你知道测试一个网站“这个布局看起来正确吗”有多混乱吗？现在将其乘以复杂的 3D 几何形状。游戏引擎所做的很多事情都是处理复杂的 3D 几何图形，其中主要验证是“它看起来正确吗”。这就是为什么游戏开发传统上需要广泛的人工质量保证和精简的单元测试。直到现在，我们才有了可以半可靠地自动测试“它看起来是否正确”的软件。
And today's records on ARC-AGI-2 are >80%. Held by LLMs that use text modality for input.
The issue with multimodal training is that it doesn't seem to bring a step-change improvement in spatial reasoning either. It helps some, but the gain is surprisingly small compared to the data and compute expended. What it helps with the most is, unsurprisingly, spatial reasoning when using image inputs.
Maybe there are gains we don't know how to extract there.
Overall, LLM performance at spatial tasks is improving, especially on things like puzzles, but that mix of "commonsense + spatial" in the same task still eludes them.
Do birds expose enough of their cognition through birdsong?
Do birds expose locomotion-relevant functions specifically through birdsong?
Do we have enough birdsong data available to start solving the inverse problem?
If "yes" on all, then we might be able to do it.
I expect "no" on most of that, for birds. But humans treat language as an interface to their higher cognitive functions, and stockpile language data. That looks an awful lot like a set of two "yes".
The last open question is: is there enough spatial reasoning reflected in the language data we have?
It's plausible that spatial reasoning is too evolutionary old and too low-level, too far removed from higher cognition, to leak into language heavily. And it's also plausible that existing LLM architecture is uniquely poorly suited to learning spatial reasoning - higher cognitive functions involved in things like writing code or even composing poetry are a better fit for the architecture. And it's plausible that we're underestimating just how complex spatial reasoning truly is - Moravec's paradox strikes again.
We know that LLMs perform poorly and improve slowly on spatial reasoning tasks, but not exactly why. And progress on things like ARC-AGI series shows that they're not completely inept.
What's your evidence of that? That AGI requires a truly novel architecture, and not just another iterative "LLM but with an extra trinket and wheels that spin ten times faster".
It's a "commonsense spatial reasoning/problem solving" kind of problem. LLMs fail at spatial reasoning forever.
What humans "easily" solve in seconds with raw spatial reasoning LLMs often find easier to solve by invoking A* or a constraint solver.
Might be that text data is particularly bad at teaching that to LLMs. Or that being good at spatial reasoning requires true recurrence, and autoregressive chain of thought is a poor substitute. Or it might be that human brain was optimized by evolution for solving spatial problems in open ended 3D environments for hundreds of millions of years, optimized for language for mere hundreds of thousands of years, and only optimized for writing computer code for a few decades at most.
The current frontier is halfway competent at benign closed 2D work, but still completely fumbles anything remotely close to open ended real world 3D work. It's getting better, but very slowly.
The classifier is about as refined as a brick to the face.
You can ask it elementary school grade biology trivia, or obscure facts about recently documented insect species, and both will downgrade to Opus 4.8 straight away.
And Opus itself was already bad with biotech questions. The fact that they somehow made it WORSE for Fable is mindboggling.
Every time things like this come up, I can't help but think of the ending of Inception.
It's less that you're convinced it's real and more that you no longer care if it is. "Feels real enough" is good enough.
I'm a technical user first, so I'm not sure if models have improved for RP the way they improved for applied STEM tasks and technical brainstorming. But if there is an improvement curve there, I wouldn't be surprised if this only grows in popularity.
That does work. Even if you drop the "specialized" part. Ensembles of the same architecture at the same scale trained on the same data do outperform a singular model of the same line - especially on corner cases. Successes of an ensemble correlate stronger than failures do.
The usual argument against is that if you have "a number of specialized models" that perform well in ensemble, you can take that ensemble, and distill it into a single larger model (dense or integrated sparse, like MoE), and get the same improvement in performance with an efficiency win.
This works because having those "specialized models" duplicates a lot of the highly conserved "low level" wiring that's required for a model to function at all. As such, you end up running a small scale version of the same "backbone" computational processes many times. "Merging" those models into a larger, denser model allows for a singular strong "backbone" to be used for everything.
Less weird unexpected failures, more innate ability to handle edge cases gracefully. Quite important when you're running high on automation and low on oversight.
它几乎就像康威定律的一面扭曲的镜子。