See the pitch memo that raised €105M for four-week-old startup Mistral

They anticipated the need for a Series A in Q3 24 in the memo "In that round (Q3 2024), we expect to need to raise 200M, in order to train models exceeding GPT-4 capacities."

I don't know, the pitch seemed generic?

Seems to me that it's the team + market that won the money.

I didn't notice that line, but my immediate thought on reading the memo was that they were going hard after European capital. Seemingly both governmental and from VC firms in Europe that are well-connected to the traditional establishment.

(Even if they don't get government funding, it looks like they've already positioned themselves to get a ton of support from the EU and from European national governments, who want local champions to compete with the American capitalists.)

My thoughts:

- it's a fairly well-written memo: clear, concise, and relatively well-structured. It's a refreshing alternative to the endless series of pitch decks which mainly induce groupthink.

- the insights aren't that original. I would guess that they are strongest on the technical level. Their product and industry insights are less original or penetrating (pushing open-source and the European angle may give them some advantages -- it almost certainly will not give them an edge over OpenAI, who seem to have a relentless focus on iterating their product, a "live player" CEO in sama, and a strong dose of healthy scrappiness and creativity.)

- my prediction: they will stick around for a time, build something impressive with their technical team, raise more funding, hire a large and strong technical team (with their enormous warchest, status as the European champion, and appeal to the open-source community), and basically become a solid "fast follower". However, I think it's less certain whether they'll be able to do anything really creative or innovative.

Their overwhelmingly strongest market is going to be large European enterprises (or maybe governments), who will have lots of specific niche problems. The downside of that is their org then moulds itself to fit the needs of enterprise customers.

More speculative: the only way they can really become dangerous will be by attracting a bunch of young top-notch 18-25-year-old devs, ML people, or other builders (basically, people with creative energy and without commitment to the "established ways of doing things"), throwing them in a room together, and giving them a remit to come up with the most crazy, ambitious, unexpected applications of AI that they can. A universal translation app customised for the needs of ERASMUS students. Some weird generative music app for Berlin nightclubs. The actual ideas don't matter, what matters is having 50-100 hyperactive GenZ builders energetically exploring the space of combinations of technology, product and market to identify the genuinely novel applications of AI.

But I predict that they will not do that. You cannot cultivate that kind of free-spirited entrepreneurial DNA while simultaneously seeking deep connections with the European political establishment.

So, my overall prediction: they will focus on the underlying technology and on connections with governments/large enterprises, and deemphasise product innovation. Basically, building an AI-powered French IBM, rather than Google+++++. (Google+++++ will be built by one or two lone young researchers following their noses and working on weird esoteric ideas.) If I am right, and product innovation is the key to harnessing the power of generative models, then these guys in Paris will be formidable but not dangerous.

Should be interesting to watch it play out!

This pretty much here.

"Oh but non-tech people don't know the value of technology" Being condescending doesn't help much, and I bet they know a lot of things the tech people don't know or don't care about.

And you might even disagree why it's not important but it is good to be aware, and not simply dismissive

Elaborating on that:

1. The founding team’s track record and credentials.

2. there is currently no major independent player in the EU of OpenAI’s scale.

3. EU entities will prefer working with an EU based partner over OpenAI.

Having the LLaMA tech lead is a win. Having someone ex-OpenAI can bolster their position even more.

Now the anti-thesis:

1. It is gradually becoming apparent that models trained on high quality data don’t need to be so big.

2. phi-1 from Microsoft is 3. In addition, it seems GPT-4 is not one model, but a collection of 8 models trained on different datasets. Along with the above, all of a sudden - training GPT-4 like models is coming within the reach of an average person.

All in - nobody has won this yet. And Meta open-sourcing their research was the cleverest strategic play that I’ve seen in a while. It has done so much to weaken their competitors while really pushing AI forward in an open way (so much for “OpenAI”).