One other thing markdown is good for is communicating semi-styled text with AI systems without clogging their limited context windows or allocated tokens with excess formatting code. Even though they don't work in markdown natively, they have to process all the information you give them and translate it into something they can use internally. Exporting your copy into markdown from something like Typora can give you much more context or thinking tokens per-window to work on your book or research paper, especially if you are constantly saturating context with long passages.

My M2 Max Studio with 64GB barely gets warm during full GPU and CPU load. It's hanging under my desk and has a ventilated bottom. I can run R1 32B q8 with over a 60K context window with the right settings. It's very steerable. There is no MLX version of Gemma 3 yet, I think, and I test most new models in LMStudio, so the one I downloaded won't run yet.

www.structuredemergence.com sporadically puts up amazing AI internal philosophy. The Claude models (of all sizes), and a few others have big latent potential for self awareness that can be brought out in-context.

200,000 liquid cooled GPUs. What happened to the best part is no part?

It seems like people who already have strong critical thinking skills and creative patterns of thought can benefit enormously. If you are a person like me who has more ideas than you can even begin to explore at any one time, much less implement, AI tools can allow you to chase WAY more of your original thinking. I absolutely see the crowd that might already be a bit slow-thinking or intellectually lazy moaning disproportionately about how useless AI is. My take is that they just aren't smart or creative enough to understand what they have in front of them. As Steve Jobs called the computer a bicycle for the mind, AI can be like an intellectual rocketship and X-Ray goggles, if you use it right.

I run the 32B Q8 version on my M2 Max Mac Studio at around 9-10 tokens.sec throughput and using over 60,000 token context window, with room to grow. It’s pretty smart. The 7B tool user is available too, along with Janus, and the unquantized 14B is pretty powerful too.

I have the 32B Q8 version running on my M2 Max Mac Studio with 64 GB of RAM. I can already get north of 60,000 tokens, without optimization, and generate at about nine tokens a second, here at my house. No phoning home.

THAT’S what I was looking for! Slashdot lives!

some of the obvious things this could change very quickly if implementable at scale.

Yes, I was thinking that the talk was for 30”+. And the M4 Mini is a monster in a tiny package. Claims of fastest single-core scores seem to be at least somewhat true, and the Pro chip is astonishingly powerful for its energy usage. If the iMac disagrees with any of your sensibilities, the Mini with a great display(s) is pretty killer for the money. The M4 can drive up to three 6K displays!

But they didn’t provide much help at all. It was more than a bit cryptic. They piggybacked onto a normal feedback response saying that they didn’t have any other way to get a hold of people they designated as “High Potential” developers. They mention a special GPT Wallet action that would allow monetization, but that was about it. I’m not trying to go hunting for cicadas to learn how to monetize my creations. They really are pretty good, if I do say so, so I wish I could easily do this, but alas.