Hype is going to cause an AI Winter
Funding is going to start expecting results to match the hype
It feels like fairly regularly we see a progression in AI that is touted as being a huge leap forwards in terms of machines truly becoming intelligent. I’m going to pick on one such piece here and explain exactly why this sort of attitude is storing up problems for the future of AI.
As someone who played Diplomacy, let me say… this is cool. This is a hard problem, a kudos for solving it. And the lead of the team did a great team explaining what they set out to do and how they did it
So, again, this is cool, this is a hard and very specific problem that this team have taken 3 years to solve. They must have dug into a huge amount of elements around how people play the game, understanding what it takes, understanding what strategies influence people, and then tying that up to something that could communicate that coherently.
This is cool, it is also specific.
This is something Gary Marcus called out, as being absolutely key in this way of creating models.
I’m an engineer, not a big brain, but in simple terms… people learned a cubic shit-ton about Diplomacy and how people play it, encoded that into some very smart models which addressed multiple aspects of the game, then linked that together with pieces that could build a strategy from that, and then added natural language on top which could help deliver against that strategy.
That is impressive, and a lot of work.
Then the wording gets woolly
So something cool has been done, for a specific game, where intents and negotiation is important.
Let’s just look at that first statement:
“CICERO, an AI agent that can negotiate and cooperates with people”
followed by
“It is the first AI system to achieve human-level performance in the popular strategy game Diplomacy.”
The implications of these two statements, from a leader in the field, is that CICERO is an AI agent that can negotiate and cooperate, and that an EXAMPLE of it’s achievements is what it has done playing Diplomacy.
This is not the same as what the researcher said, they said that their intent was to create an AI which could play Diplomacy and negotiate and cooperate with people around the game of Diplomacy.
I want to be clear, again, I think that CICERO is super cool and I can’t wait to hear more about what they did and what learnings we can take from it. However one thing I am not taking from the annoucement is that there is now an AI agent who could negotiate with people on anything other than the game of Diplomacy.
Under promise, Over deliver— not in AI you don’t
AI has lots of really great uses today. The thing is, a lot of them aren’t really very sexy, and they are certainly a million miles away from requiring some sort of fully autonomous sentient AI bot.
Sadly however this last generation of AI has been funded through insane promises to the markets, which have fallen short. In times of easy money that was fine, but we are seeing it coming how to roost.
Driverless cars, hype to reality
It is nearly 2022, and Driverless cars have already hit their winter. Remember in 2016 when it would be there by 2017?
And how is that going?
Well the latest beta was just launched
And elsewhere things aren’t better, if we take L4 as being a reasonable level of saying we have autonomous driving then here is the list of current cars that offer this as production supported options:
That is all of them. There are really good reasons why this is the case and a big part is the over-promising on what can be done, so people keep shooting for the stars and missing, when we could aim for the moon and get there.
So the result is people abandon the pursuit, they’ve been promised something that cannot be achieved, and when it can’t be delivered they aren’t overly interested in stepping back to something else.
LLMs and art generators aren’t Sentient, however much it gets implied
Large Language Models and image generators are probably the greatest area that has the least application to normal business, they’re also the things that continually get suggested as sentient.
No, it isn’t sentient, and this isn’t new. Because people almost want to be convinced of sentience they can see it. People were doing this in the 1960s with things that aren’t as smart as fish-sticks.
This is something that AI researchers know, and yet…
“And today’s large language models (LLMs) — from OpenAI’s GPT-3 to Google’s PaLM to Facebook’s OPT — possess dazzling linguistic abilities. They can converse with nuance and depth on virtually any topic”
This is from Forbes and it is completely wrong. These LLMs cannot converse with nuance and depth on any topic, not even close. They can do some cool stuff, but could they communicate with nuance on a suicide helpline? With a doctor? With a computer scientist on architecture? What about on the VAR decisions being made in the current world cup? Nope. But for specific tasks that they have been trained and tuned for they can do a good job.
The hype needs funding for those marginal gains
Deep Learning is cool, there are super cool things to be done, but it isn’t free, and today we are seeing less and less gains while the computation requirements are exploding. Exploding way beyond the means of a normal business, or indeed most universities.
The scale on the left is exponential, this isn’t a slight up-tick in computing requirements, this is a massive uptick in computational requirements. If you want to understand more about just how mad this is getting have a read of the paper itself.
The challenge here is that there is a pursuit of greater and greater scale, and a blunt force approach which assumes that if you do enough learning on enough parameters then there is a point where it flips from being context specific to being generally applicable. There is no evidence that this is true.
The doubling requirement of these models is far outstripping Moore’s Law in terms of what processors are able to deliver. Which means that this processor enabled growth it is funding enabled growth. If the funding dries up, the growth collapses.
Over-marketing is hiding the real value
The risk in all of this is that when the funding dries up we will see a collapse of the market and a backlash because it failed to match up to ridiculous marketing statements. This would be a shame, but even today I see issues with the hype impacting what people actually aim to do.
An example of this is how most organizations, despite not having a fraction of the available compute, will set off on a road of deep learning in the belief that this represents the best way to drive value and that only by doing their own deep learning can they get a ‘better’ model. The reality is, and I think the CICERO example shows it, that there are significant number of different options you can use depending on the problem statement and the deep learning alone is unlikely to be the answer.
There is also a challenge that ‘boring’ challenges such as prescriptive maintenance, areas that are well understood around failure models, scheduling and skills identification which don’t have the hype around them. I’ve seen people look at whether they could use an LLM to create break/fix work orders, and then have to be nudged towards actually solving the problem of having break/fix work orders in the first place.
When the funding dries up for the hype and marketing, there is a real risk that companies will lose track of the value, assuming that because people aren’t creating bigger and bigger models that somehow AI has failed as an industry over all.
This new generation of large scale models do cool stuff, and they are adding more tools into the AI toolbox. But there are lots of tools in there, and the odds are you can solve an awful lot of problems without requiring these new tools, and the sort of investments required to take context specific solutions to play Diplomacy and help with contract negotiation in land leasing.
