Inside | Real news, curated by real humans
Inside AI

Inside AI (Aug 25th, 2019)

Happy Sunday and welcome to the weekend edition of Inside AI.  I'm Rob May, CEO at Talla.  If you enjoy the newsletter, I hope you will check out my company as well.  Also be sure to check out our AI at Work podcast.

Let's get started with the most popular articles from the past week:

Engineer.ai CEO Sachin Dev Duggal has responded to complaints that the company exaggerated its AI abilities to attract customers and investors. In a new statement, Duggal denied the allegations and claimed that the mobile-app startup has historically preferred the term "human-assisted AI" rather than “automated software development." The rebuttal comes after The Wall Street Journal reported that Engineer.ai does not use AI to assemble code but relies on human engineers in India and elsewhere. Its chief business officer, who sued the startup earlier this year, claims Duggal “was telling investors that Engineer.ai was 80 percent done with developing a product that, in truth, he had barely even begun to develop.” Duggal claims his team never got a chance to explain the issues, even offering to meet with the WSJ six times, before the report came out. "Many of our answers, which dispelled several of the allegations, were excluded from the reporting," he wrote. - ENGINEER.AI

Scale AI's founder, 22-year-old Alexandr Wang, spoke with Business Insider about his ambitions for the company, including plans to turn the data-labeling startup into the infrastructure for ML technology. Wang's startup, which has caught the attention of some of Silicon Valley's top investors, recently raised $100 million in series C funding round, which valued the three-year-old company at $1 billion. Wang himself is a phenom, excelling at coding competitions and receiving job offers while still in high school before dropping out of MIT to build the startup. The firm's software tools do the initial task of labeling pictures before they're sent off to contractors. - BUSINESS INSIDER

During the Hot Chips conference on Tuesday, Tesla chip designers showed how the company's AI chips are dramatically better in performance compared to the earlier Nvidia chips. The fine-tuned AI chips - which have 6 billion transistors apiece - are "smart" enough to power Tesla's full self-driving abilities in the future, according to the company. Their performance has improved by a factor of 21, compared to the earlier Nvidia chips. Ganesh Venkataramanan, one of the chip designers and a former AMD processor engineer, said that in order to meet "performance levels at the power constraints and the form factor constraints we had, we had to design something of our own." The chips, optimized for self-driving cars, run at 2GHz and perform 36 trillion operations a second. - CNET

A suspect was arrested in China after trying to scan his a dead girlfriend's face using facial recognition software. The man, from southeast China, reportedly strangled his partner following an argument and then attempted to use the woman's identity to apply for a loan. The online lending company, Money Station, requires users to verify their identities via a facial recognition AI. The verification failed because the AI did not detect the woman's eye movement and detected a man's voice rather than a woman's. Workers who manually checked the failed verification reported the incident. - GIZMODO

  • Email gray
  • Permalink gray

As the readership of this newsletter has grown over the years, I realize that many of you are new to AI, and may have a fundamental question that I've never addressed here - why do we believe it is possible to build intelligent machines?  There are whole books and PhD theses written about this topic, but, today I'm going to give you 4 quick high level viewpoints on this topic.

1.  We Cannot - There are still a fair number of people who believe we can't build intelligent machines because intelligence is something special and different.  Occasionally this takes a technical form of argument but more often it's a religious or spiritual form of argument.  The idea is that humans have special stuff, most western religions would call it a soul, and that conciousness or intelligence or anything human comes from that, and since it is special and non-material, we can never replicate it.

2.  The Neurons Are Just Electrical Signals Approach - In this view of the world, we realize that at their base, neurons send and receive electrical signals.  It makes sense that some day we can create an electrical circuit that mimics the electrical signaling of neurons.  (We get closer all the time.)  In that scenario, assume you replace one neuron in your brain with this electrical neuron.  Does anything change?  Not really, if it functions the same as your neuron did before.  Now replace 10, then 1000, then 1,000,000.  Now replace them all.  What do you have?  A fully silicon based brain.  

3.  The Airplane Approach - Many researchers believe that intelligence is an algorithm, or class of algorithms, and there are multiple ways to do it, and we don't need to replicate biological neurons to build intelligent machines.  They would argue that humans don't fly the way birds fly - we invented airplanes.  It accomplished the same general goal, but, in a way that was better for us.  In this view, intelligence will come from increasingly complex algorithms, and increasing computational power that allows us to run them.

4.  The Emergence Approach - This approach could be built on top of approach 2 or 3, but, it basically says that intelligence is an emergent property.  As the complexity of a system builds, either bottoms-up like point 2 above, or from a collection of algorithms like point 3, at some point, intelligence is emergent.  If you aren't familiar with emergence as a scientific concept, I know it sounds a little magical and wishy washy but, there are plenty of examples of emergence as a prominent feature of natural systems so it is a very reasonable assumption, in my opinion.

There is another dimension to this that is important as well - physical inputs.  Some people will argue that to be truly intelligent requires a machine to be grounded in the "real" world like humans are.  These people believe that, whichever approach wins, it will win because it is embodied in a robot, not a just on a computer chip, because the robot integrates physical sensors into the AI brain and those physical inputs are necessary to be intelligent.

Now you see, at a high level, why so many people are starting to believe it is possible to build intelligent machines.  Most people are probably a 3/4 combo approach, at least based on the AI circles I run in, but, there are plenty of 2s around as well (and approaches like Koniku that mix biology and silicon).  So far, no one has been able to prove or disprove conclusively whether machines as intelligent as humans are possible, but the ideas outlined here explain why so many believe we can get there.

If you are new to AI, I hope you found this explanation helpful.  Thanks for reading and enjoy your weekend.

@robmay

  • Email gray
  • Permalink gray

Subscribe to Inside AI