Artifice of Intelligence

We live in a time of incredible technological innovation, there’s the iPhone from 2007…

…

Incredible technological innovation…

AI doesn’t work

Now despite the fact that we don’t have flying cars, VR is a stupid waste of time, and there is no computer magic to reverse global warming, we do seem to be scared of all the technological developments. Boston Dynamics robots are like the terminator, GPT-3 can write code, and Tesla has been promising full self-driving by the end of the year, every year, for the past decade. 

But, while we dream of Artificial Intelligence, it’s pretty far from our current reality of machine-learning-driven statistics.

In the 50’s some guy made a mechanical model of a biological neuron. A big old fucking computer literally motors would move to simulate the neuron firing.

And it could do linear binary classification: Just by giving it examples, it can learn to separate stuff into 2 categories. So you can say limes are green and small and lemons are yellow and larger.

–RESHUFFLE–

Actually the usual example is apples and oranges. Which is very clever, comparing apples to oranges. But I’ve changed it to other fruit because. It’s a fucking stretch. It doesn’t really work as an example, and it’s not really a joke either.

So anyway after you’ve shown it enough examples of limes being green and small and lemons being yellow and larger, it will try to create a function that will attempt to separate that data into those two categories. So eventually, you could say I’ve got something green and small here, and it will tell you you’ve got a lime.

And this was super exciting at the time, right? Full AI was just around the corner. And then obviously, it didn’t deliver. Inaccurately separating things into two categories when you already know the categories isn’t particularly useful. 

But um, they knew how it could be improved. You’d make a network of multiple perceptrons feeding into each other, and run backwards through the network as well as forwards, and use unfathomably large data sets. But obviously they were limited by the technology of the time.

Now, we’re no longer limited by the technology of that time. So we’re doing that stuff that they suggested. The end. Yeah, what we’re currently calling artificial intelligence is just that. Very basic pattern recognition.

OPTIONAL

But if you wanted to make a calculator without having to develop the underlying rules, you could just run a lot of examples through a network. Going:

“What’s 2+2?” And it’ll go “36” or whatever, it’s just random. And then you go “No it’s 4.”

And then it makes some adjustments so next time it sees the pattern “2+2” It categorises it’s answer to be less like 36 and more like 4.

Just over and over, with millions of examples, and you might get something kind of accurate. 

The exciting thing is, that’s kinda how we learn, right? They’re not teaching field theory in primary schools. But the thing is with any rules-based system it’s possible to infer the logic behind it. After you’ve show enough multiplication tables to a human, they’ll start to figure out the underlying rules.

Now the problem is, our machine learning calculator will never learn the underlying rules, the only rule it can learn is “nice patterns.”

With enough examples and time you can get it to memorise those examples, and improve the chance that it will guess other answers correctly. But it still has no idea what it’s saying and all we can do is chip diminishing percentages off of its failure rate.

That calculator will never replace an actual calculator, and accuracy in machine learning is still subhuman across the board.

That’s not to say it’s entirely useless. Of course, “nice patterns” can do a lot of things, like identify patterns, or create patterns. That’s why it’s really pretty good at image manipulation and knock off dril tweets. But for something with actual internal logic, like math, or thought, or being regular, it fucking sucks.

Pattern recognition really is the solution to everything, and nothing. It’s the solution to everything because it will always give you an answer, and nothing, because it’s never right. One of the worst human impulses. Prioritising pattern recognition over fact is why we stereotype, why we gamble, and why the girlies love idrlabs.com. Hows that for a fucking punchline, no way in hell you were expecting me to say that.

Now when the perceptron failed to be “full artificial intelligence” it lead to the first AI winter. Research in machine learning is punctuated by AI winters: hype builds, far outstrips the technological reality, then all of a sudden people realise that, so hype, massively, drops away. Only to start all over again. As the fundamental technologies remain unchanged and we find out over and over that no, the computer cannot learn to love.

So progress in AI has somewhat stalled, it’s a very popular research field with hundreds of papers coming out every day. But no one is looking at fundamentally new approaches. It’s such a popular field because you can put a new wad of data into an existing algorithm, and that’s it, you’ve just done a research paper. And that’s most innovation in AI right now, just more data, more data, more data.

OpenAI thinks that more data is the solution because as they add exponentially more data to GPT, it gets better at replicating English sentences. It seems to improve at a linear rate… Because you’re adding exponentially more sentences for it to copy.

Another way of thinking about that, is as there’s a linear increase in the amount of sentences, the improvement exponentially decreases. We don’t have an infinite supply of sentences, at this rate, they’ll soon require more examples of sentences than have ever existed, and more processing power than has ever existed.

But right now, that’s all we’ve got, just add more data, and hope improvement happens.

The best of the best right now, DALL-E 2. It’s reached a near-photorealistic aesthetic. And this is probably the perfect use-case for a neural network, in theory this is the thing that it’s best at. Producing copies of patterns it’s seen with absolutely no logic behind it. And while it is cool, it’s also kind of dogshit. Everything it produces is just a copy of its input data. Based on pattern recognition and noise. 

It has that aesthetic sense but if you ask it for a builder it’ll show you a white guy with a hard hat. A wedding is a white man and white woman holding hands. CEO is a white guy, and “a CEO” is an asian man. Whatever patterns it’s seen the most of, it will replicate, and that’s all it can do.

Forgetting about the ethics for a second, since they seem to have. It shows us that these algorithms cannot create anything new.

Like “This person does not exist.” It’s trained on a massive dataset of faces, and it tries to generate a new image based on the patterns it’s learned. Now, a lot of the time, that person does exist, it’s reproducing an image of a person, to the point that you can recognise them from the training data.

So our current artificial intelligence is just copies of the past with no intelligence. And it cannot present anything new.

No one’s got the rights to these images by the way, this isn’t a replacement for art it’s a copyright risk.

Now Tesla has the biggest, loudest dreams of Artificial Intelligence, which is interesting because they’ve never invested in AI research. Instead they rely on combining existing technologies and publicly available research. Elon Musk says that Tesla will have full AI Self-driving this year, as he has done every year for the past 8 years.

Full-self driving is what’s called level 5 automation. For reference, the best self-driving out there right now is Google’s. They have a truck that only drives one route, with cameras and sensors all around it, and has to be remotely assisted, but it doesn’t technically need to have a driver inside. And that’s at level 3.5.

Tesla is currently at level 2 and their developers don’t think that’s gonna be changing anytime soon.

But it’s in use and kind of working. The go to example for AI.

Now, there’s two halves to the self-driving system. There’s observation, which is the eyes: detecting what’s around you, then there’s driving, which is deciding when to accelerate and turn.

The observation is done by machine learning algorithms (visually demonstrate), then the driving is written in traditional explicit code. C++. “If we detect a green arrow and no oncoming traffic, accelerate to speed and turn.” That sort of thing. Actual manually written out rules for every possibility. Which is insane, the idea of doing that caused one of those previous AI winters, they we’re like “What? We have to manually write out rules for every possibility? No way, that’s insane.” So now we’re just doing that.

Now on the detection side, for every object you could encounter while driving they make a detector, and train the detector on a dataset of thousands of images of that object. But then every object has variants, you wouldn’t have trouble differentiating a stop sign from a stop sign with an unusual amount of bird shit on it, but a computer will. So they have to collect a massive data set of stop signs with an unusual amount of bird shit on them, and train that variant.

And all these variants need to have equal representation in the dataset, or else it’ll incorrectly prioritise stuff it’s seen more. Like they struggle with orange traffic lights, you don’t see orange traffic lights as often as red or green ones. Now, incorrectly thinking that the moon is an orange light is a bit of fun. But these biases in the input data mean that, due to the demographic makeup of areas where Teslas are driven, self-driving is more likely to hit a black person than a white person.

Elon Musk arbitrarily decided to remove their depth sensors, so they have no actual depth detection, everything is extrapolated from the low-res cameras on the car. And without actual depth information, their lane detection often gets thrown off by motorway bumpers, and the car will actually steer people into a crash.

So historically Tesla’s had an unusually high fatality rate when it comes to driving full speed straight into a motorway bumper. But they’ve always denied that any of their fatal crashes have involved autopilot.

Now just within the last year they’ve started to be under actual formal investigation, and right away, it turns out, a bunch of those crashes which they’d denied were caused by autopilot.

And their current plan is to work day-in-day out, to manually factor in every wild, never-before-seen possibility, and constantly re-train their classifiers from scratch. The problem with wild, never-before-seen possibilities is they’re famously hard to account for.

But they’re trying anyway. Every car is constantly providing them new data, and they have an insane amount of underpaid workers processing it. And improving the dataset is basically all they do.

So last year, comparing their training data to their testing data, they’d reached a stop sign detection rate of 99.51%. Which doesn’t sound bad, but the human equivalent is, you’re looking directly at a stop sign, can you tell you’re looking at a stop sign? You should never be answering no. We should, at the very least, be looking at a lot more nines in that percentage.

So both of these systems are sort of working but require a literally infinite process of manual coding and massaging of datasets to ever truly work. And they’re stuck with the tech that they put in those cars 8 years ago. And they’re not looking at alternative solutions, they’ve never invested in AI research.

And that’s not unique to Tesla. All the tech companies are pitching a vision of the future, while not investing in that future at all. They’re fundamentally not interested in advancing the technology. They’re interested in repackaging the old technology. Actual progress is a threat to their dominance. While pretending that it’s happening improves their image.

But look, these things will happen, the next text generation algorithm is gonna be that much better, the next image generation algorithm is gonna be that much better, they’ll keep chipping away at that self-driving kill count. But our current technologies aren’t going to magically evolve into full artificial intelligence.

What we call AI is basically spreadsheets. Weights, functions and a fuckton of data processed by a fuckton of workers. It can recognise patterns, but that’s it. And there’s interesting applications for that, you can even apply it to areas where it doesn’t have an application.

But we’re not on the brink of major technological development here. We’re in the middle of some fiddling at the edges of what came before.

Comparatively, we’re not in an age of technological innovation. Just like the algorithms, we’re stuck repeating the aesthetics of the past.