This might be pretty big. One of my biggest frustrations with smaller models (especially MoE) is their failure to track workflow state at a high level. I'm constantly reminding them what we decided on or asking them to revisit, and reminding them eats context.
Seems like this might make that a lot less painful. And if not off the bat, with some minimal tuning or even just good prompting.
Those quadratic forms loop in some nice structure for modeling all kinds of geometric problems with high level control that's hard to articulate so concisely otherwise. Conformal geometric algebra is awesome to work with, have you tried it?
But mostly the broad strokes points about the community are exactly the kind of hostility that makes geometric algebra communities so refreshing for curious young people. Geometric algebra is a welcoming pedagogy and community as much as it is a mathematical framework. If only mathematics as a whole was more welcoming.
I started out on with shaky linear algebra despite years of undergraduate education, but plenty of curiosity and intuition. The geometric algebra community schooled me and me prepared me for all kinds of "real math".
Yes the attitude that geometric algebra is the best language for everything is misguided and welcomes a lot of confusion, but most serious geometric algebra people I've met don't actually think that or say that. They're just off doing cool stuff.
I absolutely agree. You might be interested in Paulo Freire who wrote a whole book on something like what you're arguing, "the pedagogy of the oppressed".
That said, the crimes of "the cabal of wealth" are still on those responsible.
I've always considered it to be some marriage of convenience moreso than some kind of formal thing. And at any rate arguing over whether there's an explicit cabal or not very badly misses the point.
You're getting hung up on the idea of a "cabal" being some kind of formal thing that has explicit membership rules. But what we're talking about is simply the collusion of hyper-wealthy people in their various schemes. That's the cabal.
What's more dangerous, a version that's capable of actually fixing bugs well because it can identify the bugs or a version that creates more bugs because it's "not dangerously powerful" and instead just obliterates the code.
My experience is that it's not the models themselves that are limiting right now, it's the clunky alternative harnesses with weird missing features making for bad ergonomics around stuff like queue management, interruption, subagents, goals, etc.
I'm unsettled that we're all just glossing over how utterly racist this demand was to its core.
This is the same administration that's in the middle of trying ramp up revocation of natural citizenship, that's kidnapping huge numbers of melanated families in a viscerally cruel way, that's set on repealing birthright citizenship, that wants to use proof of citizenship as a voting suppression tactic all while making it more and more dangerous to register that proof in the first place. They want to throw AI safety into the same cursed pot of bullshit?
Props to Anthropic for refusing. OpenAI already has the mechanism in place since years ago. Let's see if they comply.
It's wild to me how we can have a whole conversation about this and just completely sidestep how incredibly racist this directive is, just how much it ties into the Trump administration's ongoing campaign of kidnapping melanated people. This is the same administration that denied one of the top FIFA refs a visa because they're Somali asking an AI company to gatekeep access to their on citizenship at a time when they're in the middle of attempting to challenge birthright citizenship and ramping up attempts to revoke naturalized citizenship to an unprecedented rate. Here they are attempting to throw "AI safety policy" into the same cursed pot and we're just head in the sand about that part
My system instructions tell claude not to automatically add attribution and fable ignored this. so I emphasized it again and fable decided that this was a forbidden cybersecurity topic.