Opinions on AI

[mb194dc]

The Chinese have entered the race, they've built a model for 5% of the cost of OpenAI and ChatGPT that is better in a lot of testing:

Chinese researchers just built an open-source rival to ChatGPT in 2 months. Silicon Valley is freaked out.

DeepSeek-R1, a new reasoning model made by Chinese researchers, completes tasks with a comparable proficiency to OpenAI's o1 at a fraction of the cost.

www.livescience.com

The biggest impact of this though, is that the model is open source. So we could see LLMs move towards being commoditized, that is anyone will be able to build a decent one quite easily and cheaply. So there's no moat around OpenAI or other start ups that have proprietary technology they thought would give them an edge over the others.

 

[Shihab]

The biggest impact of this though, is that the model is open source.

Does it matter?

The logic was always a fraction of the whole cost anyway. It was the training dataset that made or broke these things, plus the processing power to train (and run them). Being "open source" helps neither, for obvious reasons in the latter's case, as for the former: no one in their right mind would OSS the training set, and draw a bullseye on their backsides for every copyright lawyer from every sorry corner of this earth).

Opensource-washing is the new greenwashing!

 

[AusWolf]

Opensource-washing is the new greenwashing!

I don't know how it is with AI, but in a general sense, I disagree. Open-source lets everybody use your software for whatever purpose they have in mind. It lets the Linux community incorporate AMD drivers into the Linux kernel, for example. It should be the way to go for every software company, imo. Closed standards are only good for the company that created them.

 

[ShrimpBrime]

Voted other. Simply because I can't see the future in that sort of way. Good or Bad.

People and their intent is good or bad. The AI is a tool. Much like a farmer will tell you a gun is a tool, where the thief calls it a weapon.

And since AI seems to be implemented for mostly good intent, I can't imagine the world to get worse, which is a very broad thing to say.

What most people fail to understand, is the world, the earth is fine. It'll be here long after humans and AI wipe themselves out.

 

[Shihab]

I don't know how it is with AI, but in a general sense, I disagree. Open-source lets everybody use your software for whatever purpose they have in mind. It lets the Linux community incorporate AMD drivers into the Linux kernel, for example. It should be the way to go for every software company, imo. Closed standards are only good for the company that created them.

And "green" is not inherently bad either, but greenwashing is.

Opensource-washing != "open source."

Openwashing - Wikipedia

en.wikipedia.org

I'd rather have openly "closed-source" software that I can clearly gauge the risk of working with (legally, economically and logistically-speaking) than being entrapped with something pretending to be "open" while in reality only its facade is, with every what makes actually useful locked behind paywalls (or one you can't even pay to get through. At least not if you're anything short of a megacorp).
Good read on the topic:

The Open Source AI Definition: where's the data?

Two years into the “multi-stakeholder process” to define what constitutes as “open source AI”, we are facing a number of issues that are immense, fraught, and nuanced. Why are data excluded from the proposed definition, why are they important, and why are we even doing this?

www.juliaferraioli.com

Don't get me wrong tho, I'm not saying AI/ML *can't* be deemed open without everything being available to the net. Plenty of ML packages on Github/lab/whatever that most of the users would purposely skip training set/weights when cloning/forking/whatever, those are mostly purpose-specific packages, some image segmentation model or the like. This is not the case with the wastes-of-petajoules those big corps are peddling.

 

[SilentPeace]

It's all overhyped, a solution in search of a problem. A gigantic waste of money and energy consumption. Another dot com bubble waiting to burst. The sooner the better I say.

 

[tpuuser256]

It's all overhyped, a solution in search of a problem. A gigantic waste of money and energy consumption. Another dot com bubble waiting to burst. The sooner the better I say.

The dot com bubble did burst but you're using the fruits of the massive investments of that time even today.
There is a lot of hype and market manipulation, sure, but the potential is real. ML is going nowhere, it's literally the pinnacle of human technology. It will use a combination of cutting edge semiconductors, quantum computing, nuclear fusion (funding from high energy demand), decades of strenuous software development from probably billions of brains on the planet.... Nothing gets more advanced than this.

 

[SilentPeace]

The dot com bubble did burst but you're using the fruits of the massive investments of that time even today.
There is a lot of hype and market manipulation, sure, but the potential is real. ML is going nowhere, it's literally the pinnacle of human technology. It will use a combination of cutting edge semiconductors, quantum computing, nuclear fusion (funding from high energy demand), decades of strenuous software development from probably billions of brains on the planet.... Nothing gets more advanced than this.

What do we need it for? What does it do? Remembering that it isn't truly intelligent and never will be. It has been missold as such. It is like driverless cars, an idea which excites some people, but which is effectively useless - it will never work - and even if it did, so what? What can a driverless car do that one with a driver cannot? Given that the driver must still be present and observing what the 'AI' driver does, in order to take control if the 'autopilot' gets it wrong. So much technology is about 'look how clever this is' rather than 'this is actually useful, and it works'.

 

[Count von Schwalbe]

nuclear fusion

That would be nice. Its just around the corner, right?

What can a driverless car do that one with a driver cannot? Given that the driver must still be present and observing what the 'AI' driver does, in order to take control if the 'autopilot' gets it wrong.

If it becomes actually driverless, the advantages are myriad. A true driverless semi truck would revolutionize the trucking industry. Which could have massive knock-on effects for the whole economy. And that is just one example.

However, is AI - or more accurately, large language models - really the answer here? What is currently in use? How can it be developed further as technology improves?

 

[igormp]

The Chinese have entered the race, they've built a model for 5% of the cost of OpenAI and ChatGPT that is better in a lot of testing:

Chinese researchers just built an open-source rival to ChatGPT in 2 months. Silicon Valley is freaked out.

DeepSeek-R1, a new reasoning model made by Chinese researchers, completes tasks with a comparable proficiency to OpenAI's o1 at a fraction of the cost.

www.livescience.com

The biggest impact of this though, is that the model is open source. So we could see LLMs move towards being commoditized, that is anyone will be able to build a decent one quite easily and cheaply. So there's no moat around OpenAI or other start ups that have proprietary technology they thought would give them an edge over the others.

Alibaba has been building multiple open source LLMs for quite some time, look into the Qwen models.
Meta has also created many open source models. Both llama and qwen were used as the base models for the distilled versions of deepseek.

The "5% of the cost" figure is heavily overestimated. They did an impressive job increasing the avg GPU utilization in their cluster, but the only actual comparison we can make is to Meta, which is not the most efficient figure for training runs.

Does it matter?

The logic was always a fraction of the whole cost anyway. It was the training dataset that made or broke these things, plus the processing power to train (and run them). Being "open source" helps neither, for obvious reasons in the latter's case, as for the former: no one in their right mind would OSS the training set, and draw a bullseye on their backsides for every copyright lawyer from every sorry corner of this earth).

Opensource-washing is the new greenwashing!

Being open source is not that relevant since, as you said, almost no one will have the data and hw required to run, much less train such a big model.
However, they also published technical papers explaining all their methodology, and that was the nicest part and they indeed do some really impressive work.

They still have smaller distilled models available as well, and those are within reach of most people.

It lets the Linux community incorporate AMD drivers into the Linux kernel, for example.

FWIW, it wasn't "the community", but rather AMD themselves.

 

[Shihab]

What do we need it for?

I'm pretty sure if the internet existed back then, many people would ask Claude Shannon the same question. But then again, there wouldn't be an internet without Shannon and co's work.

The fact that Altman, Nadella, Huang and who else keep pushing the gimmicky parts of it does not fault the technology itself. Anything the can be made profitable and has the average person as an end consumer will fare the same. The internet itself is mostly wasteful; used by most for little more than endlessly scrolling down some social media, binging movies or tv shows, p*rn, or, and this is the epitome of wastefullness imo: watching others play games. Yet amongst those billions wasting "money and energy," there are millions of scientists, engineers, and other professionals doing meaningful work with it. The same holds -albiet admittedly to a much, much lesser extent- for AI/ML. There is much more to the tech than simulating bad acid trips.

However, they also published technical papers explaining all their methodology, and that was the nicest part and they indeed do some really impressive work.

This was the norm till a while ago, iirc. Wouldn't be surprised if they too pulled a bait 'n switch like their competition.

 

[Count von Schwalbe]

There is much more to the tech than simulating bad acid trips.

How much of that is based on ML that is not LLMs? Just curious, really. How diverse is machine learning once you get past the fluff of consumer-facing "AI" like ChatGPT and Stable Diffusion?

Yes, I know I could read what you linked, but I don't have the stomach for acadamia this morning.

 

[SilentPeace]

If it becomes actually driverless, the advantages are myriad. A true driverless semi truck would revolutionize the trucking industry. Which could have massive knock-on effects for the whole economy. And that is just one example.

However, is AI - or more accurately, large language models - really the answer here? What is currently in use? How can it be developed further as technology improves?

I do see how driverless trucks would be useful, I also realise it will never happen, it's a non-starter. I'm surprised anyone with knowledge of technology and its inherent limitations, and of driving and the human skills of emotional intelligence and judgement which a machine will never understand, could think driverless vehicles could ever work. The only place they could work would be in a controlled environment similar to railway tracks, closed off from non-driverless vehicles, pedestrians, cyclists, etc. Even then I doubt it would work, but it would have more chance.

AI currently seems to be about let's develop it as much as we can and then see if we can figure out something we can use it for, just because we can. Maybe I'm wrong, we'll see.

 

[dragontamer5788]

There is much more to the tech than simulating bad acid trips.

But there's also much more to machine learning than slapping down 400 GB models and bragging that your GPU setup can run it and no one else can.

That's the part of "AI" that's so frustrating. Its so obviously geared towards hyperconsumption and is moving away from engineering these days. Its mostly about "My company can afford $1 Billion". No wait, well MY COMPANY can afford $10 Billion.

Actually, my company is buying a nuclear power plant to build AI.

Actually, my company is putting $500 Billion onto AI.

What's next? Its just VCs bragging about how much funding they can throw at these things. There's almost no engineering effort on it anymore. The fucking metric is $$$ spent and literally bragging about the energy wasted (see Nuclear Power Plant purchases).

AI currently seems to be about let's develop it as much as we can and then see if we can figure out something we can use it for, just because we can. Maybe I'm wrong, we'll see.

At what cost? $500 Billion, society needs to start asking about far cheaper projects that lead to far more obvious benefits.

 

[SilentPeace]

At what cost? $500 Billion, society needs to start asking about far cheaper projects that lead to far more obvious benefits.

I agree. If you spend billions building a bridge or a hospital, you know the benefits beforehand, they are obvious and quantifiable. AI though...? Well the British government here wants the UK to become an AI powerhouse, whatever that means. We are busy throwing money at it even though it will have huge power consumption which the UK is short of already. We could have used that power to keep the lights on, heat our homes, charge up our cars or whatever, but... we're using it for AI so we can do... what? Nobody seems to be sure yet.

 

[tpuuser256]

What do we need it for? What does it do? Remembering that it isn't truly intelligent and never will be. It has been missold as such. It is like driverless cars, an idea which excites some people, but which is effectively useless - it will never work - and even if it did, so what? What can a driverless car do that one with a driver cannot? Given that the driver must still be present and observing what the 'AI' driver does, in order to take control if the 'autopilot' gets it wrong. So much technology is about 'look how clever this is' rather than 'this is actually useful, and it works'.

Intelligence lacks a solid definition but I promise you the way brains works is basically the same way as ML models. We are blessed with 5+ senses, a big brain, the social environment and language. We currently do not have the computing power of a brain with the same data input as your average human and yet we brag about how we are smarter than ML models.

Take a 2D image generation model. It lacks the physics simulation abilities of the average human brain among many other things to make (more) realistic scenes.

That would be nice. Its just around the corner, right?


If it becomes actually driverless, the advantages are myriad. A true driverless semi truck would revolutionize the trucking industry. Which could have massive knock-on effects for the whole economy. And that is just one example.

However, is AI - or more accurately, large language models - really the answer here? What is currently in use? How can it be developed further as technology improves?

It isn't and to make sure everybody understand, I never claimed that. What is true though is that when energy demand rises due to datacenters, hopefully there will be enough global funding to make nuclear fusion happen. It just needs to be economical. I speculate again but I think it will happen when hydrocarbons and natural gas get depleted almost fully, but there is a long way till there. my guesstimate is 100-200 years before commercial fusion energy is available. I wouldn't be surprised if it's before too

ML and the way it works generally is no different than human brains. Fyi I don't believe in a metaphysical 'soul' ot anything so I may be biased.

I personally don't believe LLMs alone to be the answer, the approach needs to be holistic but that's just the beginning and countless engineers and scientists are scratching their brains right now expanding the possibilities of ML.


-
To end my comment, I want to say that ML in 2025 is quite rudimentary compared to what can be done in a few centuries.

 

[TechBuyingHavoc]

I do see how driverless trucks would be useful, I also realise it will never happen, it's a non-starter. I'm surprised anyone with knowledge of technology and its inherent limitations, and of driving and the human skills of emotional intelligence and judgement which a machine will never understand, could think driverless vehicles could ever work. The only place they could work would be in a controlled environment similar to railway tracks, closed off from non-driverless vehicles, pedestrians, cyclists, etc. Even then I doubt it would work, but it would have more chance.

AI currently seems to be about let's develop it as much as we can and then see if we can figure out something we can use it for, just because we can. Maybe I'm wrong, we'll see.

I am not so negative on AI as to doubt that it can actually achieve driverless vehicles. I think this is entirely plausible in a technological POV.

But who is going to take the liability for accidents, property damage, and deaths? The manufacturer?

Yeah, I don't see that happening so the real question is a legal and sociological question on whose shoulders the liability will fall on. Until this is answered, no self-driving is going to happen on a broad scale.

 

[thesmokingman]

The Chinese have entered the race, they've built a model for 5% of the cost of OpenAI and ChatGPT that is better in a lot of testing:

Chinese researchers just built an open-source rival to ChatGPT in 2 months. Silicon Valley is freaked out.

DeepSeek-R1, a new reasoning model made by Chinese researchers, completes tasks with a comparable proficiency to OpenAI's o1 at a fraction of the cost.

www.livescience.com

The biggest impact of this though, is that the model is open source. So we could see LLMs move towards being commoditized, that is anyone will be able to build a decent one quite easily and cheaply. So there's no moat around OpenAI or other start ups that have proprietary technology they thought would give them an edge over the others.

Counter point... lmao.

According to a new article by the Financial Times, OpenAI claims to have evidence that DeepSeek, the Chinese startup that has thrown the US tech market into financial turmoil, used the company's proprietary models to train its own open-source LLM, called R1. This would represent a potential breach of intellectual property, as it goes against the OpenAI terms of service agreement.

In the article the FT writes that a source at OpenAI claims it has evidence of “distillation” occurring, which is a technique used by developers to leapfrog on the work done by larger models to achieve similar results at a much lower cost.

OpenAI says DeepSeek used its models illegally, and it has evidence to prove it, new report claims

Has there been a breach of intellectual property?

www.techradar.com

And as if on cue...

https://www.usnews.com/news/top-news/articles/2025-01-29/alibaba-releases-ai-model-it-claims-surpasses-deepseek-v3

 

[TumbleGeorge]

Counter point... lmao.

OpenAI says DeepSeek used its models illegally, and it has evidence to prove it, new report claims

Has there been a breach of intellectual property?

www.techradar.com

And as if on cue...

https://www.usnews.com/news/top-news/articles/2025-01-29/alibaba-releases-ai-model-it-claims-surpasses-deepseek-v3

DeepSeek use different approach . Different in its basement and much easier to training. Let's leave nOpen ai to prove his accusations

 

[Shihab]

How much of that is based on ML that is not LLMs?

From the niches I'm familiar with, practically none.
Some AI folks say that LLMs are capable of more than just spitting nicely worded bullshit, but I've yet to see that.

That's the part of "AI" that's so frustrating. Its so obviously geared towards hyperconsumption and is moving away from engineering these days.

Name one thing in the mass market that isn't.
We're dopamine-overdosing ourselves to oblivion, and we're lovin' it, babeh!