Dear Sergey,
Do you know that you can’t search your chat history in Gemini?
If there’s a way to kick off a letter with a gut punch, that’s got to be it.
I was once told that the best way to start a letter is with something truly impactful. And when one thinks of bizarre things in life, it’s hard to top using a Google flagship product like Gemini AI—only to realize it has no search capabilities within. This is not just surprising; it’s downright astonishing. It’s so stunning it outshines even Gemini’s habit of dodging simple questions with a shrug of “too political” or “too sensitive.”
I was tickled—honestly, relieved—when I read your leaked memo from yesterday, which is everywhere in the media. You said Google should ditch the nanny products, and I couldn’t nod harder if I tried. Someone could pen a riotous anthology on what Gemini won’t touch—whole chapters on its prim “I’m still learning” dance. One of the things that we unearthed around the time of the election was how Google Gemini, even a week after the election results were formally announced, would not answer a simple question on who won the US presidential election (https://geninnov.ai/s/FEJTLc). Not just because of the political era that we live in, Gemini not answering the simplest of questions involving the mention of President Trump and President Zelensky, let alone prompts involving Elon Musk with Carlos Slim, is nothing but a sure shot way of driving the users away notwithstanding its other tremendous abilities.
But that’s not what prompted this letter. No, it’s your other zinger from that memo: to build AGI, Googlers need to clock 60-hour weeks. As an Indian, I’ll sidestep that minefield—work hours spark more debates back home than cricket scores. Jokes aside, it’s not about the hours. It’s about the harmony. For AGI, Google needs to mesh all its capabilities internally. Let me elaborate.
Look, we love Google. Truly. It is possibly the best research organization in the world, and we are consistently amazed by the ideas and experiments that emerge from within. Last night, I was transfixed by Google Colab’s latest data science features. With your teams, this is almost a weekly affair for me: someone or the other will announce something that will leave us staggered.
What floors us more is the visible lack of cohesion across Google’s teams. Even as outsiders, one could almost detect the internal politics playing out between teams. For example, take Gemini again. It won’t search its own chats. It won’t provide Google Search results. The Search team, perhaps on the back foot, will have AI-enhanced features but would do its best not to mention Gemini! And this is not an isolated case—it’s a pattern. The wonderfully built NotebookLM has no integration without clunky links, even with the Drive, let alone any connections with Gemini. One cannot get Imagen’s capabilities without directly finding a way to its site, and the same is true for Colab, LearningLM and countless others.
Effectively, the very company that pioneered knowledge graphing has failed to apply it internally.
We’ve all heard how the Attention mechanism, that little gem that sparked a revolution, got sidelined years ago amid a dozen competing projects. Google birthed a titan of modern AI, only to let it slip through the cracks because everyone was too busy racing each other to the finish line. It’s not a tragedy; it’s a head-scratcher—innovation galore, but productizing it? The problem is bigger when one thinks about Google’s efforts at AGI.
A contrast with Apple will highlight how opposite extremes could result in similar struggles. Apple’s teams function as a single unit working on shared goals (I am sure this is not wholly true). Still, its carefully planned pursuits lead it astray in the chaotic world of GenAI. They are used to deliberative action based on detailed studies of the environment and appear flummoxed in a world where realities are changing on a weekly basis (https://geninnov.ai/s/lcNRZP). They still don't have the DNA to really keep releasing features without possibly someone going on the stage and making a big splash.
We had to wrangle four different LLMs, all flexing their fancy Deep Research muscles, just to cobble together that sprawling product table in the postscript—think “Google’s Endless AI-phabet.” Likely the longest beast we’ll ever wrestle into our notes, and not one of those brainy bots cracked even 60% of the list. We stitched their efforts together, and still—still!—we rattled off four more projects we knew cold that none of them sniffed out.
Amazon has a similarly diverse catalog of features, but they have got most of them categorized - however messily - through simple lists, all available from one landing site - their home page.
Google is the king of ad hocism. Every team is announcing something. Every feature arrives from a different part of the company. They all have nice names, logos, websites, and user flow. Try to use them together, though, and it’s like assembling a puzzle with pieces from ten different boxes.
Here is a simple project that Google products can do already but cannot do because of internal dissensions. Say I want to email my assistant the required documents for the US Visa. Currently, Gemini can create the list of required documents in a jiffy. Subsequently, I would need to search through file names in Google Drive laboriously. For the photo, the right background may require more than fetching the right image from Google Photos; editing them smoothly may require some Imagen features. After all that, I may need to combine everything using Google Contacts and Gmail to send to the right email address. God forbid if the same workflow also requires information from my Chrome history or written letter through Docs to complete the process.
I fear that Google will develop another landing site - or multiple of them - that would create different types of agents to combine features of Google’s products. They will only add to the confusion, rather than have one simple Gemini prompt, “Hey, grab all the required Visa stuff and email the assistant,” that will figure out how to do all these together with access to all my data in various Google pools.
For AGI, the point you raised in your memo is right: Google has all the tools to make it happen. But the problem isn’t 60-hour work weeks. The AGI isn’t going to be built in silos. For all its genius, Google cannot create artificial general intelligence without operating as one-stop intelligence itself. For AGI, Google needs to come together as one. AGI’s too big for splintered speedboats—it needs a mothership.
Yours in awe and a smidge of exasperation,
Nilesh
PS: A List of Ongoing Google AI Projects
The above list excludes abandoned projects like Bard, Lamda, Tailwind, Imagen Video, etc.
