It’s crazy how much fud is flying around, and legitimately buries good open research. It’s also crazy what these giant corporations are explicitly saying what they’re going to do, and that anyone buys it. TSMC’s allegedly calling Sam Altman a ‘podcast bro’ is spot on, and I’d add “manipulative vampire” to that.
Talk to any long-time resident of localllama and similar “local” AI communities who actually dig into this stuff, and you’ll find immense skepticism, not the crypto-like AI bros like you find on linkedin, twitter and such and blot everything out.
For real. Being a software engineer with basic knowledge in ML, I’m just sick of companies from every industry being so desperate to cling onto the hype train they’re willing to label anything with AI, even if it has little or nothing to do with it, just to boost their stock value. I would be so uncomfortable being an employee having to do this.
They’re all built on top of OpenAI which is very unprofitable at the moment. Feels like the whole industry is built on a shaky foundation.
Putting the entire fate of your company in a different company (OpenAI) is not a great business move. I guess the successful AI startups will eventually transition to self-hosted models like Llama, if they survive that long.
Most projects I’ve been in contact with are very aware of that fact. That’s why telemetry is so big right now. Everybody is building datasets in the hopes of fine tuning smaller, cheaper models once they have enough good quality data.
My company is realizing that hosting a model which will be private, cost-effective, and performing better than traditional algorithms is like finding a unicorn. Few months back, the top execs were jumping around GenAI like a bunch of kids. Fortunately, the Sr. research head beat some sense into them.
You’re lucky there’s a higher up that could talk down the even higher ups. Though, sometimes it’s not even about the r&d teams.
I saw company wide HR educational emails or courses telling you how to improve you work quality/efficiency, and one of them tells us to “research AI” and learn how to utilize it, talking about how great it is and improved the work efficiency by 30%. Sure, it has its uses, but I won’t go touting how great it is. And with how ChatGPT works, you have to be the biggest idiot in the world to upload all your sensitive stuff to ChatGPT just for it to make a spreadsheet faster. But without these disclaimers in the email, I doubt regular clerical staff knows about this, and it’s extremely dangerous.
What kind of use-cases was it, where you didn’t find suitable local models to work with ? I’ve found that general “chatbot” things are hit and miss but more domain-constrained tasks (such as extracting structured entities from unstructured text) are pretty reliable even on smaller models. I’m not counting my chickens yet as my dataset is still somewhat small but preliminary testing has been very promising in that regard.
What kind of use-cases was it, where you didn’t find suitable local models to work with ?
Any time you ask very domain specific questions; eg “i have collected some soil samples from the mesolithic age near the Amazon basin which have high sulfur and phosphorus content compared to my other samples. What factors could contribute to this distribution?”, both of-the-shelf local models & OpenAI fail.
The main reason is because these models are not trained on highly-specialized domains of text. Sometimes the models start hallucinating and which reduces our trust upon them.
“i have collected some soil samples from the mesolithic age near the Amazon basin which have high sulfur and phosphorus content compared to my other samples. What factors could contribute to this distribution?”
Haha yeah the top execs were tripping balls if they thought some off-the-shelf product would be able to answer this kind of expert questions. That’s like trying to replace an expert craftsman with a 3D printer.
As someone who was working really hard trying to get my company to be able use some classical ML (with very limited amounts of data), with some knowledge on how AI works, and just generally want to do some cool math stuff at work, being asked incessantly to shove AI into any problem that our execs think are “good sells” and be pressured to think about how we can “use AI” was a terrible feel. They now think my work is insufficient and has been tightening the noose on my team.
I think we should indict Sam Altman on two sets of charges:
A set of securities fraud charges.
8 billion counts of criminal reckless endangerment.
He’s out on podcasts constantly saying the OpenAI is near superintelligent AGI and that there’s a good chance that they won’t be able to control it, and that human survival is at risk. How is gambling with human extinction not a massive act of planetary-scale criminal reckless endangerment?
So either he is putting the entire planet at risk, or he is lying through his teeth about how far along OpenAI is. If he’s telling the truth, he’s endangering us all. If he’s lying, then he’s committing securities fraud in an attempt to defraud shareholders. Either way, he should be in prison. I say we indict him for both simultaneously and let the courts sort it out.
I really want to like AI, I’d love to have an intelligent AI assistant or something, but I just struggle to find any uses for it outside of some really niche cases or for basic brainstorming tasks. Otherwise, it just feels like alot of work for very little benefit or results that I can’t even trust or use.
I keep Qwen 32B loaded on my desktop pretty much whenever its on, as an (unreliable) assistant to analyze or parse big texts, to do quick chores or write scripts, to bounce ideas off of or even as a offline replacement for google translate (though I specifically use aya 32B for that).
It does “feel” different when the LLM is local, as you can manipulate the prompt syntax so easily, hammer it with multiple requests that come back really fast when it seems to get something wrong, not worry about refusals or data leakage and such.
Soldered is better! It’s sometimes faster, definitely faster if it happens to be lpddr.
But TBH the only thing that really matters his “how much VRAM do you have,” and Qwen 32B slots in at 24GB, or maybe 16GB if the GPU is totally empty and you tune your quantization carefully. And the cheapest way to that (until 2025) is a used MI60, P40 or 3090.
I receive alerts when people are outside my house, using security cameras, Blue Iris, CodeProject AI, Node-RED and Home Assistant, using a Google Coral for local AI. Entirely local - no cloud services apart from Google’s notification system to get notifications to my phone while I’m not home (which most Android apps use). That’s a good use case for AI since it avoids false positives that occur with regular motion detection.
A lot of people use them for the use case I described (object detection for security cameras), using either Blue Iris or Frigate. They work pretty well for that use case.
Wake word detection is a good use case too (eg if you’re making your own smart assistant).
The saddest part is, this is going to cause yet another AI winter. The first few ones were caused by genuine over-enthusiasm but this one is purely fuelled by greed.
The AI ecosystem is flooded, we need a good bubble pop to slow down the massive waste of resources that our current info-remix-based-on-what-you-will-likely-react-positively-to shit-tier AI represents.
After getting my head around the basics of the way LLMs work I thought “people rely on this for information?”, the model seems ok for tasks like summarisation though
I don’t love it for summarization. If I read a summary, my takeaway may be inaccurate.
Brainstorming is incredible. And revision suggestions. And drafting tedious responses, reformatting, parsing.
In all cases, nothing gets attributed to me unless I read every word and am in a position to verify the output. And I internalize nothing directly, besides philosophy or something. Sure can be an amazing starting point especially compared to a blank page.
It’s good for coding if you train it on your own code base. Not great for writing very complex code since the models tend to hallucinate, but it’s great for common patterns, and straightforward questions specific to your code base that can be answered based on existing code (eg “how do I load a user’s most recent order given their email address?”)
It’s wild when you only know how to use SELECT in SQL, but after a dollar worth of prompting and 10 minutes of your time, you can have a significantly complex query you end up using multiple times a week.
Well there is a very specific architecture “rut” the LLMs people use have fallen into, and even small attempts to break out (like with Jamba) don’t seem to get much interest, unfortunately.
Sure, but LLMs aren’t the only AI being used, nor will they eliminate the other forms of AI. As people see issues with the big LLMs, development focus will change to adopt other approaches.
There is real risk that the hype cycle around LLMs will smother other research in the cradle when the bubble pops.
The hyperscalers are dumping tens of billions of dollars into infrastructure investment every single quarter right now on the promise of LLMs. If LLMs don’t turn into something with a tangible ROI, the term AI will become every bit as radioactive to investors in the future as it is lucrative right now.
Viable paths of research will become much harder to fund if investors get burned because the business model they’re funding right now doesn’t solidify beyond “trust us bro.”
Sure, but those are largely the big tech companies you’re talking about, and research tends to come from universities and private orgs. That funding hasn’t stopped, it just doesn’t get the headlines like massive investments into LLMs currently do. The market goes in cycles, and once it finds something new and promising, it’ll dump money into it until the next hot thing comes along.
There will be massive market consequences if AI fails to deliver on its promises (and I think it will, because the promises are ridiculous), and we get those every so often. If we look back about 25 years, we saw the same thing w/ the dotcom craze, where anything with a website got obscene amounts of funding, even if they didn’t have a viable business model, and we had a massive crash. But important websites survived that bubble bursting, and the market recovered pretty quickly and within a decade we had yet another massive market correction due to another bubble (the housing market, mostly due to corruption in the financial sector).
That’s how the market goes. I think AI will crash, and I think it’ll likely crash in the next 5 years or so, but the underlying technologies will absolutely be a core part of our day-to-day life in the same way the Internet is after the dotcom burst. It’ll also look quite a bit different IMO than what we’re seeing today, and within 10 years of that crash, we’ll likely be beyond where we were just before the crash, at least in terms of overall market capitalization.
It’s a messy cycle, but it seems to work pretty well in aggregate.
Sure, but those are largely the big tech companies you’re talking about, and research tends to come from universities and private orgs.
Well, that’s because the hyperscalers are the only ones who can afford it at this point. Altman has said ChatGPT 4 training cost in the neighborhood of $100M (largely subsidized by Microsoft). The scale of capital being set on fire in the pursuit of LLMs is just staggering. That’s why I think the failure of LLMs will have serious knock-on effects with AI research generally.
To be clear: I don’t disagree with you re: the fact that AI research will continue and will eventually recover. I just think that if the LLM bubble pops, it’s going to set things back for years because it will be much more difficult for researchers to get funded for a long time going forward. It won’t be “LLMs fail and everyone else continues on as normal,” it’s going to be “LLMs fail and have significant collateral damage on the research community.”
When Mr. Altman visited TSMC’s headquarters in Taiwan shortly after he started his fund-raising effort, he told its executives that it would take $7 trillion and many years to build 36 semiconductor plants and additional data centers to fulfill his vision, two people briefed on the conversation said. It was his first visit to one of the multibillion-dollar plants.
TSMC’s executives found the idea so absurd that they took to calling Mr. Altman a “podcasting bro,” one of these people said. Adding just a few more chip-making plants, much less 36, was incredibly risky because of the money involved.
You’re absolutely right. LLMs are good at faking language and sometimes not even great at that. Not sure why I got downvoted but oh well. But AGI will be game changing if it happens.
There are people way smarter than me that claim it will be a threshold and would likely grow exponentially after it’s crossed. I guess we won’t know for sure until it happens. I do agree most people get bored easily but if this thing is possible to think for itself without interaction it won’t matter if the humans get bored.
I know nothing about anything, but I unfoundedly believe we’re still very far away from the computing power required for that. I think we still underestimate the power of biological brains.
As a fervent AI enthusiast, I disagree.
…I’d say it’s 97% hype and marketing.
It’s crazy how much fud is flying around, and legitimately buries good open research. It’s also crazy what these giant corporations are explicitly saying what they’re going to do, and that anyone buys it. TSMC’s allegedly calling Sam Altman a ‘podcast bro’ is spot on, and I’d add “manipulative vampire” to that.
Talk to any long-time resident of localllama and similar “local” AI communities who actually dig into this stuff, and you’ll find immense skepticism, not the crypto-like AI bros like you find on linkedin, twitter and such and blot everything out.
For real. Being a software engineer with basic knowledge in ML, I’m just sick of companies from every industry being so desperate to cling onto the hype train they’re willing to label anything with AI, even if it has little or nothing to do with it, just to boost their stock value. I would be so uncomfortable being an employee having to do this.
For sure, it seems like 90% of ai startups are nothing more than front end wrappers for a gpt instance.
They’re all built on top of OpenAI which is very unprofitable at the moment. Feels like the whole industry is built on a shaky foundation.
Putting the entire fate of your company in a different company (OpenAI) is not a great business move. I guess the successful AI startups will eventually transition to self-hosted models like Llama, if they survive that long.
Most projects I’ve been in contact with are very aware of that fact. That’s why telemetry is so big right now. Everybody is building datasets in the hopes of fine tuning smaller, cheaper models once they have enough good quality data.
My company is realizing that hosting a model which will be private, cost-effective, and performing better than traditional algorithms is like finding a unicorn. Few months back, the top execs were jumping around GenAI like a bunch of kids. Fortunately, the Sr. research head beat some sense into them.
You’re lucky there’s a higher up that could talk down the even higher ups. Though, sometimes it’s not even about the r&d teams.
I saw company wide HR educational emails or courses telling you how to improve you work quality/efficiency, and one of them tells us to “research AI” and learn how to utilize it, talking about how great it is and improved the work efficiency by 30%. Sure, it has its uses, but I won’t go touting how great it is. And with how ChatGPT works, you have to be the biggest idiot in the world to upload all your sensitive stuff to ChatGPT just for it to make a spreadsheet faster. But without these disclaimers in the email, I doubt regular clerical staff knows about this, and it’s extremely dangerous.
What kind of use-cases was it, where you didn’t find suitable local models to work with ? I’ve found that general “chatbot” things are hit and miss but more domain-constrained tasks (such as extracting structured entities from unstructured text) are pretty reliable even on smaller models. I’m not counting my chickens yet as my dataset is still somewhat small but preliminary testing has been very promising in that regard.
Any time you ask very domain specific questions; eg “i have collected some soil samples from the mesolithic age near the Amazon basin which have high sulfur and phosphorus content compared to my other samples. What factors could contribute to this distribution?”, both of-the-shelf local models & OpenAI fail.
The main reason is because these models are not trained on highly-specialized domains of text. Sometimes the models start hallucinating and which reduces our trust upon them.
Haha yeah the top execs were tripping balls if they thought some off-the-shelf product would be able to answer this kind of expert questions. That’s like trying to replace an expert craftsman with a 3D printer.
As someone who was working really hard trying to get my company to be able use some classical ML (with very limited amounts of data), with some knowledge on how AI works, and just generally want to do some cool math stuff at work, being asked incessantly to shove AI into any problem that our execs think are “good sells” and be pressured to think about how we can “use AI” was a terrible feel. They now think my work is insufficient and has been tightening the noose on my team.
This. Exactly.
Seriously, I’d love to be enthusiastic about it because it’s genuinely cool what you can do with math.
But the lies that are shoved in our faces are just so fucking much and so fucking egregious that it’s pretty much impossible.
And on top of that LLMs are hugely overshadowing actual interesting approaches for funding.
I think we should indict Sam Altman on two sets of charges:
A set of securities fraud charges.
8 billion counts of criminal reckless endangerment.
He’s out on podcasts constantly saying the OpenAI is near superintelligent AGI and that there’s a good chance that they won’t be able to control it, and that human survival is at risk. How is gambling with human extinction not a massive act of planetary-scale criminal reckless endangerment?
So either he is putting the entire planet at risk, or he is lying through his teeth about how far along OpenAI is. If he’s telling the truth, he’s endangering us all. If he’s lying, then he’s committing securities fraud in an attempt to defraud shareholders. Either way, he should be in prison. I say we indict him for both simultaneously and let the courts sort it out.
“When you’re rich, they let you do it.”
I really want to like AI, I’d love to have an intelligent AI assistant or something, but I just struggle to find any uses for it outside of some really niche cases or for basic brainstorming tasks. Otherwise, it just feels like alot of work for very little benefit or results that I can’t even trust or use.
It’s useful.
I keep Qwen 32B loaded on my desktop pretty much whenever its on, as an (unreliable) assistant to analyze or parse big texts, to do quick chores or write scripts, to bounce ideas off of or even as a offline replacement for google translate (though I specifically use aya 32B for that).
It does “feel” different when the LLM is local, as you can manipulate the prompt syntax so easily, hammer it with multiple requests that come back really fast when it seems to get something wrong, not worry about refusals or data leakage and such.
Attractive. You got some pretty solid specs?
Rue the day I cheaped out on RAM. soldered RAMmmm
Soldered is better! It’s sometimes faster, definitely faster if it happens to be lpddr.
But TBH the only thing that really matters his “how much VRAM do you have,” and Qwen 32B slots in at 24GB, or maybe 16GB if the GPU is totally empty and you tune your quantization carefully. And the cheapest way to that (until 2025) is a used MI60, P40 or 3090.
I receive alerts when people are outside my house, using security cameras, Blue Iris, CodeProject AI, Node-RED and Home Assistant, using a Google Coral for local AI. Entirely local - no cloud services apart from Google’s notification system to get notifications to my phone while I’m not home (which most Android apps use). That’s a good use case for AI since it avoids false positives that occur with regular motion detection.
I’ve been curious about google coral, but their memory is so tiny I’m not sure what kinds of models you can run on them
A lot of people use them for the use case I described (object detection for security cameras), using either Blue Iris or Frigate. They work pretty well for that use case.
Wake word detection is a good use case too (eg if you’re making your own smart assistant).
The Coral site lists a few use cases.
The saddest part is, this is going to cause yet another AI winter. The first few ones were caused by genuine over-enthusiasm but this one is purely fuelled by greed.
The AI ecosystem is flooded, we need a good bubble pop to slow down the massive waste of resources that our current info-remix-based-on-what-you-will-likely-react-positively-to shit-tier AI represents.
Agreed that’s why it’s so dangerous. These tech bros are going to do damage with their shitty products. It seems like it’s Altman’s goal, honestly.
He wants money/power, and he is getting it. The rest of the AI field will forever be haunted by his greed.
After getting my head around the basics of the way LLMs work I thought “people rely on this for information?”, the model seems ok for tasks like summarisation though
I don’t love it for summarization. If I read a summary, my takeaway may be inaccurate.
Brainstorming is incredible. And revision suggestions. And drafting tedious responses, reformatting, parsing.
In all cases, nothing gets attributed to me unless I read every word and am in a position to verify the output. And I internalize nothing directly, besides philosophy or something. Sure can be an amazing starting point especially compared to a blank page.
It’s good for coding if you train it on your own code base. Not great for writing very complex code since the models tend to hallucinate, but it’s great for common patterns, and straightforward questions specific to your code base that can be answered based on existing code (eg “how do I load a user’s most recent order given their email address?”)
It’s wild when you only know how to use SELECT in SQL, but after a dollar worth of prompting and 10 minutes of your time, you can have a significantly complex query you end up using multiple times a week.
That and retrieval and the business use cases so far, but even then only if the results can be wrong somewhat frequently.
Ya, it’s like machine learning but better. That’s about it IMO.
Edit: As I have to spell it out: as opposed to (machine learning with) neural networks.
I mean… it is machine learning.
It’s also neural networks, and probably some other CS structures.
AI is a category, and even specific implementations tend to use multiple techniques.
Well there is a very specific architecture “rut” the LLMs people use have fallen into, and even small attempts to break out (like with Jamba) don’t seem to get much interest, unfortunately.
Sure, but LLMs aren’t the only AI being used, nor will they eliminate the other forms of AI. As people see issues with the big LLMs, development focus will change to adopt other approaches.
There is real risk that the hype cycle around LLMs will smother other research in the cradle when the bubble pops.
The hyperscalers are dumping tens of billions of dollars into infrastructure investment every single quarter right now on the promise of LLMs. If LLMs don’t turn into something with a tangible ROI, the term AI will become every bit as radioactive to investors in the future as it is lucrative right now.
Viable paths of research will become much harder to fund if investors get burned because the business model they’re funding right now doesn’t solidify beyond “trust us bro.”
Well you say that, but somehow crypto is still around despite most schemes being (IMO) a much more explicit scam. We have politicans supporting it.
Sure, but those are largely the big tech companies you’re talking about, and research tends to come from universities and private orgs. That funding hasn’t stopped, it just doesn’t get the headlines like massive investments into LLMs currently do. The market goes in cycles, and once it finds something new and promising, it’ll dump money into it until the next hot thing comes along.
There will be massive market consequences if AI fails to deliver on its promises (and I think it will, because the promises are ridiculous), and we get those every so often. If we look back about 25 years, we saw the same thing w/ the dotcom craze, where anything with a website got obscene amounts of funding, even if they didn’t have a viable business model, and we had a massive crash. But important websites survived that bubble bursting, and the market recovered pretty quickly and within a decade we had yet another massive market correction due to another bubble (the housing market, mostly due to corruption in the financial sector).
That’s how the market goes. I think AI will crash, and I think it’ll likely crash in the next 5 years or so, but the underlying technologies will absolutely be a core part of our day-to-day life in the same way the Internet is after the dotcom burst. It’ll also look quite a bit different IMO than what we’re seeing today, and within 10 years of that crash, we’ll likely be beyond where we were just before the crash, at least in terms of overall market capitalization.
It’s a messy cycle, but it seems to work pretty well in aggregate.
Well, that’s because the hyperscalers are the only ones who can afford it at this point. Altman has said ChatGPT 4 training cost in the neighborhood of $100M (largely subsidized by Microsoft). The scale of capital being set on fire in the pursuit of LLMs is just staggering. That’s why I think the failure of LLMs will have serious knock-on effects with AI research generally.
To be clear: I don’t disagree with you re: the fact that AI research will continue and will eventually recover. I just think that if the LLM bubble pops, it’s going to set things back for years because it will be much more difficult for researchers to get funded for a long time going forward. It won’t be “LLMs fail and everyone else continues on as normal,” it’s going to be “LLMs fail and have significant collateral damage on the research community.”
deleted by creator
It is. It’s that plus an important process for living organisms rather than just burning something.
It’s selling the future, but nobody knows if we can actually get there
It’s selling an anticompetitive dystopia. It’s selling a Facebook monopoly vs selling the Fediverse.
We dont need 7 trillion dollars of datacenters burning the Earth, we need collaborative, open source innovation.
The first part is true … no one cares about the second part of your statement.
What’s the source for that? It sounds hilarious
https://web.archive.org/web/20240930204245/https://www.nytimes.com/2024/09/25/business/openai-plan-electricity.html
Yep the current iteration is. But should we cross the threshold to full AGI… that’s either gonna be awesome or world ending. Not sure which.
Current LLMs cannot be AGI, no matter how big they are. The fundamental architecture just isn’t right.
You’re absolutely right. LLMs are good at faking language and sometimes not even great at that. Not sure why I got downvoted but oh well. But AGI will be game changing if it happens.
Based on what I’ve witnessed so far, people will play with their AGI units for a bit and then put them down to continue scrolling memes.
Which means it is neither awesome, nor world-ending, but just boring/business as usual.
There are people way smarter than me that claim it will be a threshold and would likely grow exponentially after it’s crossed. I guess we won’t know for sure until it happens. I do agree most people get bored easily but if this thing is possible to think for itself without interaction it won’t matter if the humans get bored.
I know nothing about anything, but I unfoundedly believe we’re still very far away from the computing power required for that. I think we still underestimate the power of biological brains.
Very likely. But 4 years ago I would have said we weren’t close to what these LLMs can do now so who knows.
What makes you think there’s a threshold?