Microsoft joined the other major firms announcing a significant strategic realignment spurred on by artificial intelligence.
Key amongst them:
💥 Fantastic story from Cade Metz describing Microsoft’s custom AI hardware, field-programmable gate arrays which can be optimised for machine learning tasks. These are being used not just in Bing but across Azure. Remember, Google, Facebook and Apple are all making their own custom hardware (or even silicon) to accelerate machine learning performance.
- A reorganisation of the company around a 5,000 person AI division. The primary goal appears to be to reduce the latency from new research innovations making their way into products. AI is a special-case technology which doesn’t require a huge amount of end-user retooling for tech companies to deliver value to users. The growth of the Internet required consumers to buy PCs, get ISP subscriptions, acquire modems. And they arrival of mobile as a platform required us to buy smartphones. A consumer AI enhancement can work with your buying new hardware because the AI improvement can be delivered as a lightweight runtime via an app upgrade or managed on the back end in the server farm.
I’ve made a rough estimate of the sort of scale of human investment into AI and machine learning
by major internet companies. It is pretty mind-boggling. (Several of these firms also combined to form Partnership for AI
, an interesting group to address ethics issues.)
Indeed, the cost of more powerful models is rising. Eliot Turner tweets that in some recent papers the computational cost to train some Google models ran to $13k
. This is significant as you may have to run multiple experiments before you get to a model you like and on-going learning will be expensive. Once trained models are relatively cheap to run and execute in their ‘inferencing’ mode.
But it demonstrates the increasing barriers to entry for smaller firms, and perhaps a challenge for financial controllers handling budgeting. I’m not an accountant but one thought is that you should be able to amortise the upfront investment in building a machine learning model. But you could make a case for depreciating it over time (and treating ongoing training as 'maintenance’). I suspect in many smaller firms this isn’t the case and it is all treated as an operating expense. Any accountants want to weigh in, lmk.