[Animats]

The distance between "sort of works" and "works" for AI is considerable. Not infinite.

Look at self-driving cars. The first tries were in the late 1950s, with GM's Firebird 3, guided by wires in the road. By the 1980s, the first self-driving vehicles were moving around CMU, very slowly. By the early 1990s, experimental highway driving had been demoed. In the early 2000s, we had the DARPA Grand Challenge, which had off-road driving on empty roads working. Then there were a few experimental self-driving cars that sort of worked on general roads. Many startups, most went bust.

Today you can take a driverless cab in San Francisco. 64 years since GM's Firebird III. (Which still exists, in driveable condition, in GM's in-house collection.)

It may take a while to get from GPT-3 to Microsoft Middle Manager 2.0. But the path is clear now.

[logifail]

> The distance between "sort of works" and "works" for AI is considerable. Not infinite. > Today you can take a driverless cab in San Francisco [...]

From the outside, it sure does look like driverless is still firmly at "sort of works":

"After California regulators approved the expansion of driverless taxi services in San Francisco earlier this month, it took only a little more than 24 hours for a series of events to begin that seemed to justify the taxis’ detractors.

The day after the vote, 10 autonomous vehicles operated by Cruise, a subsidiary of General Motors, abruptly stopped functioning in the middle of a busy street in the North Beach neighborhood of San Francisco. Posts to social media showed the cars jammed up, their hazard lights flashing, blocking traffic for 15 minutes.

A few days later, another Cruise vehicle drove into a paving project in the Western Addition and got stuck in freshly poured concrete.

And then last week, a Cruise car collided with a fire truck in the city, injuring a passenger in the car.

So it was that last Friday Cruise agreed to a request from the California Department of Motor Vehicles to cut in half the number of vehicles it operated in San Francisco, even though regulatory approval for more remained in place. The company, which has had 400 driverless vehicles operating in the city, will now have no more than 50 cars running during the day and 150 at night."[0]

[0] https://www.nytimes.com/2023/08/22/us/california-autonomous-...

[emptysongglass]

That's Cruise. Waymo has driven 1 million miles as of last January with only two incidents that have met the government's reporting criteria and no injuries. Those stats are impressive.

https://waymo.com/blog/2023/02/first-million-rider-only-mile...

[Animats]

Cruise started as a "fake it til you make it" operation. The tradition continues.

[qsort]

Does Microsoft Middle Manager 2.0 stop working in the presence of traffic cones?

More seriously, the distance between "sort of works" and "works" might not be infinite, but it most likely involves fundamentally unpredictable future developments of the current technology. There is no straight line of incremental improvements that gets us there.

It's fairly straightforward to imagine that if you have a 4.77mhz CPU and 64kb of RAM you will soon have a 3ghz CPU and 64gb RAM.

Bigger number going brrr is no guarantee of anything, here, so much so that models using a fraction of GPT4's resources are somewhat competitive.

By all means continue developing the technology, but claims that we are within arms' reach of X for disparate values of X are not exactly supported by anything.

[vineyardmike]

> Does Microsoft Middle Manager 2.0 stop working in the presence of traffic cones?

I think if you put a traffic cone on my IRL managers head she’d stop working too. Maybe a bit different… but maybe not.

I don’t think the poster was saying we’re within arms reach of it, but that there’s a path that takes us from here to there. A MSFT AI manager obviously wouldn’t behave exactly like a real human, but a tool that aggregates and summarizes information from many reports for a high level manager, and helps negotiate priorities is something potentially doable with some prompt engineering and advancements in models.

I respect many of my past managers, and some of them were great mentors and materially improved my life but 50+% of the value managers provide to an organization can be done with a bit of glue between GPT-x and Jira. It’d free up a lot of their time for the remaining 50% too.

[drawkbox]

> Does Microsoft Middle Manager 2.0 stop working in the presence of traffic cones?

That could be a benefit. Imagine being able to go dev shields up with just cones around your desk.

When it is time to innovate, surround yourself with cones.

"You haven't checked on your 'resources' in an hour, how is resource #4233 (Johnson) doing on that new feature?" -- funding borg consultcult front man

"I don't know, his cone shields are up, we can't get our needed hourly metrics" -- Microsoft Middle Manager 2.0

[veltas]

> Does Microsoft Middle Manager 2.0 stop working in the presence of traffic cones?

Yes, but it integrates well into Outlook.

[pjc50]

> Microsoft Middle Manager 2.0

I think we've found the common thread between AI and crypto: the dumping of externalities and increase in energy usage.

Do people really want to work for Microsoft Middle Manager? Have we not seen enough horror stories about metric-slavery in Amazon warehouses and the gig economy? It might be cheaper, but it's also worse, for a class of people who don't get any input in the decision. Similarly self-driving unleashes a new class of poorly behaved "learner" drivers on the road, who may be less aggressive but are also capable of causing problems from that very timidity and lack of general competence.

[arcticbull]

The other common thread is it’s all the same grifters. The same people who hyped up the worlds slowest coal powered linked list as the future of money are now hyping up AI. I do think AI is somewhat different as we will be left with more to show for it than “PayPal for Cambodian human traffickers running pig butchering scams.”

My banal take is: notwithstanding above, all the risk premium is gone for investors and employees. Everything in the AI space is priced assuming flawless execution over a 20 year time frame. As a potential employee - or investor - that equity is just not interesting to me.

[Animats]

> The other common thread is it’s all the same grifters.

Yes, the same grifters are migrating to AI. But there's a real industry in AI, unlike blockchain, where it was almost all grifters.

The pure scam industries, day trading, binary options, cryptocurrency, contracts for difference, retail FOREX, etc. do have many of the same people. If you want to follow this, there's Offshore Alert.[1]

[1] https://www.offshorealert.com/

[Animats]

> same grifters are migrating to AI

This just in:

"MetaQuiz is the brainchild of MetFi DAO, a true trailblazer, investor, and incubator in the realms of AI, Metaverse, and Web3. With its unwavering commitment to innovation, MetFi DAO is reshaping education and entertainment in unprecedented ways. The platform aspires to redefine how individuals from diverse backgrounds learn, grow, and earn rewards in a secure and engaging environment. ... MetaQuiz introduces an unprecedented pathway, where the allure of learning seamlessly merges with tangible opportunities for earning."

AI, Metaverse, Web3, DAO, education, and gambling! What could possibly go wrong?

[vGPU]

Just you wait until it realizes that firing you would result in a net carbon usage savings for the company.

[naillo]

That's true for driverless cars but there's many areas where you don't need 100% to be useful. Stable diffusion and LLMs are 90% there but still very useful and cool.

[Tanjreeve]

Most of those areas don't normally have a lot of value or you're marginally improving on existing recommendation systems. It's a coin toss on how much of an improvement this wave of LLM powered ML systems will have over the previous wave of "enterprise knowledge collation" systems and if that will be enough for people to buy them.

[cstejerean]

People seem to be buying them already, looks like OpenAI is now pulling in $80M per month in revenue.

[intended]

Self-driving is never going to happen, and its also sitting in some sort of informational blind spot for the people working on it - I have no idea why.

There is no such thing as "driving" there is no physical force, or particle. There is no force preventing you from driving in the opposite direction of traffic, or through glass panes.

"Driving" is entirely a social phenomnon, the confluence of societal self impositions and engineering.

If you have a car on fire in front of you, you will need to reverse in the wrong direction of traffic.

In many countries - you have to regularly deal with drivers going full tilt, on the wrong side of the road.

Or You have to deal with theft, and people trying to rob you at every red light.

Im underscoring that this is a social issue. You would need to create models for each country and region, to truly improve self driving.

Self driving assumes a far narrower problem space than reality gives a fig for.

Self driving theory currently works in the same way any theory that assumes spherical cows works.

[esperent]

> In many countries - you have to regularly deal with drivers going full tilt, on the wrong side of the road.

I don't agree with most of your comment but this point is worth examining. I think it's not an argument against self driving for a few reasons:

1. There's lots of places in the world where people driving on the wrong side of the road is uncommon. We can start with self driving there. You can apply this to many other situations as well - AI can't handle ice yet? Well, let's start with non-icy roads. Even when you apply all the stipulations like these that you need, you'll still be left with a large enough percentage of the world to make self driving useful. Especially since a lot of the places that are suitable will be rich cities in developed countries.

2. Driving habits change. Thailand is a good example of this, it's a country in transition from the "developing country driving style" to the "developed country driving style", for want of better terms. Driving there 15 years ago was an extremely different, far noiser, far more dangerous experience than driving there now. It's still got a long way to go, of course.

3. If self driving becomes the norm, well then, problem solved. We don't even have to get to FSD. Partial but always on assisted driving that nags you whenever you drive on the wrong side of the road or go over the speed limit would probably be enough to cause a shift if most people have it.

[intended]

Therein lies the rub

Movement, is animal like - perceiving self, environment and understanding movement through it. - Movement is universal. Its the application of physics.

Thats not driving.

Driving is a social construct. It is the application of physics while navigating a social world.

Driving is observing the law, observing social constructs (that differ regionally), adapting to new constructs based on location and environment.

Thats the blind spot for self driving proponents. They conflate the two things, but talk primarily about the first.

As a result, you will never get self driving - the assumptions are wrong.

Let me put it this way - you get self driving when a car decides its best course of action is to reverse in traffic, because it perceives a tsunami coming from ahead of it.

[esperent]

> Driving is a social construct. It is the application of physics while navigating a social world.

Interesting way of looking at it. In the countries where people tend to drive at high speed on the wrong side of the road, I'm inclined to agree with you.

In countries where people follow the rules of the road, I think it's the opposite. There, it's primates who evolved to follow and react to a complex social and physical environment being forced to follow a simple and rather tedious set of rules, which require constant vigilant attention. This is something we humans are not good at - we're designed for short bursts of attention in situations where rules need to be interpreted intuitively, not driving for three hours while strictly following the rules of the road.

> observing social constructs (that differ regionally)

Well, the social constructs around driving also different temporally and are changing all the time. Self driving, even partially implemented, is sure to have profound effects on these cultures all over the world.

>perceives a tsunami coming from ahead of it.

In extreme cases like tsunami, floods, or fires, there are solutions. In the short (and probably medium) term the cars will still have human controls, so the human can instantly take over. In the long term, well maybe there can be a "panic button" that allows a human to shout voice controls. Either way, it's just another solvable technical issue, and probably much easier than many other issues that need to be solved.

BTW it's not like researchers haven't considered to these things

https://www.thecarconnection.com/news/1122534_self-driving-c...

[jstx1]

I don't understand the leap from "it's difficult and there are many special circumstances" to "it's never going to happen". I don't get why you think that any of these problems are unsolvable given enough time.

[abainbridge]

Here's one way to get to the "it's never going to happen" outcome:

The grand-parent comment argues that we need AI that works in a much broader set of circumstances to "solve" self driving. In particular it would need to understand how other humans would react to its actions in novel situations. That is approximately a description of having a Theory of Mind. Some argue that you can't have a Theory of Mind without being conscious. We might ban captive conciousnesses for ethical reasons.

I think you could invent lots of other scenarios that yield the "never going to happen" outcome. They probably all sound ludicrous, because having an AI that understands the workings of human minds sounds ludicrous (and frightening).

[Animats]

Actually, I'm not saying that. I'm saying that self-driving is nearing success, and that it took a while. LLMs are at the beginning of that process.

[neuronic]

To me this is a technocratic stance - believing that technology will always find a solution and absolve us of all problems.

Maybe. Maybe not. What I read from his post is much closer in spirit to "we cannot simulate real physics down to the atom" when modeling protein folding so what we do instead is building models (good enough estimations) and AI approaches to recognize complex (but still top level) patterns.

These reduced views on reality help our problem space to a large degree but they can never account for the full scope of the physical reality and they largely work on assumptions which might be proven wrong at any point in time.

[intended]

Thats not the point.

The point is people who are proponents of self driving have a blind spot. Its construed as a code problem, ignoring entirely the massive social aspect of human behavior.

Driving isn't locomotion. You can't assume that the people around you are going to be reasonable, sane or predictable. More data doesn't fix that, unless you start building a "civilzed behavior" model and then add that to your movement model.

I really want to see how honest people are when they build a fair representation of human civilization, warts and all.

Self driving as has been defined assumes some absurd things about the world in which humans live. No one is going to buy a car which wont know to run when they are about to get robbed.

To be blunt, the blind spot assumes you live in America (and probably california), not that you live in Brazil or India or Egypt.

[salawat]

Because no matter how much time you give it, there are social problems that need to be solved first; and said social problems are not the kind that are amenable to mere technical solutions.

Literally the case of: If the technically perfect implementation existed tomorrow, we still would not be ready to flip the switch because of how drastic the reorg of societal norms would be. It'd be a complete refactor.

[ryanwaggoner]

This comment is pretty funny in the context of a world where there are currently cars driving millions of miles without human intervention in a variety of conditions. I don't think anyone cares whether it meets your personal definition of "self driving".

Reminds of the Chinese proverb: “Man who say it cannot be done should not interrupt man doing it.”

[intended]

Cmon, - you are the one misrepresenting the actual definition of self driving - aka Level 5.

No vehicle is at level 5.

Calling it “my definition” and then switching the actual definition is unfair.

[ryanwaggoner]

Who decided that only level 5 qualifies as "self-driving"?

[intended]

This comment is pretty funny in the context of someone starting out by claiming no cares “whether it meets your personal definition of "self driving".”

[ryanwaggoner]

No, you said that I changed the “actual definition” of self-driving from Level 5 to something else. I’m asking what makes Level 5 the “actual definition” and not just an arbitrary point that you’ve personally decided means self-driving?

[tim333]

I guess like heavier than air powered flight will never happen. I mean there would be so many problems.

[pferde]

Indeed, it would probably require the invention of a true Artificial Intelligence, and of high quality as well. Not the glorified autocomplete the scammers are trying to pass as AI now, but the real thing that can sense and understand the world around it.

[gremlinsinc]

it'll happen but the car won't really be self driving, I think a literal ai robot cheaufer is more likely. AGI will be able to do everything a human can do, including driving. We're IMHO less than a decade, if not less than half a decade from that.

If you don't believe me name a scatter chart showing ai papers since 2017. Notice how this year dwarfs all previous years and last year blew the water out of the previous ones but the ones before generative art models were a bit slower and more regular.

If we can perfect ai agents and automated LLMs, we could create personified agents of varying backgrounds in a virtual computer lab tasked with doing ai research 24/7. We could put 60 ai scientists in the lab and just watch them work.

Maybe the lab is vr such that human scientists could enter and collaborate with ai counter parts. imagine applying this to cancer research, etc.

When an LLM can train itself, program children(next version), and submit orders or plans for new hardware fabs to a factory to breach current limitations, then we basically have von Neumann probes that multiply, grow, learn, and don't need human intervention.

[simiones]

LLMs are nowhere near being able to research other, better LLMs. It's not even known if existing LLMs are at 1% or at 99.99% of what an LLM can in principle achieve.

[drawfloat]

Being able to take a cab in a single tiny part of the world remains firmly in “sort of works” camp.

[Ekaros]

Also when you start to get in more extreme but not actually extreme scenarios. I wonder how well will these do with let's say 5cm of fresh snow. With no other traffic yet meaning no road markings are visible, and in bad case signs could be covered. Probably a yearly scenario in many parts of the world.

[TMWNN]

>It may take a while to get from GPT-3 to Microsoft Middle Manager 2.0. But the path is clear now.

I was intrigued by the announced Business Chat feature for Microsoft Teams <https://www.reddit.com/r/singularity/comments/11swyeu/introd...>, but learned that it is just summaries of conversations. That's not quite what I'd imagined, which is something like this:

----

A: ... and that is why I think we should go with option 1.

B: No, the points you mentioned support my case for option 2.

C: Nothing you guys have said changes my mind about option 3 being best.

D: Business Chat, what do you think?

BC: Based on this discussion, and my research, option 1 seems more realistic but option 2 would be more profitable if possible. My reasons are ...

C: Business Chat and you guys all don't understand point N, which is the main reason why option 3 is best.

B: Higher profit is exactly why I think option 2 is the way to go.

A: No, our rival is going to hit the market next month. We need to get something out there ASAP. Option 1 can do this.

D: You've all given me things to think about. Thank you for coming. Business Chat, email me a summary of the meeting, and set up a followup meeting for Tuesday 3pm.

----

That is, AI used as a colleague/assistant, not necessarily subordinate but not seen as omniscient, either; another viewpoint to consider. Like you said, Middle Manager 2.0. When do you think the above will be feasible? This year? Five years?

[yMMe2WYE_D]

It might be more an issue of "if" than "when". Can LLM even go beyond being "interactive textbook"? On the other hand, looking at constant complaints from data scientists, ChatGPT can replace them here and now. It might not be able to understand the problems, but supplying boss with arguments that support their decision is the matter of weeks, maybe just an extra preamble/prompt will be enough.

[tome]

> It may take a while to get from GPT-3 to Microsoft Middle Manager 2.0. But the path is clear now.

Is there a way you could make the path clear to me too? I just don't get it. Sure, I can imagine in principle it's possible. But I don't see how you can already see it's certain. Is there something you can share that will allow me to see why it's certain?