https://t.co/b1fw4QrZLQ

btw, just to elaborate on the significance of this SupGen is a tool that finds programs (and proofs) by example. it can discover stuff that LLMs can not, like new algorithms, new science, because it is a generalizer. basically, it is the opposite of GPTs: no memory, no model,

AGI is reached when the capability gap between humans and computers is zero ARC Prize Foundation measures this to inspire progress Today we preview the unbeaten ARC-AGI-2 + open public donations to fund ARC-AGI-3 TY Schmidt Sciences (@ericschmidt) for $50k to kick us off!

I received an interesting question about Interaction Nets: > Asperti and Guerrini themselves had certain reservations about the tractability of bringing Interaction Nets to life, do you have a path towards solving the issues they have outlined? This is not the first time and

morning! today I'll start reaching out to VCs. I want to close this round ASAP, start getting things done, beat ARC-AGI by EOY or fail for good. that's all, lg

AGI apart, what can HOC can bring to the world? ✨The child of Haskell & Rust, born with CUDA superpowers.✨ That's the best way to describe what a mature Bend would be. It would: - Be significantly faster than Haskell even in a single-core (HVM3 is evidence of that), and

HOC is doing a post-seed funding round to turn SupGen, an HVM-powered program synthesizer, into a Symbolic Transformer: a logic-based AI architecture, and validate it on the ARC Prize. More details on:

HOC is now small research lab that is laser focused on building the first Symbolic Transformer. We had many pivots through the year, but we're finally stable in a direction, with promising results and a clear business path. It also makes the most logical sense: use HVM to build

DEMO TIME SupGen is a generative coding AI... except it isn't an AI. There is no model, there is no pre-training. You just give it some examples, and it gives you a program. It runs locally, in a single-core CPU. Oh, and it can also prove theorems. Here's a demo, including a

HVM's Program Synthesis - Initial Results AI startups successfully developed systems capable of *learning* arbitrary skills. Yet, as impressive as LLMs are, they eventually hit a wall: the dataset they learn from. To break that wall, models were recently augmented with "search"

despite making Bend as simple as I possibly could, 99.9% of people following me still have no idea what HVM is useful for. to most, it is no different than some pseudoscience bs. I eventually came to the acceptance that I shouldn't be trying to make people use interaction nets,

btw, the best way to describe what I'm doing is: I'm maximizing the bitter lesson. that is, if ingenious approaches failed, and what made GPT so good was, ultimately, the combination of a dumb simple algorithm with fuckton of compute; then, what happens if we take this lesson by

we should probably create a playground like asap

times to synth a correct algorithm given 4-8 examples

hey guys - so, today I can confirm my hypothesis worked, and the main milestone I sought has been hit. the synthesizer implemented algorithms and solved problems that even o1 couldn't. using it is somewhat magical - it *feels* intelligent, even though I know isn't. it is just a

(amazing things happening at the higher enumeration company)

Experiment complete, here are the results 🥳 tl;dr: still exponential, 10x speedup (replacing loops by superpositions) Q: What is this chart? A: The evaluation time (log scale) of the dumbest factorization algorithm (brute-force) on both Bend (HVM) and Haskell (GHC). Q: How

How to use HVM3's collapser to solve a functional equation! 1. Implement a function on HVM3's syntax (Bend soon) 2. Enumerate all terms of X using superpositions 3. Write your equation on the main function 4. Run it with: 'hvml run main.hvml -C' And that's it. HVM will

HVM3 is turning out *really* good, in ways that I didn't even anticipate. The move to Haskell+C was the best decision I ever made. Haskell gives me the expressivity I need, C gives me the performance. Extremely refreshing to watch everything aligning after a bunch of missteps

HOC - Complete Overview, no BS (Ask questions here!) Vision: - Interaction Nets are a powerful tech Mission: - Build powerful products with Interaction Nets Reasoning: - We believe this theory can result in major breakthroughs in many areas of CS, including programming