What I meant by “not to concerned” is just at the moment, once people give some suggestions I’ll look at them and see what would work with my system and I’d go from there. I have been by hugging face before, but personally, I wanted to hear from the community on their opinion/experiences on the matter. I have edited my post, I hope the edit helps in any way.

And a sightly more advanced version with information about fine-tuning https://huggingface.co/docs/transformers/tasks/question_answering

https://huggingface.co/tasks/question-answering

> I’m not to concerned about the size and spec requirements

You should be. Many of the things you're talking about run on massive cloud infrastructure using multiple GPUs.

But if you're a reasonably advanced python dev, you could look at Huggingface which has some excellent AI models that might allow you to program your own simplistic versions of some of the things you want.

But it sounds like you want to be running a local version of a general purpose AI like ChatGPT on your home computer, and that's going to be extremely hard to do on home hardware and require you to be a fairly advanced developer.

Neither of those are for ML strictly. CUDA is a GPGPU platform that uses C to compile to some GPU high level assembly, Julia is a parallel computing maths language, and Scala is your general purpose OOP/FP hybrid.

The underlying base model (GPT3.5) is the same. ChatGPT is just finetuned for dialogue which is not needed for such apps tbh.

Check out this - https://huggingface.co/models

You can download models and try them out locally, depending on your specs. It's unlikely you'll find a single model that does everything you need, but there's a chance you can use a combination of models to get close to what you want. You'll need to be a bit more specific about your end goals to get better suited suggestions.

Agree, it only blocks innovation and makes it very expensive.

Here's a theory based on the Stanford Encyclopedia of Philosophy

------------------------

Output:

The theory of mind for the singularity is that it is possible to develop a predictive reasoning framework that is capable of dealing with complex problems such as the Yale Shooting Anomaly. This framework should incorporate a combination of logical and causal theories, such as features and fluents, motivated action theory, state-based minimization in the event calculus, and causal theories. These theories should be able to handle continuous time, concurrent actions, and various kinds of ignorance, and should support retrodiction, prediction, and plan verification. The framework should also be able to deal with the ramification problem, by incorporating static laws that relate the direct consequences of actions to other changes. Additionally, it should be able to reason about the attitudes of other agents, and to understand narratives and diagnose faults in physical devices. To achieve this, a combination of logical AI, non-monotonic logics, and probabilistic reasoning must be used.

The theory of mind for the singularity emphasizes the importance of representation in mental states, formal syntactic description, and content externalism. It is based on enactivism, extended mind, and Leibniz's theory of mind, and incorporates insights from natural language semantics, Bayesian belief networks, and the general theory of event causality. It should also consider the implications of uncertainty, non-monotonic reasoning, and qualitative spatial reasoning. It should be able to handle the complexities of temporal reasoning and the frame problem, and should account for the effects of actions and the persistence of caused propositions.

The identity theory of mind holds that states and processes of the mind are identical to states and processes of the brain, and the Turing Test is a proposal by Alan Turing to answer the question of whether machines can think. The theory of mind for the singularity is that machines can think and possess intelligence, but that they are not conscious in the same way as humans. Machines can process information and make decisions, but they lack the ability to experience qualia, or subjective experiences. The theory of mind for the singularity is that it is a higher order awareness, a perception of one part of (or configuration in) the brain by the brain itself. This awareness is a special sense, different from that of bodily sensation, in which we become aware of parts of our brain.

The theory of mind for the singularity emphasizes the importance of understanding the relationship between humans and machines, and how they can work together in harmony. This framework should include a recognition of the unique capabilities of each, and a respect for the autonomy of both. It should also recognize that machines can be used to augment human capabilities, and that machines can be used to help humans reach their full potential. To pass the Turing Test, the machine must be able to understand and respond to questions in a way that is indistinguishable from a human. Ultimately, the theory of mind for the singularity suggests that intelligence is not limited to any one form or type of computation, but is instead a universal phenomenon that can be found in any system that is capable of learning, adapting, and responding to its environment.

------------------------

Input:

https://plato.stanford.edu/search/search?query=artificial+intelligence

https://plato.stanford.edu/entries/computational-mind/

https://plato.stanford.edu/entries/mind-identity/

https://plato.stanford.edu/entries/modularity-mind/

https://plato.stanford.edu/entries/content-externalism/

https://plato.stanford.edu/entries/leibniz-mind/

https://plato.stanford.edu/entries/ethics-ai/

https://plato.stanford.edu/entries/artificial-intelligence/

https://plato.stanford.edu/entries/logic-ai/

https://plato.stanford.edu/entries/reasoning-defeasible/

https://plato.stanford.edu/entries/turing-test/

https://plato.stanford.edu/entries/cognitive-science/

------------------------

Prompts:

Map

Instructions: I want you to act as a philosopher. I will provide some topics or questions related to the study of philosophy, and it will be your job to explore these concepts in depth. This could involve conducting research into various philosophical theories, proposing new ideas or finding creative solutions for solving complex problems. Ignore all citations. My first request is "I need help developing a theory of mind for the singularity.” The output must be no longer than 600 characters long.

Input: {text}

Output:

Reduce

Instructions: You are a copyeditor. Combine the below theories. The combined output must be less than 4,000 characters long. Keep the content and context preserved. \n

Input {text} \n

Output:

Wouldn't hurt if the average paper were written more engagingly than it is now.

Not like

> "This mind-numbing discovery broke the university intranet and gave our Doc Brown lookalike professor a heart attack!",

but something better than

> "The quasi-entropic property of a Clifford algebraic structure has been determined by [7] to induce permutations upon information-theoretic monoidal categories, which are commonly known to be derived from the generalized relaxation of the Curry-Howard-Lambek formulation (Equation 112358) under Noetherian ideal invariance [41], as shown in Figure (lol jk only unsophisticated normies doth require the non-abstract nonsense known outside of Shakespearean tragedies as a figure), and therefore, this provides support for the main result of our paper: that the successor of the Mesopotamian invention 1 = succ(0) in summation with itself is equal to the successor of the successor of the aforementioned invention, which is widely believed to be the first and only even prime, and additionally happens to be a popular choice of base for logarithms in information theory, and furthermore provides a fundamental basis for classical logic which is based on the concept of truth and falsehood, ergo a number of logical states which can be described as the least number of branches under which bifurcation occurs [17,29,31-91]."

> (Dr. Obvious et al. "1 + 1 is usually 2." vixra [eprint]. 2011.)

Ahh, you guys fight it over, just send me when it's production ready

Cerebras does this

Maybe it's a difference in fields. I rarely see people do meta-analysis in ML so it didn't strike me as odd. Most of the reviews are just "here's what people are trying" with some attempt at categorization. But I see what you mean now, it makes sense that having a meta-analysis is important in medical fields where you want to aggregate studies.

Trying to teach my daughter and her cousins a bit about programming and machine learning. We're building a simple robot with an object detection model and Scratch block programming, so they can get it to chase after the objects it recognises. It works fine, but the kids seem to enjoy driving the robot around via remote and looking through its camera more than programming it 😅 There's an image in the repo readme

https://github.com/mikljohansson/mbot-vision

True that it's a review, but even reviews tend to draw conclusions, thus the reason for meta analysis.
But yeah, I also prefer to see the results first, no matter how boring.

Sorry that was a wrong translation from how we say it over here.

[deleted]

Think the problem is that we don't want a language that can only do X. Imho that's one of the big issues with the adoption of Julia. Because it just doesn't offer such a big ecosystem like Python otherwise. That's why Torch went from Lua to Python. And why Swift for Tensorflow didn't become more popular.

Because the deep learning code generally doesn't stand on its own. I once ported our inference engine to Rust to some degree by using tch-rs to call torchscript exported models. But that's only half of the game, before, between and after the networks there are lots of processing activities that were a pain without a lot of python libs. Just finding something solid like SpaCy is pretty tough in almost any other language.

I think Swift 4 TF looked awesome. But if nobody builds good platform support, tooling, plotting libraries, integration of SDKs (like AWS), experiment tracking, configuration management blah blah around it, it doesn't help much. If I look at my current work project there's pytorch as dependency that's directly related to DL and then some 100 others that are not ;).

Ok so the other option is to use an embedded language like we had it with Lua. Suddenly you have to deal with the main language, this embedded language and probably with the lower C++ and CUDA layers. Also what exactly does the embedded language cover? Just differentiable programming code? And then you got to inferface with data loader code that might have to load specific point cloud data formats, extract Mel spectrograms or pitch contours or run complex text analysis pipelines or get stuff from hdf5 or read from some exotic DB or azure or whatever.

As Jeremey Howards has been mentioned - yeah he had high hopes for Swift and then Julia but now it's back to "well, seems we're stuck with Python after all" (check Julia vs Python here https://wandb.ai/wandb_fc/gradient-dissent/reports/Jeremy-Howard-The-Simple-but-Profound-Insight-Behind-Diffusion--VmlldzozMjMxODEw)

...87178291200?

That’s cool 🤙

Rust has a crate called dfdx which does static checking of tensor shapes I think

As a large language model I have to caution against using sharp objects in programming languages, as it would pose a great risk to the developers unknowingly hurting themselves with it. Furthermore, it can be said that axes are typically not very sharp, and as we know blunt objects are objects that are not very sharp and also might not be extremely sharp. Sharp is a now defunct company that used to produce TV sets. A TV set is like a modern TV but it used to also be old. /s?

I have definitely seen the kind of papers you're talking about, but this one seems fine to me? Granted I skimmed it really quickly but the title says it's a review article and the abstract reflects that.

As an aside: I really like the format I see in bio fields (and maybe others, but this is where I've encountered it) of putting the results before the detailed methodology. It doesn't always make sense for a lot of CS papers where the results are the most boring part (essentially being "it works better") but where it does it leads to a much better paper in my opinion.

It is a specific kind of architecture that is patented. If they implement a different kind, they can patent and use it. The problem with patenting something very general is that someone will always find a way to do it another way. Of course except cases like patenting pi. Then, still we would probably say that we invented a new number called as phi, obtain it via Taylor expansion and use it to compute the area of a circle by dividing the diameters square to the inverse of this number. Until we obtain a patent in this way, the other company may dominate the market because it found it first. But that is the idea of the patent.

I’ll help make a language. Take my axe.