He was a clever king, but no one good could be king of England for more than about 18mths in those days.

I think that the parent probably meant Henry VIII.

Yep sorry fixed typing too fast missed an I

How to know if one should fine tune/pretrain or RL / reasoning train given some data set?

i honestly dont think there's a simple y/n answer there - i think considerations include mostly like 'how costly it is to do so', 'how often do you think you'll need it', and so on. traces are not as "ephemeral" as FT models - since you can use those to guide agent behaviour when a newer model is released (but still, not as evergreen as other assets - traces generated using say GPT4 would seem pale and outdated compared to ones created on the same dataset using Opus4.5 i reckon)

Interesting - ages ago I used SML to look at the relationships between Shakespears plays.

https://medium.com/gft-engineering/using-text-embeddings-and...

Validation is a problem here - you find relationships, but so what? Is it right.... I can't say. It is interesting though.

Colin Ritman is so Demis Hassabis.

There's a loop between adoption of a technology and adaptation of a technology. For some domains that loop is fast, the adaptations prove to be easy and the feedback from adopters is easy to get. For other domains it's slow, especially towards the end of the process of going from something interesting to something useful. A good example of a slow loop is self driving, it's hard to get feedback about self driving in safety critical real world situations... another example is medicine.

The other issue is that the value is more or less all in the LLMs (at the minute). For example, I built a data engineering toolkit using LLMs, it created synthetic data from examples, it created ingestion pipelines given different source filed and a target, it created data test rules. I liked my little toolkit and some people were impressed, but the value was all in the models that underpinned it. The crust of clever bits that added value was thin, very thin. Ok, we used the llms to generate some python that then created the synthetic data and testing rules to reduce costs, we had three or four "agents" that worked together to create the pipelines. We decorated target code with open provenance code to create provanance... But just by saying these things or letting you use the toolkit and you seeing what it made - that's enough for any half competent person to relicate it (with AI assistance) in an afternoon, or maybe a couple of afternoons. Maybe.

So, to create a viable company is going to take significant effort (if you can think of a value add) because the value add still has to be real.

can you give a bit more information on 100's of qubits below threshold? I wasn't aware of 100's...

https://www.nature.com/articles/s41586-025-09848-5 performs CZ gates on up to 256 qubits with fidelities of 99.5%, which is good enough to run surface codes below threshold.

What would SOA be?

Claude Code is perfectly able to access your git/jj history directly. Just ask it to review a commit.

Good for you George E Collins.

I think that's a rather conspiratorial way of framing it.

I think it's more about someone trying to do the most good that was possible at that time.

I doubt he cares much about prizes or money at this point.

It's hardly a conspiracy to use strategy and intelligence to maximize the probability of achieving the outcome you desire.

He doesn't have to care much about prizes or money at this point: he won his prize and he gets all the hardware and talent he needs.