Happy Superbowl Sunday! Welcome to the weekend edition of InsideAI. I'm Rob May, CEO of Talla. I also host the AI at Work podcast. Our latest episode is with Bloomberg Beta investor James Cham. Our conversation actually sparked the commentary I'm writing about today - about economic incentives and AI adoption.
The most popular articles from our daily newsletter this week are below:
Summit, the world's most powerful supercomputer, broke a record for the fastest-running machine learning experiment. The experiment used algorithms to detect extreme weather patterns such as hurricanes out of a huge dataset of climate simulations. The project was run by the National Energy Research Scientific Computing Center at Lawrence Berkeley National Lab and is one of 19 projects chosen for the Summit Early Science Program, which grants access to the Summit supercomputer. — WIRED
The chip industry continues to move toward specialized chips for AI, according to both Microprocessor Report and Communications of the ACM. There is a flood of new custom ASICs (application-specific integrated circuits), including Amazon's Graviton chip and Huawei's Kunpeng 920 in response to the market dominance of Intel's Xeon processor for inference and Nvidia's GPU cloud-based training. John L. Hennessy and David A. Patterson, winners of the A.M. Turing award for their chip design work, say that the rise of domain-specific languages and architectures "will usher in a new golden age for computer architects." — ZDNET
TAE Technologies is collaborating with Google and using the Optometrist algorithm to find the ideal conditions for fusion. Fusion power is a theoretical concept of using nuclear fusion to generate electricity. The process requires fuel and extreme pressure and temperatures to create plasma in which fusion can happen. The Optometrist algorithm is being used to manage all the variables in the fusion testing being conducted at the Lawrence Livermore National Laboratory. — THE VERGE
The Verge asked AI experts to name their favorite books, stories, or blogs and compiled a list. The list includes "Profiles of the Future," by Arthur C. Clarke, "The Book of Why," by Judea Pearl and Dana Mackenzie, "Franchise" by Isaac Asimov, "The Diamond Age," by Neal Stephenson, "Machine Learning for Humans," by Vishal Maini and Samer Sabri, "Sorting Things Out," by Geoffrey C. Bowker and Susan Leigh Star, "The Master Algorithm," by Pedro Domingos. — THE VERGE
Researchers from MIT and Microsoft developed a model to improve autonomous systems. The model uses human input to help uncover "blind spots" in a system's training and to provide feedback. Humans can provide data by means of demonstration or correction. The researchers will present a pair of papers on the new model at the upcoming Association for the Advancement of Artificial Intelligence conference in Honolulu, Hawaii. — SCIENCE DAILY
-- Commentary --
My interview with James Cham for AI at Work led us to discuss, both on and off the podcast, some ideas about AI adoption. James made an excellent point which is - AI in business applications is really about the model. What model can you build with the data you have, and how good is it? But if you think about machine learning models at work, what you really need is a detailed model of what a front line employee is doing. In knowledge focused companies, this is a junior person who is dealing directly with real data, executing a task, as an analyst, marketer, sales rep, programmer, etc. What you really want is a model of what a single person does in a day, and combine that with other models of what their peers do every day. But, what economic incentive do they have to do that? If they build a model of their job, their job goes away.
At Talla, we have tried to work around this by designing a user interface that captures much of the work a person does, for the purpose of automating it, and allowing them to be more effective and deal more with exceptions and interesting problems than mundane and routine tasks. But one reason we focus on Support and Service organizations rather than Sales or other groups (where automation could have a similar impact) is that Support workflows tend to be dictated top down, whereas Sales workflows tend to grow bottoms up based on what works. We sold to some Sales organizations and saw mixed adoption and usage.
But in general, your options are to have business unit leaders dictate this, or, users have no incentive to do it. And even today, most software is not like ours, and is not designed to capture and learn from regular workflows. And then there is the ethnical issue of job loss - it's not a problem now when we are making people more productive but, what about when we automate higher and higher levels of work?
The best idea to get out of this, that I have heard, is from Robin Hanson's Age of Em. In this book, Hanson argues that each of us will eventually have an emulation of ourselves, and that emulation can run once, or many simultaneously, and we will get paid for it. At that point, training and growing our emulations and what they know, their areas of expertise, their views on the world, will be valuable. We will all become trainers of our own person robot models, rather than workers executing any tasks.
Most buyers still look at AI software like a tool, and expect to buy it like software, rather than like labor. I always encourage people to take AI tools from the headcount budget instead of the tools budget. Changing this mindset will take a while but, AI adoption at the level it needs to happen won't occur until the mindset shift becomes the mainstream view. If you've read anything like Age of Em that addresses this economic piece of AI, it's a topic I'm deeply interested in and would love your recommendations.
Thanks for reading and see you next week.