The racing dynamic is very scary from an Alignment perspective, did not expect that so soon.
I think it’ll be while before we see a LLM integrated into Search the inference cost is still too high. Maybe when they have TPU v5/6 or some other big cost reduction. Good commentary here:
I wonder if Gmail (rather than search) might be the first place we see a LLM used - it seems a better fit.
That mail was v good, Dylan actually just put out a follow up to that today:
I think the lightweight model they are alluding to is probably something like LaMDA following Chinchilla scaling which probably a ~50% reduction in parameters vs GPT3 so probably quite a significant inference cost.
I had about 50% of my modest holdings in Google. I have sold about a third of it today for other ventures to reduce my risk and lock in some profits. Made about 20% gain in 2 months.
Consensus seems be building that Google has woken up and is actively deploying that huge cash stack on compute.
Dylan has been making comments about the number of TPUv5s being made for a while and emphasised it with this piece today (which unfortunately is mostly paywalled)
It does feel a bit weird that 2 of the top 3 models (arguably #1, #2) were both trained on TPUs yet Google is worth less than Nvidia, especially when presumably OAI will ditch the GPUs and join the ASIC club in 2025 if their Microsoft partnership holds.