Kurzweil Applied Intelligence

~ the goodbye ~

I work on machine vision now, which means the ancestors of my own trade are companies like this one: a 1982 outfit selling speech recognition to hospitals at a price and a speed that only made sense if the alternative was paying a human typist. I was two years old when it was founded and Turkey was not a market it ever thought about. I read Kurzweil Applied Intelligence the way you read the previous generation of the thing you do for a living, which is with more sympathy than the numbers deserve.

Ray Kurzweil had already shipped a reading machine for blind people in 1976 and started a synthesizer company, and in 1982 he started this one in Waltham, Massachusetts, to work on speech. The technology of the period was discrete recognition. You paused between words, the vocabulary was bounded, and the system had to be trained on your particular voice before it understood anything. That is a bad general-purpose product and a decent specialized one, which is why the company aimed at medicine.

Radiology and emergency medicine turned out to be close to ideal. A radiology report has a small vocabulary and a formulaic shape, the person dictating it is the same person every time, and the existing process was a doctor talking into a tape that somebody else typed up hours later. KurzweilVoice for general dictation and the medical products underneath it did not need to understand English. They needed to understand one doctor describing one kind of image, and that was reachable in 1990 in a way that open dictation was not for another decade.

The company went public in August 1993 and raised about $24 million. Less than a year later its auditors worked out that a meaningful share of the sales behind that offering had never happened.

~ the chronicle ~

~ the warehouse ~

The scheme started small, in the way these always seem to. Revenue got booked a few days before a deal actually closed, then a couple of weeks before, and then, in the words of trial testimony, the policy went out the window entirely. Sales were recorded against customers who had not signed. Salespeople were pressured into forging customer signatures. The product was boxed and shipped to a rented warehouse where it sat, so that the paperwork would have something physical behind it.

Coopers & Lybrand found it during an audit in the spring of 1994. Bernard Bradstreet, the president and co-chief executive, was convicted in May 1996 of conspiracy and securities fraud, and was sentenced to two years and nine months in prison plus a $2.3 million fine. Thomas Campbell, the vice president of sales, got eighteen months. The treasurer pleaded guilty and testified against them, and another vice president was acquitted. Investors in the 1993 offering lost roughly $24 million.

~ what it was actually good at ~

The fraud is what the company is remembered for and it is not the interesting part. The interesting part is that this was real working speech recognition running on early-nineties hardware, deployed in hospitals, saving actual hours, at a moment when the general public's idea of talking to a computer came entirely from films. The constraint was not that the researchers were bad. It was that continuous large-vocabulary recognition needs compute and data that would not exist for fifteen more years.

Everything the field did in that period was a workaround for that gap. You paused between words so the system knew where the boundaries were, and the vocabulary stayed small so the search space did too. Every one of those constraints is gone now, and they went away because the hardware moved rather than because somebody had a better idea about language.

~ where it went ~

Lernout & Hauspie bought Kurzweil Applied Intelligence in 1997 for about $53 million in cash and stock, folded it into a dictation division, and lined up an alliance with Microsoft. For a couple of years that looked like the right home. L&H was Belgium's biggest technology company, it was buying everything in speech it could reach, and it had a market value near ten billion dollars.

Then L&H turned out to have invented roughly $377 million of revenue of its own, most of it routed through Korea, and a Belgian court declared it bankrupt on 25 October 2001. Kurzweil Applied Intelligence had already stopped existing as a company by then. It is a specific kind of ending: a firm brought down by faked sales, sold to a firm that was faking sales on a scale ten times larger, and buried by the second collapse before anyone had finished writing about the first.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »