You’re totally misunderstanding the context of that statement. The problem of classifying an image as a certain animal is related to the problem of generating a synthetic picture of a certain animal. But classifying an image of as a certain animal is totally unrelated to generating a natural-language description of “information about how to distinguish different species”. In any case, we know empirically that these LLM-generated descriptions are highly unreliable.
From the appendix:
TOTAL, COMPLETE, AND ABSOLUTE QUANTUM TOTAL ULTIMATE BEYOND INFINITY QUANTUM SUPREME LEGAL AND FINANCIAL NUCLEAR ACCOUNTABILITY
This week the WikiMedia Foundation tried to gather support for adding LLM summaries to the top of every Wikipedia article. The proposal was overwhelmingly rejected by the community, but the WMF hasn’t gotten the message, saying that the project has been “paused”. It sounds like they plan to push it through regardless.
The actual pathfinding algorithm (which is surely just A* search or similar) works just fine; the problem is the LLM which uses it.
I like how all of the currently running attempts have been equipped with automatic navigation assistance, i.e. a pathfinding algorithm from the 60s. And that’s the only part of the whole thing that actually works.
levels of glazing previously unheard of
The multiple authors thing is certainly a joke, it’s a reference to the (widely accepted among scholars) theory that the Torah was compiled from multiple sources with different authors.
I’m not sure what you mean by your last sentence. All of the actual improvements to omega were invented by humans; computers have still not made a contribution to this.
Yes - on the theoretical side, they do have an actual improvement, which is a non-asymptotic reduction in the number of multiplications required for the product of two 4x4 matrices over an arbitrary noncommutative ring. You are correct that the implied improvement to omega is moot since theoretical algorithms have long since reduced the exponent beyond that of Strassen’s algorithm.
From a practical side, almost all applications use some version of the naive O(n^3) algorithm, since the asymptotically better ones tend to be slower in practice. However, occasionally Strassen’s algorithm has been implemented and used - it is still reasonably simple after all. There is possibly some practical value to the 48-multiplications result then, in that it could replace uses of Strassen’s algorithm.
I think this theorem is worthless for practical purposes. They essentially define the “AI vs learning” problem in such general terms that I’m not clear on whether it’s well-defined. In any case it is not a serious CS paper. I also really don’t believe that NP-hardness is the right tool to measure the difficulty of machine learning problems.
As technology advanced, humans grew accustomed to relying on the machines.
honestly the only important difference between them is that emacs’s default keybindings can and will give you a repetitive stress injury (ask me how i know…)
Apparently MIT is teaching a vibe coding class:
How will this year’s class differ from last year’s? There will be some major changes this year:
- Units down from 18 to 15, to reflect reduced load
- Grading that emphasizes mastery over volume
- More emphasis on design creativity (and less on ethics)
- Not just permission but encouragement to use LLMs
- A framework for exploiting LLMs in code generation
i might try writing such a post!
When people compile compilers do they actually specialize a compiler to itself (as in definition 3 in the paper) as one of the steps? That’s super interesting if so, I had no idea. My only knowledge of bootstrapping compilers is simple sequences of compilers that work on increasing fragments of the language, culminating with the final optimizing compiler being able to compile itself (just once).
I’ve been using Anki, it works great but requires you to supply the discipline and willingness to learn yourself, which might not be possible for kids.
(From AI 2027, as quoted by titotal.)
This is an incredibly silly sentence and is certainly enough to determine the output of the entire model on its own. It necessarily implies that the predicted value becomes infinite in a finite amount of time, disregarding almost all other features of how it is calculated.
To elaborate, suppose we take as our “base model” any function f which has the property that lim_{t → ∞} f(t) = ∞. Now I define the concept of “super-f” function by saying that each subsequent block of “virtual time” as seen by f, takes 10% less “real time” than the last. This will give us a function like g(t) = f(-log(1 - t)), obtained by inverting the exponential rate of convergence of a geometric series. Then g has a vertical asymptote to infinity regardless of what the function f is, simply because we have compressed an infinite amount of “virtual time” into a finite amount of “real time”.