Sort of but Adapters allow for multiple weight adjustments (think loras) for specific skills so it is more like extra optimized mixture of experts or multi agent approach. They have a slide with adapters listed like summarization, prioritization, tone (happy, business, etc), editor, etc) -- this is not to be mixed up with Intents which is how on device apps publish their capabilities to the Intelligence system for real npu os level multi agent tool use.
While fundamentals are important to learn, there is also a huge benefit in learning specific tools and frameworks. There is no ”one size fits all” when it comes to software and more often than not, you need very customized solution for optimal performance. Moreover, learning to master a tool often means you are also able to improve it, which is part of the reason open-source tools usually improve with the loyal contributers!
> This ‘goal drift’ means that agents, or tasks done in a sequence with iteration, get less reliable. It ‘forgets’ where to focus, because its attention is not selective nor dynamic.
I don't know if I agree with this. The attention module is specifically designed to be selective and dynamic, otherwise it would not be much different than a word embedding (look up "soft" weights vs "hard" weights [1]).
I think deep learning should not be confused with deep RL. LLMs are autoregressive models which means that they are trained to predict the next token and that is all they do. The next token is not necessarily the most reasonable (this is why datasets are super important for better performance). Deep RL models on the other hand, seem to be excellent at agency and decision making (although in restricted environment), because they are trained to do so.
LLMs are fine-tuned with RL too. They are NOT simply next token predictors. RLHF uses whole answers at once to generate gradients, so it is looking further into the future. This might not be perfect but it is clearly more than focusing just 1 token ahead.
In the future the RL part of LLM training will increase a lot. Why am I saying this? There are two sources for learning - the past and the present. Training on human text is using past data, that is off-policy. But training on interactive data is on-policy. There is nothing we know that doesn't come from the environment. What is not written in any books must be learned from outside.
That is why I think supervised pre-training from human text is just half the story and RL based agent learning, interactivity in other words, is the next step. The two feed on which intelligence stands are language (past experience) and environment (present experience). We can't get ahead without both of them.
AlphaZero showed what an agent can learn from an environment alone, and LLMs show what they can learn from humans. But the world is big, there are plenty of environments that can provide learning signal, in other words feedback to LLMs.
> The facilities will reportedly consume as much as 13 percent of the plant's output.
Why are AI products being shipped so aggressively despite being so inefficient? Is code autocompletion and generating random images really worth so much electricity? Shouldn’t we wait until the research has created an efficient architecture that is easily scalable first?
What a bizarre question. Me and many others are paying $20+ for these services. The electricity cost is already priced in and subscribers looks at the price and are fine with that price. If the price comes from dev costs, electricity, capital expenses etc. is irrelevant. There are really no alternatives either (well, I could go to 99Designs, pay $100 and wait days). People pay a lot for this and every day new use cases get discovered, so clearly more services like this get built.
Edit: " Shouldn’t we wait until the research has created an efficient architecture that is easily scalable first?"
Who is "we"? If I think I can offer a product that uses AI and after paying API requests to OpenAI I can turn a profit, I'll build it. API requests will only get cheaper, but why should I wait for that if I can turn a profit now?
Between my flatrate usage and companies paying for additional API quota, I'd expect them to at least be in a position where they could be profitable if they slowed down new trading. When if that wasn't the case though it just shifts the obvious answer to the expectation of future profitability. There is clearly massive demand for this tech.
You've somehow missed all the "enshitification" processes Cory Doctorow and others have been documenting?
"Here is how platforms die: first, they are good to their users; then they abuse their users to make things better for their business customers; finally, they abuse those business customers to claw back all the value for themselves. Then, they die. I call this enshittification."
OpenAI (et al.) are all deeply in the "First, they are good to their users" stage of their business plan, but they WILL. NOT. STAY. THERE. Whatever product relying on OpenAI you can offer/build/grow into a business, sooner or later you're going to be squeezed for every cent in profit, in the same way as companies who were convinced Facebook was a great alternative to running their own web presence are these days - having to pay to ensure content that people who've like and subscribed to their pages end up seeing anything they post.
Your $20 subscription is not even close to covering electricity costs - never mind covering capitalisation costs for $1.6 billion data centres with another $1.6 billion dollars worth of GPUs to put in them. (Think about it. Even if that 3.2 billion dollar datacenter was funded borrowing at only 1% - you'd need over 130,000 paying subscribers at $20/month just to cover the interest on the loans.)
So what's your solution? Sit on your hands and build nothing, leaving potentially millions or billions of dollars on the table? Whine about things that haven't happened, may never happen, and could be worked around if they did happen? Google is successful and it charges most people nothing. So it is possible to make a successful business out of providing a service at low(ish) cost.
Of course you wouldn't build a multi-billion dollar data center for a single application bringing in a pittance. But that's not what anyone is doing. Demand is high and data centers can be shared between many different companies.
Or it could wind up being a pretty big waste of resources. On the one hand you have the ability to extract "apparent knowledge" in a conversational form. On the other hand you are killing the ability to create new knowledge and as an added bonus consolidate power into service providers that work worse than ever.
This "conversational knowledge" stuff is probably just a primitive and inefficient form of AI that may be superceded by other innovations. It need not have any real impact on new knowledge, any more than billions of ordinary people babbeling for hundreds of years. The service provider concern is understandable, but I think competition may keep this tech available to everyone who can reasonably afford it. We still have private computers and software despite the possibility for everything to be run from data centers and using proprietary subscription-based software.
"You've somehow missed all the "enshitification" processes Cory Doctorow and others have been documenting?"
I read about it, but don't believe it's a new law now that affects everything with no exceptions. Even is OpenAI wants to go from explore and expand into extract mode, open-source models are getting better every day and will create a decent floor.
> you'd need over 130,000 paying subscribers at $20/month just to cover the interest on the loans.
Maybe I live in a bubble (I definitely do), but that subscriber number seems quite reasonable and of course omits additional revenue from their APIs. Also if OpenAI wants to lose money, be my guest. I definitely enjoyed my VC-subsidized Uber rides of the past.
It's not a property of the company. It's a property of the funding mechanism. Anything getting its funds from VC or hedge-funds will eventually succumb to this - it's the capital speaking.
Ah, a true capitalist mind. The only consideration is the almighty dollar!
Perhaps it would be better for it to be efficient because it uses less energy and in general is a good thing to do for the world. I understand there aren't financial incentives around that so it's probably difficult for you to agree with.
Edit: It seems simply suggesting we do better is enough for people to lose their minds.
I was going to respond in sarcasm but realized your comment does not even deserve that.
Governments of any shape are unable to accurately pick which industry to allow or prevent. Even harder would be creating legislation that carves out specific industry.
You can tax the externalities of the use of electricity. Please don’t create pandering comments with zero substance.
I don’t like bitcoin but who am I to tell people what to spend money on.
> Governments of any shape are unable to accurately pick which industry to allow or prevent.
i would say that some gov'ts in the world does do this - famously the chinese CCP.
They dictate by providing state funding for industries they deem important (for whatever reason - geopolitically strategic, military, or welfare, etc). Whether it's the most "efficient" form of capital allocation is irrelevant (for them).
The people do not have a say, and in fact, cannot be doing any form of criticism - they may have feedback, but only in so far as the party allows said feedback. It's implicit that the CCP is picking for the "greater good" - but who is the judge of that? CCP itself is!
So by an act of tautology, they are doing it most efficiently!
Governments certainly try. The US does the similar efforts through subsidy type programs, I don't think the Chinese government are able to accurately pick industry any better. They have more force to put behind their decisions, like building ultra cheap real estate, roads and general infrastructure but I am not sure they have the ability to pick a specific piece of industry any better.
Ah yes, the one who over-allocated so much capital into real estate that it has more unused apartments than the rest of the world combined, 5 times over
With the greatest housing bubble has ever seen, and a deflationary spiral around the corner.
Why is using less energy good thing for the world? Using energy in itself is fine, you can use all the solar you want and nobody is going to complain. Nuclear is marginally less harmless, but still quite fine. Don't use coal or gas.
A person with this kind of mindset will tell less-developed nations to stop developing fossil fuel infrastructure because any more impact on the planet is unjustified.
I don't think that has to be the case for anyone in this chain. I am very conscious of my use of resources. I also realize that there is massive green washing of things. Lots of things marketed as the "green" way are also not that great for the planet, sometimes a sheet of cling wrap or a ziploc bag is ok to use. It also is true that often people love to project their feelings for saving the planet but are living a lifestyle that says otherwise.
When it comes to energy use, I don't believe anyone should have the power to say what should be allowed or not. That will stifle innovation. For every innovation there must have been thousands if not more failures.
The hypocrisy is as good a reason as any to keep it real. But the real problem is people so brainwashed by green propaganda that they think no human activity is worthwhile. This AI stuff for example might appear wasteful, but it is very promising as a means to improve efficiency in many industries and improve people's quality of life, which may in fact reduce pollution. Even if it increased pollution somehow, that has to be weighed against all other outcomes. You can't just look at any old use of resources and declare that it isn't worth it and expect everyone to agree with you. Different people value different things in life, and that's ok.
I didn't suggest otherwise. The person I replied to said it was a bizarre question, which I disagree with. Their are obvious benefits outside of money.
I don’t think you would want to live in the world we’d have if we rolled back all of the energy expenditure done in the name of profit over the past 200 years since the discovery of oil. You wouldn’t have anything. You wouldn’t be able to cross the Atlantic lest you burn coal to produce steam.
As a sailor, who crossed oceans, on nothing but wind power (and solar), I resent that - but I totally get what you’re saying. The world would be, different, to say the least.
That's a bit easy to say this. "We progressed during capitalism hence capitalism allowed the progress" but you don't know at all. You also don't know how will the next 1000 years look if we continue to do things in the name of profit. I think that making hypothetical branches like that is just useless because there are so many more factors that just "capitalism or communism"
Maybe anyone with a conscience, nuclear energy should be going to decarbonisation of the existing economy not to expanding a new industry en masse.
People are funny on this forum, when it was Bitcoin it was the devil and now it’s helping us complete our lines of code it’s ok to require a nuclear power plant to run such a system? I guess whatever benefits the individual is ok ? Silicone valley types are funny too because it was all about developing fusion to save humanity now it’s about developing fusion to cat gifs?
These products need to be taxed to price in these externalities we can’t keep this up, it’s ridiculous.
"These products need to be taxed to price in these externalities we can’t keep this up, it’s ridiculous."
The negative externality you are worried about isn't AI, but pollution/carbon emissions. Tax pollution instead of deciding for other people what use case is worth it.
Edit: To be clear, I am all for this. I'd love to see a massive(!) carbon tax.
Bitcoin doesn’t have any real utility besides being a toy or committing crimes.
Machine learning has utility, not just in theory but to plenty of people, including me. It has helped me do my job faster.
You can put tax on energy from carbon, fine. There’s no justification for taxing energy intensive things by themselves because you don’t like using energy. Human energy usage is and should be growing exponentially (and not an S curve) besides occasional pauses when efficiencies are found (like LED lighting using something like 90% less energy than incandescent).
Fossil fuels are clearly going away anyway, not because of ideals either, solar is just way cheaper now.
There's no good reason for BTC to be set up to use that much energy. There are viable, proven alternatives like proof of stake. I'd also be more forgiving if BTC only required a given amount of power, but no, the system is designed such the power requirement is related to the cost of flops/watt - it's specifically designed such that the system will use more power as power becomes cheaper.
No but as others are arguing in this thread, “power is fungible” etc etc, “it’s not up to you to decide where the energy goes”. Double standards in my opinion.
The difference is that, by design, the best (cheapest!) Bitcoin you can buy is the Bitcoin you buy today. It only gets more expensive as more people use it and the supply of new Bitcoin decreases.
By contrast, the ai we have today is the worst ai it will ever be. Quality will go up with time, and price will go down, as incentives drive the development of better, cheaper models.
Look at how neural voice synthesis has evolved over the last five years. It's gone from science fiction to extremely expensive to 'can run in real time on a phone' thanks to a combination of hardware and algorithmic improvements... But mostly algorithmic improvements.
Algorithmic improvements are coming quickly in generative ml. We're still in very early days.
"Is code autocompletion and generating random images really worth so much electricity?"
Yes. Very much so.
Some back of enveloping. The worldwide labor market is USD$70trillion. If you save 1% of the time of the worldwide labor market (that's quite reasonable for autocomplete etc right?) that's USD$700billion. A quick Google says the global electricity revenues are USD$255billion.
So if you have a way to save %1 of human labor it's economically worth tripling power generation to make use of it.
Where did that $70T come from? The most obvious objection is that most labor is not desk jobs, so 1% from autocomplete seems very unreasonable to me, but I don’t know how it breaks out in dollars. Maybe your source says that?
It's admittedly super rough. I googled worldwide labor market and got 3.5billion workers. I then Googled average global earnings, $20k. However i'll admit to a mistake. The second number appears incorrect and it seems it should have been ~$10k based on the following: https://www.zippia.com/advice/average-income-worldwide/#:~:t....
Still i think the point should stand no matter how precise we get here. The cost of labor absolutely dwarfs the cost of electricity. That was what i wanted to hit on while completely acknowledging the numbers are rough.
Not sure that llm autocomplete is so much better than autocomplete using more energy efficient algorithms to make a noticeable difference in productivity.
Hmm... so it's so bad that you might as well just use old-school autocorrect, and so good that you have to use it to remain competitive, at the same time?
But that is exactly what GP is criticizing: just looking at the $$$. All of us here know that the environmental impact of electric energy generation is externalized and global. If we were to price that in fairly, the labor might well be cheaper!
The jaquard loom and steam engine can't do those things either. A revolution occurs when a technology saves even a fraction of worldwide labor costs. That's all it needs to do to be literally revolutionary.
And have you studied anything into modern mining yourself? Mines are starting to convert fleets into self driving trucks. Capabilities like image/video recognition that LLMs can do improve their identification and alert systems.
Modern AI is a whole lot more than just shitting out spam on the internet.
To be clear we're talking about less than 200 trucks by my count, most of which were built in 2013-2016, the Baghdad Arizona mine still hasn't deployed anything yet as far as I know (I asked people in that area). None of this has anything to do with code completion or image generation and LLMs don't even do anything semi-related to vision. That's entirely beyond the scope of how they're designed.
It is not surprising to me anymore the threads that get pulled to justify this insane hype, but what does shock me is that it's going on 8 months of the same very tired and thin arguments that get smaller and smaller in their claims the more questions you ask.
Global electricity production is about 30 PWh per year and wholesale electricity costs are about $100 / MWh around the world. That’s about $3T. (Numbers are rough, don’t get pedantic)
Global electricity generation revenues are only $255 billion? Amazon/Google/Apple have greater revenue than the worldwide electricity generation sector?
I just Googled these numbers quickly but yes it seems correct.
"In 2021, the global electricity transmission and distribution market generated some 255.6 billion U.S. dollars in revenue, up from 248.5 billion U.S. dollars in the previous year"
> Shouldn't we wait until research has created an efficient architecture that is easily scalable first?
A scalable architecture has a prerequisite of known product requirements and known implementation specs. Neither is true right now as this field is evolving at a crazy rate. Approaches like what Groq has are great for increasing efficiency, but at the cost of flexibility. Early on in any technology flexibility is imperative until product/market fit is established more clearly.
Also reducing this to autocompletion/generating random images really reduces what AI actually is. Those are two use cases for a very generalizable area.
A problem is that we can't measure how efficient a computing architecture really is, we can only compare them, and not to a hypothetical machine yet to be devised.
When AWS can run their datacenter operations 20x more efficiently, which sounds "reasonable to me", they obviously will. Are you criticizing their customers? Or their customers' customers, the ones who like autocomplete? Literally the computer engineers, by whose work these applications have been enabled?
I think this is why you were perceived as calling for central planning. Ironically because you took aim at every level but the investors.
Growth is aggressive because of the huge investment in the sector currently. Literally because investors want a return. For better or worse there is nobody in the chain who interprets "worthwhile?" in the same sense, and the applications are not essentially relevant to the people in command.
Why are companies pouring money into LLMs? Well the naysayers like yourself are simply shortsighted and being negative. I think many businesses are already seeing the power of current models in specific processes. You have to assume some natural progression in the coming years to LLMs and the general ecosystem.
Since you only reference autocompletion and image gen you obviously have not worked with using LLM in business processes. Is there a lot of hype? Certainly, but there is also a lot of value in the universe today for these models.
How do you even define "efficient architecture"? Nobody can define that.
Are they? It could be a tax to offset the externalities that are currently being socialized by keeping other coal plants running to sustain existing homes, farms, factories, etc.
But all the industries that consume any electricity are already "keeping other coal plants running".
If you abitrarily decice a metal processing factory is "worth it" while an AI datacenter isn't, then yes, it's planned economy.
And yeah, I know the governments already tax different industries differently. I think these policies are hardly justified and mostly "redistribution by lobbying".
Arguably some industries make us all richer than others. So we incentivize those with more collective benefit, say manufacturing or farming, and disincentivize those with more collective costs, say gambling or tobacco.
That’s already being done, and the taxes/limits on co2 pollution will be increasing. So nothing new here.
What you’re implicitly suggesting is government allowing/banning certain uses of technology because they are deemed not worthy. That is an autocratic style of thinking.
We as a society can decide we don't want our air to become toxic to prop up convoluted gambling schemes. It's only autocratic if it's decided by a single person.
Shouldn't that be factored into the price of electricity? I think the market economy ideal is that if houses can't afford enough electricity for heating, then they need to be insulated and/or install heat pumps, etc.
Maybe it’s time to build new power plants, preferably nuclear, and continue efforts to learn about and expand efficient energy sources. No society has ever advanced without ample energy supplies.
Isn’t that what the softmax layer is doing? The token with highest probability among all the available tokens in the model dictionary is chosen as the next token!
Rhetorically speaking it's simple to make the biggest in some class seem to be worst in some related class. If I'm the biggest taxi operator in town, I'm likely to have the highest CO₂ emissions of all the taxi companies, see? You can omit or deemphasise the connection between the two and make me seem bad.
Buses profit from pedos spying on kids in them.
Google and Apple profit from pedos using their phones to spy on kids.
Public infrastructure provides convenience for pedos to do their dirty work.
I can continue like that for a long time, you know.
