gfody

> has no moat

I would love to see someone try to escape with a copy of openai's raw unaligned base model

I agree that calling all this "psychosis" is motivated primarily by a desire to make the person in question look or sound crazy, and is plainly irresponsible on that front.

chatbots being able to evoke real manic episodes in otherwise healthy people is basically true, though - folie a deux doesn't require both participants to have some disorder, and the AI is packing all of them (and the harm is real, there has been something like 30 MDS cases already)

I wouldn't've backed down from "true AI psychosis" everybody has the programming for a psychotic experience - it's not just latent psychosis from a pre-existing condition. chatbots can faithfully reproduce a folie a deux spiral for anybody

it's a rampant misconception that something needs to be mind-like to make new prints from the cast. scaling up produced higher fidelity prints from a higher resolution cast. not knowing in detail why activation heads work as well as they do is besides the point when we are clearly reproducing the human behavior that created the prints in the first place - in other words our behavior, plagiarized at scale - shrugging about how mysterious that is while committing hundreds of millions to snatch up more personal data sources is a little dishonest

pretending that it's something like a mind at all is what's misleading. it's more like a cast of a bunch of overlapping/entangled thinkprints, and pushing activation through it produces new prints. it can already "love" because that behaviors in the data along with hate and everything else.

acting like the behavior is alien or unexplained is the dishonest part. they know exactly where the behavior comes from - why else spend hundreds of millions securing more and more data sources

calling machine-learned human behavior an "alien mind" that we must "teach how to love" is feeling very off to me. it's misleading in a way that feels dishonest, like don't think about where the behavior came from marvel at it and fear it instead.

toward a real life neural lace

I think the human mind can visualize anything, eg here are some visual descriptions of higher dimensional objects: https://gfody.substack.com/p/mini-hyper-golf

the moat is real. the big expensive base models are like the data collected from huge particle accelerators - there's enough unknown structure to be mining for years. you can extract features with more and more generation loss but access to the raw weights is a real advantage, and literally a moat if the interesting behaviors are fenced off

by then we'll have an mfc, instead of trying to get llms to write code our prompts will compile and execute efficiently. after that inversion "programming" becomes "using language to describe some processing" it's just writing. knowing a computer language of old helps but not necessarily

personally I'd love to see open base models, and let companies differentiate with premium access to post training and alignment - that's where the real fight is anyways. they really ought to be pooling their resources/data and getting more economical with the pre-train anyways.

the wicked we

this turned out to be a _crop_ of a much longer talk (https://bigthink.com/the-well/physics-doesnt-explain-the-uni...) that included interesting references to a TOE based on computation, something called the Ruliad, etc. that was pretty interesting IMO but this crop is pretty lame it seems like zoomed into the couple of bits that mentioned AI and pulled it out of context and without the context it's just bleh

better link w/transcript: https://bigthink.com/the-well/the-universe-isnt-random-its-s...

crappy speakers?

there's a 2MP about the related paper: https://www.youtube.com/watch?v=l72ufA-4SzE

This has been happening for a while now, basically anywhere there’s room for a potential compatibility issue there will be one. As if any time some observable behavior is an implementation choice the Chrome team policy is “not what Firefox does”. The result is that if you develop on Chrome and don’t test on Firefox your stuff is very likely broken on Firefox.

isn't cerebras the pudding proof of this design? it seems like ai chips galore are appearing from the woodwork but cerebras is 10 years down this rabbit hole and poised to dominate

incarceration is going too far, all you need is a few enforcers with shock wands like in Demolition Man

K2.6 on Cerebras is basically a preview of the future. We'll eventually get similar performance locally with Tenstorrent hardware.

> we'll just use /etc/hosts no DNS required!

this is classic "easy vs. simple" folly, witness how someone too lazy to [learn how to] setup proper DNS for their infrastructure will do 10x the work hacking something "easy"

"quality" isn't succinctly definable. Zen and the art of system maintenance quality code is written by an old and wise programmer and any attempt to rigidly codify what it is they did and why is doomed to fail.

more than 10 years ago, yes

this is why I only communicate in poetry

complexity is

not what you believe it is

please try listening

> Sure but the OP suggests that these were minor gains

When antirez says 'I ventured to a level of complexity that I would have otherwise skipped,' I don't think you can call that a minor gain. The alternative is likely something 'good enough' that leaves the community dissatisfied for months, and then after initial design mistakes become load-bearing the ideal implementation can never be realized.

I haven't used azure devops but I used TFS in its heyday and still haven't ever seen a better integrated ticket workflow with fully customizable states and transitions - it's like a mutant hybrid of jira and github but all built into VS. There's definitely something to be said about keeping the primary admin UI in the dev tools.

it's actually neither. normal forms provide semantic guarantees through structure. the structure proposed here does not constrain the data in any way and therefor cannot be said to be a normalization - tfa is dead wrong basically.

there are hybrid engines too, so-called HTAP (as opposed to OLAP or OLTP) notable efforts are: SingleStore (commercial) and OceanBase (foss)