bodies are just dynamical systems, they could be in other substrates too. More on this later.
Conclusion
AI could “wake up” much sooner than we expect. Money talks, and humans aren’t as unique as we rationalize.
