How powerful is AI becoming?

How powerful is AI becoming? Does AI make technology companies stronger?

There is some truth in this, but at the same time ML also sees a lot of diffusion of abilities – there can be as much decentralization as centralization.

https://aidude.io/

First, what does it mean to say that machine learning is about data? 

Based on the academic culture of ML, the entire basic science is quickly published after its formation - almost everything new is a paper on which you can read and build. But what are they building? Well, in the past, software developers wrote logical steps ("rules") when they wanted to create a system for detecting something. To identify a cat in a picture, write rules for finding edges, fur, legs, eyes, pointed ears, etc., wall together and hope it will work. The problem is that while this works in theory, in practice it is more like trying to build a mechanical horse - which is theoretically possible, but establishing the required complexity is impractical. 

Statistics Engine:

We can not describe all the logical steps that we use to walk or recognize a cat. In machine learning, instead of writing rules, you pass examples (many of them) to a statistics engine, and the engine creates models that can differentiate. You give him 100,000 photos with the inscription "Cat" and 100,000 photos with the inscription "No cat", and the machine calculates the difference. ML replaces manually written logical steps with automatically determined data patterns and is much more effective for a very wide range of problems - simple demonstrations lie in computer vision, speech and speech, but the use cases are much broader. 

How much data you need is a moving target: 

There are research paths that make it possible to use smaller data sets, but at the moment (more) data is rapidly getting better and better. Ai

Here's also the question: if ML allows you to do new and important things, and ML is better, the more data you have, then it means that companies that are already big and have a lot of data are getting stronger. 

How far?

 How far does the winner-take-all effect go? It is easy to imagine a positive cycle that strengthens the winners: "more data = accurate models = better product = more users = more data". From here it's a simple step from statements like "Google / Facebook / Amazon owns all the data" or "China owns all the data" – fear that the most powerful technology companies will become more powerful and the most populous countries will become stronger.“ Attitude to the central use of data.

Well, something like that:

First, while you need a lot of data for machine learning, the data you are using is very specific to the problem you are trying to solve. GE has a lot of telemetry data from gas turbines, Google has a lot of search data, and Amex has a lot of credit card fraud data. You can't use turbine data as an example to detect fraudulent transactions, and you can't use a web search to detect a gas turbine that is about to fail. 

However, ML is a generalizable technique – you can use it for fraud detection or face detection - but the applications you create with it are not generalizable. Everything you build can only do one thing. 

This is similar to all previous automation waves: 

Just as a washing machine washes clothes but does not wash or cook, chess programs are not recorded and machine learning translation systems cannot recognize cats. Both the application you created and the dataset you need are very specific to the task you are trying to solve (although this is a moving target and research continues to try to make learning more transferable between different datasets).

This means that machine learning:

Implementations will be very widespread. Google will not "own all the data" - Google will own all the Google data. Google will have relevant search results, GE will have better motor telemetry, Vodafone will have better analysis of call patterns and network planning, all different things developed by different companies. Google is better at being Google, but that doesn't mean it's good at anything else.

Next, one could argue that this simply means that the big players in each industry are getting stronger – Vodafone, General Electric and Amex each own "all the data" no matter what they do, thus forming a rivalry against the competition. 

But here it is more complicated: 

There are all sorts of interesting questions about who owns the data, how unique they are and how unique they are, and about the right aggregation and analysis points.

So: 

As an industrial company, do you keep your own data and build ML systems to analyze it (or do you pay a customer to do it for you)? Do you buy ready-made products from sellers who have been trained with other people's data? Do you mix in your data or the training derived from it there? 

Do suppliers need your data or do you already have enough? 

The answer will vary in different parts of your company, in different industries and in different use cases.

On the other hand, if you are starting a business to use ML to solve real problems, there are two basic data questions: 

How do you get your first data to train a model to attract your first customers, and how much data do you need? do you really need it? The second problem, of course, breaks down into many problems: will the problem be solved with a relatively small amount of data that you can easily get (but many competitors), or do you need more, which is difficult to get, and if network effects can also benefit from this, will the winner also take over all the dynamics? Does the product receive unlimited more data or is there an S-curve?

It depends on whether:

Some data is unique to a company or product or has strong proprietary advantages. GE engine telemetry may not be widely used to analyze Rolls-Royce engines, but if it were, you wouldn't share it. This may be an opportunity for starting a business, but it is also the place where many internal large IT and contractor projects of the company take place

Some data applies to use cases in many companies and even in many industries. "Kind of weird about this call" is probably a common analysis for all credit card companies – "The customer sounds annoyed" probably applies to everyone who has a call center. This is a "mixed" problem.

Many companies are founded here to solve problems for many companies or different industries, and there are network effects on the data here.

However, there are also cases in which the provider no longer needs the data for each incremental customer after a certain point in time – the product is already running.

In fact, a startup might see some of them, since machine learning is spreading in almost everything. 

Our portfolio company Everlaw manufactures legal discovery software: 

If you're suing someone and they send you a truckload of papers, that helps. Machine learning means that you can do a sentiment analysis for a million emails ("show me anxious emails") without having to train the model with data from the case, because the emotional examples to train the model are not required to come from that particular lawsuit (or a lawsuit). Conversely, you can also group your data.

 ("show me emails with this identity") 

and do not do this anywhere else. Drishti, another portfolio company, uses computer vision to inspect and analyze production lines - some of these functions are trained based on your data, others are not specific to your company at all, but cross-industry.

In one extreme case, I interviewed only one manufacturer of very large vehicles that uses machine learning to obtain accurate breakdown detectors. Of course, these are the training data.

(many, many examples of signals from flat tires and non-flat tires)

But it's not so difficult to get this data. It is a feature, not a moat.

I have also already said that there are two problems with ML startups:

 How do you get the data and how much do you need? But these are just technical questions: they also ask how they get to the market, what their addressable market is, how valuable the problem they are solving is for their customers, etc. However, there will soon be "AI" startups - these will be companies for industrial process analysis, companies for legal platforms or companies for sales optimization. 

In fact, the proliferation of machine learning does not mean that Google is stronger, but startups of all kinds can build things much faster than before with this state-of-the-art science.

How powerful is AI becoming?

Forces of AI:

An Introduction to 10 Key points 1 - How to define AI? 

The term artificial intelligence was introduced in 1956 by American researchers Marvin Minsky and John McCarthy during a summer research school at Dartmouth College (Hanover, New Hampshire, USA). Although there is no strict consensus definition, "AI" is commonly referred to as the ability of a machine to perform tasks that are usually performed by humans. 

However, this definition remains too inaccurate for experts who generally use these computer programs with a more specific vocabulary according to the type of algorithms used (for example, expert, hybrid, neural, adversarial, generative systems, etc.).) use case (conversational agent, face or character recognition system, recommendation system, etc.).). The legislator is also trying to reach a consensus on the definition of AI. 

For example, the definition of AI in the proposal for the Regulation of AI Systems by the European Parliament ( IA Law ) has been constantly changing to distinguish this disruptive technology from traditional computer software.1. It is difficult for institutions to stabilize a taxonomy and definitions to designate systems with rapidly changing capabilities. 

AI Systems:

For example, one of the definitions of the OECD2 does not mention generative systems that are able to generate new text or image content - in contrast to so–called predictive AI systems that specialize in pattern recognition or curve fitting to make predictions.

2 - Some important dates in the history of AI: 

Since the late 1940s, the history of AI has interwoven periods of technological brilliance and disinterest. Periods of disinterest are called "KI-winters". Thus, the glorious narrative around AI in the mid-1970s and late 1980s did not reflect the concrete achievements of the technology actually used in the industry at that time, which led to a drastic decrease in funding and a general disinterest in technology.

Although there is no strict consensus definition, "AI" is commonly referred to as the ability of a machine to perform tasks that are usually performed by humans. As in the literature, technological heydays are accompanied by a number of streams of ideas on how best to develop powerful AI systems. This is how cybernetics and the adaptive machine were created in the years 1940-1950. 

System:

In particular, Wiener applies control theory and dynamical systems to develop anti-aircraft systems to correct trajectory prediction errors in real time and direct missiles. These ideas will be taken up much later to develop deep neural systems.

https://aidude.io/