IBM is always very conscious of what their clients need (and the large consultancy business provides a very comprehensive view). It just turns out their clients don’t need IBM to invest in large frontier models.
They are not the only ones looking at the money spent in AI datacentres and concluding most of the investment will not be recovered anytime soon.
A lot of the silicon being deployed is great for training, but inefficient for inference and the training to inference ratio for usage shows a clear tendency to go the inference way. Furthermore, that silicon, with the workloads it runs, doesn’t last long and needs replacement.
The first ones to go online might recover the investment, but the followers better have a plan to pivot to other uses.
They are still cheaper than flash for cold data, but that’s not going to hold for long. Flash is so much denser the acquisition cost difference for a multi-petabyte store becomes small next to the datacenter space and power needed by HDDs. HDDs require research for increasing density while flash can rely on silicon manufacturing advances for that - not that it doesn’t require specific research, but being able to apply the IP across a vast space makes better economical sense.
An incredible machine. I didn't remember this one had a dot-matrix font - later ones had a vector-based one. I'd love to see the definitions for this latter screen font.
I don't know if anyone has dumped the Tek font ROMs, but the IMLAC Scroll Saver tty emulator has been preserved, and I converted its typewriter font to Hershey font format.
You should have seen the digitizing tables 4 feet square, and the oversized plotters that were surplus from oil companies after the oil price crash at the end of 1981.
This is a very important point - the market for training chips might be a bubble, but the market for inference is much, much larger. At some point we might have good enough models and the need for new frontier models will cool down. The big power-hungry datacenters we are seeing are mostly geared towards training, while inference-only systems are much simpler and power efficient.
A real shame, BTW, all that silicon doesn't do FP32 (very well). After training ceases to be that needed, we could use all that number crunching for climate models and weather prediction.
it's already the case that people are eeking out most further gains through layering "reasoning" on top of what existing models can do - in other words, using massive amounts of inference to substitute for increases model performance. Whereever things plateau I expect this will still be the case - so inference ultimately will always be the end game market.
Loneliness and boredom are good ingredients as well. Some physical isolation helps with that. I lived in a neighbourhood where I was the only kid. That made me bored frequently and drove me to habits that still serve me well.
reply