oreghall

There should be a version of ‘earn to give’ that’s something like ‘get pretty to persuade’ I can’t think of a way to say this that doesn’t feel mean, but I think if you have a cause you really care about, and are lacking in basic hygiene practices, you’re leaving money on the table We’re wired to take fitter & more hygienic/presentable people more seriously than someone who eg clearly needs a shower and a hair cut. You may not want the world to work that way, but it just does I’m a little worried I’m going to get torn apart for posting this, but I genuinely want to see the people working on important causes have the best fighting chance to achieve their goals

- labs already have and will continue to build application layers on top of their own models, eg claude code and codex - they have strongly signaled that they're going to just keep doing this in many other verticals, likely folding in capabilities into their "everything apps" (claude cowork etc, openai's upcoming consolidated app), allowing them to do shit like... law, and bio research - enterprise users are very very happy to pay large sums to use these specific apps because they do huge amounts of very valuable work, they don't need direct API access for these, and going forward as the labs get more into harness engineering etc the direct API access will just be less useful. people pay to solve problems! - increasingly exposing the newest models in public APIs just allows companies to 1. trivially build and maintain competitors to frontier labs and 2. distill rollouts at scale - as models become significantly more capable the labs need more and more control over what people are doing with them - so if their revenue mostly comes from first party non api offerings, and exposing the APIs just leads to competition + distillation + safety concerns... they'll just stop adding their best models to the API - openai has already effectively experimented with this with periods of the -codex suffixed models only being usable through codex!

"who cares if Cursor used Kimi 2.5 as a base, starting with a commoditized pretrained model was always the right move anyway" nah, sorry, what it proves is Cursor is still fundamentally reliant on frontier labs. Kimi 2.5 is only as capable as it is because it's a distill of Opus 4.5. the only open model that ever showed it was capable of trading blows w the frontier was deep seek, and it really seems that moment has passed. the question was whether Cursor could really break the dependency chain and start building improvements based entirely on their own expertise and data. and Composer 2 shows that they *can't*, that they need the general model quality and intelligence from 4.5 to get anywhere, and that really what they're doing is laundering culpability through Chinese labs so they don't have to get their hands dirty doing distillation themselves. when Opus 5 and GPT 6 are significantly more capable along many dimensions, more RL with coding rollouts aren't going to be enough to save Composer 3, they'll either need to have caught up with whatever the frontier labs are doing internally, which right now we have pretty strong evidence they just don't have the research capacity for or... wait for another distill. and how much longer do you *really* think OpenAI and Anthropic will continue to serve their frontier models through publicly accessible APIs? that was always a revenue and data bootstrap. it's ending within the next two years.

Gad Saad asks Joe Rogan: “How many Jews in the world?” Joe: “A billion… uh, half a billion conservatively.” Reality: ~15.8 million. That’s 0.2% of humanity. Still FEWER than pre-Holocaust (16.6M in 1939).

You are close but you are missing critical theory (won’t spoil, enjoy)

This has been solved! (or so it seems?) I wanted to give a context on how amazing that is, and how many this should unlock - things that puzzled me for 10 years - but I'll just explain the solution itself, because I'm so excited about it! Turns out it is really simple, elegant, and even obvious - which is embarrassing, because I should probably have realized it years ago... but that also makes it more likely to be correct, so, I have no complaints! To arrive at it, I isolated the simplest example that, intuitively, "should fuse" but doesn't, and evaluated it manually, using an new, hypothetical "mov node", which allows a variable to be used more than once across different branches. I then asked: which interactions are missing, to get to the expected output? Turns out, we just needed two interactions: - MOV-SUP: exactly like DUP-SUP (just commutes). - MOV-MOV: just compacts two MOV's into one. And... that's it. See the image below. But... why *these* interactions? Because *MOV nodes are just unordered DUPs*! (Yes, this is something that I have tried before, and even posted here, but it didn't work back then. Turns out my *implementation* was wrong, but the idea is solid.) This key insight justifies everything: - MOV-SUP is just DUP-SUP, which is fair enough. - Since MOV is unordered, x₀ and x₁ is just x and x. - This allows MOV vars to be used more than once. - It also explains MOV-MOV: it just compacts MOVs! This new interaction, MOV-MOV, is the missing piece: it prevents fan node accumulation, allowing us to dispatch linear variables to different branches, while still letting the function "fuse", in the sense the normal form of F^N(x) has constant size, which, in turn, drops the complexity of their iterated application from O(N) to O(log2(N)) in an optimal λ-calculus evaluator such as the HVM. So, I quickly added MOV nodes to HVM4, and the result speaks for itself: applying clr() 2^65535 times is down from 7,343,093 to just 4,064 interactions; a measly 1806x speedup! Obviously, that is just a silly way to convey that we observe there is, indeed, an exponential speedup. This should unlock so many things. For many years I've attempted so many promising algorithms on HVM, only to get blocked by this very issue. Not being able to use a variable on different branches is VERY limiting. Now that blocker is gone, I have so many old ideas to revisit. Even SupGen was profoundly affected by this... Below is the full fusion of `F.F` where `F = const False`. Also thanks Lorenzo, LeeeeT and others who posted insights and partial solutions that helped me get to this.

more of this

In technical terms, this is joint self-supervised compression. To justify this, I rely on the Minimum Description Length principle (MDL) I take MDL to its logical conclusion by compressing every possible source of information

if you gave any random poor person 1 billion dollars they would probably spend it in much more reprehensible ways than current billionaires do