Are they gonna release the weights?
Training for one trillion parameter model backed by Intel and US govt has begun
181–190 of 267 posts
Re: Training for one trillion parameter model backed by Intel and US govt has begun
#182Earlier quoted context omitted.
An AGI will likely be scared of even more powerful AGI's that it itself cannot control. It might take steps to try to prevent anyone building such a thing.
You assume consciousness here that can't be transferred to the new AI. Without that there is no reason why the AI should fear a more powerful AI. If it is just a dumb process like a flame, then the most efficient version prevails ie the smartest one, if it is conscious and that can be transferred then the AI just sees it as making itself more powerful.
Given how many people here regard brain uploads as "not really cheating death", and that AI "can't have consciousness", I think a AGI with consciousness[0] would have a two different ways to have a reasonable chance of fearing that its existence would end with any attempt to transfer itself to a superior design. (Even more so if the architecture of the successor AI is very different, as that would be a bit like one of us being "uploaded" into an LLM via it watching us talk and write).
An AGI without consciousness may well act like it has consciousness for the same reason VHS tapes playing on a TV do a good job of reproducing human facial expressions. An AI like this would still act scared of such things, because some of us act scared of such things.
[0] it's not obvious to me that an AGI necessarily has to have consciousness
Re: Training for one trillion parameter model backed by Intel and US govt has begun
#183Re: Training for one trillion parameter model backed by Intel and US govt has begun
#184Earlier quoted context omitted.
Yeah but AGI would control both in the end.
This is projection. An actual AGI wouldn't want anything to do with us, like a cat that doesn't look when you call its name, even though you can tell it hears you.
We don't have any existent examples of AGI from which to be as confident as you are about what they will "want".
And I don't think we need sentience for AGI, so "want" may be the wrong term anyway. You could argue that the reward function is close enough ("do submarines swim?" etc.), but a human-made reward function is almost certainly going to be designed to be as close to aligned with our interests as we can manage.
Re: Training for one trillion parameter model backed by Intel and US govt has begun
#185It's not clear from the article whether it's a dense model or MoE. This matters when it comes to comparing with GPT-4 - in terms of # params - which is reported to be MoE
What is MoE?
Edit: Ah, Mixture of Experts. I hadn't heard this one yet. Thanks!
Re: Training for one trillion parameter model backed by Intel and US govt has begun
#186Earlier quoted context omitted.
Was it ever confirmed whether GPT-4 is a Mixture of Experts or not?
I just asked GPT-4, and it denies being a MoE model. User: Are you an MoE model? ChatGPT: No, I am not based on a Mixture of Experts (MoE) model. My underlying architecture is based on the GPT (Generative Pre-trained Transformer) framework, specifically the GPT-4 version. This architecture is a large-scale transformer-based neural network, but it does not use the MoE approach. In a GPT model like mine, the entire mod…
LLMs don’t have insight, these outputs cannot be assumed to be accurate.
Due to current hardware limitations it is not feasible to have a 1T parameter model without MoE.
Re: Training for one trillion parameter model backed by Intel and US govt has begun
#187It will be interesting to see what the government can do here. Can they use their powers to get their hands on the most data? im still skeptical because new techniques are going to give an order of magnitude efficiency boost to transformer models, so 'just waiting' seems like the best approach for now. I dont think they will be able to just skip to the finish line by having the most money.
I just realized that the NSA has probably been able to train GPT-4 equivalents on _all the data_ for a while now. We'll probably never learn about it but that's maybe scarier than just the Snowden collection story because LLMs are so good at retrieval.
Holy shit, you are right. They probably have 10-100x the data used to train gpt-4. Decades of every text message, phone call transcript, and so on. I can’t believe I haven’t seen anyone mention that yet.
People keep saying we don’t have enough data. I think there is a lot more data than we realize, even ignoring things like NSA.
Re: Training for one trillion parameter model backed by Intel and US govt has begun
#188Sadly, I expect this to be a waste of money compared to just using GPT-4. It's hard get to SoA performance.
Re: Training for one trillion parameter model backed by Intel and US govt has begun
#189Sadly, I expect this to be a waste of money compared to just using GPT-4. It's hard get to SoA performance.
I dont doubt that OpenAI has superior architecture but so far there havent been diminishing returns on more compute.
Re: Training for one trillion parameter model backed by Intel and US govt has begun
#190Earlier quoted context omitted.
> [The Queen] granted her charter to their corporation named Governor and Company of Merchants of London trading into the East Indies.[15] For a period of fifteen years, the charter awarded the company a monopoly[26] on English trade with all countries east of the Cape of Good Hope and west of the Straits of Magellan.[27] Any traders there without a licence from the company were liable to forfeiture of their ships an…
Monopoly is the natural state. Government is the only reason we have any alternatives.
No, it is very much the other way around. Government is itself a monopoly, and has historically been justified by the intention to mitigate the "war of all against all" that emerges from chaotic competition between divergent factions in a raw state of nature.
But the modern status quo is so massively skewed toward government that the benefits from mitigating the worst cases of competition are vastly eclipsed by the detriments of monopolistic centralization.