Basic search with a lot of compute power has worked better than expected in machine intelligence.

Judged as of February 2015, against the state of machine intelligence research then. True: by early 2015, published results and expert commentary showed that simple or general methods scaled with large compute (for example deep neural nets trained on GPUs, as in ImageNet results, or search-based game playing) performed better than many experts had expected. Mostly true: this was broadly the view, though the 'basic search' framing fits only some of the leading results. Mixed: the evidence is split, with some fields showing compute-driven gains beyond expectations and others not. Mostly false: the gains came mainly from new algorithmic ideas rather than simple methods plus compute, and this is overstated. False: the evidence shows compute-heavy simple methods had underperformed expectations by 2015.

Set before the research began.

1. E1 Microsoft researchers say their newest deep learning system beats humans -- and Google VentureBeat · 10 February 2015

   “The Microsoft creation got a 4.94 percent error rate for the correct classification of images in the 2012 version of the widely recognized ImageNet data set , compared with a 5.1 percent error rate among humans, according to the paper.”

2. E2 CMU School of Computer Science Carnegie Mellon University (hosting the NIPS 2012 paper by Krizhevsky, Sutskever and Hinton)

   “We also entered a variant of this model in the ILSVRC-2012 competition and achieved a winning top-5 test error rate of 15.3%, compared to 26.2% achieved by the second-best entry.”

3. E3 Microsoft, Google Beat Humans at Image Recognition EE Times · 19 February 2015

   “Microsoft did use Nvidia GPUs, but bought and configured their own supercomputer using them to simulate parametric rectified linear neural units to become the “1st to beat a human” at image classification.”

4. E4 Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification - Microsoft Research Microsoft Research · 1 February 2015

   “Second, we derive a robust initialization method that particularly considers the rectifier nonlinearities.”

By early 2015, published results showed that large deep networks trained on GPUs were doing very well. AlexNet won ILSVRC-2012 with a 15.3% top-5 error, against 26.2% for the second-best entry (E2). In February 2015, Microsoft reported a 4.94% ImageNet error, below the 5.1% human rate, using GPU hardware it bought and configured itself (E1, E3). However, the Microsoft paper also credits algorithmic work, including a new initialization method (E4), and the system used parametric rectified linear units (E3), so the "basic search plus compute" framing fits only part of the leading results.