A simple paper worth a trillion dollars
39,176
Published 2024-07-17
Web Search Engine (1998) infolab.stanford.edu/pub/papers/google.pdf
Attention Is All You Need (2017) arxiv.org/pdf/1706.03762
Thank you to Yossi Matias, Google's head of research, for speaking with me.
Subscribe to my channel to see more videos like this: youtube.com/user/tibees
Support me with a monthly donation on Patreon and receive a gift from me: www.patreon.com/Tibees
Buy me a coffee (one-off donation): ko-fi.com/tibees
Website: tobyhendy.com/
TikTok: www.tiktok.com/@tibees
Twitter: twitter.com/TobyHendy
Second channel: / @tibees2
Editing by Noor Hanania
All Comments (21)
-
Has that guy from Google recently tried to actually....use Google?
-
Google today is everything this paper criticized.
-
they predicted quite correctly how their search engine would become unusable.
-
Oh the irony of 'cellular phone' searched on google now returning 90% 'cellular' companies advertising their products with only one wikipedia link not being an ad. Google still works well but you've once again got to be more specific to avoid being sold something.
-
AI used in search.............. Alexa, I need emergency medical assistance. Alexa responds, I have place emergency medical assistance on your shopping list.
-
What gets my goat is I search YouTube for one video outside of my normal range, and the next day every fifth recommended video is related to that. There's no accounting for desire of variety. I miss the old days when the front page had recommended videos. Back when that meant something.
-
Yes AI has been ruining out searches for some time now. We have noticed. Clearly it's been sending more advertised results than the wanted results.
-
"Don't be evil" Google code of conduct! (Now removed)
-
"AI Search", powered by LLMs, Isn't even search; it uses a language model to respond to a prompt by generating a likely response based on the training data. The training data sets are not public, nor are the parameters given to the model generation along with the data, nor are the specific model generation algorithms, so the gaps and biases in the model can only be assessed through the responses it gives. For mimicking the general characteristics of text in the training data they're pretty good, but for factual information they're only any good at facts which are repeated consistently in the training data with a context similar to nearly every possible prompt which would be constructed as a query for that fact; for anything less common the response can only be correct by coincidence. Of course, not all "AI" is LLMs, but as far as has been made public "AI Search" is an LLM trained on the crawled pages, and is not trustworthy for generating correct answers or even factual statements
-
We need a new paper. Now the internet is only 200 websites and finding good information is harder than ever. 😢
-
The last 3 minutes hits differently!
-
Always remember: Your privacy is important to us.
-
Fritz Haber's 1908 "Synthesis of ammonia from its elements" has been worth hundreds of trillions.
-
“You can judge a man by the books in his library.” — Mark Skousen
-
I very much like your comments about libraries. Research librarians are angels.
-
As a great philosopher once said, "money over everything"!
-
That scrap book journal looks wonderfully maxamillist
-
Advertisers are paid for the attention they get, but we are rarely, if at all, rewarded for the attention we give. Are y'all paying attention?
-
Regardless of their paper's premise, here we are.
-
I thought you were going to be looking at Claude Shannon's Master's paper