Notes taken by Jay Goldman. Apologies for transcription errors and omissions. Feel free to use for whatever you'd like!
What will the Web2.0 World look like? Suggest your definition of Web2.0?
Nova: The labels have no inherent meaning. People come up with a consensus of what they mean. People had a backlash against Web 3.0 at first because it also has no inherent meaning. I think of them as decades (Web 1.0 is 1990 - 1999, Web 2.0 is 2000 - 2009, etc.). What was the focus of each era? Web1.0 was inventing the backend of the web and the infrastructure and core business models. Web2.0 is about the frontend and usability and things like AJAX. Web3.0, where we're headed, goes back to the backend. The innovation dollars and media will be around the semantic web or dataweb and making the web less like a file server and more like a database. It's not that we're going to invent something new in AI, but that the web itself will become richer and more interoperable. We'll be able to do more with the data that's on the web.
Elizabeth: My company tries to solve a much smaller problem than Web3.0. We analyze corporate data, sometimes in the millions of records, but it's a much smaller problem. How do you categorize information when people are trying to make it transparent and communicate in as many ways as possible?
Barney: From plumbing to meaning. We're doing increased levels of interop at the plumbing level. There's a new focus that's coming to add meaning and to make the plumbing meaning aware. All this content published thus far is designed for human consumption. When you want to compose, you compose based on the plumbing (insert a 'page', use a 'widget'). When you want to build an application, you have to struggle at the person level. We're seeing this shift with tagging where people are now starting to add a new level of meaning and now applications are emerging on top of it.
Powerset is building a new search engine based on natural language understanding. Two levels: content and query. If you can have meaning in the documents or queries, you can do a meaning to meaning match. Human users can use content at the level of thought rather than the level of machines. You'll be able to do searches with an assumption that you can use some language. The doors for PowerLabs will open in September. Social Networking community which will invite you to come in and play with us and the new technology and co-create. Right now you have to think about what documents you're looking for and what words will be in them, possibly with some advanced operators like quotes around phrases. If you wanted to find out what Hillary Clinton has said about global warming, you have to figure out the keywords that will give you that ("say", "told", "stated", "warming", "greenhouse", etc.). Right now we get "Did you mean?" but it's a spelling correction. We'd like it to say "Your query was ambiguous" and when you say "environment", do you mean "political environment"?
Nova - how does Radar Networks differ from that?
Nova: We're building community networks that help people add and augment the semantic understanding of the web. We're helping users do certain things which are normally a lot of busy work. Instead of forcing people to think in trees or hierarchies, we're helping them to think in ways that the brain actually works. This will likely be a subscription model, but we're not sure yet.
Barney: We're not really sure about pricing either, but today it's mostly advertising.
Elizabeth - I was teasing you earlier to say that your company gets rid of lawyers but you said that you augment lawyers.
Elizabeth: Discovery used to mean the guys who came in uniforms and name tags and emptied your filing cabinets. Now it means tens of millions of records in databases and the lawyers have no hope of finding it other than through keywords. Our system looks for things which aren't there like records which have been deleted or places where a phone call was had when you'd expect an email. We're in a different world because we charge for software. We have no content on our website but a good chunk of the Homeland Security folks show up every month. A lot of the things we've learned how to do won't help you avoid litigation but are just generally good things to do. It's not just the email that lands you in jail. We're mentioned a few times in the book Send about email etiquette. We do way better than a room full of lawyers using some of our own modelling techniques, but even things like automatic categorization techniques are much better. Most of what we do is in context - what does "Yes, let's do it." mean? Let's have a beer? Let's commit securities fraud? We look for where the context is, even if it's in a different record. We sometimes sell automated review where we're replacing lawyers, but usually only when there's very low exposure - customers know they haven't done anything but are being sued anyway. Generally we augment their lawyers and provide a service which hasn't existed before.
Barney - there are a few new search engines popping up like Wikia and Mahalo. What's with the new wave of 'human search'?
Barney: One of the elements is that the work one human does makes life better for another one. There's also human created content like Wikipedia, benefiting from a whole community of work. And there's also benefiting from someone else's small efforts, like tagging a search result. Mahalo is taking results which might not be great, like results from Google, and supplementing with new pages. Unlike Yahoo! Directory, this is in addition to search results. Human work to make good designed pages. I'm not entirely sure what Wikia is - they've talked about making better Open Source search tools and things like "I've done these searches which are better" to build on communities. Wikipedia is becoming a very interesting international asset embracing all human knowledge. Powerset is able to take advantage of all of this infrastructure that's out there. If people have already done work, we can use it. If you look up Einstein in Wikipedia, you'll find that he's already been categorized as a physicist, so if you came to a natural language search engine and asked which physicist had been asked to be a prime minister, we'd know.
Sarah Brown (Google): Elizabeth - seems like you've got an automated system that works really well. Have you ever compared it to human results?
Elizabeth: The lawyers don't like that answer. Humans are only right about categorization 60% of the time. The categories in the supine don't necessarily match up to the categories people filed things in. We've been assessed at 99.6% accuracy by an outside statistician.
Clay Shirky: What are you doing to bridge the gap between a human deciding that Einstein was a physicist and asked to be a prime minister and getting a machine to do it?
Barney: That info is actually in Wikipedia as a table of German physicists so we don't have to extract it. We're based on 30 years of research into Natural Language. We read a sentence (about 1 CPU second per sentence) and extract out the relationships. If the sentence was "IBM acquired Lotus in 1995.", the system determines that there was an acquiring event, IBM was the actor, Lotus was the object, and 1995 was the time period. If you asked "Who did IBM acquire in 1995?", we anaylze that and look through the graph of relationships for buying events where IBM is the actor and the time is 1995.
Joseph: Do you guys have an intuition into what will be game changing for a wide variety of people in this space? We can disambiguate that Jaguar is a car fairly easily.
Barney: There are some obvious ones and some less obvious. We play with our system a lot and you come across unthinkable queries that you would never even try. We're identifying these big projects whenever we think of one and trying to solve it. The "who said something" problem is basically impossible at this point, but of huge merit to journalists among others. THe question "Who had a complaint about the Treo 650?" is really difficult to solve. If you start with a vertical search engine, you have to start with the vertical and then construct a graphical interface around it and the user has to do that whole process in reverse to solve it. That goes away with Natural Language, which will become a killer application.
John: I have a Wikipedia entry and when it first emerged, I decided to an experiment. It was largely written by the Free Kevin movement, so you can imagine that it was a bit critical of me. I added a quote which came from a book which was quite critical of us and it stayed that way. Back in April, two words were added and I became a "Jewish American" writer. This was a bit troubling and none of my other Jewish American friends were identified, so I took it out and it came back and I took it out and it stayed that way. It turns out that the author was my 13 year old nephew who wanted to help me out and I was, for a time, one of the top 50 Jewish American authors. So, the question is, what about gaming the system?
Nova: So far nothing we've said here is really new. This has been going on for decades and none of this is really new. In a finite set of documents where you know the authors, it's much easier. In a much larger set where people might be intentionally distorting the truth, you get a very dirty knowledge base with assertions which aren't true and mistakes, so it turns out to be a pretty hairy problem with issues of trust. How do you assess the value of the different sources and change the weight of the graph? If somebody had written a page somewhere that said Einstein was named the Prime Minister of some country, your knowledge-base would have that fact in it. There are some statistical techniques you can apply based on, for example, how many times was a fact expressed? That doesn't help with weak signals which may be true but not known by many people. That's where you have to leverage people and communities of experts. When we talk about the semantic web and Web3.0, we have to make sure not to leave those techniques out of the discussion. Some of the tools from Web2.0 will carry over, like AJAX, and can help us build interfaces for rating sources or adding tags. I think the system will have a lot of errors in it, but the big ones can be detected and weeded out. People will garden the pages they care about which will become more and more accurate, and some less well-loved pages will remain incorrect for years.
Barney: If we compare these technologies to what would be perfect, we'll be very disappointed. If you compare them to what would be possible otherwise, you'll be much happier. If you think about doing this all by hand, you'll realize that this is pretty good. If you can drill down through how the system concluded a fact, you can adjust it. We need to set user expectations to not being perfect but being much better than what you used to be able to do.
Nova: For those applications which use reasoning, you can build some intelligence in to alert people when there might be a problem. The issue is that the current generation of reasoning systems live in memory and when something changes, you have to reason across the whole system again. That's computationally expensive and impossible right now.
Elizabeth: Some sources might be right about a whole bunch of things but intentionally wrong about others, so you can't really have a trust model at the source level but have to go further down and say that you trust a source for certain types of information. We have other issues because gaming our system tends to be more about things like switching to a different language for a brief period and switching back so that traditional searches through emails won't find that part.
Barney: We can do things like analyzing coherence, which looks for parts of a page which actually doesn't make sense and ignore it as spam.
Nova: We're moving from a statistical approach to search to a linguistic approach. Most of the current spam is an attempt to mess with Baysian filters and sneak things through. We can detect if sentences don't make sense or aren't grammatically correct, but the next generation of spam will be linguistic spam which is correct linguistically but not actually true.
Audience question: How easy is to to adapt your systems to other languages?
Nova: There is some research out there for major European, Arabic, and Slavic languages. Ancient languages are lacking but catching up. We're not reinventing the wheel but leveraging work that's already been done. Without respect to the language itself, you can look at the structure without knowing the meaning and be able to do some analysis. Numbers are generally language independent as well. You can also just look at patterns of strings.
Barney: We're leveraging some work specifically on multi-lingual technology. There's a theory that all languages are based on one deep root human language and all languages are just variations of that. PARC has gotten linguists together for 15 years and made them do the same tasks in different languages and it's proven the theory. A small team can come up to speed at a very high level in 1 - 2 years based on this.
Elizabeth: Our model is different because we look for things which stand out. If you always write to your Grandmother in German, than it doesn't stand out. If you never write to someone in German then suddenly do, it stands out.
Clay Shirky: Authoritativeness is obviously a big issue. Do you have an authoritativeness metric? Is it intrinsic like Technorati's is, or is it imposed?
Elizabeth: We don't measure it in the same way that Nova and Barney. Some can be authoritative and understand a small amount and it's authoritative as far as it goes. It's a problem but we know who the people are so it's less of a problem for us.
Nova: There's been some research into hubs and authorities and authorities have a lot of links pointing to them and hubs have a lot of links pointing away. That's what Google looked at to figure out their algorithms. Bibliometrics. You can look at authority on a global level but people will have their own opinions about who to trust. Your friend may not be the best authority, but you trust them anyway. Authority and trust aren't the same metric but they do intersect. We look at proximity to me as one measure. Did I create it? Is from a friend or colleague? Related to one of those people? A member of a site I participate in?
Elizabeth: In our case, it's sometimes the opposite of what you might think. That document that goes out to everyone is often less authoritative than the one shared by a small number.
What's happened to the metaphor of the Agent?
Nova: Remember General Magic? Their vision of the semantic web predates all of this by at least a decade. They thought there would be autonomous agents which would go out into the ether and do things and report back. That grand vision of agents running all over the web is amazing and inspiring because the web is so distributed you have to have a solution scaled the same way. It's like the way an ant colony solves a problem. It didn't happen, but why? Tim Berners-Lee's vision of the semantic web was basically a restatement of the agent vision. The more intelligence you add to the web, the less the agents need to know and the dumber they can be. The precursor to the semantic web was from DARPA can called DAML. We'd need an agreement on an infrastructure so that all the agents could be hosted on different sites and do their thing. The model that exists today is that the agents aren't autonomous but basically just processes that run outside like the Googlebot. We're 10 - 15 years away.
Barney: Agents are coming back so get ready for it. We'll have lots of computing power and infrastructure and you'll start expecting the tools to know more about you and your relationships. It doesn't matter if the architecture is agent-based or not - just that you come to expect that the agents are more than just simple processes and go out and do things.
Nova: What I mean by agent and what Barney means are very different, so this is becoming a semantics argument, which is why we need the semantic web.
Question from the audience: You want to be able to go out on the web and say certain things and have them be attributed to you. That's the real value of identity - not to make comments on people's blogs through OpenID which is fun but not the real point.
Nova: That's mostly already here. There are portals you can go to and download agents but can they get smarter than they are now? There's a trend or paradigm of software that can anticipate what your intent is and help you achieve that more easily. What's not going to happen anytime soon is the notion of truly autonomous agents in the vision of some of the founders of the semantic web. The model of the web today is client/server and the model of agents is that they're off on their own.
Question from the audience: It strikes me that the number of options in terms of widgets and platforms like mobile and desktop and Facebook is pretty close to agents. We're not quite at a client/server model anymore and have shifted to something different with the rise of cleaner web services and with objects moving around via AJAX or RSS and ways to exchange structured data objects and not just HTML.
Barney: I think you're right - no one has really imagined the next step. Right now we're at the point where there's a lot of objects published and a programmer can do a mashup but an average person can't go out and do it when they have an ad hoc task to take care of. When we get there you can gather this and do it on your own. Yahoo! Pipes is a bit like this.
Audience Member: Proto is doing this now to allow you to grab data and do something with it. Simple Windows application.
John: does anyone have a progress report on Psych?
Barney: It was a long term project to build an intelligent system with the knowledge of an encyclopedia to have a basic common sense knowledge. The DOJ just did a review recently and found that it had some very deep knowledge in some areas but that they couldn't easily be combined together. They continue to exist and work and have funding and projects and have recently released Research Psych and other tools.
Nova: Seems like Wikipedia is doing the same sort of thing. Psych is like a cathedral and Wikipedia is a choral reef, but I think it will do a better job of keeping up with culture.
Barney: They're very different though. Wikipedia is a big block of text and Psych is structured knowledge.
Nova: True, but Wikipedia has some structure and keeps up better with the cultural shifts. Psych has a lot of holes in it.
John: Psych is able to answer questions like "Which US city is able to withstand a nuclear strike?"
Nova: But not things like which "Pop star is likely to have a top 10 hit?"
Clay Shirky: The problem isn't just that there are holes. There's a "Critique of Pure Reason". We're not often doing pure reasoning but usually guessing. If I leave the room and come back and the glass of water I was drinking is missing, what do I assume? Martians took it? More likely, my wife put it in the dishwasher. Psych lacks the ability to make those human assumptions.
John: I wrote an article a while ago about AT&T deploying a very basic voice recognition system that understood five words and it saved them 100s of millions of dollars in phone service costs. Go to United Airlines and make an experience to go from point a to point b and your experience will be better than Expedia. It works for a narrow domain space. The common wisdom in the 60s was that the economy always creates more jobs than it destroys. Is that still true?
Nova: It will always create more jobs, but they might not be in the US.
John: When will I be able to write an article about a call center which moves to the US and then comes back here to write software?
Elizabeth: I've heard that they're moving back because the efforts to train Indians to look, sound, and feel like Americans failed.
Nova: I'm not sure why they cared. I just want someone to answer my question. I think systems like Firefly, the Personal voice-based assistant, were ahead of their time. Mobile devices with smaller screen real estate are pushing us towards voice based interfaces, which adds a challenge to the already difficult problem of the meaning of text by laying on the meaning of voice changes. There have been some efforts here like VoiceML, but it's limited. There are some systems out there which are as good or better than humans at transcription, being used by the intelligence services. It's great to have all this semantic metadata in your data, but where does it come from? Our AI might be good enough to insert it, but there's a good chance that it's going to have come from people who get hired to put it in by hand. You may have more accurate metadata that way. Voice may not create jobs, but the metadata needed for it will. If it's good enough, it might start removing entry level jobs like receptionists or entry level tech support, but it won't be able to handle issues like "Windows is crashing and giving me this error."
Barney: Within a decade you will expect that all of your devices have some amount of voice-based interaction. This is like ATMs, which eliminated some number of tellers but really changed our entire interaction. There were no tellers at 3am who lost jobs.
Elizabeth: If you look at your experiences with voice systems, the most irritating part is where you want to do something that isn't within their system and they want to keep you in it for economics.
Nova: The best part is that you can say "operator".
Question from the audience: What are the downsides?
Barney: If it results in better search, there aren't any. Our biggest challenge is in managing user expectations. It's not going to be perfect and this is the next 20 years of innovation, so we have to make sure that people come in with the right attitudes and get value out of it and help make it better. If they come in with the wrong attitude, they'll get disappointed and not come back.
Nova: People are much more tolerant of their own mistakes than when you make mistakes for them. Tagging and blogging are full of mistakes but we tolerate them because we know the human mind is fallible. There are privacy issues and concerns we haven't addressed. When you're making inferences across different data sets, not all of which should be shown to everyone, how do you determine what you can answer? Are there privacy concerns that arise as a result of chaining across data sets?
Audience member: I don't know if anyone has used ???? Networks, the free 411 system, but it's very good. they've just received a patent on serving targeted ads based on a voice system searching a database. People are using it and their cost structure has to be that there are no people involved.
John: When I look at Microsoft's attempt to put WiNFS under Vista to create relationships between documents. I've been thinking about Steve Jobs' dictum last week that the design point for apps on the iPhone will be Web2.0. Is that thing that MS tried to put under the OS moving off into the cloud and we're looking at the Web OS?
Nova: The desktop is a file server. Oracle would like the platform data to be their database. There's a tension between the two models. The semantic web is somewhere in between the two. RDF is a big opportunity because it's a lot simpler than a relationship database and it's much better for mashups because it's easier to mush together and you don't have to deal with relationships. I think that may become the platform for the future. With WinFS, MS wasn't trying to do something that simple. More like a database. They missed an opportunity with WinFS but it's not the end of the game. They have such a huge install base, so if they introduce a new storage system on the desktop in the future, they'll have vast numbers of people using it. The question is, really, will the data even be on the desktop? I think the data is moving out into the cloud and everything will be done through the browser with the desktop as a thin client. Google is pushing heavily for that. In a mobile world with limited compute and data resources, you'll need it to be remote. How relevant is the desktop in the future?
Question from the audience: You mentioned a lot of things but didn't talk about a dataweb where elements are more linkable than today. Pulling together elements from different databases to join them together.
Barney: If you think about it right now, we have URLs but we don't have URLDBs. We have URIs, but not UDBRIs. If your path can create a path to the data and the query in one, then you can share them and it becomes a network.
Nova: One of the relevant technologies is Sparkle, which is a query language for RDF which can do a query across data in many locations and build a single result. Some of the building blocks for this worldwide database are coming but it will be a year or two before they're really here. This is the big movement of the next ten years.
Barney: I have a paper on my website about this kind of thing.
Question from the audience: Let's look forward and assume we have a worldwide database with the problems worked out. What kind of applications or changes can you imagine? Planning a vacation it's somewhat underwhelming to think of Expedia 2.0 as the end point of this whole process.
Barney: Travel is exciting but basically any task where you'd have an assistant do things for you. Planning a trip or a meeting or a workshop. Any thing with multiple moving parts assembled together. A product with parts and vendors and specs for example. What I think of as an order in my app is an order in your app.
Elizabeth: I look at it a little differently. There's never one truth so if you can align who said what with their motivations, from sources with some authority and trust, it will make my life much easier.
Nova: If you want a really far reaching, risky vision, I would take the 100 or more than 100 year view and say that what we're really all doing is participating in the evolution a new form of intelligence. It's a new and better form of collective intelligence than we have today, which is highly fragmented. Organizations are collective intelligences but aren't very smart. We've talked about adding metadata to knowledge and a bit about reasoning. We'll get higher forms of AI that live out on the web, but they won't be that important. More interesting is large groups of human beings, motivated by larger goals of intelligence, being more than the sum of their parts. There will be a new level of intelligence that will emerge within these groups, with a level of intelligence at a software level. Are we going to become the Borg? I don't think so. More like social insects, or super organisms, but maybe with lest of a caste structure.
John: Hopefully it will have a sense of intelligence. If Google is the new Microsoft, will they perform that function for the semantic web? Sometimes they seem to play in this space and give you two different search results for the same query.
Barney: Google is definitely the new Microsoft in this space. You've got a company like Google who is doing many different things. They have lots of things which make money right now. Is it worth addressing these advanced technologies with any serious intent now or do lightweight efforts and wait and see? There are innovator's dilemmas for them and Microsoft and Yahoo!. Google's interface has barely changed over the last ten years and Microsoft and Ask are doing more innovation in that area. Don't bet against the big players having a big role in the future.
Elizabeth: It's a question of motivation. Simple as that.
Nova: Google has the resources to do anything we've been talking about, but culturally they won't. Their DNA comes from a statistical approach to information retrieval rather than linguistic based.