Hacker News new | past | comments | ask | show | jobs | submit login
The function that gives AI value is the ability to make predictions (forbes.com)
53 points by myinnerbanjo 8 months ago | hide | past | web | favorite | 39 comments

This article is weird. I work with ML (AI is overloaded term) and don't recognize myself in this article at all. It seems to be written for managers, politicians, or economists or something.

It is like stating: "The function that gives software value is the ability to create if-then statements." Both remotely true and meaningless.

Conflating analysis with predictive modeling, pretending self-driving cars are a thing of the last decade (and not fully operational since the 80s) and this:

> “So what’s going to happen is that these prediction machines are going to make predictions better and faster and cheaper, and when you do that, two things happen. The first is that we will do a lot more predicting. And the second is that we will think of new ways of doing things for problems where the missing bit was prediction.”

If using ML or DL qualifies as a subset of AI, then AI qualifies as a subset of software and IT. Turning above statement into:

> “So what’s going to happen is that these computer are going to run code better and faster and cheaper, and when you do that, two things happen. The first is that we will do a lot more coding. And the second is that we will think of new ways of doing things for problems where the missing bit was software.”

Then you are still correct, it is a safe bet, but you are correct about a very insignificant thing.

Yeah. Article summary: AI is about prediction. So if you want to use AI, reformulate your problem as predicting the answer.

Cheap dopamine hit of false comfort for someone distressed about not understanding AI. And also a book ad maybe.

statements like that typically come from best-selling author and keynote speaker. This kind of articles seem valuable to some audience, but it's probably something I can never understand.

Like the world cup finals that was shared some weeks ago? Man, I've been coding since I was a kid. I have a deep love for technology but the hype age we are living in, is just too much.

Bet on the World Cup using output from the same algorithms that got it wrong this time over 20 World Cup events, and you are likely to still come out way ahead. Real-world events have chaotic properties to them, but overall these predictions tend to be quite accurate.

Casinos allow you to bet on games like roulette or blackjack because they have a mathematical expectation - a prediction - that they will come out ahead overall. Would you say that this prediction is wrong because individual players sometimes win a lot of money?

You can’t look to a single event that differs from a prediction and use that as a basis for an argument that the entire model is flawed.

I was just at the international institute for forecasters conference on Colorado as an applied practitioner.

There was a lot of hype on how to best use ML/Neural-nets in time-series forecasting. Well, it was well-founded hype. It was hype by people who know there is potential there, but who also consider a 20% accuracy improvement over naive methods that were developed in the 1980s to be a cause of great success.

And even then, it doesn't always work. NN/ML has had true breakthrough success in classification type problems, and fitting situations of nonlinear high-fidelity. For example, self-driving cars.

But in terms of economics/demand/weather dynamics, which include much much less data, and often deal with more chaotic macro-patterns (i.e. Your data is a time-series of a few megabytes over years, rather than gigabytes over minutes from cameras), it offers much less.

20% is pretty significant actually, I've read papers where they report baseline beats in the 0.1% to 1% range.

Improvement of 99.0% to 99.2% is a 20% improvement.

a 20% reduction in mis-classification or a .2% absolute improvement in success, or a 0.02% relative improvement to success.

What methods are those?

Then you should know a prediction about a game with a ton of variance with many entrants is almost always wrong. But assuming you have a perfect model, simulated over millions of attempts, it's the most correct.

That's not just AI; the function that gives science value is the ability to make predictions. This is where the replication crisis came from -- when nobody's judging you on the ability to make predictions, your predictions wind up being worthless.

The function of every decision and every input into a decision is to make predictions. The value of everything anything does is a judged by how well it serves a purpose of some kind to someone.

If a tree falls in the forest and no one is around to hear it, does it make a sound? Outside of Occam's razor and (measurable) knock-on effects, it's impossible to judge in any meaningful way.

I'm with you, it not just AI, it is basic math and all science that makes predictions, and that is why we pursue them at all. This article strikes me as an article/advertistment for a "Dick and Jane" first reader type of explainer, for business people to alleviate some of that hype. Maybe that is what the book they are hawking is about...

Or to optimize, or to replace human operators, or to entertain, or to create....

All of which can be reformulated as predictions:

* Predict which strategy will arise given a metric to optimize

* Predict the next action a human operator would perform here

* Predict which action yields the most likes / smiles / upvotes

* Predict which output will have the most citations if it was an article in a scientific journal

Your remark isn’t a rebuttal but a reaffirmation. You have fallen prey to the bias of not thinking in sufficiently high generality.

True but vacuously so. Any system that has output can be rephrased the way you just phrased it:

* Photoshop predicts the output of a given a set of buttons, filters, UI states, etc.

* Your car's steering system predicts the wheel outputs given steering wheel inputs.

* The abs( ) operator predicts the absolute value of a number.


If we want to make a funny analogy with ML using my first example: this model -- i.e. the entire Photoshop software -- is trained in a slow and manual (not automated) iterative process against the cost function "whether Photoshop engineers and managers will ship it as the next version of Photoshop". It's repeatedly tested against it (or through whatever training algorithm its designers want, including waterfall. The training algorithm doesn't have to be good - but anything they use to design the software by definition is the training algorithm for the model - under this stretched analogy / way of thinking about it.)

I just mention this to show the absurdity of this way of thinking about it - like the entire Adobe campus is just one giant training algorithm for the "next version of Photoshop" model which predicts "what will the output of pressing these buttons be".

If you'd like a second example: any simple pocket calculator going back to 1970 is a system that "predicts the result of its operations and operands". The cost function is the happiness of the engineers who designed it, and a human is involved in the iterative method of training the model, whose cost function is the human's happiness with it. Kind of an absurd way of thinking about the system.

So while these (and anything else with an output) can be formulated as "predictions", my examples going back to 1970 aren't machine learning, since a human is involved in this training loop. So this sense of "prediction" is kind of specious.

Sure, you can call them all predictions but you don't gain any insight by doing so. And you lose OP's point, which I thought was insightful.

Ok that is clever. I take your point!

Actionable predictions are tricky though because acting on the prediction before the event happens influences the system.

For example, my AI says stock X will go up 10%. My act of buying stock X drives the price up, therefore reinforcing the prediction regardless of it's original accuracy. Or, if I predict stock X will go up 10% before the earnings can and don't tell anybody until after the event is over and don't act on the information, what's the point?

So? Just reformulate it in a way that allows you to apply the Kakutani fixed point theorem and call it a day. That’s what John Nash did.

Good luck finding a Nash Equilibrium in the stock market.

I could probably generalize: the value of every model is in its ability to make predictions.

This includes, and not limited to:

* mental models: if your mental model doesn't predict reality, then it probably needs to change

* financial models: if it doesn't forecast well, it is wrong

* even AI models :-)

I would add though: "All generalizations are dangerous, even this one." ― Alexandre Dumas

This is why, as a machine learning engineer, my skepticism about most AI / ML work is rooted in the sociology of predictions / forecasting.

Humans seldom care about actually improving forecasts beyond the level given by very simple models. We are drawn to a very, very tiny set of domains where people might care, like epidemiology or meteorology, while in the vast majority (investing, customer analytics, national security, energy, climate, politics) we totally don’t care. Advanced algorithms only help to the extent they offer marketing hype or drive recruiting.

Robin Hanson alread wrote a great summary about this:

< https://www.cato-unbound.org/2011/07/13/robin-hanson/who-car... >

I think things like computer vision would be more accurately labeled as "drawing conclusions" rather than "making predictions". And if you can get your prediction success rate close to 100% that seems more like drawing a conclusion as well.

I would say it's making a prediction because to me the phrase "drawing conclusions" implies a conscious mental model of information from which a result is established. Making a prediction is not necessarily reliant on these mental models - especially conscious ones - it's about making an inference based on available information.

Another way to put it is a person would look at a person and notice fur, eyes, paws, ears, and the specific shapes and colors of these things and conclude it's a cat. Take away a leg, or an ear, or have it half out of frame, and most likely a person would still recognize it as a cat. The idea of "cat" exists in the mind of the agent in this case, but a computer may predict cat only if the animal is fully in frame and not missing any parts. The machine is entirely reliant on features whereas a person is reliant on a mental model that has more elasticity in what it defines.

Computers can and do label objects missing most of their features. This is especially important in computer vision work for cars, as inaccurately labeling an arm and a head as "not human" could lead to tragedy.

Now, I am sure you know this, so I don't know why you chose to use an example that's inaccurate in practice.

I chose this example because, in practice, sometimes changing a single feature does ruin the prediction, especially in computer vision. Often, the systems are somewhat resilient to these kinds of errors, but often not also.

The fact that a computer can label an object missing many features does not imply that it cannot also make a mistake doing so. Like the Tesla that couldn't recognize a truck right in front of it.

Then there's Google's Deep Dream, which did silly things like think that all hammers had arms attached to them.

Then there's also this: http://www.evolvingai.org/fooling

and many other examples like it. I chose a simple example that would be maximally relatable and still accurate even with respect to state of the art algorithms and datasets with billions of samples.

There exists some way to state this formally that shows these formulations are isomorphic, like a statistical Church-Turing thesis.

Does a tree that falls alone in the forest make a sound? Potato, potatoe. shrug

Show me!

Wouldn't you say that a conclusion is a prediction about the results of further investigations?

Say you have a photo, and your image tagger says it's a photo of an apple. What that really means is "if I were to go to the place and time where this photo was taken, there would be an apple there in front of the camera".

hmm. This might just be a clarification of terms, but couldn't you say that 'predicting' something is merely a type of conclusion where the information used in the process of reasoning is incomplete or noisy?

So all predictions are conclusions but not all conclusions are predictions. I.e, a deductive argument comes to a conclusion, but, as we know, inferential methods aren't deductive.

Humans are not predicting what a photo depicts when they look at it. They more or less immediately come to a definite conclusion.

No, we are absolutely predicting when we are looking at a photo; firstly, we really do not see the ‘whole thing’, our eyes make quick, stereotyped (cough predictive cough) movements to a few spots on the image and mentally imagine the rest. Secondly, that’s assuming we recognized something in the image, if it’s unclear, like a grainy photo or lots of shadows, we start to make informed guesses about what could be there.

All of our sensory systems work this way - we use an enormous amount of context to bubble up predictive categories and the collect marginal amounts of information until one prediction dominates over all others.

When one doesn’t, we get gestalt like illusions; like the cubes that are projected both forwards or backwards - or dresses that are blue and black and white and gold - or laurel and yanni in perfect harmony, etc.

Some people disagree: https://www.quantamagazine.org/to-make-sense-of-the-present-...

This is a pop-sci article, so take it with a grain of salt, but there is a growing body of research that attempts to frame many cognitive processes as inherently predictive (inferential?). I think this view arises pretty naturally when you start thinking about us having a 'model' of reality that we is updated based on sensory information, but I digress.

I think I see where you're coming from.

A human might not be 'predicting', since the term has some kind of temporal element connotation. But we could say that it's 'inferring' what the photo represents?

Ya that's what I was basing my comment on mostly. I do understand your parents reasoning though, although it is comming from a more logical perspective than I was.

I was commenting from the perspective of an analogy from personal experience of my own conscious.

At the root of this article is a complete lack of insight into what it means to compare apples with apples and oranges with oranges.

One of the points made by Mr Gans is: When I’m going to catch a ball, I predict the physics of where it’s going to end up. I have to do a lot of other things to catch the ball, but one of the things I do is make that prediction.

You do yourself a disservice sir - there is no AI on earth that could possibly match your ability to determine where a ball will land and work out how to catch it (one hand or two, over or under), whilst teetering on a pair of legs or perhaps diving. You may also be about to land in water and be working out how to deal with that as well at the same time. If you are diving you will also be making some horrifically complicated calculations that will ensure that you don't snap your neck or ribs and land in one piece. You may do that whilst daydreaming about something else.

"AI" is making some wonderful advances. Doing most of the stuff that we do routinely is not one of them. For starters an "AI" doesn't have a body!

What about optimisation : everything is known, the challenge is the optimal distribution...

I remember when fusion was only 25 years away. I predict the same for AI.

ctrl-f "softmax"; 0 results

Applications are open for YC Summer 2019

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact