Long story short... e-means produced better results than my dyf algorithm for the majority of the datasets I used as a comparison. But, dyf still managed to be much faster (which matters more for the embedding work that I do).
I did a bit of open heart surgery and transplanted the recursive partitioner from dyf into e-means, and found that it can work really well to auto-select the k-parameter (at least in everyone's favorite mnist example).
I worry that we as a species may end up cargo-culting ourselves into a pattern of fake work that produces real value without understanding the underlying mechanisms. But, it doesn't have any points in this sprint, so I'm going to spike a backlog and bring it up in the retro.
Amen to that, AI sometimes doesn't even choose the right answer even when a probe of its internal state shows that it has it. It won't explain itself correctly either. We need to burn our hand in order to learn, and for there's no instruction manual for something so emergent.
The word "understands" does a lot of heavy lifting here... there's a difference between true understanding and what people believe they understand. There's also the fact that many people don't understand the majority of the code at their company anyways.
Tmux is a good way to orchestrate sessions and keep them observable and interactive. I can easily adapt it to whatever mental model I need to get the job done.
Yeah, the whole thing seems like a human centipede of rug pulling. Probably the same as it's always been. Curating AI knowledge should be something that we put our best researchers towards, but realistically I think we wind up with 2-3 highly biased nationalistic models that are constantly copying off each other's notes.
I think neuro-symbolic AI has a lot of potential here, since small models can handle a lot of conversational inputs, while relying on wired-in solvers for more complex symbolic math/computation needs. https://jjd.io/posts/swollm-bbh-leaderboard.html
I like directed acyclic graphs and/or DAG because it's a succinct description and contract. Trying to change the name of it makes me quiver with uncertainty.
(Pardon the self promotion) Libraries like turnstyle are taking advantage of shared representation across models. Neurosymbolic programming : https://github.com/jdonaldson/turnstyle
It's a different kind of fatigue, but it's something a felt I got stronger at over time. Beats waiting IMHO, but be sure to give yourself a chance to rest.
People here always want to run the software themselves first, but then the next day they want to pay someone else to host it. If you're running into people throwing security flags, the silver lining is you're also a stone's throw away from offering a hosted option.
Honestly, yeah, makes sense to me. If we value freedom at all we value the right to tell the government to stuff it with their legally binding nothingburger deals.
I'm quite sure that's what's already going on. LEO is big and empty and pretty boring, but it's also resilient to increasing trends of natural disaster, war, and civil unrest. Having a foothold there matters more and more every day.
I thought that book was pretty dull, and had a terrible romance subplot. But I do have to say it pops in my head more and more often, just like Idiocracy. Someone should definitely do a movie so the illiterates know what they're missing at least.
While that sounds like a dealbreaker, I can't help think "yeah but if a decoding method took advantage of prefix in a similarly scalable way, one would reap the same benefits".
I'm an organ donor, and I have no problem giving my organs away after I'm quite positive I won't feel it. I think as long as there is no sensation, there is no pain, and certainly no formal concept of "self". But still, with matters of exploitation of our bodies, we should tread very very lightly.
The issue isn't that there's a lot of change coming from inside the markets, it's that there's a lot of change coming from outside the market, and it's all interconnected.
I always chalked that up to being around people that were new to the country and hadn't yet wrapped their head around the latest fashionable American cynicism. Generally takes a few decades to kick in, but when it does it's usually worth the wait (see Kumail Ali Nanjiani).
I also recommend Color Brewer, which is integrated already in a lot of visualization tools, and has great options for display variants and accessibility options : https://en.wikipedia.org/wiki/ColorBrewer
Humans make mistakes too. Case in point, the hallucination engine didn't tell the person to ingest bromide. It only mentioned that it had chemical similarities to salt. The human mistakenly adopted a bit of information that furthered his narrative. The humans touting and bigging it up are still the problem.
Nix always felt like an OS on the blockchain. It's fine if you value verifiability above everything, but becomes very burdensome if you find yourself tweaking your dev environment often. Still, I think it probably teaches good discipline.