HackerTrans
TopNewTrendsCommentsPastAskShowJobs

borapdx

no profile record

Submissions

Show HN: Yottaanswers – get direct answers to queries from billions of vectors

yottaanswers.com
3 points·by borapdx·vor 4 Jahren·1 comments

comments

borapdx
·vor 4 Jahren·discuss
No you are not, because you do not know for sure if the device is really competitive. Most people assume such judgments are clear and obvious but that is completely wrong. It is really about other people think, including of course the patent examiner but others too. If they think FAANG co. patent is not, or partially competitive, they you are fine.

The entire patent system is so obscure, non-obvious and surprising that I would almost NEVER assume any patent is competitive to another just that may seem to be the case at the moment.

In addition, it is basically impossible that what you have is identical to their published patent. Even if they do overlap some of your stuff will be good for another patent, based on differences and you will be in effect piggybacking on that one.

The short moral of the story is one should NEVER assume ANYTHING is obvious in the patent system. And keep in mind I am not saying it does not work.
borapdx
·vor 4 Jahren·discuss
All good points but what matters to them is positive correlation with increasing short term revenue and showing they are in a phase they are extracting (some) profits from what they got billions from investors through the years.

On top of that, their relationships with gazillions of Chinese suppliers are heavily disrupted by the political situation but that is a subject nobody wants to talk about right now.

The moment when Amazon is going to reap what they are sowing now is not here yet but it is rapidly approaching. It will be marked by a BIG negative earnings surprise (see Netflix for future guidance).
borapdx
·vor 4 Jahren·discuss
Yes, there are cases users want links and there is no problem displaying them in the results with direct answers e.g as sources. It is really about the change of focus, to providing direct answers, in both short and longer form, as opposed to just the snippet. Snippets have been around from the very beginning and they served us and Google well but it is high time to move on from line and a half with bunch of dots in between words.
borapdx
·vor 4 Jahren·discuss
No, it is not just you, the quality of Google results has gotten dramatically worse. BTW, you can ask Google 'is Google in decline' and see for yourself (DISCLAIMER: my answer on Quora is highly ranked where you can go into more details).

In short, Internet as we have known it has fallen apart because there are hardly any true organic links left. They used to be the lifeblood of the Web, the links based SOLELY on the relevance and quality of the linked page. Such a notion is quaint these days.

The Web used to grow exponentially in the early days, from say 1994-2008. Google launched with 100M webpages in their index, then they grew it to 1B in 2000 but they were NOT the first to reach that milestone, they were beaten to it by a Norwegian search company FAST that launched alltheweb.com with 1B before Google in 2000.

FAST was acquired by Yahoo in 2003 and you can guess what happened to them. For a few years Google and Yahoo were playing off against each other in terms of the size of their indexes, with Google always in the lead. But around 2003 the game of numbers stopped as they both announced they would stop publishing numbers. The Web continued to grow, still basically exponentially, and the next big milestone was announcement of Cuil in 2008. Cuil was a competing search engine created by a top team of Anna Patterson and Tom Costello, also with Luis Monier from Altavista. Their claim was that they would launch with an index of 120B pages, with an index bigger than Google.

That was widely considered an outrageous claim, as the notion that Google knows practically everything was already firmly entrenched. But they did manage to stir things up a bit, to the point of Google issuing a vague release in their official blog claiming they knew about 1 trillion urls. Of course they did not mention anything about indexing all that but the damage was done.

Shortly upon launch, the quality of Cuil results turned out to be far worse than expected which is really a shame as their basic premises were spot on, apart from index quality. Cuil then promptly fizzled out.

Note that projecting exponential growth (doubling every 18 mo i.e. quadrupling every three years) since 2008 we would expect 4^4*120B or more than 3 quadrillion(!) for the size of the Web index, with Google knowing 8 times that.

Such an expectation is plain silly, especially having simple queries such as e.g. 'Novak Djokovic', or 'Roger Federer' on Google returning less than 100 results.

But all this is only a (smaller) part of the story. Indexing is now a LEGACY technology, more than 20 years old. Users expect much more than returning bunch of blue links with matched keywords in response to their queries. They want much of the time direct answers to their questions.

The technology to do it has been known for 10 years now, in terms of dense vectors also known as word, sentence and other types of embeddings. Direct answers would be then found by nearest neighbor search. The scale of the system would of course have to be in the billions. BTW, it is a very interesting open question how many direct answers Google can provide now, in terms of infoboxes/featured snippets. Google has been coy about the issue but in my professional opinion, as a founder of multiple search engines, the answer is no more than around 20B. Feel free to shed more light on the subject and challenge this number.

In summary, the time has come to have a system based on vectors and nearest neighbor search with billions of vectors, giving direct answers to queries, with no ads nor tracking,and hopefully with API too.

One more DISCLAIMER, such systems are online for all to try and play with, at https://qaagi.com (for causal queries about causes and effects of things with billions of ranked answers) and https://yottaanswers.com (for factoid and and general questions what/how/where etc. with billions of answers). Both of the projects are led and principally funded by me, Borislav Agapiev.
borapdx
·vor 4 Jahren·discuss
The principal step will be to move on from keywords and indexing, which are now a legacy technology, almost 26 years after Google started it all.

Returning blue links is a thing of the past, as the Web of yesterday is long gone. Blue links always were about surfing i.e. following hyperlinks just for the sake of it since the main premise was most of them were of high quality and quickly proliferating.

All that is gone now and the links are a promotional thing how to get paid in one way or another. This is why Google results have been deteriorating, regardless of tens of trillions of archived pages on the Web. Google has lost the principal ranking signal years ago.

The next huge scale smart information system will be based on dense vectors (a few hundred dimensions) such as in AI but the key will be much bigger scale, of (tens of) billions of vectors. Contemporary AI works won datasets 4-5 orders of magnitude smaller, getting bogged down in gigantic transformer models such as GPT-3 with 175B+ parameters that take weeks and millions of dollars just to train. One might wonder what is innate knowledge of such a huge model, and it is not much as one can see for themselves as GPT-3 is now open (until Apr 1).

The future will be based on embeddings that are NOT contextualized i.e. no separate vectors for different senses in superpositions. Such systems will not be based on ads nor tracking as the resources required will be orders of magnitude less than what is currently required at Google.
borapdx
·vor 4 Jahren·discuss
Good thread, there are more and more alternatives as Google has been in decline for years in terms of quality of results. This decline is out in the open as can be easily seen on Google itself.

DuckDuckGo is very good and they have been steady in terms of their results and growth for years. Kagi is nice too (I know the team as a disclaimer but their effort is a good one). You.com and Neeva, as well as others are getting in there too.

I founded a distributed search engine Wowd.com more than 10 years ago but we were a decade too early, not that it really matters now.

In terms of shameless plugs I also suggest YottaAnswers (I am a founder) which is a smart AI system capable of returning billions of direct answers. Direct answering is key as the open Web with high quality organic links is long gone and search results with blue links and keyword matching are now a legacy technology.

A key question is then how many direct answers (in infoboxes, bolded) Google is currently capable of. I think the answer is surprisingly small, around 15-20B in my professional opinion. That is not a small number per se, but is so compared to Google's enormous resources ($135B+ cash, 1M+ servers, 16K+ PhDs, 60T+ webpages, 10T+ user clicks ...) Of course I invite all parties to shed more light my estimate above as Google keeps it a closely guarded secret.

Voltaire famously said more than two centuries ago that all fiat currencies inevitably tend to their intrinsic value, which is zero. I propose a modified variant for ads, that also tend to the same intrinsic value eventually.
borapdx
·vor 4 Jahren·discuss
Yottaanswers came about as an effort to build a scalable AI vector search engine capable of billions of direct answers to questions, with no ads nor tracking.

The question of how many direct answers Google is currently capable of (e.g. 'how high is mount everest') is of particular interest as that number IMO is not nearly as high as many think. My (professional) estimate is 10-30 billion. To be clear, by direct answers I mean distinct infoboxes with clearly distinguished direct answers as opposed to blue links with snippets matching keywords from the query. I certainly invite all feedback on this question and the number above as Google keeps it as a closely guarded secret.

Yottaanswers is in that respect very large, containing billions of vectors and continuing to grow. Considering the power law of distribution of questions and answers, Yottaanswers already has a significant chunk of the fat head of the distribution (about 15% of the questions generating 85% of the answers).

Of course, a key goal has always been to demonstrate that building such a system with no tracking nor ads is very much possible with resources (many) orders of magnitude fewer than Google.