norvid_studies Before Twitter Reddit was my most used social media. For a brief time I posted a lot about Mars, here's a sample of some things I wrote there

You have no excuse to not have a Substack now

Most interesting bit of the new K2 Thinking release to me is that the experts (except the shared expert) are quantized to INT4

It feels like the consciousness I'm experiencing is the only one in my brain. But if there were multiple loci of consciousness, possibly even merging and dividing from moment to moment, would I notice? I think I wouldn't, and that we shouldn't be sure we're alone in our brains.

Theia has done some really interesting work and I expect funding this has very high returns in terms of LLM understanding per dollar.

This simple subliminal learning experiment provides moderate evidence against GPT-5 and 4.1 having the same base model. Subliminal learning only works when the teacher and student have the same base model. 4.1 can transfer to 4o, but seemingly not to 5. alignment.anthropic.com/2025/sublimina…

Exciting to get parameter counts for closed frontier models like this. Not sure if "based on" means the final Grok 3/4 are the same size, or are pruned. Would love to know the active parameter counts too.

A friend of mine has early access to cutting edge corporate jargon, I heard the phrase "let's double-click on that" from him long before anywhere else. I asked him what's new these days and he says it's "the shark closest to your body" for the most urgent issue.

I'm surprised that AI lab revenue growth rates have remained steady as long as they have. I expect these lines to bend down a bit soon though, both OpenAI and Anthropic are estimating growth factors of more like 2-3x for 2026. epoch.ai/data/ai-compan…