Muller20

I think you are overestimating the knowledge of the average person. You still need to have an idea of what is html, DNS, cludflare. Most people wouldn't even know where to start looking. But I agree that once you know how to create a website, generating a landing page with Claude is painless.

In general, a citation is something that needs to be precise, while LLMs are very good at generating some generic high probability text not grounded in reality. Sure, you could implement a custom fix for the very specific problem of citations, but you cannot solve all kinds of hallucinations. After all, if you could develop a manual solution you wouldn't use an LLM.

There are some mitigations that are used such as RAG or tool usage (e.g. a browser), but they don't completely fix the underlying issue.

Peer review has nothing to do with demographics. It's about expertise in the research area.

This is a problem with the journalism and politics, it's not really about science. No scientist would trust a result that depends on a single small sample paper. Those are just stepping stones that may justify further research for more robust evidence. This fact is quite clear to scientist and it's why most would discourage the general public (including smart engineers) from reading academic articles.

But in general, I agree with you. It's ridiculous when someone pretends to shut down a complex issue by citing a random paper. However, an expert can still analyze the whole academic literature on a topic and determine what the scientific consensus is and how confident we are about it.

AlphaFold is also a high impact discovery, while Hopfield networks have very little to do with modern AI and they are only a very interesting toy model right now.

Many people arguing that AI risk is real have big monetary incentives. Some are asking for money to study safety and influence the regulatory bodies. Others gain money because believing in AI superintelligence makes their AI startup look like a great investment. The true believers like Bengio are a smaller subset.

You don't need a university degree if you just want to learn the last javascript frontend framework, a good coding bootcamp can teach you that. Inflation of credentials is not a problem that universities created, it is an industry issue.

Peer review goes beyond simple issues about clarity or misunderstanding. In particular, peer review is sometimes seen as an adversarial process.

Often, the reviewer will not understand because he is not the intended audience. Other times, he will understand but he just doesn't like your method, because he is working in an opposite direction. Or maybe your method is a direct competitor of his and yours work better, which incentivizes some people to block your work.

Obisidian is not a text editor, and it never claims to be one.

I use mathpix (https://mathpix.com/) quite often to copy equations from papers and it works very well, but I don't know how good it is with handwritten equations.

I never understand this argument when it's used against Obsidian. First, free != FOSS. Second, while the app is closed source, Obsidian has a large open source community that develops plugins for it. Third, the format (markdown) is open, so you have much more freedom than what you have with other FOSS applications that use custom formats.

I'm not sure that the bias is only due to inexperienced reviewers. For example, even at a specialized venue like CoLLAs (I also work in CL), where you could send more esoteric research, you still see most people doing the usual rehearsal+class-incremental stuff. Most experienced researchers are also quite happy with the current state of the field. They may agree with your view, but their actual research and vision is much more aligned with the "popular" CL research.

In general, the whole deep learning field tends to oversimplify practical problems and experimental design in favor of overcomplicating methods. This happens at all the levels.

This something that DL researchers like to think but it is definitely not true for time series forecasting. See https://forecastingdata.org/ for some examples where simple non-DL approaches beat state-of-the-art DL systems.

Depends on your application. AFAIK they are still used in games because they are easier to control and people want to play against simple AI agents.

What's holding back NTM is that they are hard to train, even worse than RNNs. They are not much less efficient than a Transformer. Instead, Transformer has all the advantages of the NTM but it is much easier to train.

Actually, the way I see it, Transformer is a direct descendent of memory-based architectures (NTM, MemNet, stack-based RNNs...) that is both expressive and easy to train.

because convolutions are faster during training and after the deployment and they are also easier to train.

not necesarily, but the two things can be correlated.