Google AI tools are surprisingly underrated
Apparently a 2.2 trillion dollar company tools has not enought visibility
Google has a problem releasing: they start by announcing the product, which generates a lot of hype, but all we get is a landing page and a paper. After a few months, when people stop caring, Google quietly releases the tool… but it is a slow rollout in the US only. 🤦
Compare this with the OpenAI strategy. They create hype long before releasing, then casually drop it with a closed beta, with a public release right after.
It's also Google's fault that nobody follows its tools—the names change occasionally: They invested millions in Bard just to change to Gemini.
Gemini also has multiple tiers and versions with caveats, like “1.5 Pro” is better than “1.0 Ultra”. There are multiple tiers and versions, like 1.5 Pro 002—the last number being extra padded with 2 zeros means more will come to confuse everyone.
So, why should you care about them? Their AI is far from capable like ChatGPT 4o's — Gemini feels more like GPT 3.5. Well, because it excels in the needle-in-haystack problems.
ChatGPT fumbles a lot of its tokens. It theoretically gives you 16k tokens in GPT-4o, which seems like a lot, but it tends to forget the earlier tokens, the more tokens you feed it. This is probably because it might use a “rolling-window” approach, so it does not consider those earlier tokens as much as it should, but you still pay for all of them.
If you want to use Gemini tools, you must also use different websites with generic names that will probably change twice before the project gets killed. The tools are Notebook LLM and AI Studio. Not to be confused with LLM Studio, a popular FOSS tool not made by Google — I told you it was confusing.
Notebook LLM
Notebook LLM sells itself short. It claims to be a tool for studying and brainstorming ideas using your documents. They also claim they can turn documents into an AI podcast(?). At first, it seems like things a student would use to cheat help with their exams. But the tool shines by being an excellent “Google Search for your documents.”
If you feed it with large documents, it can retrieve information and make interesting critiques. For example, I am using this tool extensively for my new engineering manager book.
Having the AI do extensive critiques with sources to back them up helps me write better and not lose my agency as an author by turning things into slop. I don't want to use GenAI since it only gives generic advice from SEO-hungry websites. Using NotebookLLM means I am still in control of my writing.
AI Studio
What Notebook LLM can do for large documents, AI Studio can do for large videos. It is very useful to extract details from any media you have and seems to be the only tool to do so.
I imagine this tool being a game changer for people who use video to document things, like scouts for filming locations or real estate agents looking to write a pitch.
I sent 3 videos I made for later reference when I visited a house. I asked them to write a pitch for those videos. One thing that amazed me is that it saw a grapevine and used it for the pitch!
I've been using this tool for transcribing and annotating video evidence. My house inspection is 17 minutes long, and there was no way I was ever going back to watch the whole thing again. This transcription helps a lot to get the action points.
Something is very strange with it, though: Since the Lite mode worked so well. I thought the “Pro” model would work even better, but surprisingly, it was way worse. I don't have an explanation for this.
Another limitation from Google is that the model is fine-tuned towards safety. Some very casual videos I've uploaded triggered the warning that I might have been breaking the ToS.
It also uses the word “diverse” a lot; it probably means that, in its internal token, it should focus on diversity, but it starts using this word for everything from “diversity of rooms” to “diverse problems.” It uses this word like ChatGPT uses “delve.”
Gemini is far from beating ChatGPT and perhaps even Claude Sonnet. They are playing safe with what they have, but we should not count them out of the game. Those tools are currently free, and NotebookLLM does not train on your data. With ChatGPT, segmenting these documents/videos into tokens/images is prohibitively expensive. The thing left for us to decide is how long "free” will last…
If you are interested in the engineering manager book, I will send you the first chapters for free. These chapters will include the best and worst parts, how to succeed in the manager interview, and how the role differs in every company. If you provide feedback, I will send you the digital version of the book for free once it is released!