Being familiar with Dileep's past research: the algorithm used by Vicarious here is very likely based on the Hierarchical-Temporal Memory (HTM) technology developed by Hawkins, Dileep and the folks at Numenta [1].<p>HTM is documented in a Numenta white paper [2] and in an older form in Dileep's PhD thesis [3]. Numenta has also open-sourced an implementation of HTM, NuPIC [4].<p>It is important to note that, while HTM is a very interesting approach to machine learning, it has never been shown to beat (or even concurrence) the state of the art on any problem.<p>I have been following Numenta/Grok's progress closely because they seem to me to have the most potential for pushing this tech forward. Also I like their open-source strategy compared to the 100% secret approach of Vicarious. Vicarious has an impressively bombastic PR (that you can see in action in the above article) but basically nothing to show for it (until now, that is, if this breaking of CAPTCHA turns out to be real and the be more reliable than existing state-of-the-art CAPTCHA-breaking algorithms).<p>Long before they had any working technology, Vicarious founders had been throwing around quotes such as: "[our goal is] in five years to make a human-level vision system. Anything a human can recognize visually our algorithm would recognize visually." (2011) [5]<p>And today, after creating a CAPTCHA-breaking program (which has already been done in the past; I'd like to see a benchmark comparison), they're basically claiming they've created a human brain.<p>They also claim they will have a strong AI by 2026 [6]. Bombastic much?<p>[1] <a href="https://groksolutions.com/landing-page.html" rel="nofollow">https://groksolutions.com/landing-page.html</a><p>[2] <a href="http://numenta.org/resources/HTM_CorticalLearningAlgorithms.pdf" rel="nofollow">http://numenta.org/resources/HTM_CorticalLearningAlgorithms....</a><p>[3] <a href="http://alpha.tmit.bme.hu/speech/docs/education/02_DileepThesis.pdf" rel="nofollow">http://alpha.tmit.bme.hu/speech/docs/education/02_DileepThes...</a><p>[4] <a href="http://numenta.org/nupic.html" rel="nofollow">http://numenta.org/nupic.html</a><p>[5] <a href="http://blogs.wsj.com/venturecapital/2011/02/10/vicarious-systems-says-its-artificial-intelligence-is-the-real-deal/" rel="nofollow">http://blogs.wsj.com/venturecapital/2011/02/10/vicarious-sys...</a><p>[6] <a href="http://www.artificialbrains.com/vicarious" rel="nofollow">http://www.artificialbrains.com/vicarious</a>
I point you here:<p><a href="https://news.ycombinator.com/item?id=6633515" rel="nofollow">https://news.ycombinator.com/item?id=6633515</a><p>Quoting:<p>The popularity of this story evidences how universally hated are CAPTCHAs. here are some of the submissions:<p><a href="https://news.ycombinator.com/item?id=6625245" rel="nofollow">https://news.ycombinator.com/item?id=6625245</a> (forbes.com)<p><a href="https://news.ycombinator.com/item?id=6625247" rel="nofollow">https://news.ycombinator.com/item?id=6625247</a> (kurzweilai.net)<p><a href="https://news.ycombinator.com/item?id=6625351" rel="nofollow">https://news.ycombinator.com/item?id=6625351</a> (technologyreview.com) <- Main discussion<p><a href="https://news.ycombinator.com/item?id=6626405" rel="nofollow">https://news.ycombinator.com/item?id=6626405</a> (vimeo.com) <- video of process in action<p><a href="https://news.ycombinator.com/item?id=6627848" rel="nofollow">https://news.ycombinator.com/item?id=6627848</a> (vicariousinc.tumblr.com)<p><a href="https://news.ycombinator.com/item?id=6628086" rel="nofollow">https://news.ycombinator.com/item?id=6628086</a> (cbc.ca)<p><a href="https://news.ycombinator.com/item?id=6628092" rel="nofollow">https://news.ycombinator.com/item?id=6628092</a> (wired.com)<p><a href="https://news.ycombinator.com/item?id=6629173" rel="nofollow">https://news.ycombinator.com/item?id=6629173</a> (wired.com)<p><a href="https://news.ycombinator.com/item?id=6629559" rel="nofollow">https://news.ycombinator.com/item?id=6629559</a> (newscientist.com)<p><a href="https://news.ycombinator.com/item?id=6629656" rel="nofollow">https://news.ycombinator.com/item?id=6629656</a> (dailydot.com)<p><a href="https://news.ycombinator.com/item?id=6629708" rel="nofollow">https://news.ycombinator.com/item?id=6629708</a> (mashable.com)<p>========<p>There are two other comments worth reading.<p><a href="https://news.ycombinator.com/item?id=6629173" rel="nofollow">https://news.ycombinator.com/item?id=6629173</a>:<p><pre><code> That's a lot of fancy words to say that they overfit
to their training data.
That makes this sound like a very typical result in
supervised machine learning (if it's a result at all).
They have used an algorithm to learn a brittle heuristic
that works in the cases it was trained to work on.</code></pre>
Is this really some master algorithm that can crack all kinds of CAPTCHAs?<p>CAPTCHA is pretty much an umbrella term for a variety of anti-spam protection and human verification procedures, many of them having nothing in common beyond the name.<p>Thus "cracking CAPTCHA" is meaningless unless you specify what type(s).<p>You'd need textual analysis, sentiment analysis, expression evaluation, OCR, full-scale image recognition, puzzle solving and a variety of other techniques to scratch most implementations. So it can't be one algorithm. It'd have to be many.<p>Then in most cases it'd probably be simpler to figure out some client-side exploit in the way CAPTCHAs handle and send data. Find a weak link, much like DC949 did with reCAPTCHA.