Note: this article from the well-known technology investor Chris Dixon's personal blog cdixon.org
Visualize business processes, in a thinking puzzle idea maze/, a qualified entrepreneurs can predict to a parting of the ways how to turn to, it will lead the company into the Fortune top; the opposite way, is doomed to decline. A failed entrepreneur, but only knows to follow catch up streaming, photo sharing, P2P wave, but their industry's history was unknown, they will not fail before lessons learned experience, and they can't predict the next new technologies will cause a change.
I think that gives more specific instances of this question would be more interesting. So I choose a case of AI start-up companies, here is my idea for artificial intelligence entrepreneurs drawn maze sketches.
Accuracy of 80%-90% MVP
Machine learning circles there is a popular saying is "machine learning is really good at partially solving just about any problem/machine learning can solve most of the problems". Say yes, to make a general model that can be applied in the 80%-90% case is not difficult. But after that, that is takes time, energy and financial resources put in to maintain the normal operation of the model. According to General experience, achieving accuracy and versatility of 80% only takes a few months to, but the rest of the 20% part is able to draw on many years of time, or are implemented throughout his life.
User experience tolerance
In the first step you have to make hard choices:
1, or continue to study it takes accuracy to 100%
2, suitable for most situations, but part of the right product
If you select the 2nd, that experience commonly referred to as the "fault tolerant tolerance UX/user experience"
In our daily lives have been unwittingly exposed to too many examples of tolerance. AutoCorrect feature the iOS system, for example, or in the Google search results "did you mean x? "。 Or broad said, you can use the Google search engine itself as a model of good user experience tolerance, because it does not go directly to the first search result, but 10 results are pushed to the user to choose from.
But when you decide to implement 100% accuracy, then it is a new approach. After all, the algorithm was invented by humans, of course, you can not simply rely on algorithms to help you achieve your final 10%-20% the accuracy of the target, you can only "rote way"-train with as much data as possible to set-up your model. Data is a key element of AI, because:
In algorithms and computing resources to complete cases, data is only missing the key point and the most important ingredient;
Algorithm is made most public resources, and good data is either not yet born or were privately owned.
Continuous segments
Data you have to do only one thing: constantly broken down. Even if you live in a subdivision in the field, and also try to further subdivided. If you want to build a robot that can solve all the problems, than to build a robot that can help schedule.
If your goal is to build an x, then it needs to be subdivided areas established under the MVP, to achieve some of the x, to conquer it is possible to achieve overall x. One suggestion is that to try to keep subdividing until you can't segment limit, after all, when after the success of the project, and then gradually expand the market but is more easy for me.
How to get the data
Broad said the data came in two directions:
1, establish their own data
2, crowdsourcing
Of course we can cite two vivid examples, Google map and Waze. Google map of huge databases is established in thousands of employees amount of surveying in the field prospecting, Waze is called crowdsourcing with the baiwanhao people in the world to achieve this objective. So unless you like Google with deep pockets can afford to go it alone access to resources, or still more suitable.
So we went to the next key nodes, how startups get data?
1, public resource data "into their own"
2, crowdsourcing
Of course, the first method is the most simple and crude example from Wikipedia "pull data". In fact there are many startups have tried to hold Wikipedia good legs, but did not have much success.
For a small team, the least expensive and most feasible way of course, is crowdsourcing. All the key points for the success of the package is, how to design the perfect incentive systems, allowing users to retrieve data from the database, able to consciously data returned?
In this connection, I share with you their experience. Last year I invested in a company called Wit.ai, Wit was developed to provide voice to text service or a natural language. Frank said that the Wit.ai version 1.0 error rate is not low, but the company provides API interfaces and developers are invited to test and correct the bug Control Panel. In the process, many users not only to use the free services, but the majority of users "crowd wisdom", the Wit.ai system is becoming more perfect. To celebrate, then Wit the company acquired by Facebook, but this whole process really can be called a model for many entrepreneurs learn from.
Tips
Above is just my idea of AI start-up companies draw maze rough interpretation of the sketch, but there still are some tips, Kenzo france
1, draft conclusions offensive references, can become a discussion here, but are not able to become business creed
2, the success of new technology is mostly "business as unusual", just as the Internet, smartphones, cloud computing, and the currency, at first we thought it was a little small things, but finally have set off waves of technological change.
So don't take any of you so-called "minor details".
viacdixon
1087 votes
Millet router iPhone case Kenzo
As millet router upgrade, new millet router global launch Broadcom Broadcom 4709C dual-core 1.4GHz CPU, with separate 512MB system Flash memory with 4 Gigabit wired network interface, built-in Seagate/Toshiba professional supervisory-level hard drive, normal 1TB, optional 6TB, support for mobile phones and cameras automatically upload.
https://www.youtube.com/watch?v=o3Ah8RAVZyo
View details of the voting >>
No comments:
Post a Comment