LessWrong 2.0 Reader

You mention in another comment that your kid reads the encyclopaedia for fun, in which case I don't think The Martian would be too complex, no? 

I'm also reminded of how I started perusing the encyclopaedia for fun at age 7. At first I understood basically nothing (English isn't my native language), but I really liked certain pictures and diagrams and keep going back to them wanting to learn more, realising that I'd comprehend say 20% more each time, which taught me to chase exponential growth in comprehension. Might be worth teaching that habit. 

“My experience may not be applicable to you.”

Thanks for the note - my experience has been exactly the opposite. A classic case of the law of equal and opposite advice :)

Had something of a similar reaction but the note about far-UV not having the same problems as other UV serilization (i.e., also harmful to humans) I gather the point is about locality. UV in ducks will kill viri in the air system. But the spread of an airborn illness goes host-to-target before it passed through the air system.

As such seems that while the in-duct UV solution would help limit spread, it's not going to do much to clean the air in the room while people are in it exhailing, coughing or sneezing, talking.... 

I suspect it does little to protect the people directly next/in front of a contagious person but probably good for those practicing that old 6 foot rule (or whatever the arbitray distancing rule was).

Just my guess though.

Apologies for the soldier mindset react, I pattern-matched to some more hostile comment. Communication is hard.

Quick comment regarding research.

If far-UV is really so great, and not that simple, I would assume that any company that would be selling and installing might not be some small Mom and Pop type operation. If that holds, why are the companies that want to promote and sell the systems using them and then collecting the data?

Or is would that type of investment be seen as too costly even for those with a direct interest in producing the results to bolster sales and increase the size of the network/ecosystem?

I think perhaps a first one might be:

On what evidence do I conclude what I think is know is correct/factual/true and how strong is that evidence? To what extent have I verified that view and just how extensively should I verify the evidence?

After that might be a similar approach to the implications or outcomes of applying actions based on what one holds as truth/fact.

I tend to think of rationality as a process rather than endpoint. Which isn't to say that the destination is not important but clearly without the journey the destination is just a thought or dream. That first of a thousand steps thing.

If the kid is enjoying the robot stories then that's definitely the place to start. Foundation goes well after robots.

Took an ML approach, got radically different results which I'm choosing to trust.

(source epistemic status: mostly experiential and anecdotal from a lay lucid dreamer who knows a few other lucid dreamers)

The common negative effects from my lucid dreaming experiences:
- If I'm not careful with how I exert the "influence" I have in the dream, I can "crash" the dream, usually resulting in me waking up and having trouble getting back to sleep for a bit
- When I use a lot of influence in a lucid dream, especially to extend the length of a dream, I find that I end up seeming way less rested than normal (but that has proven hard to try and quantify beyond "when in the day do I hit a point of exhaustion")

A somewhat less common negative effect I keep in mind:
- Some people I know have had issues where their nightmares became far more unpleasant after trying to learn lucid dreaming to "fight back"

Model/FinetuneGlobal Mean (Cosine) Similarity

Gemma-2b/Gemma-2b-Python-codes0.6691

Mistral-7b/Mistral-7b-MetaMath0.9648

As an ex-Googler, my informed guess would be that Gemma 2B will have been distilled (or perhaps both pruned and distilled) from a larger teacher model, presumably some larger Gemini model — and that Gemma-2b and Gemma-2b-Python-codes may well have been distilled separately, as separate students of two similar-but-not-identical teacher models distilled using different teaching datasets. The fact that the global mean cosine you find here isn't ~0 shows that if so, the separate distillation processes were either warm-started from similar models (presumably a weak 2B model — a sensible way to save some distillation expense), or at least shared the same initial token embeddings/unembeddings.

Regardless of how they were created, these two Gemma models clearly differ pretty significantly, so I'm unsurprised by your subsequent discovery that the SAE basically doesn't transfer between them.

For Mistral 7B, I would be astonished if any distillation was involved, I would expect just a standard combination of fine-tuning followed by either RLHF or something along the lines of DPO. In very high dimensional space, a cosine of 0.96 means "almost identical", so clearly the instruct training here consists of fairly small, targeted changes, and I'm unsurprised that as a result the SAE transfers quite well.