> It’s very early to talk about what Mistral is doing or will be doing — it’s only around a month old — but from what Mensch said, the plan is to build models using only publicly available data to avoid legal issues that some others have faced over training data, he said; users will be able to contribute their own datasets, too. Models and data sets will be open-sourced, as well.
Unless I missed it elsewhere in the article this part really glosses over something I have a hard time understanding. 113M seed round at a 240M valuation for a company that has no product yet and a plan to build something like something else that already exists. This feels insane to me.
They're investing in the team, which I think is smarter than investing in an idea. Those people built llama at meta and flamingo/chinchilla at deepmind.
"When I was an intern, in one of the training presentations, a senior banker told us to distinguish between the process and the results. He said that we should focus on the process, which we can control, rather than the result, which is subject to luck. And here at Goldman, he said, we don’t punish people for losing money for the right reason. I have always loved asking questions, so I asked him, was anyone ever punished for making money for the wrong reason?"
Just like that old Dropbox comment that people love to bring up, you can make a gazillion dollars off a bad bet, and you can lose following a good strategy, that's the case everywhere probabilities are involved. Outcomes don't determine if the original evaluation was sound or not.
I agree and just to clarify: I don't think/know whether this startup will be successful or not, in fact I don't even know if buying DeepMind was a good investment for Google (financially speaking). I only mentioned it because the comments were very similar.
I get that. Call me old fashioned but if I were investing a hundred million dollars I'd like the team to build _something_ for me first. Maybe that's why I dont have a hundred million to invest lol.
It costs like $50M to build a next gen model, so if what they want to pitch is training next gen models it makes sense they get $100M to try. Similar to Hippocratic AI $50M seed raise. If anything this founder team seems more legit.
I talked to Munjal and team when they were ideating on Hippocratic, even he acknowledges that healthiq was a shit show. AFAIK it was 30 people running a 1k person call center and failing. This motivated Hippocratic as they think they can do AI-based call center replacement because they were so burned out of running human-staffed call centers.
I think Hippocratic is interesting because their leadership is business focused and not very technical, while a lot of other new foundational model cos are led by technical people. I’m curious how this shakes out.
"the team" here seems to be a trio of research scientists/data scientists? (Some famous in their research community bubble)
Competent product/sales person, and engineering person seems missing for this to succeed. Especially if they used to rely on the eng infrastructure of giants.
They seem smart folks so probably can cover the software engineering gap, especially with $100m to outsource some of the hard problems to some cloud infra.
But I'd be more confident if they had started with a co-founder that's more product/sales oriented.
Google, Facebook and others spent billions buying the best ML people and giving them the best, and got scooped badly by OpenAI. And I know it's a cliche but lots of better engineered "iphone/ipod" products existed that got destroyed by apple. Unless they can invest in the man himself - a Jobs or Altman, I don't think investing in a good team is an automatic road to victory.
I don't think you can even remotely compare Jobs and Altman here, especially given how badly he missmanaged the enormous advance OpenAI had with Dall-E. In fact, he was lucky to have midjourney and stable diffusion kicking his ass on image generation and pushing him to actually release a product, otherwise Anthropic and Llama could have done the same on the LLM side.
Somewhere in their offices they be putting quote from Neils Bohr to pat on their back :)
We are all agreed that your theory is crazy. The question which divides us is whether it is crazy enough to have a chance of being correct. My own feeling is that it is not crazy enough.
Maybe. But there seems to be a reasonable investment case. Especially if the EU is making stupid laws, being the "European AI" company seems quite valuable.
To be frank my first instinct is "slush fund" from the start just because they are appealing to nationalism first. Starting off with a thought terminator is a mark of a deceiver.
Edit:
I mean would you expect anything less than a slushfund or grift from an "America AI" project?
> that really graduated from the very best french schools
I find it shocking when people refer to what school someone went to as a credential in just about any context. I usually consider it a red flag, an indication that there is no functional understanding being applied in the evaluation
i don't think you realize what it means to have a polytechnic or even better normal sup ulm diploma.
those schools don't draft people based on bullshit application. It's a super super hardcore science exam that only a few dozen people per year pass. It's a bit like the putnam, with bachelor-level math and physics curriculum.
The mistake people make in France is to conclude that people coming from those schools will be good at just any position, management, running a huge company, etc.
But in that specific case, they'll actually be doing hardcore science and ML all day long. It makes perfect sense.
I do agree those people are good in hard sciences. I did those exams and they're hard.
About your comment about putting people in positions because of those diplomas, it's also because of the network you acquire at those kinds of schools, and cronyism. Granted, this kind of cronyism is less worse than in others countries where it can be more pernicious and less official :-)
not any context... the school they went to is specifically selecting math wiz.. Which seems a bit related to the field of ML (but please correct me if i'm wrong).
Maybe as related as "this product was produced in Germany" would be related to a product's quality? In that there's a loose correlation there but if that's one of the main points in a product review it means the reviewer isn't doing their job.
It feels like a typical "young & hyped" European business - fanfare in the beginning, capital from old money networks that want to stay relevant, then fizzling out slowly without showing anything worthwhile later.
I've worked at a couple of mid level undertakings of the kind. Smart people do it on the down low and milk it for years. The inner circle draws fat engineering salaries, and they keep surfing on the latest tech hype wave.
I just listened to one of the team members next to Macron at a conference in Paris and I sense that French politicians really want to make this work - for protectionist reasons or just to bring back French R&D in the game. With that in mind though it's interesting to see that 40% of it is already owned by foreign capital.
Didn't the EU just pass an AI law that prohibits generative AI?
> In a bold stroke, the EU’s amended AI Act would ban American companies such as OpenAI, Amazon, Google, and IBM from providing API access to generative AI models. The amended act, voted out of committee on Thursday, would sanction American open-source developers and software distributors, such as GitHub, if unlicensed generative models became available in Europe. While the act includes open source exceptions for traditional machine learning models, it expressly forbids safe-harbor provisions for open source generative systems.
This reminds me of Quaero and later Qwant. Both were French attempts to create national champion companies to compete with Google and establish digital sovereignty. Quaero totally failed. Qwant is still hanging on but nearly dead.
Oh, definitely, but keep in mind that a lot of huge companies came out of the dotcom bubble. There was a lot of nonsense, but quite a few well-planned products as well, some of which even succeeded.
I think AI's going to be the same way, at least at first. Long-term there's of course a possibility of takeoff, slow or fast; I don't think we've excluded that yet.
Crypto was a massive innovation unlike AI - I'm referring to the ability to liquidate investments to retail by completely bypassing securities laws. I don't think this AI wave can get to the same level since it's way harder to sell the bubbly assets to the next lemming. Public markets via a traditional IPO won't be able to sustain this bubble.
I guarantee that AI will have it's own best fraudulent company like the rest of the other startups in each of their industry sectors: Theranos (for Healthcare), WeWork (in general), FTX (for crypto) and Nikola (for electric vehicles).
All thanks to the continued unjustified and extreme seed funding rounds and over-valued companies like this especially with no product or any proof other than 'the team'.
Fan of wework here. As a longtime happy customer I do not believe they belong in a list of fraudulent companies.
They still enable me to walk into any major city and have a nice office to work out of and I appreciate them for that. They changed the real estate game for startups and solopreneurs and should be commended for this!
Yeah, the wework hate is getting a bit out of hand. They (well, co-working spaces) added uniformity, predictability, and an international footprint to a legacy business of renting office space.
'In the end today’s AI boom is an evolution. It evolves from the huge databases built by the Cloud Czars, from the technology used to search those data stores, and from the ability of Graphics Processing Units to turn out results quickly.'
The French like to go in early in tech, minitel is a good example, but they almost invariably get leapfrogged by subsequent waves.
In the AI era who owns the data and data sanctity is an upcoming legal quagmire that the big players will win. I don't see France ever being a big player in this world.
French innovation is quirky but brilliant IMO - minitel was years ahead pre internet.
For some reason French innovation doesn't spread much beyond France. Concorde (I love that plane so much) was an Anglo French project. A lot of my friends parents in the UK midlands were involved with at massive project, but it never scaled to fleets and eventually withered away after an astonishing amount was spent developing it in the '60s ($2.8 billion in that era's valuations).
Not denegrating French innovation here at all, I just don't think the rigid ENA/Ecole Polytechnique elites are practical or competent in understanding how to exploit and build on great ideas by their countrymen.
> Not denegrating French innovation here at all, I just don't think the rigid ENA/Ecole Polytechnique elites are practical or competent in understanding how to exploit and build on great ideas by their countrymen.
I don't disagree that these schools still need to evolve a bit, but you'd be surprised at how much they've already moved in that area. I know multiple people from Polytechnique are heavily involved in entrepreneurship and/or AI.
I think my point is that ENA / Ecole Polytechnique people are arguably harming French innovative abilities. Many US entrepreneurs are grass roots college dropout types...
Well... We're talking tech unicorns here. France ain't exactly know for it's tech companies. There's what, eight EU (not french, but entire EU) tech companies in the top 100 tech companies worldwide?
The biggest EU one is german (not french), it's SAP. At least two french ones are basically french militaro-industrial complex / fake-private actually state-owned ones, being "big" by virtue of sucking french taxpayers money.
I don't know what France's culture is (although I'm a native french speaker and typing this from France) but "producing tech unicorns" certainly ain't part of it.
Tech unicorns are those start-ups with valuations above 1 billion, not the top 100 (top by what what measure, revenue, profit, market cap, number of employees?).
Also, a nations economy doesn't depend on the number of VC backrd tech unicorns. That France economy has quite some issues ain't news, the perceived lack of tech unicorns (if there ever was a pseudo-SV centric view of things, that's a front runner), is none of them.
@hef19898 Not 'better' just different culturally. The US installs ivy league types at well funded startups just as the french install their ENA/Ecole Polytechnique elites in their state funded 'startups' but somehow the French dynamic doesn't work nearly as well as the US one.
Also, the US Navy is buying a French/Italian frigate design to replace, sorry, I meant support the littoral combat ships that don't work. The only reason Australia cancelled the contract for French is that, all of a sudden, they got nuclear ones.
Agree on France, the EU is particularly rigorous in its oversight of AI also. In terms of the boom/bust in the article, historical patterns are useful to keep in mind, but they don't dictate what will happen.
I can categorically say I'm a productive ChatGPT user as of today. This might not apply to every industry, but for developing/refactoring there is already a GDP-boost with ChatGPT's name on it. Come to think of it, the AI hype-cycle has been in effect for some time already no? Maybe this is the productivity stage...
>I can categorically say I'm a productive ChatGPT user as of today. This might not apply to every industry, but for developing/refactoring there is already a GDP-boost with ChatGPT's name on it.
Really? I asked ChatGPT to code review my code and it halucinated issues that did not exist.
I also had the idea to give it 100 lines of code to refactor but after I got the response I realized that will can't trust that it did not fucked up so I decided to do the cleanup myself.
I can use it as a faster Google search to answer me soem questions, but I need to always check the answers.
Both GPT-3.5 and GPT-4 have gotten considerably worse in quality for me since the May 24 update. Yesterday I asked GPT-4 to refactor some code I'd written in haste with a lot of duplication etc - a typically perfect task for a LLM - and it just gave me the same code back, without comments
My hope is that Intellij guys will move faster with their tools now that there is presure from copilot. There are still plenty of code refactorings that could be automated and are not and also there are obvious bugs that their code analysis do not catch(though I understands JS is a shit language to analyze).
When this open LLM could run on my local machine, I could at least use them to find stuff for me, I work on a big old project, I know there is soem functin or code that does X somewhere but I no longer remember th e files or function names. So a smarter search would be helpful.
Maybe someone at this small office software shop in Seattle, the one I constantly forget the name of, you know the that invested some table change into an unknown AI thing, should read that, too.
What I didn't read was a business plan. How do they plan to capture a fraction of the value they create, and pass a fraction of that fraction on to investors?
Perhaps there is more cultural willingness to surrender 50% of a company when you live in France and are already used to sending 50% of your income to the government. But even as an American, I think I could be convinced to give up 50% of my newly formed company for $113 million of runway...
Unless they are devoting that $113M to making and releasing open source models a consumer can run, which seems unlikely, I see no way they can succeed in building a better proprietary AI.
Google is already feeling the threat to their $100 Billion/year business and is willing to throw massive amounts of money at AI.
OpenAI isn’t so much a startup but should be seen as Microsoft’s strategic vision and as such has all the Microsoft resources to push it forward. Microsoft has poured billions into Bing for years in vain hoping to cut into Google’s search dominance. Now comes the chance to use AI to outcompete Google and they will pour their resources into it.
With neural networks, massive amounts of data tend to overwhelm any secret sauce. Google and Microsoft have access to massive amounts of data as well as the hardware required to process these massive amounts of data to train neural networks
I see this as mostly a way for a European company to ride the hype wave of AI, combine that with European paranoia and envy with regards to the US and make a lot of their investors rich.
Indeed. With all the costs we need to deal with here in Europe when running a company, there's zero chance anyone can afford to roll their own hardware for this without cash from the USA and that's not going to happen.
This company is dead before it got started. With the AI Act looming over it in the future even more so.
What is this hostility towards non-US companies I see on HN??
EU has a lot of small-medium companies in both AI and cloud and they are doing just fine. If anything, they are better at adapting to regulations compared to, say, openAI...
Ahhhh the French, tech grift, has always been celebrated for its excellence. There is a California tech grift, by Marcus Andreessen, inspired by that same French excellence. It's built on generative AI and like the best French tech grift, it's highly speculative.
Unless I missed it elsewhere in the article this part really glosses over something I have a hard time understanding. 113M seed round at a 240M valuation for a company that has no product yet and a plan to build something like something else that already exists. This feels insane to me.