> a cryptographically secure, decentralized ledger is the only solution to making AI safer.
This article manages to say that a consensus algorithm is the answer to problems with AI without using the word "consensus" a single time. Probably because it doesn't even consider that's what blockchains are.
The article talks about using "proof of gradient" to do inference instead of crunching hashes. But this is nonsense, because inference takes inputs and produces a deterministic output. There's no mining. Checking the work takes the same resources as doing the work. Proof of work output can be checked with essentially a single hash.
As much as this would be wonderful in a universe where it's possible, it's simply not possible. The author throws out a bunch of buzz words for things that sound similar in AI and crypto and tries to make them sound like they are interchangeable. They're not.
It’s standard for white papers to write equation and theories that takes effort to prove wrong and it usually is(get the lazy guys to invest), even just by skimming the few parts that the analogy to finding the right hashes is to calculate gradients does not make any sense, I just stop there.
> Checking the work takes the same resources as doing the work.
> As much as this would be wonderful in a universe where it's possible, it's simply not possible.
We do live in that universe, under some currently believed assumptions. An NP-complete problem is an example of something where checking a solution is (thought to be) easier than finding one.
Zero-knowledge proofs make it such that checking that a computation (such as inference) has been done correctly is easier than doing the computation (even keeping some parts private). A great reference is here: https://people.cs.georgetown.edu/jthaler/ProofsArgsAndZK.pdf
That's pointless work because there's no incentive to do the "mining" (because nobody "wins" ). If you implement a system to choose a "winner" then you didn't need the computations in the first place because you've invented proof of stake.
This seems half-baked and there are numerous faulty assumptions in this article.
For example, Bitcoin miners cannot computer gradients. Their ASICs can only calculate double-sha256.
Additionally, the premise of sending gradients of models trained on private data while retaining privacy seems problematic. While you likely can't reverse it to calculate the batch's contents, it is leaky.
Further, gradient calculations are not a good proof of work. A good proof of work is difficult to calculate and easy to verify (i.e. an extremely low hash value).
The core premise, using blockchains to store terabytes of models and datasets, doesn't make any sense whatsoever.
The problems highlighted in this article however are valid, and it would be great to see something like IPFS for datasets and models.
This was my reaction too. How do you “prove” my gradient is valid?
Well perhaps one way is you could have another LLM take a look at the data you are submitting and have it predict p(useful|not useful) , and create an incentive for users to generate authentic data
This must be quite an insult to the inventor of fully decentralised AI. In 2012 Prof. Jelasity invented gradient mining with SGD, https://arxiv.org/pdf/1109.1396
Works reasonably well, you probably also need a ledger like Trustchain for accountability and Sybil protection. Consensus is too costly. This problem is unsolved and actively worked on academia for decades.
> A good proof of work is difficult to calculate and easy to verify (i.e. an extremely low hash value).
That's a specific Proof-of-Work system known as Hashcash [1].
There are many others, such as Cuckoo Cycle, in which the solution is a fixed-length cycle in a random bipartite graph [2].
Very important topic, but this is a bad article whose authors have their heads too far into cryptocurrency which IMHO is a mostly failed approach to decentralization.
The decentralized AI that is working is just sharing models. Splitting the inference across the net for cooperative execution is possible.
The thing that really needs to be solved is distributed training. We need to be able to train base models in a manner more like Folding@Home so we’re not captive to organizations with big expensive GPU farms to make any innovation.
In order to solve inequality as a result of AI, you don't need to decentralize AI, you can just decentralize the government of AI (and the AI remains centralized). And of course people who would work with that AI would be subject to the law, so if they did a different thing that people democratically decided, they would be penalized.
And that is safer because if everyone has access to the model, then people can use it in unsafe way. And it would be much more difficult to enforce the laws, when you have to control what everyone in the world does, than just one or few entities (companies, organizations, governments, whatever...). Unless it talks about decentralizing it in such way that different nodes store different parts of the model, instead of all nodes having the entire model (in that case, it's fine).
It's valuable to examine the challenges in machine learning without assuming decentralization as a solution:
> High Cost and Resource Requirements
For training and local inferencing use, quantization may help. Problem becomes local via quantization vs. remote full tensor use. Solution may involve distributed inferencing. Techniques like model distillation can help create smaller, more efficient models for inferencing.
> Data Privacy
For training, some private datasets may be needed. For local inferencing use, determining what needs to be inferenced locally vs. what needs to be run remotely may be useful. Problem becomes privacy scope mapped onto a marketplace to mitigate high cost and resource requirements. Techniques like model explainability (versioning) and robustness testing can help build trust in AI systems.
Complying with data privacy regulations and ensuring that AI systems adhere to legal and ethical standards can be a challenge, especially in international contexts.
> Incentives
Instead of assuming the solution when considering the problem, we assume there is an incentive to either simply train a model or use one. Problem becomes financial rewards, data access agreements, or even altruistic motivations.
> Stale Data and Reproducibility
Both the code and datasets for training the model need to be updated. Inferencing needs RAG, so the augmented reference data needs to be updated as well. Anything updated might need some type of revision control, especially if that data (or code) results in poor output. Labeling data and knowledge transfer are other problems that needs revision control.
> Interoperability
We can assume a marketplace for a ML train/inference platform is needed. We have HuggingFace, for example. The problem here is likely based on the tendency for datasets to be private, such as in the case of Llama 2. Models contain the "essence" of the dataset, but we still need RAG to ground the responses.
The use of the Lightning Network combined with a proposed 402 response code is an interesting concept for addressing some of these challenges: https://github.com/lightninglabs/aperture
It could provide a decentralized and efficient way to facilitate payments for dataset access, training, and inferencing to incentivize data sharing and model usage.
> All of this may sound a little ridiculous but it’s not.
Actually it is. But in perverse way it isnt.
"AI" is currently being developed and commercialized in what is essentially an old-fashioned way, with a whiff of theft of intellectual property of the training material never too far.
Decentralized AI is a distinct destination but the technologies that might realize it need not have anything remotely related to blockchain (a solution that was conceived for a different - perceived - problem, the centralization of money creation)
> The academic literature is ridden with examples of state of the art (SOTA) models that weren’t reproducible
Probably too late, but Elsevier could fix its bad reputation if they required relevant academic papers to submit reproducible code and subjected them to automated testing.
although zero knowledge proofs can compress the "proof of gradient" so it takes much less time to verify than recompute, it take orders of more work to generate the proof, moreover can we merge multiple "improvements"? if not then upon the selection of the "winner" all those are lost?
Do you want Terminator? Because this is how you get Terminator. A blockchain-based distributed giant AI could go rogue and it would be very difficult to stop it.
Wouldn't it be reliant on the nodes of the blockchain continuing to run it?
And if the crypto-economy is doing well enough that node-runners are economically incentivized to continue running, then is the rogue AI even that harmful?
Chaos engineering may sound cool and cyberpunky but it'd not what we need right now for AI safety. Unless you guys want to have your kids tortured by moloch
A way for the humanities to get some of the CS STEM funding by writing sci-fi.
Edit: The number of AI safety sessions I’ve joined where the speakers have no real AI experience talking about potentially bad futures, based on zero CS experience and little ‘evidence’ beyond existing sci-fi books and anecdotes, have left me very jaded on the subject as a ‘discipline’.
I believe it comes down to three groups:
AI researchers/organisations wanting to make their work sound very important/scary
Humanities researchers wanting STEM funding
AI organisations trying to bring in legislation to slow down competition
Yes, there is a lot of bunk AI safety discussions. But there are legitimate concerns as well. AI is close to human level. Logically they become dangerous, specially if given autonomy and bad goals. Many of the accredited researchers recognize this.
There is some level that you can discuss AI safety without AI expertise (specially as of a few years ago where everything as so uncertain), but I think currently you need a lot of awareness of physical and computational limits. Taking those limits into account, we're clearly very close to human level intelligences that can scale in unpredictable ways (probably not "grey goo" ways), but potentially dangerous ways under various scenarios, including manipulating our digital lives if there are humongous AI systems controlling everything as we are in danger of getting into as a society.
I think there's also a lot of elitism toward humanities implied that you should try to get past too. Humanities have a lot of insights about human nature, even if not all of it is reliable. See philosophers like Derek Parfit.
(in case you're wondering, I've implemented a few AIs mostly RL algorithms)
The thing I always get caught up on, when making comparisons between computers and humans wrt autonomy, is that the computer reaches the output state from the input by a clock cranking the CPU forward, ie it's a function that runs when the environment around it forces it to run. To put it in the LLM context, between words and after a Stop token, the "intelligence" is dead - frozen - suspended until the next function call.
How can a machine, then, possess anything like self-directed behavoir, when it never has a sense of self-preservation? Basically this is my axiom, that sense of self requires fear/awareness of mortality and the good sense to avoid those things that end you.
Perhaps you could concoct a machine that runs in an infinite loop with no off switch, I guess my question for you is, in what way can a machine have autonomy?
And my distinction between living and dead might be, a living system acts out of self preservation, consuming or modifying its environment to survive/thrive, while a dead system is simply acted upon by the environment its embedded in, like a crystal growing due to molecular force and temperature gradient - or an adding machine being cranked by a higher being.
It is very obvious you have never read the book Superintelligence or any literature on the subject because you try to post here shower thoughts. But here is the thought of the most highly rated h-index computer scientist in the world to help out:
https://yoshuabengio.org/2023/06/24/faq-on-catastrophic-ai-r...
Perhaps you'll categorize this in the first group, but I've found Robert Miles' arguments/videos to be convincing, and he was talking about it before it became particularly fashionable to do so:
AI safety for a lot of the movie-style concerns is pretty easy: don't connect it to the f*king nuclear launch system you moron. Ensure that any output on the launch decision path is checked by a human that actually engages their brain.
(Alas if you wanted to avoid "killer robots" you're already too late, see that whole episode where a certain large military power decided that target selection of cellphones to kill due to "metadata" could be achieved by a system not unkind to the facebook algorithm, with the only safeguard some wet-behind-the-ears kid who's piloting the drone from half way around the world and a "computer says die" attitude).
most of the most prominent AI researchers including 2 out of 3 Turing award winners + Ilya Sutskever are very very sold on AI safety. Imagine being a rando on the internet calling their research a joke. Imagine the audacity and ego. Human ego is such a weird thing.
Imagine thinking I should care what unproven doom 2 random AI Researchers have speculated on. Imagine being so afraid of the unknown that you can't think for yourself.
This article manages to say that a consensus algorithm is the answer to problems with AI without using the word "consensus" a single time. Probably because it doesn't even consider that's what blockchains are.
The article talks about using "proof of gradient" to do inference instead of crunching hashes. But this is nonsense, because inference takes inputs and produces a deterministic output. There's no mining. Checking the work takes the same resources as doing the work. Proof of work output can be checked with essentially a single hash.
As much as this would be wonderful in a universe where it's possible, it's simply not possible. The author throws out a bunch of buzz words for things that sound similar in AI and crypto and tries to make them sound like they are interchangeable. They're not.