[HN Gopher] Meta AI: "The Future of AI Is Open Source and Decent...
___________________________________________________________________
Meta AI: "The Future of AI Is Open Source and Decentralized"
Author : alexandercheema
Score : 151 points
Date : 2024-09-18 17:40 UTC (5 hours ago)
(HTM) web link (twitter.com)
(TXT) w3m dump (twitter.com)
| jsheard wrote:
| Decentralized inferencing perhaps, but the training is very much
| centralized around Metas continued willingness to burn obscene
| amounts of money. The open source community simply can't afford
| to pick up the torch if Meta stops releasing free models.
| leetharris wrote:
| There's plenty of open source AI out there that isn't Meta.
| It's just not as good.
|
| The #1 problem is not compute, but data and the manpower
| required to clean that data up.
|
| The main thing you can do is support companies and groups who
| are releasing open source models. They are usually using their
| own data.
| jsheard wrote:
| > There's plenty of open source AI out there that isn't Meta.
| It's just not as good.
|
| To my knowledge all of the notable open source models are
| subsidised by corporations in one way or another, whether by
| being the side project of a mega-corp which can absorb the
| loss (Meta) or coasting on investor hype (Mistral,
| Stability). Neither of those give me much confidence that
| they will continue forever, especially the latter category
| which will just run out of money eventually.
|
| For open source AI to actually be sustainable it needs to
| stand on its own, which will likely require orders of
| magnitude more efficient training, and even then the data
| cleaning and RLHF are a huge money sink.
| exe34 wrote:
| if you can do 100x more efficient training with open
| source, closeAI can simply take that and train a model
| that's 100x bigger/longer/more tokens.
| bugglebeetle wrote:
| AKA why Unsloth is now YC backed for their even better
| (but closed source) fine-tuning.
| moffkalast wrote:
| https://huggingface.co/datasets/HuggingFaceFW/fineweb
|
| The #1 problem is absolutely compute. People barely get
| funding for fine tunes, and even if you physically buy the
| GPUs it'll cost you in power consumption.
|
| That said, good data is definitely the #2 problem. But
| nowadays you can just get good synthetic datasets from
| calling closed model APIs or just using existing local LLMs
| to sift through trash. That'll cost you too.
| citboin wrote:
| >The main thing you can do is support companies and groups
| who are releasing open source models. They are usually using
| their own data.
|
| Alternatively we could create standardized open source
| training data like wikipedia, wikimedia as well as public
| domain literature and open courseware. I'm sure that there
| are many other such free and legal sources of data.
| KaiserPro wrote:
| but the training data _is_ one of the key bits that makes
| or breaks your model 's performance.
|
| There is a reason why datasets are private and the model
| weights aren't.
| Der_Einzige wrote:
| Compute is for sure the number one problem. Look at how long
| it's taking for anything better than Pony Diffusion to come
| out for NSFW image gen despite the insane amount of demand
| for it.
|
| Look at how much computer purple AI actually has. It's
| basically nothing.
| cynicalpeace wrote:
| One area that's interesting, but easy to dismiss because it's
| the ultimate cross-section of hype (AI and crypto) is
| bittensor.
|
| AFAICT it decentralizes the training of these models by giving
| you an incentive to train models which will mine the crypto if
| you're improving it.
|
| I learned about it years ago, mined some crypto, lost the keys
| and now kicking myself cuz I would've made a pretty penny lol
| jsheard wrote:
| Does it actually work? AIUI the current consensus is that you
| need massive interconnect bandwidth to train big models
| efficiently, and the internet is nowhere near that. I'm sure
| the Nvidia DGX boxes have 10x400Gb NICs for a reason.
| cynicalpeace wrote:
| I have no idea. The idea is certainly interesting but I've
| never actually understood how to run inference on these
| models... the people that run it seem to be unable to just
| talk simply.
| CaptainFever wrote:
| I've seen bittensor before. I think it makes sense, as a
| way to incentivise people to rent their GPUs, without
| relying on a central platform. But I've always felt it
| was kind of a scam because it was so hard to find any
| guides on how to use it.
|
| Also, this doesn't seem to actually solve the issue of
| fine tuners needing funding to rent those GPUs? One
| alternative is something like AI Horde, which pays GPU
| providers with "labour vouchers" that allow them to get
| priority next time they want GPU. Requires a central
| platform to track vouchers and ban those who exchange
| them. Basically a sort of real-life comparison of
| mutualism (AI Horde) vs capitalism (bittensor).
| bloatedGoat wrote:
| There are methods that make it feasible to train models
| over the internet. DiLoCo is one [1] and NousResearch has
| found a way to improve on that using a method they call
| DisTro [2].
|
| 1. https://arxiv.org/abs/2311.08105
|
| 2. https://github.com/NousResearch/DisTrO?tab=readme-ov-
| file
| numpad0 wrote:
| Centralized production, decentralized consumption.
| pjkundert wrote:
| The future of _everything_ you depend on is open source and
| decentralized.
|
| Because all indications are that the powers over you cannot abide
| your freedoms of association, communication and commerce.
|
| So, if it's something your family needs to survive - it has
| better be distributed and cryptographically secured against
| interference.
|
| This includes interference in the training dataset of whatever
| AIs you use; this has become a potent influence on the formation
| of beliefs, and thus extremely valuable.
| caeril wrote:
| It's not the training dataset.
|
| All of these models, including the "open" ones, have been
| RLHF'ed by teams of politically-motivated people to be "safe"
| after initial foundation training.
| pjkundert wrote:
| And I'm not even _remotely_ interested in the "corrections"
| supplied by some group of right-thinking meddlers!
|
| This corruption _must_ be disclosed as assiduously as the
| base dataset, if not more so.
| pjkundert wrote:
| Or, at _least_ package them up as "personnas" and give
| them an appropriate name, eg. "Church Lady", "Jr. Marxist
| Barista", "Undergrad Philosophy Major", ...
|
| Actually, those seem like an apt composite description of
| the PoV of the typical mass-market AI... 8/
| Der_Einzige wrote:
| Not mistrals. Mistral large is willing to tell me how to
| genocide minorities or NSFW without any kind of
| orthogonalization or fine tuning. Please actually try models
| instead of pontificating without evidence.
| pjkundert wrote:
| I wasn't aware that there was any publicly accessible
| interface to the Mistrals (or any other) models without
| training-wheels!
| Qshdg wrote:
| Great, who gives me $500,000,000, Nvidia connections to actually
| get graphics cards and a legal team to protect against copyright
| lawsuits from the entities whose IP was stolen for training?
|
| Then I can go ahead and train my open source model.
| riku_iki wrote:
| you can pick existing pretrained foundational model from corp
| (google, MS, Meta) and then finetune it(much cheaper) with your
| innovative ideas.
| monkeydust wrote:
| Curious but is there a path where llm training or inference could
| be distributed across the BOINC network:
| https://en.m.wikipedia.org/wiki/Berkeley_Open_Infrastructure...
| alecco wrote:
| * pre-trained models
|
| * does not apply to training data
| exabrial wrote:
| only when it's financially convenient for them...
| dkga wrote:
| Well, yes. They are a company, with shareholders and all. So
| while not breaching any law, they should indeed pursue
| strategies that they think would be profitable.
|
| And for all the negativity seen in many of the comments here I
| think it's actually quite remarkable that they make model
| checkpoints available freely. It's an externality, but a
| positive one. Not quite there yet in terms of the ideal - which
| is definitely open source - and surely with an abuse of
| language, which I also note. But overall, the best that is
| achievable now I think.
|
| The true question we should be tackling is, is there an
| incentive-compatible way to develop foundation models in a
| truly open source way? How to promote these conditions, if they
| do exist?
| nis0s wrote:
| I like the idea of this! But is there any reason to be concerned
| about walled gardens in this case, like how Apple does with its
| iOS ecosystem? For example, what if access to model weights could
| be revoked.
|
| There is a lot of interest in regulating open source AI, but many
| sources of criticism miss the point that open source AI helps
| democratize access to technologies. It worries me that Meta is
| proposing an open source and decentralized future because how
| does that serve their company? Or is there some hope of creating
| a captive audience? I hate to be a pessimist or cynic, but just
| wondering out loud, haha. I am happy to be proven wrong.
| candiddevmike wrote:
| Stop releasing your models under a non FOSS license.
| hyuuu wrote:
| the view of the comments here seems to be quite negative for what
| meta is doing. Honest question, should they go to the route of
| openai and closed source + paid access instead? OpenAI or Claude
| seem to garner more positive views than llama open sourced.
| naming_the_user wrote:
| The models are not open source, you're getting the equivalent
| of a precompiled binary. They are free to use.
| RealStickman_ wrote:
| Free to use with restrictions, so you maybe get 1.5/4 FOSS
| freedoms.
| meiraleal wrote:
| Not much would change if they did. Meta intentions and OpenAI
| intentions are the same: reach monopoly and take all the
| investment back with a 100x return. Anyone that achieves it
| will be as evil as the other one.
|
| > OpenAI or Claude seem to garner more positive views than
| llama open sourced.
|
| that's more about Meta than the others. Although OpenAI isn't
| that far from Meta already.
| troupo wrote:
| They use "open source" to whitewash their image.
|
| Now ask yourself a question: where does Meta's data come from?
| Perhaps from their users' data? And they opted everyone in by
| default. And made the opt-out process as cumbersome as
| possible:
| https://threadreaderapp.com/thread/1794863603964891567.html And
| now complain that the EU is preventing them from "collecting
| rich cultural context" or something
| https://x.com/nickclegg/status/1834594456689066225
| KaiserPro wrote:
| > Perhaps from their users' data?
|
| nope, not yet.
|
| FAIR, the people that do the bigboi training, for a lot of
| their stuff cant even see user data, because the place they
| do the training can't support the access.
|
| Its not like openAI where the lawyers don't even know whats
| going on, because they've not yet been properly taken to
| court.
|
| at Meta, the lawyers are _everywhere_ and if you do naughty
| shit to user data, you are going to be absolutely fucked.
| VeejayRampay wrote:
| people are just way too invested in the OpenAI hype and they
| don't want people threatening that in any way
| rkou wrote:
| And what about the future of social media?
|
| This is such devious, but increasingly obvious, narrative
| crafting by a commercial entity that has proven itself
| adversarial to an open and decentralized internet / ideas and
| knowledge economy.
|
| The argument goes as follows:
|
| - The future of AI is open source and decentralized
|
| - We want to win the future of AI instead, become a central
| leader and player in the collective open-source community (a
| corporate entity with personhood for which Mark is the human
| mask/spokesperson)
|
| - So let's call our open-weight models open-source, and benefit
| from its imago, require all Llama developers to transfer any
| goodwill to us, and decentralize responsibility and liability,
| for when our 20 million dollar plus "AI jet engine" Waifu
| emulator causes harm.
|
| Read the terms of use / contract for Meta AI products. If you
| deploy it, some producer finds the model spits out copyrighted
| content, knocks on Meta's door, Meta will point to you for the
| rest of the court case. If that's the future for AI, then it
| doesn't really matter whether China wins.
| foobar_______ wrote:
| It has been clear from the beginning that Meta's supposed
| desire for an open source AI, is just a coping mechanism for
| the fact that got beat out of the gate. This is an attempt to
| commoditize AI and reduce OpenAI/Google/Whoever's advantage. It
| is effective, not doubt, but all this wankery about how noble
| they are for creating an open-source AI future is just
| bullshit.
| CaptainFever wrote:
| I feel the same way. I'm grateful to Meta for releasing libre
| models, but I also understand that this is simply because
| they're second in the AI race. The winner always plays dirty,
| the underdog always plays nice.
| KaiserPro wrote:
| but they've _always_ released their stuff. Thats part of
| the reason why the industry uses pytorch, that and because
| its better than tensorflow.
|
| In the same way that detectron and Segment anything is an
| industry standard.
|
| Sure, for LLMs openAI released a product first. but its not
| unusual for meta to release useful models.
| ipsum2 wrote:
| You're wrong here. Meta has released state of the art open
| source ML models prior to ChatGPT. I know a few successful
| startups (now valued at >$1b) that were built on top of
| Detectron2, a best-in-class image segmentation model.
| Calvin02 wrote:
| Doesn't Threads and Fediverse indicate that they are headed
| that way for social as well?
| redleader55 wrote:
| The last time we had a corporate romance between an open
| source protocol/project, "XMPP + Gtalk/Facebook = <3", XMPP
| was crappy and it was moving too slowly to the mobile age.
| Gtalk/Messenger gave up on XMPP and evolved their own
| protocols and stopped federating with the "legacy" one.
|
| I think the success of the "Threads + Fediverse = <3" relies
| on the Fediverse not throwing the towel and leaving Threads
| as the biggest player in the space. That would mean fixing a
| lot of problems that that people have with Activity Pub
| today.
|
| I don't want to say the big tech are awesome and without
| fault, but at the end of the day big-techs will be big-techs.
| Let's keep the Fediverse relevant and Meta will continue to
| support it, otherwise it will be swallowed by the bigger
| fish.
| bee_rider wrote:
| For some reason, this has made me wonder if we just need
| more non-classical-social-media fediverse stuff. Like of
| course people will glom on to Threads, it means they can
| interact with the network while still being inside
| Facebook's walled garden...
|
| I wonder if video game engines could use it as an
| alternative to Steam or Discord integration.
| LtWorf wrote:
| The problem was not that it was not evolving. The problem
| was that they decided they had trapped all the users of
| other networks they could trap.
|
| Slack did the same killing xmpp and irc bridge. I don't see
| them making a matrix bridge.
| bee_rider wrote:
| > Read the terms of use / contract for Meta AI products. If you
| deploy it, some producer finds the model spits out copyrighted
| content, knocks on Meta's door, Meta will point to you for the
| rest of the court case. If that's the future for AI, then it
| doesn't really matter whether China wins.
|
| As much as I hate Facebook, I think that seems pretty...
| reasonable? These AI tools are just tools. If somebody uses a
| crayon to violate copyright, the crayon is not to blame, and
| certainly the crayon company is not, the person using it is.
|
| The fact that Facebook won't voluntarily take liability for any
| thing their users' users' might do with their software means
| that software might not be useable in some cases. It is a
| reason to avoid that software if you have one of those use
| cases.
|
| But I think if you find some company that says "yes, we'll be
| responsible for anything your users do with with our product,"
| I mean... that seems like a hard promise to take seriously,
| right?
| rkou wrote:
| AI safety is expensive, or even impossible, by releasing your
| models for local inference (not behind API). Meta AI shifts
| the responsibility of highly-general highly-capable AI models
| to smaller developers, putting ethics, safety, legal, and
| guard-rails responsibility on innovators who want to innovate
| with AI (without having the knowledge or resources to do so
| by themselves) as an "open-source" hacking project.
|
| While Mark claims his Open Source AI is safer, because fully
| transparent and many eyes make all bugs shallow, the latest
| technical report makes mention of an internal, secret,
| benchmark that had to be developed, because available
| benchmarks did not suffice at that level of capabilities. For
| child abuse generation, it only makes mention that it
| investigated this, not any results of these tests or
| conditions under which it possibly failed. They shove all
| this liability on the developer, while claiming any positive
| goodwill generated.
|
| It completely loses their motivation to care for AI safety
| and ethics if fines don't punish them, but those who used the
| library to build.
|
| Reasonable for Meta? Yes. Reasonable for us to nod along when
| they misuse open source to accomplish this? No.
| bee_rider wrote:
| I think this could be a somewhat reasonable argument for
| the position that open AI just shouldn't exist (there are
| counter arguments, but I'm not interested enough to do a
| back and forth on that). If Facebook can't produce
| something safe, maybe they shouldn't release anything at
| all.
|
| But, I think in that case the failing is not in not taking
| the liability for what other people do with their tool. It
| is in producing the tool in the first place.
| rkou wrote:
| Perhaps Open AI simply can't exist (too hard and
| expensive to coordinate/crowd-source compute and
| hardware). If it can, then, to me, it should and would.
|
| OpenAI produced GPT-2, but did not release it, as it
| couldn't be made safe under those conditions, when not
| monitored or patch-able. So it put it behind an API and
| owned its responsibility.
|
| I didn't take issue with Meta's business methods and can
| respect its cunning moves. I take issue with things like
| them arguing "Open Source AI improves safety", so we
| can't focus on the legit cost-benefits of releasing
| advanced, ever-so-slightly risky, AI into the hands of
| novices and bad actors. It would be a failure on my part
| if I let myself get rigamaroled.
|
| One should ideally own that hypothetical 3% failure rate
| to deny CSAM request when arguing for releasing your
| model still. Heck, ignore it for all I care, but they
| damn well do know how much this goes up when the model is
| jailbroken. But claiming instead that your open model
| release will make the world a better place for children's
| safety, so there is not even a need to have this
| difficult discussion?
| eli_gottlieb wrote:
| If it was really open-source you'd be able to just train one
| yourself.
| nicce wrote:
| Only if you were a billionaire. These models are starting to
| be so out of reach for single researchers or even traditional
| academic research groups.
| echelon wrote:
| This sort of puts the whole notion of "open source" at risk.
|
| Code is a single input and is cheap to compile, modify, and
| distribute. It's cheap to run.
|
| Models are many things: data sets, data set processing code,
| training code, inference code, weights, etc. But it doesn't
| even matter if all of these inputs are "open source". Models
| take millions of dollars to train, and the inference costs
| aren't cheap either.
|
| edit:
|
| Remember when platforms ate the open web? We might be looking
| at a time where giants eat small software due to the cost and
| scale barriers.
| bschmidt1 wrote:
| It's especially rich coming from Facebook who was all for
| regulating everyone else in social media after they had already
| captured the market.
|
| Everyone tries this. Apple tried it with lawsuits and patents,
| Facebook did it under the guise of privacy, OpenAI will do it
| under the guise of public safety.
|
| There's almost no case where a private company is going to be
| able to successfully argue "they shouldn't be allowed but we
| should" I wonder why so many companies these days try. Just
| hire better people and win outright.
| doe_eyes wrote:
| No major tech corporation is interested in openness for the
| sake of it, period. It's useful when it undermines your
| competition. It's why Google was pouring money into open source
| to hurt Microsoft while not particularly interested in open-
| sourcing their own flagship services. And it explains the bulk
| of Facebook-Google-OpenAI dynamics right now, where one group
| of actors is desperately trying to secure a moat through
| exclusive content deals and regulation... and the other side is
| trying to ruin their plans for the lolz.
|
| Facebook is pretty universally hated by techies, but as it
| happens, their incentives align with the incentives of non-
| commercial internet. Why make them pass purity tests? It's OK
| to root for what they're doing, it doesn't mean we have to
| marry the company.
| abetusk wrote:
| This is the modern form of embrace, extend and extinguish.
| "Embrace" open source, "extend" the definition to make it non
| open/libre and finally extinguish the competition by shoring up
| the drawbridge to the moat they've just built.
| troupo wrote:
| I've had as a comment to a comment, but I'll repost it at the top
| level:
|
| They use "open source" to whitewash their image.
|
| Now ask yourself a question: where does Meta's data come from?
| Perhaps from their users' data? And they opted everyone in by
| default. And made the opt-out process as cumbersome as possible:
| https://threadreaderapp.com/thread/1794863603964891567.html And
| now complain that the EU is preventing them from "collecting rich
| cultural context" or something
| https://x.com/nickclegg/status/1834594456689066225
| rkou wrote:
| Also known as https://en.wikipedia.org/wiki/Openwashing
|
| > In 2012, Red Hat Inc. accused VMWare Inc. and Microsoft Corp.
| of openwashing in relation to their cloud products.[6] Red Hat
| claimed that VMWare and Microsoft were marketing their cloud
| products as open source, despite charging fees per machine
| using the cloud products.
|
| Other companies are way more careful using "open source" in
| relation to their AI models. Meta now practically owns the term
| "Open Source AI" for whatever they take it to mean, might as
| well call it Meta AI and be done with it:
| https://opensource.org/blog/metas-llama-2-license-is-not-ope...
| dzonga wrote:
| the reason - i'm a little bearish on AI is due to its cost. small
| companies won't innovate on models if they don't have billions to
| burn to train the models.
|
| yet when you look back at history, things that were
| revolutionary, it was due to low cost of production. web,
| bicycles, cars, steam engine cars etc.
| rafaelmn wrote:
| > yet when you look back at history, things that were
| revolutionary, it was due to low cost of production.
|
| Nuclear everything, rockets/satellites, tons of revolutionary
| things that are very expensive to produce and develop.
|
| Also software scales differently.
| zwijnsberg wrote:
| yet if the weights are made public, smaller companies can
| leverage these pretrained models can't they?
| miguelaeh wrote:
| The first cars, networks, and many other things were not
| unexpensive. They became so with time and growing adoption.
|
| Cost of compute will continue decreasing and we will reach that
| point where it is feasible to have AI everywhere. I think with
| this particular technology we have already reached a no return
| point
| farco12 wrote:
| I could see the cost of licensing data to train models
| increasing significantly, but the cost of compute for training
| models is only going to drop on a $/PFLOP basis.
| CatWChainsaw wrote:
| Facebook promised to connect the world in a happy circle of
| friendship and instead causes election integrity controversies,
| bizarre conspiracy theories about pandemics and immigrants to go
| viral, and massive increases in teen suicide. Not sure why anyone
| would trust them with their promises of decentralized-AI and
| roses.
| atq2119 wrote:
| Good. Now compare to OpenAI. Clearly what Meta is doing is
| better than OpenAI from the perspective of freedom and
| decentralization.
| CatWChainsaw wrote:
| Cool. What Meta is doing is better than cigarettes from the
| perspective of addiction. If Meta is the best we have, then
| we'd better create something better, or prepare for the
| inevitable enshittification.
| jazzyjackson wrote:
| I love it. TOBACCO PRODUCTION MUST BE DECENTRLIZED.
| menacingly wrote:
| Decentralized on centralized hardware?
| latchkey wrote:
| Evidence is showing that AMD MI300x are proving to be a strong
| contender.
| mrkramer wrote:
| Yea, I believe you Zuck, it's not like Facebook is closed
| centralized privacy breaking walled garden.
| pie420 wrote:
| IBM Social Media Head: "The Future of Social Media is OPEN SOURCE
| and DECENTRALIZED"
|
| This must be a sign that Meta is not confident in their AI
| offerings.
| jmyeet wrote:
| Two things spring to mind:
|
| 1. Open source is for losers. I'm not calling anyone involved in
| open source a loser, to be clear. I have deep respect for anyone
| who volunteers their time for this. I'm saying that when
| companies push for open source it's because they're losing in the
| marketplace. Always. No companiy that is winning ever open
| sources more than a token amount for PR; and
|
| 2. Joel Spolsky's now 20+ year old letter [1]:
|
| > Smart companies try to commoditize their products' complements.
|
| Meta is clearly behind the curve on AI here so they're trying to
| commoditize it.
|
| There is no moral high ground these companies are operating from.
| They're not using their vast wisdom to predict the future.
| They're trying to bring about the future the most helps them. Not
| just Meta. Every company does this.
|
| It's why you'll never see Meta saying the future of social media
| is federation, open source and democratization.
|
| [1]: https://www.joelonsoftware.com/2002/06/12/strategy-letter-v/
| uptownfunk wrote:
| It's marketing to get the best researchers. The researchers want
| the meta pay and they want to hedge their careers to continue to
| publish. That's the real game, it's a war for talent. Everything
| else is just secondary effects.
| croes wrote:
| Also Meta: The future is VR
| stonethrowaway wrote:
| I'll link to my comment here from approx. 52 days ago:
| https://news.ycombinator.com/item?id=41090142
|
| This is chess pieces being moved around the board at the moment.
___________________________________________________________________
(page generated 2024-09-18 23:01 UTC)