I like to think of intelligence, knowledge, taste, and comprehension as 4 axes.
For example, Astra has lots of knowledge and intelligence but mid taste and bad comprehension.
Smart is a combination of all 4 traits.
With LLMs, we like to think of "smart" and "dumb" as one axis (because we think of humans this way).
I'd like to argue against this framing. Instead, try to think of "smart" and "dumb" as two different axes. A model can be incredibly smart AND dumb at the same time.
For a good example of this, look at a Gemini model. They are incredibly smart: you can bench them and see the capability. The amount of knowledge Google bakes into their models is incredible. Yet when you ask it to do work, the amount of stupid things the models will do is similarly incredible. They are "smart" and "dumb" at the same time.
Similarly, look at a model like Fable 5.1. It is not quite as smart as Astra, but it is significantly less dumb. Astra is still the smartest model available today, but it is also one of the dumbest, regularly doing things that make literally no sense whatsoever.
Finding the right balance of both "smart" and "dumb" is a challenge that everyone needs to figure out, from users to researchers. If we stop thinking of it as either-or and instead realize both can exist together, understanding model behavior gets much easier.
1
10
1,641
Introducing: ✨ 𝓩𝓮𝓻𝓸 ✨
- 1 bit native bit
- Built with 𝕣𝕒𝕞
- 0 lines of code, 0 dependencies
- Nothing, less than one, round
Introducing: Athas Browser
- 2.4 MB native macOS browser
- Built with SwiftUI and WebKit
- 1.9k lines of Swift, zero third-party dependencies
- Bookmarks bar, tab previews, vertical tabs
2
30
This might go down as OpenAI’s biggest fumble, 6 Sol is so unreliable that I’m convinced they just renamed 6 Terra to save on compute
12
Ok but do you realize how much money @anomalyco would make if we could use @OpenCodeJr in opencode go?
It just needs to be deepseek flash being the scenes
It should also be free but the free version is lazy and won’t do anything, perfect funnel
2
36
Based on my vibes (I have no evidence to support this lol) P=NP for everything except finding the polynomial time solution via ML
In other words, RL/training isn’t polynomial time but through RL you can find a polynomial time solution to anything you can check in polynomial time
Imagine if the solution to P vs. NP is just: you make the NP problem into an RL environment, then the model is the polynomial time solver
18
How long until the fly can code agentically?
Imagine if we had a thing that:
- doesn't need alignment. it's useful without being a threat
- was completely stateful, with no "context window"
- could continuously learn from experience
- had insane specialized spatial / sensorimotor reasoning
- responded to real stimuli through circuitry grounded in actual physics
- ran on consumer hardware and barely used any RAM
- required no datacenters and ran on low single-digit watts
- was completely open source, with no subscription or API bill
- wasn't another next-token predictor
- could be connected to an LLM when useful, but didn't have to be
- was inspectable down to individual neurons and connections
OH WAIT. We do.
It's the "stupid fly brain" you all keep making fun of.
PS: it's not sentient. It's a connectome-derived model of a real nervous system, and you can keep it that way.
male-cns.janelia.org/
1
1
21
Why does every single model release exceed projected usage by a very significant amount?
It happened with K3, Fable, Astra, SWE-2, and presumably more.
Does nobody think to consider this in their projections?
36
Apple didn’t have to make it ugly just to prove the leaks wrong
The leaked concepts and dummies looked 1000x better
21
Archo retweeted
Nobody can reply with an economically valuable task they can complete on a computer that GPT-6 Astra can’t. And I mean nobody
2
1
2
58
Nobody can reply with an economically valuable task they can complete on a computer that GPT-6 Astra can’t. And I mean nobody
2
1
2
58
