[HN Gopher] Kyutai AI research lab with a $330M budget that will...
       ___________________________________________________________________
        
       Kyutai AI research lab with a $330M budget that will make
       everything open source
        
       Author : vasco
       Score  : 242 points
       Date   : 2023-11-19 11:48 UTC (11 hours ago)
        
 (HTM) web link (techcrunch.com)
 (TXT) w3m dump (techcrunch.com)
        
       | treprinum wrote:
       | $330M is like 33 training runs of a large LLM. Not sure that
       | would bring open source anywhere near GPT-3.5.
        
         | abiraja wrote:
         | Have you tried Mistral?
        
           | dannyw wrote:
           | Mistral is genuinely groundbreaking, for a fast, locally-
           | hosted model without content filtering at the base layer. You
           | can try it online here: https://labs.perplexity.ai/ (switch
           | to Mistral)
        
             | anonzzzies wrote:
             | Are more companies/teams than the creating team working to
             | get this to copilot/chatgpt standards?
        
             | js4ever wrote:
             | Wow I was not expecting this, It's really something else in
             | terms of speed, and results are not bad! Will test it more
        
             | eropple wrote:
             | It's very fast, but it doesn't seem very good. It doesn't
             | take instruction well (acknowledges and spits back the same
             | wrong stuff) and doesn't seem to have much of a corpus or
             | it's dropping most of it on the floor because it
             | successfully answers zero of my three basic smoke-test
             | questions.
        
               | sigmar wrote:
               | >doesn't seem to have much of a corpus
               | 
               | what do you mean by 'corpus'? It is only 13GB so
               | questions that require recalling specific facts aren't
               | going to work well with so little room for 'compression',
               | but asking mistral to write emails or perform style
               | revisions works quite well for me
        
               | mark_l_watson wrote:
               | Are you running mistral-7B or mistral-7B-instruct?
        
             | audessuscest wrote:
             | Thanks for the link, do you know any other similar services
             | that support fine-tuning ?
        
         | gwervc wrote:
         | Maybe at today cost. But this is a lab about AI not just LLM.
         | Also given that French research is underfunded since decades,
         | our researchers are accustomed to producing good results at a
         | budget cost.
        
           | nwoli wrote:
           | Also has incredibly good math institutions
        
           | umanwizard wrote:
           | "French research is underfunded since decades" is a French-
           | ism, just so you know. Normally native English speakers would
           | say "French research has been underfunded for decades".
        
             | Obscurity4340 wrote:
             | Depuis ?
        
               | umanwizard wrote:
               | That can be used with either a duration ("depuis des
               | decennies") or a point in time ("depuis 1995"). English
               | "since" can only be used with a point in time.
        
               | Obscurity4340 wrote:
               | Merci, francais ami!
               | 
               | Edit: is that correct or is there an apostrophe Im
               | missing? It would better if you were francaise, n'est-ce
               | pas?
        
               | umanwizard wrote:
               | I think "merci, mon ami francais" would be more correct,
               | but at any rate, I'm not French!
        
             | trealira wrote:
             | I've noticed a lot of comments on Hacker News that use this
             | construction. I don't point it out because often people
             | dislike unsolicited advice about their English.
             | 
             | Spanish speakers also make this mistake. I haven't met
             | German speakers, but they also say the equivalent of "is
             | (adjective) since (time)" in German. It makes me wonder if
             | English is just unusually strict about this distinction
             | compared to other European languages.
        
               | vunderba wrote:
               | I am also not a huge fan of unsolicited grammatical
               | advice, because I feel like it's both low effort and
               | derails the conversation. If the grammar error is more
               | syntactical in nature and doesn't obscure the meaning of
               | the thoughts intended to be conveyed, then let's just all
               | _move on_.
               | 
               | This is an international forum, we're well accustomed to
               | being able to parse meaning despite a few inconsequential
               | grammatical issues.
        
               | umanwizard wrote:
               | I pointed it out because when I lived in France, I
               | personally appreciated people pointing out grammatical
               | mistakes I made in French. If the OP doesn't care about
               | sounding more like a native English speaker, that's
               | perfectly fine and he can just ignore my comment.
        
               | bee_rider wrote:
               | People can just not engage in conversations they don't
               | like. This forum has a branching structure, it is easy to
               | hide a topic that you don't like.
               | 
               | If someone was to be really odious and hateful, the
               | presence of that sort of thing could be harmful, but this
               | seems like a polite correction of a common problem. And
               | it is an international forum, there are plenty of people
               | who might get something from language-chat.
        
               | lucubratory wrote:
               | As someone learning a language that's not my native one,
               | people who correct my Chinese kindly are the absolute
               | gold standard. People who correct it rudely are rare but
               | still more helpful than people who say nothing and just
               | move on. This is very common in language learning
               | communities. My default assumption is that other people
               | who are learning a language that I know natively probably
               | feel similar, and would appreciate kind, contextual
               | corrections. I am happy to adjust if someone lets me know
               | they don't want that or if it's against the rules of a
               | specific forum.
        
               | umanwizard wrote:
               | I think most European languages lack the distinction
               | between "he did X" and "he has done X". Actually, a lot
               | of languages (including Spanish, German and French) have
               | a distinction like this syntactically but it doesn't
               | really mean the same thing as it does in English.
        
               | SamBam wrote:
               | Huh, I just had a fairly long and fruitless conversation
               | with ChatGPT 4 trying to understand whether this was
               | true. ChatGPT kept insisting that romance languages
               | distinguish between the preterite ("he did," simple past
               | tense) and the present perfect ("he has done"), but every
               | time I asked it to give me specific examples it would
               | start translating them, find that they were the same, and
               | then say "well, in Italian [or French] it's actually more
               | about context..."
               | 
               | Finally it was able to give me the Spanish "El hizo esto"
               | vs "El ha hecho esto," but admitted that whether one used
               | one vs the other was quite regional.
        
               | bee_rider wrote:
               | I wonder if it is just the result of applying some
               | transformation to "was underfunded for decades" in an
               | attempt to make it cover the present.
        
             | boisgerault wrote:
             | A Gallicism :)
        
         | nmfisher wrote:
         | If they're researching avenues other than LLMs, that money goes
         | a lot further.
         | 
         | We can only hope.
        
         | espadrine wrote:
         | During the conference where it was announced, they indicated
         | they have a partnership with Scaleway to access their Nabu 2023
         | supercomputer[0]. I expect it will be similar to the
         | relationship between OpenAI and Microsoft Azure, with free
         | credits as part of their investment in exchange for being at
         | the forefront of AI datacenter design. Indeed, Xavier Niel is
         | one of three investors in Kyutai, and founded Scaleway.
         | 
         | [0]: https://www.scaleway.com/en/news/scaleway-releases-the-
         | detai...
        
         | qwertox wrote:
         | Maybe related:
         | 
         | EU is starting a program where AI startups will get access to
         | their supercomputers and also get a chance to win prizes.
         | 
         | 16 November 2023: _Commission opens access to EU supercomputers
         | to speed up artificial intelligence development_ [0]
         | 
         | - Launch of the Large AI grand challenge: This competition -
         | launched today, is a collaboration led by the EU funded project
         | AI-BOOST, with access to the European Supercomputers being
         | facilitated by the EuroHPC JU. It encourages the wide
         | participation from European start-ups with experience in large-
         | scale AI models. The winners are expected to release the
         | developed models under an open-source license for non-
         | commercial use, or through publishing their research findings.
         | The challenge will select up to four promising European AI
         | start-ups that will be given access to EuroHPC supercomputing
         | facilities to foster the development of large-scale AI models
         | in Europe and a EUR1 million prize will be distributed among
         | the winners.
         | 
         | - Opening up European supercomputer capacity: Access will be
         | established for ethical and responsible AI start-ups, enabling
         | them to efficiently train their models using European
         | supercomputers.
         | 
         | - Enhanced activities and services: the EuroHPC JU will advance
         | activities and services powered by High-Performance Computing
         | to foster trustworthy AI in Europe. These efforts will aim to
         | facilitate increased accessibility for AI communities and
         | promote the optimal and efficient use of HPC technologies for
         | scientific and industrial innovation.
         | 
         | [0]
         | https://ec.europa.eu/commission/presscorner/detail/en/ip_23_...
        
           | konschubert wrote:
           | First they set up massive barriers such that you'd be insane
           | to start an AI company in Europe. And then they dole up
           | subsidies.
           | 
           | Makes no sense. That money will go to waste.
        
             | zirgs wrote:
             | What massive barriers are there?
        
               | JumpCrisscross wrote:
               | > _What massive barriers are there?_
               | 
               | Starting and scaling a business in Europe is harder than
               | in America. This is true for fundamental reasons, like
               | language and national borders. It's also true due to
               | regulation and the culture of work and towards commerce
               | in general.
               | 
               | Some of those come with reasonable tradeoffs, _e.g._ in
               | respect of employee protections. Many do not,
               | particularly when it comes to licensing, bureaucracy and
               | the peculiar way most European tax law and regulation
               | tries to compensate for its licensing and bureaucracy by
               | adding more bureaucracy in front of a balancing subsidy.
        
               | carstenhag wrote:
               | Opinion of a german here: Build & break things (without
               | caring for laws or implications) is what the US and their
               | companies do. Often it works, and sometimes breaking the
               | law makes a new law pop up to legalize it, but that's not
               | a good thing.
               | 
               | Yes, it's a big annoying to start a company in some EU
               | countries. No, it's not as bad as HN claims.
        
               | 8n4vidtmkvmk wrote:
               | Working for a big American company.. we have to abide by
               | a lot of laws, including the EU ones.
               | 
               | Small companies can get away with it for awhile, but not
               | forever
        
               | imjonse wrote:
               | Requiring companies to offer more than two weeks of paid
               | time off per year.
        
         | gentleman11 wrote:
         | Sorry, could somebody show a source for that? An $11M bill,
         | that's nuts and just need a reference lol
        
         | tyingq wrote:
         | There's some compute also, which may not be counted in the 330M
         | 
         |  _" Co-founded by the Iliad group...To do this, the laboratory
         | will use the computing power made available to it by Scaleway,
         | an iliad Group subsidiary. Scaleway's supercomputer has the
         | highest-performance computing power for AI applications
         | deployed to date in Europe."_
         | 
         | https://kyutai.org/CP_Kyutai_AI_EN.pdf
        
         | coding123 wrote:
         | Seems like it could be unlimited if people just bought hardware
        
         | logicchains wrote:
         | > $330M is like 33 training runs of a large LLM. Not sure that
         | would bring open source anywhere near GPT-3.5.
         | 
         | Open source is already quite near GPT-3.5; it's reaching GPT-4
         | level that is the challenge.
        
       | eschluntz wrote:
       | Hopefully it's commercial open source, not for research only
        
         | kaliqt wrote:
         | Yeah, very tired of everyone trying to pass off "source
         | available" licenses as "open source" licenses and getting away
         | with it.
        
           | yieldcrv wrote:
           | the "and" in FOSS is actually doing alot of heavy lifting, as
           | open source really does only mean source available
           | 
           | yes, I want FOSS too
        
             | babyeater9000 wrote:
             | I just want free software. The oss is redundant to me.
        
               | gchamonlive wrote:
               | Why is it redundant? Redundant to what? What would
               | account for what opensource brings without opensource
               | itself?
        
               | meiraleal wrote:
               | The F, free software. opensource(r) isn't needed but it
               | became a bigger brand than free software and everything
               | with source available is called open source nowadays
        
               | thinkmassive wrote:
               | Free Software is inherently open source
        
               | meiraleal wrote:
               | Yes. The problem is that open source isn't free software.
        
               | babyeater9000 wrote:
               | You have free software. Free software is pretty rigidly
               | defined. You also have open source software, which people
               | also seem to think is defined. I'm my opinion, the
               | concept of open source software is vague enough that its
               | definition is open to interpretation. Look at the people
               | claiming that source available software is open source.
               | Source available software is, in fact, open source
               | software, even if it's not compatible with copyleft. Free
               | software is not open to interpretation. Open source
               | software can be free software, but some software can
               | rightfully be called open source software even if it
               | isn't free software. So, if we are using the terms
               | interchangeably because they are the same thing, then
               | open source is a redundant term. If open source software
               | and free software are not the same, which might be the
               | case sometimes, then I want free software. I'm not a
               | programmer. I don't care to make money from software and,
               | frankly, I don't care about the money making aspect of
               | software. Open source stuff, to me, reeks of corporate
               | capture. I don't want telemetry, or to be bled
               | financially to use a product. I don't believe that
               | software is or can ever be a product. Algorithms
               | shouldn't be copyrighted even if they are wrapped in a
               | programming language. I don't care about implementation.
               | I think this is a case of A is B and B is sometimes A.
               | It's the sometimes case that really bothers me.
               | https://www.gnu.org/philosophy/open-source-misses-the-
               | point....
        
               | JumpCrisscross wrote:
               | > _I don 't believe that software is or can ever be a
               | product_
               | 
               | You must mean should be, since we have decades of
               | evidence to the contrary.
        
               | babyeater9000 wrote:
               | I mean that software, in it's written form, is the
               | documentation of knowledge from software development, a
               | service. I view sciencing as another service that
               | produces knowledge. Knowledge has zero cost of
               | duplication and, as such, cannot be considered a product.
               | Artifacts that are produced by the application of
               | knowledge are products because they have a non-zero cost
               | of duplication. Computer hardware is an example of a
               | product. I don't view intellectual property as property
               | either. Software, in my opinion, isn't a product.
               | Software is knowledge. I don't claim to be correct. I'm
               | attempting to share my point of view. Anything with zero
               | cost of duplication isn't a product in my mind because
               | these things are infinitely copyable once created. Once a
               | mathematician discovers a math they don't retain rights
               | to it. Charging money for software is, in my view, no
               | different than trying to make people pay for secret
               | knowledge. You might be able to keep the secret locked
               | down for a while, but it will get out eventually.
               | Knowledge is the closest thing we have to magic, and if
               | we choose to view it through the zero sum lens of
               | capitalism, I think that does society a disservice in the
               | long run. If I were a wizard I would share the magic, not
               | try to charge money to teach people a spell or two. It
               | might be the case that all products are knowledge given
               | form, but keep in mind that knowledge exists before and
               | after discovery and its fruits/artifacts must be created
               | with work.
        
               | ben_w wrote:
               | The open source[0] is the only reason anyone has time to
               | make most of the valuable free software.
               | 
               | We can't all be like Donald Knuth or Simon Tatham making
               | TeX and PuTTY as personal projects.
               | 
               | [0] specifically the freedom to fork, to develop further,
               | and to make new releases that others can also build upon,
               | which means I aver that many of the public AI models are
               | sufficiently open that they're de facto open source even
               | if the licensing isn't there.
               | 
               | Even if it's a de jure violation of the copyright to make
               | a derivative, I'm not sure you could prove that had
               | happened when all the weights are floating point numbers
               | you can randomise slightly as a first step -- if training
               | just happens to move them back to the original values,
               | well, that's just evidence the optimiser was working.
        
               | gentleman11 wrote:
               | Money is important to find ventures, but the open source
               | aspect is important for guaranteeing user freedom in the
               | long run in our society
        
             | Lapha wrote:
             | >as open source really does only mean source available
             | 
             | The definition and history of the term as a licence is
             | unambiguous in that the only restriction on redistribution
             | is that it contains the source code under the same licence.
             | There are senior engineers alive today that weren't even
             | born when this was the commonly understood meaning of the
             | term, it's not a new concept.
             | 
             | The term and usage is being co-opted these days but that's
             | bound to happen when it's not a legally protected
             | definition. Give it another 10-20 years and I'm sure we'll
             | be having the same argument over whatever term ends up
             | replacing 'open source'.
        
               | yieldcrv wrote:
               | I'm all for "language evolving"
               | 
               | Language exists to convey a shared concept
               | 
               | In this case, the evolved version of open source fails to
               | convey a shared concept in comparison to the prior term,
               | "free and open source software" or FOSS for a shorthand
               | adjective
               | 
               | Here, people with knowledge of the lexicon are using it
               | accurately, and people without knowledge of the lexicon
               | or its etymology are complaining when they should be
               | pushing for FOSS instead of getting surprised everytime
        
           | ekianjo wrote:
           | you can thank the "open source" movement for muddying the
           | waters
        
             | eropple wrote:
             | The OSD is plenty clear.
        
               | ekianjo wrote:
               | except in the name
        
               | Lapha wrote:
               | 'Free' software has problems with its name too. The ones
               | muddying the waters are people and companies releasing
               | source code with a proprietary licence while trying to
               | latch onto the open source branding.
        
           | Kiro wrote:
           | Getting away with what? The default is closed source.
        
             | anonzzzies wrote:
             | Which is basically the same for most companies. Too
             | dangerous to even _read_ the source.
        
             | sp332 wrote:
             | Getting away with calling it "open source".
        
           | ctxc wrote:
           | Source available but with a pricing for commercial use seems
           | to be the most sustainable way for open source (pun intended)
           | though.
        
             | api wrote:
             | But that gets in the way of open source as free labor for
             | giant companies.
        
               | lsaferite wrote:
               | I feel like the license used on Llama covers that, right?
               | Carve out an exclusion for large companies so they have
               | to obtain a commercial license and let all the ones below
               | that access it for free.
        
               | JumpCrisscross wrote:
               | > _Carve out an exclusion for large companies so they
               | have to obtain a commercial license_
               | 
               | And this is why open source means source available, look
               | at the license. There are infinite variations between
               | free software and classic source available.
        
             | KolmogorovComp wrote:
             | You can (and most likely should) still use a OSI-approved
             | license if you want, such as AGPL/GPLv3, while providing
             | the possibility of case-to-case commercial licenses.
             | 
             | gifski [0] for example is a successful open-source project
             | doing that.
             | 
             | [0] https://gif.ski/license.html
        
               | BrianHenryIE wrote:
               | I just see now, there are two Gifskis, one Rust, one
               | Swift. It's worth looking at the maintainer of the Swift
               | one, he has 1000+ repos, many with 1000+ stars.
               | 
               | * https://github.com/ImageOptim/gifski
               | 
               | * https://github.com/sindresorhus/Gifski
               | 
               | * https://github.com/sindresorhus
        
               | nativeit wrote:
               | He appears to be the original creator of the "Awesome X"
               | repo: https://github.com/sindresorhus/awesome
        
               | KolmogorovComp wrote:
               | Both gifski are maintained by @sindresorhus, as you can
               | see by commit counts. (Even if he's not the main
               | maintainer of the swift version).
        
           | JumpCrisscross wrote:
           | > _very tired of everyone trying to pass off "source
           | available" licenses as "open source"_
           | 
           | This train has passed. Sort of like crypto/crypto. Open
           | source functionally means source available; the rigid
           | definition remains free software.
        
             | edouard-harris wrote:
             | I've heard the licenses Meta releases its AI models under
             | described as "open access", to differentiate them from the
             | stricter "open source" definition.
        
         | gumballindie wrote:
         | If it's out there then it's fair use isnt it? Or does that only
         | apply to non ai people's ip?
        
           | MacsHeadroom wrote:
           | ML models are probably not actually copyrightable, for the
           | same reasons their outputs aren't. So you can probably just
           | ignore model licenses, yes.
        
         | nwoli wrote:
         | I mean I'm happy with just the weights if that's all they can
         | do. If it's actually useful then you should be able to use it
         | to build something important, not just make middlemen "shovels
         | during goldrush" saas apps
        
           | Obscurity4340 wrote:
           | Can you provide a small snippet of what is meant by
           | "weights", I assume that refers to some probabillistic ratios
           | or something?
        
             | IanCal wrote:
             | The weights refers to the trained model.
             | 
             | Lots of neural nets are pretty standard, and work by
             | multiplying lots of numbers by other ones. The key is to
             | figure out _which_ numbers to use. Those are the weights of
             | each connection between the artificial neurons in the
             | model.
             | 
             | With them you can run the model yourself.
        
             | layer8 wrote:
             | https://deeplizard.com/lesson/ddd2drizla
        
           | lukeschlather wrote:
           | They've clearly said they are truly making this open source,
           | the hope you're replying to is not about data vs. weights
           | it's about licensing. Having all the weights and training
           | data isn't really useful if the license prevents you from
           | using it. In fact it's a problem since you might open
           | yourself up to trouble just by reading it.
        
         | YetAnotherNick wrote:
         | I would want an open source company which has clear one time
         | pricing for commercial uses. I think if a company could release
         | state of the art open source model, 1000 compnaies could
         | definitely pay $5000, making it much more sustainable.
        
           | elorant wrote:
           | If there was a black box of ChatGPT that doesn't need to
           | connect to the Internet companies would gladly pay
           | $50k/yearly to use it. There's huge demand but privacy
           | concerns keep a lot of corporations away.
        
       | modernpink wrote:
       | What does the name mean? Presumably it's Japanese but not seeing
       | anyone Japanese in the company
        
         | thibaut_barrere wrote:
         | I read somewhere else:
         | 
         | > the answer is called Kyutai (pronounced "Cute AI" and meaning
         | "sphere" in Japanese),
        
         | lowkey_ wrote:
         | Maybe not Japanese, but the French love Japanese culture.
        
           | jpgvm wrote:
           | The love appears mutual from my time in Japan.
        
       | mrits wrote:
       | That money will go fast with trying to poach out of industry ,
       | buying nvidia hardware, and no clear direction. Hopefully it is
       | just the fault of the author and not actually another case hoping
       | a bunch of phds will come up with something given time and money.
        
       | amelius wrote:
       | Is the complete code open source? Or only the models?
        
       | htrp wrote:
       | How are they going to continue past the first set of model
       | releases?
        
       ___________________________________________________________________
       (page generated 2023-11-19 23:01 UTC)