There have been a great many successes in the AI field. It's easy for people to forget though; once something works, nobody calls it AI anymore.

Personally I find I need sophisticated domain-specific heuristics to evaluate consumer reviews in particular spaces. For instance, when I buy a lens for my camera, I'm going to look at reviews, but it's tricky because there are always good and bad reviews for any lens. Some of the people who get bad reviews had a camera with a screwed up AF or they never really understood how to use the lens or what its limitations were. Then you look at say, a Sigma lens for the Canon platform and you'll see that different people are having wildly different results and you'd better just forget about it.

Dashboards and stuff like that is a waste of time. What I really want is something that creates a super-expert opinion that's maybe 3 sentences to a paragraph wrong. A bit beyond the state of the art.

----

More generally, I think Doug Lenat had the right idea with Cyc, but he went about it the wrong way. Had Doug not been able to make a comfortable living doing work for the government, he would have been forced to produce a revolutionary product, but he wasn't.

I think that the linked data space around the semantic web is going to explode and ultimately produce the "commonsense" knowledgebase that it takes to build real NLP systems.

Most of the people I know in the knowledge-management space are trying to develop expensive projects for government, pharma, legal discovery and such. I think, however, they are in the "pay a lot, get a little" businesses that are going to be disrupted by the next wave. On one hand you've got Google, Microsoft and a few biggies that are going to develop large-scale but low-margin products. The other side is going to be a vast, largely low-margin market of operators who pull from and add to the great linked data pool... which is going to grow like a Katamari ball until we reach the Singularity, maybe around 2025 or so.

I have decades of experience with old style AI and a decade of the new kind :-). I specialize in language understanding algorithms and near-pefect sentiment analysis is something I expect we'll be able to do eventually using the methods I've invented.

The top level bit to worry about is whether you are attempting to re-do something that is already known not to work. Litmus tests: Are you using models of language such as grammars that are explicitly programmed in? Is your system specific to a single language so that switching to another would require complete re-coding? Do you employ linguists? If you answer yes to these, then you are in trouble.

I discussed how to get a modern AI education in an early blog entry at http://monicasmind.com

I propose a shift in direction of AI research in the second video at http://videos.syntience.com and explain why that's needed in the first one. Three more videos discuss details.

I have a theory/motivational site (6 pages or so) at http://artificial-intuition.com

I'm available for high level consultation on these issues. I worked at Google (I quit 2006) and although I cannot talk about what they do, I certainly will have an idea about what will be required to outperform them, both short term and long term.

If anyone wants to support Syntience Inc. in our effort to get true understanding to computers, please get in touch.

  - Monica Anderson
  http://syntience.com

I am sure someone can write an AI program that is comparable with human intelligence on paper at this moment in time. But, the complexity of the algorithm would probably be exponential, hence massive parallelization is necessary.

I think the ibm blue brain project has the right approach for AI. http://en.wikipedia.org/wiki/Blue_Brain_Project

Or at least that is the way I heard it.

In my experience, the complexity of most NLP applications work against them in the sense that they're competing with simpler approaches that are less computationally intensive. Good technology will let you do things other people can't do, but if you want to make it a business you'll need to know how to compete against Google doing 80% of what you can do with simple pattern matching. Your results will need to be orders of magnitude better before you have something people will use, let alone for which they'll pay. And be careful not to get trapped in a field where you'll be competing against companies paying serious cash for commercial databases to which you do not have access.

I'm personally skeptical there is much of a market for sentiment analysis incidentally, but the same tools are pretty useful for search (preprocessing, etc.). I think you'll find it difficult to get third-party adoption unless your product drives direct revenue for someone or can very visibly improve their product. But the problems are important and worth solving!

2) I would build the technology and a cool showcase, but I would not directly target the consumer, but rather someone who can use your technology to make money.

Your technology might be interesting for advertisers to ensure that ads are only displayed next to articles with a positive sentiment towards to product. E.g. to avoid Toyota ad next to the latest news on stuck gas pedals. Maybe show a Volvo ad instead. I don't know if something like this is done at present.

Yes, there are already advertising networks that do semantic-targeted contextual advertising, e.g. peer39.com, collective.com. Not sure how they are doing though. Either way, it is something that we will look into. Thanks.

I've been working on alternative pricing algorithms for 8 years. Most people call this a waste of time. Recently, one of my algorithms started showing great promise. (I'll know within two months.) Do what fascinates you.

I work in NLP, my company has a sentiment analysis product. It's a very small part of what we do, and it's focused on a particular application. NLP itself has a very wide range of applications, some theoretical, some practical, some already in use all day long at very popular websites. Two months is a very short time for complex problems in NLP, believe me.

Regarding tech risk, I think you're taking your particular product niche and casting it to something unnecessarily broad. Polarity detection is just one possible task in NLP, which is itself a sub-domain of AI. It can be done reasonably well without using any real NLP techniques even (hence the crowded social media monitoring scene). That paragraph made it sound to me like you are stressing yourself out unnecessarily about 'solving AI'. It's sentiment analysis (which can be a challenge to get right), focus on that.

Regarding market risk, marketing new technology is a challenge for any new technology, and any given startup will have the same questions that must be asked and answered. The application of your idea may be marketable or it may not be, that has little bearing on the viability of AI/NLP. The best you can do is make sure the company has done due diligence and has a plan, hopefully with some market research to back it up.

1) You need to find out who the paying customers are in this space and what features they really want. It seems very possible that average consumers would have no interest in your service but that power users, marketers or some other segment might. Once you know where the interest is you can make better feature decisions. For example, you might find that reporting, not AI, is what drives marketing sales.

2) Fundemental improvements in AI are really hard. Look at the results of the first Netflix prize. Lots of world class teams worked on that problem and the winning solution only produced something like a 10% improvement in the results. If that's the only edge your business has over the competition I doubt consumers would even notice. On the other hand, AI has matured to the point where it's pretty easy to produce good results. I'd rather bet my business on predicatble, good results and treat it as a pleasant surprise if we make a fundemental breakthrough.

Good luck!

Listen attentively to your intuition. Your "worry" can stop you faster than any statistical or third opinion about whether AI is a hard problem. Doesn't mean you should stop worrying and continue. Spend time imagining your ideal solution in great detail --- touch, smell and taste it if you can. Judge based on how you feel about that end point.

I heard someone said "AI is what hasn't been done yet." You don't need to prove your worthiness to anyone or any community by "solving a hard problem". Instead, as chasingsparks put it, "do what fascinates you".

Market risk is more of a concern to me. If you're presenting results directly to users they'll probably care greatly about the quality of your algorithm. If you're just using aggregated data to drive, for example, marketing campaigns then the noise will be washed out given enough data.

Did you know, that large retail sites increasingly employ people to post merited positive, and no-merit negative reviews to funnel consumers into a buying decision? And, by economy of scale, they do this in volume -usually massively overwhelming genuine on-site reviews. Working out the impact is left as an exercise for the reader.

http://techcrunch.com/2008/06/30/pluribo-is-cliffsnotes-for-...

Not sure how to react to this but sigh.

There's a paper I'm trying to find for you about analyzing affect in news articles for the purpose of trading stocks - I think they got things working fairly well.