AI Ecosystem– Google Gemini

Google gets its act together.

• Google has stopped panicking (see here) and launched a new large language model (LLM) which it claims takes multimodality to a new level (true) and greatly improves the model’s ability to reason (dubious).
• Most importantly, Google has productised Gemini meaning that there are different versions of it which are already being targeted at users, customers and different use cases.
• This is Google’s first foray into the AI Ecosystem, and we can be very sure that at some point soon, there will be a Gemini Kit to build AI services on top of it as well as a store to distribute and sell these services.
• Google has launched its latest LLM, Gemini but it has been deliberately vague regarding some of the details of what it is and its key characteristics.
• Google has not said whether Gemini is a foundation model like GPT-4 or a finished product like ChatGPT or Dall-E which are versions of GPT-4 that have been further trained.
• Given the benchmarking that it has done, I assume that this is a foundation model and that software development kits (SDK) will soon be available so that customers and developers can fine-tune it for their specific use cases.
• Google has launched Gemini in three sizes but has not said how large each one of them is yet another sign of how the once open and collaborative world of AI has become a hotbed of fierce competition.
• There are three sizes available:
  • First, Gemini Ultra: which will almost certainly be cloud-based and suitable for the most complex tasks.
  • Ultra is the largest model (hundreds of billions of parameters)
  • Second, Gemini Pro: which Google says is the best model for scaling across “a wide range of tasks”.
  • In practice, Google is admitting that to do a range of different tasks users will have to train and run multiple copies of Gemini and run them side by side.
  • If it was one model that could do everything, then depending on usage, a user would only need one copy of the model to be present.
  • I would guess that this model is 50bn to 100bn in size.
  • Third, Gemini Nano which has been optimised for efficiency to run on-device meaning smartphones.
  • There has already been plenty of noise in this space and given that the cutting edge is around 10bn running at INT4, this is roughly where I would put it.
• These products are ready for the market now and Gemini Pro has already been incorporated into Bard although the Bard entity is now very cagey about whether it is using PaLm 2 or Gemini Pro to generate its responses.
• However, in a very quick initial test, I have found Bard to be much more accurate than it was before even on some really obscure information as well as retrieving information from the internet accurately.
• These three versions signal that Google is ready to go to market with these models and I suspect we will now see Google set out its claim to the AI ecosystem with these three at its heart.
• This is the first sign for some time of what I have long believed which is that Google remains a force to be reckoned with when it comes to AI.
• It made a complete hash of its message in H1 2023 when it was badly spooked by the success of ChatGPT (see here and here) and Gemini is a sign that Google has calmed down and done what it is best at.
• Google has a huge installed base to which it will now roll out these products via Google Search, Gmail, Chrome, Android etc.
• Gemini Nano will be made available on Pixel devices, but I suspect that discussions are already underway with Qualcomm and MediaTek to ensure that their processors are optimised to run Nano.
• The net result is that Google’s new model looks like a big step forward, but it is crucial to remember that like every other model out there, it has no causal understanding of what it is doing.
• It is also incapable of reasoning just like GPT-4 no matter how many times Google and OpenAI use the reasoning word in their marketing materials.
• If these could properly reason, then the problem of hallucination and edge cases would be dramatically reduced and the benchmark figures provided indicate that this is not the case given how similar Gemini Ultra is to GPT-4 and how much GPT-4 can be induced to do stupid things.
• Google has finally got its act together and has laid a viable claim to the AI ecosystem and if the products are as good as it says they are, then it will be in a very strong position given that 4bn users interact with Google at least once a month.
• I think Google is finally justifying its position as the best AI company in the world, but the valuation of the company is pretty much reflecting this position already.
• I do not have a position in Google as the valuation is not yet compelling but if there is a sharp correction or signs of life in its growth, that could change.