What are you using as your training data out of interest? Very impressive generation.

I'm always wondering that. Where do people get 660 million labelled images from?

I was under the impression a compact resurrected the session anyway so it is not a good idea on a cold session. Is that true or is it actually cheaper?

Yes, it's better to compact before it gets out of cache. I think new versions of Claude Code can do that automatically. It can be cheaper if you switch to Haiku or something

Newer versions of CC do in fact compact when the cache is close to expiry. I haven’t seen it work 100% of the time but it does it most of the time

You could use a dumber model (luna or haiku) for compact to get around this

I do this + caveman and my allowed usage lasts longer.

That’s forced refilling the cache under a different model.

You are correct, the first turn costs about the same either way, the difference comes after. Without compacting, every later turn re-reads the non-compacted context, and after compacting it reads a short summary.

So it only pays off if the session goes on. In my own dogfooding tests, I compacted 47 times out of 50 it had suggested.

Neal Stephenson - Anathem is great and not an easy reader. I enjoyed most of his other books but that one was the complexity sweet spot for me.

Umberto Eco and Ursula Le Guin I agree on.

His latest one, polostan, may be the worst book I’ve ever read on the other hand.. No idea what happened.

If it's only proofreading text you've written, the changes will be minimal enough that watermarking seems impossible to me.

> Turn the lights down low in the bedroom

Sets lights to 30% but also off

> Turn the lights low in the bedroom

Sets lights to on

This is a cool idea but I think humans assume more than 14MB of intelligence. This is like the unhelpful guard in the swamp castle of Monty Python's Holy Grail

Haha, its not a generalist model, cheap devices cannot run LLM in production, so we brig on-device intelligence to them in the size they can. Needle is niche, not a Gemma or GPT replacement.

Right, but I don't think consumers would consider this "intelligence". It's too dumb.

I'd suggest writing by hand as a proof of work, for now at least!

Sounds interesting but the blog seems hijacked by some advert to me, anyone else?

If that study was of submariners, is it possible long-term high-level exposure causes the body to adapt?

I am suspicious of 0.1% having a significant effect though, given oxygen is around 20% and we naturally exhale a couple of percent CO2.

I mean that can't be right, as the body's breathing response is triggered by that amount of CO2 buildup. It's not about what's in the air. It's about what the body can take up. Maybe submariners are self-selected to be more physically fit, e.g. larger heart, lung capacity etc. to compensate.

I think he's doing 15% every month, not every year. It's not an implausible growth rate for a unicorn startup; it is implausible to expect it.

I've been thinking about this, and I think software is uniquely knowledge work that has the most defined structure and least personally interaction. Hell, some of the software I write is for machine to talk to other machines. It's not surprising such a closed system is so amenable to AI, and other knowledge workers are not getting the same benefits.