The 55th Anniversary of the Internet

Congratulations to UCLA on the 55th anniversary of the Internet. My fifty-five-year UCLA career ran alongside the evolution in computing. My computational roots include a father who taught slide rule at San Mateo College in the 1950s, and a graduate-school mentor who learned to do factor analysis and matrix inversion on a 20-dial Monroe calculator, wired punch boards for matrix multiplication, and used large graphs to find the tetrachoric correlations corresponding to different proportional cuts. Computing had advanced greatly when I entered Cal as an undergrad in ‘62 and Illinois as a graduate student in ’66. I joined the UCLA faculty in July 1969, just a few months before the first Internet message from UCLA to Stanford from Node 1 pictured above. UCLA’s IBM 360-91 did the computations for my dissertation involving both a new multidimensional-scaling method and a 3-mode analysis of the object-to-object dissimilarities over people. I learned FORTRAN 2, but moved to stat packages as soon as possible. I ditched my Illinois stat package when it became clear in the early 70s that SAS would be the standard for quite a while. I remember when my first large market-model reported degrees of freedom in scientific notation having run out of fixed-width space. That was one inflection point.

?

I was part of a small cadre that pioneered market-response modeling: market-share models, market-mix models. Scanner data provided commercial-scale problems from the late 1970s on. The academic swap was data access in exchange for published methods. The computational infrastructure available at UCLA and other top universities could handle the scale of those times. When I foresaw Internet 1.0 data in the later 1990s, I felt UCLA didn’t have the infrastructure to support that scale. I co-founded Strategic Data Corp with Giovanni Giuffrida. He and I published the first data-mining article in the major management literature, Management Science. Prior to our work, datamining was considered by most management scientists to be a collection of heuristic techniques with little statistical foundation. We took the residuals from a published market-response model and looked for (and found) stable patterns in what statisticians considered white noise. Giovanni’s specialty was rule-generating dataminers, which gave understandable rules for managers to implement. The alternatives were neural nets where explanations are opaque. Areas such as FICO credit scoring were provided a natural incubation environment for neural nets since opacity was not a liability, but an asset when most readable rules were blatantly discriminatory. In a time when 24 hours of Internet-log data took 36 hours to analyze, Giovanni’s dataminer and the segmentation system I developed from the US Census scaled to 9 billion optimized decisions a day by the time Fox Interactive Media acquired SDC in 2007 and moved to scale in support of MySpace and the family of companies in Fox Interactive Media. ?By then Hadoop was mature enough to enable neural-net learning over distributed databases, rather than the matrix style, multi-modal arrays that dominated prior analytical systems. That was a second inflection point for AI systems development. I was emeritus by that point, and while still active in campus-wide endeavors, no longer was doing analytical modeling.

?I taught in the 1990s about the process of disruptive change in technology landscapes. Most disruptive technologies start as inferior products. Steam ships in the early 1800s were inferior to the clipper ships that dominated Atlantic trade. Steam ships were slower and carried smaller loads at higher cost per ton. Clipper-ship companies listened to the voices of their best customers and did incremental innovations, such as putting enlarging new ships a bit and putting up another mast. Steam ships incubated in the river trade where their ability to go against the prevailing wind gave them an advantage. In that milieu, steamers innovated at a more rapid pace than clipper ships. When steam ships emerged from the river trade to take on the clipper ships, they toppled the former leading companies in that space and wiped out clipper ships as competition. Some analogous thing happened with AI models. Large-Language models surprised me when they became publicly available, but didn’t shock me. In the late 1980s and 1990s, I’d developed and written about CHAINs, Combined Human and Artificial Intelligence Networks. I still think that is a better solution, avoiding this dangerous issue of autonomous systems making life-critical decisions.

?We dealt with models with hundreds or low thousands of parameters and hundreds of thousands of degrees of freedom. Those were models you could cross validate with fresh data. LLMs have from billions to trillions of parameters, and I have no idea about formal validation procedures. Another inflection point came when Tesla recently replaced over 300,000 lines of C++ code managing Full Self Driving with an end-to-end neural net.

?It's been a long a glorious trip from slide rules to contemporary AI efforts. While dangers lurk around autonomous AI systems, everyone involved in this computational revolution deserves appreciation for the progress we now see every day.

#UCLA #Internet #Innovation #UCLAAnderson

?

?