Why bet on Google in the generative AI race
I don’t want this post to be interpreted as an advertorial in favor of Google, just a cold analysis of what I see around me and on our clients’ projects at Coddity. It will inevitably evolve with the progress and the maturity of OpenAI or Mistral (and their API 🇫🇷).
[DISCLAIMER1] I’m taking Meta and Mistral out of the equation, my point being about the use of LLMs via API.
Why Google?
The main reason is that running a GenAI project at the scale of a company is, ABOVE ALL, running a software project, where you’re going to need to call a service that is available and stable.
Google Cloud is the only one among the GAFAM to offer, all at once:
An infrastructure that holds up
I’ve lost count of the number of 50x errors returned by the OpenAI APIs on one of our production projects, and yet without a heavy load. At Google, no error logs came back on a client’s project, even though it was under a heavy load. All of it with an LLM in the europe-west9 region.
I haven’t run tests of GPT on Azure, but I consider OpenAI and Microsoft (until further notice) as 2 distinct companies. Remember the November drama.
Proprietary LLMs that perform
If nothing equals GPT-4 (*), Google’s models have nothing to be ashamed of, especially since the release of the Gemini suite. Experience shows me that we ultimately need the power of a GPT-4 in a fairly limited number of use cases, and there, Google’s models are good.
Besides, the fact that Gemini is available in a “nano” version says plenty about the projection of uses that Google makes of it. Cherry on the cake, if you have a precise use, fine-tuning a model is very easy, deployable in a Europe region.
Besides, I’m surprised that few people noted that GCP was the first to offer LLMs deployed in the europe zone.
Stable products, designed for “production”
If the pace of releases is less sustained than OpenAI’s, I see in that a will to take care of what gets shipped. I have yet to come across bugs that are identified by the whole community and left untreated for months apart from an “ok. seen”, the way it is at OpenAI.
A point that should please: you can set the level of “safety” of what the model will accept and generate. If Mistral offers an equivalent switch, Google has built in several types and levels of control. Enough to make a whole legal department happy.
[DISCLAIMER2] I reaffirm my support for open source generative models, which we use as much as possible at Coddity. These are, in my view (and that of many others), the key to security tied to AI, an accelerator of innovation, and (last but not the least) the real counterweight to a desire for excessive regulation on the part of our politicians.
(*) I haven’t tested Gemini Ultra yet!