Google just released the holiday gift all of the techies have been waiting for: their new AI model, Gemini 🎁
And it's screaming two words to me: REASONING and PRODUCTION.
(Also, for those of you who don't know this, when I speak with someone for an hour, I get a weird Sims-like visual of words over their head of name values, so when I say that words stand out to me, I literally mean the words pop up from the page. And the same thing happens with AI news. Is that weird? Am I alone? Anywho, back to Gemini...)
Here's what my followers need to know:
1) We're moving from 'knowledge' to 'action' - AI is going to move further upstream and will be tasked with taking action, and guess what, actions require reasoning. Something I've stressed heavily for the last several years is the need for critical reasoning in this process. That's the 2024 focus—AI moving into proactive interventions.
2) We're getting more efficient - AI has become 16500x more efficient in the last 11 years, and we're starting to see that for state-of-the-art large language models. Say what you will about GPT-4 performance, but it is too expensive for most companies (including OpenAI itself!) to run at scale. Assume 2024 brings lower AI OpEx.
3) It's multi-modal from here on out - ChatGPT is multi-modal and doesn't require the user to preselect what modality they want to use, Gemini seems to be the same. There will be no lines between images, music, text, voice (always a big part of Google demos!), and video. Just one pile of "data". Assume that 2024 is multi-modal.
4) There's no one model to rule them all - we've know heard this from Andrej Karpathy, Google DeepMind, and more, but there will be different models for different use cases. Gemini has 3 tiers: Pro, Ultra, and Nano. Pro is for Bard and Google products, Nano is for Pixel, and Ultra is for "I have deep pockets and want the highest grade possible." Performance will be a consideration: with a score of 90.0%, Gemini Ultra is the first model to outperform human experts on MMLU (massive multitask language understanding). Assume that 2024 has more multi-model orchestration and delegation.
That's it. I rushed to get this out. And I need to eat some lunch.
Is Google back in the game? What else stood out to you from this release?
Give this post a save for your 2024 AI planning, and drop a comment below ⬇