Comment by theptip
20 hours ago
> Google has world-class research teams that have produced unbelievable models, without any plan at all on how those make it into their products beyond forcing a chat window into Google Drive.
NotebookLM is a genuinely novel AI-first product.
YouTube gaining an “ask a question about this video” button, this is a perfect example of how to sprinkle AI on an existing product.
Extremely slow, but the obvious incremental addition of Gemini to Docs is another example.
I think folks sleep on Google around here. They are slow but they have so many compelling iterative AI usecases that even a BigTech org can manage it eventually.
Apple and Microsoft are rightly getting panned, Apple in particular is inexcusable (but I think they will have a unique offering when they finally execute on the blindingly obvious strategic play that they are naturally positioned for).
Google was the absolute king of AI (previously "ML") for at least 10 years of the last 20. They are also an absolute behemoth of tech and have consistently ranked among the most valuable companies in the world for multiple years, valued at trillions of dollars today. Hell, they're on version 7 and production year 10 of their custom AI ASIC family.
When considering the above, the amount of non-force-fed "modern AI" use they've been able to drive is supposed to be shown by things to the level of a question button on YouTube and some incremental overlaying of Gemini to Docs? What does that leave the companies without the decade head start, custom AI hardware, and trillions to spend to look to actually do worth a damn in their products with the tech?
I'm (cautiously) optimistic AI will have another round or two of fast gains again in the next 5 years. Without it I don't think it leaves the realm of niche/limited uses in products in that time frame. At least certainly not enough that building AI into your product is expected to make sense most of the time yet.
> YouTube gaining an “ask a question about this video” button, this is a perfect example of how to sprinkle AI on an existing product.
lol if this is the perfect example, "AI" in general is in a sad place. I've tried to use it a handful of times and each time it confidently produced wrong results in a way that derailed my quest for an answer. In my experience it's an anti-feature in that it seems to make things worse.
The best and latest Gemini Pro model is not SOTA. The only good things it has are the huge context and the low API price. But I had to stop using it because it kept contradicting itself in the walls of text it produces. (My paid account was forced to pay for AI with a price hike so I tried for a couple of months to see if I could make it work with prompt engineering, no luck).
Google researchers are great, but Engineering is dropping like a stone, and management is a complete disaster. Starting with their Indian McKinsey CEO moving core engineering teams to India.
https://www.cnbc.com/2024/05/01/google-cuts-hundreds-of-core...
It was the best model according to almost every benchmark until recently. It’s definitely SOTA.
There are problems with every model, none of them are perfect. I've found Gemini to be very good but occasionally gets stuck in loops: it does, however, seem to detect the loop and stop. It's more cost effective than the Claude models, and Gemini has regular preview releases. I would rate it between sonnet and opus except it's cheaper and faster than both.
For whatever reason there are tasks that work better on one model compared to another, which can be quite perplexing.
No amount of big context window can stop the model from context poisoning. So in a sense, it's a gimmick when you start having the feel of how bad the output is.
> when they finally execute on the blindingly obvious strategic play that they are naturally positioned for
What's that? It's not obvious to me, anyway.
inference hardware, especially starting with on device ai for the mac. I think they should go as far as making a server chip, but that's less obvious today.
My guess would be local AI. Apple Silicon is uniquely suitable with its shared memory.
Yeah exactly. The MacBook Pro is by far the most capable consumer device for local LLM.
A beefed up NPU could provide a big edge here.
More speculatively, Apple is also one of the few companies positioned to market an ASIC for a specific transformer architecture which they could use for their Siri replacement.
(Google has on-device inference too but their business model depends on them not being privacy-focused and their GTM with Android precludes the tight coordination between OS and hardware that would be required to push SOTA models into hardware. )
3 replies →
Embrace the vibe, man
The biggest counterexample would be that dead-ai-autotranslate-voice sucking every gram of joy out of watching your favourite creators, with no ability to turn it off.
> YouTube gaining an “ask a question about this video” button, this is a perfect example of how to sprinkle AI on an existing product.
I remember when I was trying to find a YouTube video, I remembered the contents but not the name. I tried google search and existing LLMs including Gemini, and none could find it.
It would also be useful for security: give the AI a recording and ask when the suspicious person shows up, the item is stolen, the event happens, etc. But unfortunately also useful for tyranny…
Yeah to be clear, I think Google is the strongest in AI product development of the FAANG companies. I included them in the list because the most complaints I see about AI product integration among FANNG comes from Google products; the incessant bundling of Gemini chatboxes in every Workspace product.
Those examples are interesting and novel, but don't anywhere near live up to the promise of the next great technological revolution, greater than even the internet. I'm fairly sure if an all-knowing genie were to tell Google that this is the best AI gets, their interest in it would drop pretty quickly.
I think for most people, if NotebookLM were to disappear overnight it'd be a shame but something you can live with. There'll be a few who do heavily rely on it, but then I wouldn't be surprised to hear that at least one person heavily relies on the "I'm feeling lucky" button, or in other words, xkcd 1172
If its really useful, how long do you think it will take Google to kill it ? ;-)
> Apple in particular is inexcusable
This isn't me defending apple, but, let me play out a little scenario:
"hey siri, book me tickets to see tonight's game"
"sure thing, champ"
<<time passes>>
"I have booked the tickets, they are now in your apple wallet"
<<opens up wallet, sees that there is 1x £350 ticket to see "the game", a interactive lesson in pickup artistry>>
You buy apple because "it works" (yes, most of that is hype, but the vertical integration is actually good, not great for devs/tinkerers though.) AI just adds in a 10-30% chance of breaking what seems to be a simple workflow.
You don't notice with chatGPT, because you expect it to be the dipshit in your pocket. You don't expect apple to be shit. (although if you've tried to ask for a specific track whilst driving, you know how shit that is. )
I mean Microsoft hosts key AI models in their AI Foundry, I don't think they're hurting.
https://ai.azure.com/catalog
> YouTube gaining an “ask a question about this video” button, this is a perfect example of how to sprinkle AI on an existing product.
> Extremely slow, but the obvious incremental addition of Gemini to Docs is another example.
These are great examples of insulting and invasive introductions of LLMs into already functional workflows. These are anti-features.
The Ask button in YouTube is a game changer for the use case of "what timestamp in this hour-long video talks about topic x?".
What's the existing functional workflow for that? Downloading the captions and querying with a local LLM or a very fuzzy keyword search?
Perhaps this is a difference in terminology, but in no way do you need a LLM for fuzzy search. Semantic search, fuzzy keyword search, and text to speech have existed for years and predate the technology for an LLM. In your use-case, do you really need a chatbot to "ask the video" about this, wouldn't a "search in video" function that does the same thing be better?
I guess I’m using the product wrong if I find them useful?