What kind of bubble is AI? | Hacker News Reader

What kind of bubble is AI?

pluralistic.net

43 pointsHoasi37 comments

I'm biased, but I think the answer is semi structured problems.

Automating a tractor that has to till or seed is much easier than a self driving car, and alleviates some very low margins high labor intensive activities.

(I work at a place that is automating logistic yard operations, which is fixed cost, well structured, and predictable problem/environment)

I think "everywhere all the time" AI for chat, driving, coding, or medical image analysis etc is a pipe dream absolutely. And what we will lose when it all pops is the pragmatic solutions to solveable problems that sound more like "algorithms for assistance occasionally or part time".

> ... semi structured problems.

i think i agree with this,

> I think "everywhere all the time" AI for chat, driving, coding, or medical image analysis etc is a pipe dream absolutely.

but i'm not sure i'd put medical image analysis in this group. i mean, aren't there already several examples of software outperforming humans in this domain?

There are also several examples of software outperforming drivers. Yet here we are with a growing suspicion of self-driving cars.

There's no extreme correct. The correct future is a blend of software and human control, with, I suggest, a bias towards software in well-structured areas, and a strong bias towards huamns elsewhere. And it's humans after all that would structure the world for software. That's fine, it's how it's always been with machines. We just don't want to throw the baby out with the bathwater when we lose trust in "AI".

Where I work, we use AI for tasks that are similar to analyzing medical imaging (but not medical). It easily outperforms humans. Not in terms of accuracy in identifying features -- it's about the same as people in that -- but in terms of speed.

That seems true for both chatgpt and ai art programs. The output is unverifiable, often crap, and very fast to get. I'd argue we're losing something if we make it impossible for human experts to get trained in professional settings by outcompeting them with ai.

> semi structured problems

What's the state of the art with A.I. digesting RDF, OWL, and the like ?

Even if the subsidies that drive increasingly sophisticated models evaporate, we'll still have the existing models we can run on commodity hardware. Some of those models will be good enough for the risk-tolerant low-value applications. Low value doesn't mean no value.

I also think that machine vision for agriculture can be a little more risk tolerant than for cars (doesn't matter if your weed burner occasionally torches a stalk of corn the way that it does if your robotaxi occasionally runs over a pedestrian).

Compared to what the institutional investors are chasing, that will not have been worth it. But it'll be nonzero.

>Even if the subsidies that drive increasingly sophisticated models evaporate, we'll still have the existing models we can run on commodity hardware.

Exactly. So much discussion in the media and elsewhere about generative AI assume that AI-as-a-service from big companies like OpenAI is the only way forward, and if they die, so does AI in general. But we already can run quite powerful models locally. For example, I don't think Cory's example of a $10/month service to draw D&D character portraits makes sense now. Surely anyone geeky enough to play D&D could download and install InvokeAI (or other similar open-source program) and create their portraits for free.

The inherent problem I'm seeing is that training the base models is incredibly expensive and computationally intensive. You can't do it without institutional capital. Gpt-4 I believe cost 100 million to train. I'd bet they're also losing money on a lot of customers because inference isn't that cheap either. If investors realize something like investing in this stuff isn't paying off, it becomes pretty hard to make new models. Maybe someone would have to develop a distributed system like folding at home to get it done.

Well, the various models on sites like civitai are definitely not created by institutions -- especially not the many, many NSFW ones. Yes, you can argue that they are relying on existing models from StableAI and the like, but even if those were never updated, that wouldn't stop the creation of new models on top of them. We kind of see this already; most checkpoint models and loras on civitai are still based on the older Stable Diffusion 1.5 rather than the newer Stable Diffusion XL.

I mean that's kind of my point. Derivative models are all dependent on models built by institutional players like stability. Is there a limit to how good civitai models can be? If there is, you can't move past that without higher quality base models, which requires a lot of capital. That's just what my intuition says.

I don’t think the autonomous vehicle example and the radiology example are really analogous. I’ve agreed for a long time with the premise that an “almost self-driving” car isn’t much good - certainly at the consumer level the whole value proposition is that you don’t have to pay attention. But that’s a result of it being a real-time application constrained by physics. Medical diagnosis is a more deliberative process that already incorporates inputs from multiple tools.

Various good arguments along the way, but I don't buy the overarching premise.

It can't both be useful and valuable enough to become pervasive and then suddenly disappear because it isn't valuable enough.

Sure some of the free VC money startup will go bust & there is hype but that doesn't make the entire thing a bubble that implodes into nothing.

Also - the fact that I derive personal value and am willing to pay to me indicates that this is less bubbly than say crypto or dot com.

For some reason the author thinks that skills you learn developing in PyTorch will go away if Facebook does. Wouldn’t the open source community just fork the source or develop a new framework? It’s not like backprop will stop working.

Same here. I couldn't resist rereading that multiple times. It was a head scratcher. Gives me the impression that the author doesn't understand the details, which further undermines their arguments.

Keep reading:

"Perhaps the communities who've invested in becoming experts in Pytorch and Tensorflow will wrestle them away from their corporate masters and make them generally useful. Certainly, a lot of people will have gained skills in applying statistical techniques."

It's not just about PyTorch, but the whole ecosystem (code, cloud, and data) that these tools depend on to work, and what happens when they get integrated into an economy then disappear.

The author is Cory Doctorow, who has a very good intuition about these kinds of issues. His presentations about open source and the war on general purpose computing are very important and prescient.

There is already a lot of real value in scientific domain-specific models. One such example is weather forecasting. We can suddenly create forecasts with astonishing accuracy on commodity hardware in a matter of seconds, versus several hours with traditional numeric models. This is life-saving.

And it's open source, e.g.: https://github.com/NVlabs/FourCastNet

(2023)

Discussion: https://news.ycombinator.com/item?id=38698190

What kind of bubble is AI hardware as useful/commerical models shrink in requirements.

What type of AI? AGI? Textual based AI? Computer vision AI? What kind of bubble - the kind that has happened before in the history of AI…a slow evolution of progress just like now with fits and bumps but always progressing over time. It will never fit a business or persons agenda but it will always get better.

I would like to have a reading assistant that highlights persuasive techniques like rhetorical questions and - probably even better - neutralizes them for me if needed.

Yet another catastrophist haphazardly connecting dots that shouldn't be connected. The article "lost" me here:

> More than $100b has been incinerated chasing self-driving cars, and cars are nowhere near driving themselves

Erhm, ..what ? An over-eager Cruise rushing to market does not negate 20 years of Google-quality research. I have driven Waymo cars multiple times in SF and by user testimonials, they have already avoided accidents (possibly even saved lives). I am finding myself thinking WWWD (what-would-Waymo-do) when I'm driving.

A lot of people (including Cory) have decided that because the initial overoptimistic estimates a decade ago that predicted true (as in no steering wheel, not as in just helping you stay in your lane and avoid collisions) self driving cars by 2020 or whatever were wrong, that this means that the entire concept is impossible. I don't get the argument myself -- it's not like this is the first example of a technology that took longer to turn into a usable product than expected.

it's a dotcom bubble with the potential to not really be a bubble but turbulences

What did I just read...

It's as if every second person on earth became a Scientologist spouting correct sounding but flawed ideas and conjectures ultimately aiming for your pocket book. But every once in a long time it discovers a useful and important new protein so you can't discount it entirely.

Total nonsense article.

How does one talk about self driving and just gloss over the existence of Waymo?

Everything he says is relevant. Waymo is the exception, not the rule, and I'm sure even they are working hard to operationalize this stuff and make it actually cheaper than humans.

There are several irrelevant/incorrect statements.

PyTorch doesn't need facebook to survive. Even if pytorch did disappear tomorrow, it's wrong to suggest that all folks competent at building/using models using pytorch would have a set of useless skills. No one who understood the situation would write that. The author clearly hasn't built a model or used any of the (several) ML libraries which all have similar constructs. The skill is to be able to keep straight in your head large tensors and their shape as they go through a bunch of operations - that's not framework dependent.

"Risk intolerant" is a fancy term made up that is then used to paint every high value application as not a fit for the technology. There are many applications ( customer support, technical support, litigation research, mortgage underwriting, marketing copy, natural language queries of DBs) where LLMs are already being put to use and reducing the number of people required to do the job.

This is written by a journalist who isn't actively involved in actually building or using the technology. They misunderstand things to the point of nonsensical conclusions.

The term "bubble" in AI refers to inflated investments and valuations based on speculative rather than intrinsic values. While the AI sector experiences significant hype, its broad applicability and real-world utility across various industries offer a stabilizing factor against bubble dynamics. Continuous technological advancements in AI contribute to its tangible value creation, countering speculative pressures. The future of AI's market stability will hinge on aligning expectations with actual capabilities, amidst navigating regulatory and ethical landscapes.

(edited for brevity)

What's up with this ChatGPT generated comment?

It’s making a statement. AI commenting on the AI bubble.

Seems more like a joke than a statement to me.

I did not know what kind of Bubble AI is so I asked it and posted its answer to get a second opinion and also so it won't feel like we are talking about it behind its back.

it's against the rules here to post AI-generated comments.

see https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...

That should have been the first comment. Could have deleted it. Won’t happen again.

What ChatGPT missed here is that when bubbles collapse, the damage can extend beyond its edges. So even some perfectly viable AI use case that doesn't require the bubble's financial largess may suffer damage from the pop whether it is "tangible value creation" or not.

Reply on news.ycombinator.com