In this talk I give the results of some experiments we have done with Internet of Things at the extreme edge, on the sensor boards themselves. There are around 30 billion installed IoT devices in 2020 and this number is growing exponentially for the foreseeable future. Because IoT market size is growing much faster than cloud market size (14% per year versus 5% per year), future edge computing systems will do more computation directly at the edge. We have done two experiments at the extreme edge: a resilient sensor fusion system using sonars to track person movement in real time and a MapReduce framework for resilient big-data-style computations. Both experiments are built on the GRiSP edge computing platform. Each GRiSP board directly supports computation, communication, stable storage, Pmod sensors and actuators, and natively hosts a complete Erlang OTP on the board over an RTEMS real-time hardware interface layer. We demonstrate that substantial parts of edge applications can run reliably at the extreme edge. This is highly cost-effective: no other computers nor cloud connections are needed. We have released software for both experiments under an open-source license and we intend to continue developing this approach as part of our research on practical IoT.