Thanks! I hadn't tried it. I've started using it today; I limited a couple options about dired-local and images in kitty due to concerns from upstream security audits, but otherwise looks fantastic.
Thanks! Agreed ghostel looks fantastic. I still wish there was an in between a full TUI and a full non-interactive version, but switching to ghostel means I can fully live within emacs again.
Similar story here (user since early 90s). I wish more AI harnesses could also have a dumb-terminal mode in addition to the fully non-interactove mode, so they could all stay naturally searchable/copyable in simple comint buffers. Even the best current terminal emulators within Emacs are not great to recursively handle Emacs (or modern AI TUI) well enough, so I occasionally also use a shell putside Emacs.
You now fully support version 1.0. IIRC, the first version of Mathematica I used back in early 1993 was 2.1 soon followed by 2.2, which was useful for a couple of years. How much extra work is it to reach full compatibility with that version? (I had used MathLink at the time.)
Alzheimers has a very long development stage compared to almost all other dementia (Lewis body also has a long timeline, but none of the other forms has anywhere close to the known length for Alzheimer’s disease). The original study article has no causal analysis, so a simple explanation is that super early Alzheimers symptoms (possibly even undetectable ones outside the context of intense taxi driving) drives people to abandon their taxi- or ambulance-driving jobs and seek other forms of employment.
Google’s trajectory unfortunately starts to feel similar to when yahoo lost everything in the early dotcom bubble except it owned a huge chunk of Alibaba. Google’s internal decision to stop publishing useful research probably started the slow but sure process of internal decline. I still hope they can recover, given their amazing position in terms of data and infrastructure, but it no longer feels like an obvious or easy task, when only 6 or 7 years ago they felt untouchable.
My point is exactly that we cannot even begin to start the process that will lead to “know everything” there is to know unless we make it a science first. Ad hoc statistical models are different than understanding or complete knowledge of a subject.
The early physics background is messy and incorrect. I didnt read the full position paper, but from its start: The Lorentz transformations were by Lorentz, well before Einstein’s paper on special relativity; the principle of relativity also existed before the Einstein paper. The math was all there, with steps taken by Maxwell, Voigt, Larmor, Lorentz, and Poincare. Einstein supplied a clean physical interpretation, making all inertial frames equivalent, making simultaneity frame dependent, and explaining length and time deformations without the need of the concept of ether. Skimming the end of the paper with the arguments about lack of abduction or inability to make the analogy without sensory experience, I see that this paper is unfounded speculation rather than solid/hard philosophical logic. As a position paper it is OK to appear, but i think it misses the point of how LLMs or other autoregressive learners of future states can build analogies and intuition that can help them formulate new theories of the world. Soon it will be more obvious to everyone, so I am not very worried about these writings.
Agreed in principle for today’s new ones, though with all the buying frenzy it quickly becomes hard to procure anything but the latest hardware and the economics of older generation hardware makes little sense for new data center deployments.
I agree these are central components, but to avoid oversimplification and the mistaken belief that modern LLMs do a lot of search during inference: If it was so simple, the traditional computer algebra systems would have reached similar breakthroughs when deployed at large supercomputer centers. This didnt happen because the search space is huge. You definitely also need a fancy learning algorithm. Although these ingredients would suffice (depending on what the learning algorithm is), you probably also want to learn in the absense of a strong verifier at every step, to allow building a fuzzy/erratic sense of the search space that can lead to planning/intuition and allow distant jumps in a targetted direction.
I wish we knew everything about how LLMs work! We only know very basic elements related to their construction and traning dynamics, and pretty much every major question we would like to address still has unknown or vague heuristic answers. This is expected for such a young field of study. In physics, we know the Schrodinger equation, but we dont know everything about how the world works or how to create new materials even though we know that these materials are composed by atoms and we can simulate small collections of them. In cell biology, we know the sequences that make up the DNA of a cell and we approach the time we can build minimal synthetic cells with pieces we understand, but we only scratch the surface of our level of understanding of how the cells actually work and new discoveries are added every day. In biology at large we still keep finding new types of tubes inside human brains—not sure what you mean by plenty, but we certainly have an extremely limited understanding of human cognition compared to what we might have in 50 years from now. It is not just anout LLMs and intelligence—I would like us to be able to answer practical questions about how LLMs work in order to improve general or specialized LLMs even faster than today. We “know” about scaling in an empirical sense, and it certainly has a long way to go, but it does not feel close to a complete understanding.
I agree boring problems exist; bounds may have a fare share of them. None of the bounds problems in this set are even close to this category; many of them are closer to the type of contributions that in the past got recognized by special awards. Your initial replies were misleading.
Use nvidia hardware instead and use a larger cluster serving many more users concurrently. Easily 10x–20x higher token rate per GPU with public solutions like dynamo and sglang.
Your inverted logic does not hold. The fact that such useless problems for bounds exist does not mean that improving bounds is useless. 9 fields medals in the last twenty years, including the one to Terrence Tao, were for improvements on bounds. 3 of the 4 medals in 2022 were for bounds; 2 of these medals were in combinatorics.
It is unfair to dismiss contributions to decades old open problems as equivalent to calculating more digits of pi. It missed the mark by a lot—as does the two bucket simplifaction.
> Whilst current models can't 'intuit' and come up with conjectures
I disagree. I routinely let LLMs speculate or generate hypotheses along the way of helping with technical research. Sometimes they can prove the correctness of a concrete math idea but other times even an unproven conjecture helps with the numerical algorithm implementation and the result is then simply supported by additional data. I guess that any autoresearch-adjacent application has LLMs intuiting and coming up with hypotheses/conjectures—as do the steps/lemmas along a complex proof. In my opinion the modern LLMs are powerful intuitive thinkers that generate lots of conjectures of varying quality or importance.
Not sure what your first sentence means, or why you are quoting AI. Many of these problems individually were math at a level approaching the highest possible for expert human mathematicians. These are not simple combinations of existing ideas, or following of human intuitions, or implementing something following specific human instructions. Then again maybe you mean that math is not knowledge and that all math is simply extending the basic axioms using known patterns, to which I would not agree.
It is much more subtle than your specific example, which is strictly speaking a bug, though ofc it has been used during pretraining for efficiency purposes. Sglang and miles have been working towards full determinism in open source codebases, so the LLMs can help explain the subtleties encountered in actual projects if you point them to these repos.
In the simplest possible case, a distributed addition of floating point numbers is not deterministic if you don't specify the order of the addition operations. If floats are added in a first-come first-serve fashion (the simplest implementation of a reduction operation) you already lose determinism. These could be activations from multiple experts, but it could also be adding chunks of a matrix operation that uses multiple CUDA cores.
If you manage to achieve determinism in temperature zero, it is possible to extend it to determinism at higher temperatures, because at that point you only have to keep the pseudo-random-number generator state in sync across parallel instances and this problem has been solved.
You can do whatever you want with the model within your own organization. If you use it commercially—either as a model-as-a-service business or in a very large-scale product—you should check the additional license terms, which go beyond MIT. My interpretation is that Moonshot cares about the exact inference behavior and accurate representation of their model or derivatives, and perhaps also about capturing some additional value despite their own GPU limitations, so the extra license terms focus on those large-scale commercial deployments.