4 ms·

Hi Trevor and Jared, My startup is developing technology for natural language understanding (in contrast to NLP). We believe we can approach human-level unders

Hi Trevor and Jared, My startup is developing technology for natural language understanding (in contrast to NLP). We believe we can approach human-level understanding for standard texts (email, blogs, online articles) in 2.5-4 years. We are currently self-funded and can comfortably do so for about a year--by the end of which, we believe we can develop a fairly advanced demo that surpasses existing techs in some, but not all, areas. What kind of demos do you think would impress best-in-class recruits/investors to join our team? The current options we have in mind are: 1) Solving a subset of Winograd schema (commonsense reasoning challenge) with a general approach (i.e., easily extensible with additional investment in knowledge acquisition/data sources). No systems known to public can currently solve all of them (or even a subset generally). http://commonsensereasoning.org/winograd.html http://commonsensereasoning.org/winograd.html 2) A conversational agent capable of conversing with humans at the level of a 4- or 5-year-old (without resorting to typical chatbot tricks). 3) Surpassing the state-of-the-art systems in a couple of standard AI conference tasks or on datasets released by leading companies (Google, Facebook, etc). 4) Other tasks we have not thought of... Because of resource and time limitations, we likely need to focus the initial efforts on one, or at most two, of the options above. (A mature system should be able to do all those and beyond, but this is only for about one year from now.) A couple extra questions if you have the time: - Given the startup's long-term timescale before monetization and its technical nature, what sorts of investors should we focus on talking with? - Are there chances of IP leaking and causing problems with patenting our tech later on? What should we do to prevent that? Sincerely appreciate your time to answer these questions. -- Ken Noppadon

tlb 11y ago
What kind of demos do you think would impress best-in-class recruits/investors to join our team All the options you mention would be technically impressive (so good for recruiting hackers), but none of them are obviously on the path to building a huge business. You should lean towards the option where you can imagine having a near-monopoly in a huge market. There's a big danger of getting sucked into some application where better NLP isn't the differentiator. Most interactive agents have small enough domains (getting cable TV installed, parsing S-10s, triaging support emails, ...) that general reasoning isn't critical, and the companies that succeed in this space will be defined by other factors like sales and integration with the customer's IT stack. I wouldn't worry about IP leakage. I've seen a lot of companies, but none that failed because of IP leakage. They fail for a long list of other reasons: slow execution, bad product-market fit, founder breakups, etc.

tlb 11y ago

What kind of demos do you think would impress best-in-class recruits/investors to join our team All the options you mention would be technically impressive (so good for recruiting hackers), but none of them are obviously on the path to building a huge business. You should lean towards the option where you can imagine having a near-monopoly in a huge market. There's a big danger of getting sucked into some application where better NLP isn't the differentiator. Most interactive agents have small enough domains (getting cable TV installed, parsing S-10s, triaging support emails, ...) that general reasoning isn't critical, and the companies that succeed in this space will be defined by other factors like sales and integration with the customer's IT stack. I wouldn't worry about IP leakage. I've seen a lot of companies, but none that failed because of IP leakage. They fail for a long list of other reasons: slow execution, bad product-market fit, founder breakups, etc.