• 0 Posts
  • 182 Comments
Joined 3 years ago
cake
Cake day: July 24th, 2023

help-circle
  • True, that might happen at some time in the future, I’ll believe it when I see it. But at the moment most of the generated images or texts are only good enough if a human sloperator does some quality control and regenerates or edits the results until they look good. All the models i’ve seen so far have specific quirks/preferences which repeat as patterns. This might not be immediately apparent when a new model comes out, but it takes a lot less time to identify the quirks than it takes to train the model in the first place.

    To be clear, I’m not implying that I couldn’t be fooled by a genAI text or image, but the majority of uncontrolled zero-shot slop is easy to spot, because it is just that, slop. Look at any random slop readme.md in a github repo for example. The structure is always the same, it is filled with a collection of words that tries to mimick what a readme would look like, including weird jargon imitations but it lacks the clarity, personal touch, developer ego and soul it would have if a real author was writing about his own cherished creation. Look for the things that are absent instead of the things that are there. Same for identifying slop images… Image Gen focuses on the subject and gets sloppy on the “less important” bits, like backgrounds.




  • I don’t entirely agree with the doom and gloom of that article. Humans are particularly good at adapting to changes. And humans are particularly good at detecting patterns and anti-patterns. Patterns are signals, and while adaption will take a while and will be painful, I think we’ll work it out. For example, detecting AI slop reliably and intuitively is a skill that more and more people learn. Once the pattern is learnt, it’s a new useful signal. Other new signals emerge for quality detection. In a world of crap software, the few that work well will stand out. Anything that can be “created” by anyone with a few prompts loses its value. What’s valuable will be what stands out from the slop.




  • Ok, I’ll bite. Assuming that the article means Large Language Models exclusively when mentioning “AI”.

    Good specialized models aren’t that easy to build for a lot of use cases.
    Coding is one of the more problematic cases. People tend to think that coding is about code, so if we just create a model that is specialized in writing code, we’re golden. However it turns out this isn’t true. Coding is not only about code itself, but about creating an application that has a specific purpose and features in a specific domain. The LLM needs an understanding of not only the plain code, but needs knowledge about the subjects the application is about, and these are mostly in the domain of general knowledge. Another issue is that the model itself needs to communicate with the user in an appropriate way, and this needs general knowledge as well.

    “Summarize documents, draft memos, analyze spreadsheets”

    All these examples require an LLM to have a good general knowledge to grasp what the original text is about. You won’t get a decent summary when the general knowledge of the subject at hand is subpar. My point is, bigger LLMs with more general knowledge handle these tasks better than smaller ones. I’m not saying smaller models can’t do the job, but there will be differences in quality.

    It’s not that hard to let existing LLMs use specialized knowledge.
    The actual problem is to prevent this specialized knowledge from being stolen by US/Chinese AI companies. So instead of focusing on building own “specialised” models, let’s work on preventing the ongoing theft of knowledge, and using the right tools for the job.





  • While that is partly true, it oversimplifies a few things. And these are the things that make LLMs a dead end for intelligence:

    • Evolution means constant learning and adaption. LLMs can’t be adapted on the go, they are stuck with their initial training data and retraining is a slow and externally controlled process. Apart from training, every additional information you give it while running must be loaded into the context, the space for context is limited and must be flushed before supplying different context. That’s a hard limit for evolution, at least with current technology.
    • The physical world is more important than you might think because everything hinges on this realm, including the electricity, cabling and hardware involved in running a model. The idea of an intelligent entity that exists purely as an ethereal cloud of logic unbound by physics is hard to justify, because the shape of logic itself is defined by the physical ground we are standing on. Therefore, an AI would need a first-hand access and understanding of the physical world to be successful. LLMs can’t have any of that, because they are purely language-bound.

    I personally think Yann LeCun (a pioneer of machine learning and AI) is on the right track with his research and his opinion that LLMs are a dead end for artificial intelligence. The reasons are more on a technical than a philosophical side.


  • Not the OP, but any software developer who has spent some time with the inner workings of LLMs can tell you that. Open weight models are a good opportunity to explore how LLMs work on the inside and to experiment with them a lot more than you can do with the so called Frontier models. The technology however is the same. These are just very fancy statistical prediction machines, and the reason why you are so impressed with them is that the statistical average of what sounds reasonable in a conversation is exactly what people expect to hear on average, given a specific context. It’s a neat little trick that works only only as long as the context fits (contentwise and spacewise), and once you realize that you’ll see LLMs as what they truly are. There is no magic, and the “intelligence” is the intelligence of the people who built those systems, the systems themselves are not intelligent in any commonly used sense of the word.





  • I don’t have an internal monologue, it works more like a blackbox releasing thoughts when they are finished, the reasoning stays hidden more or less. I can reject the thought or send it back for further refinement, but it does not at all resemble a monologue or train of thought thing. Fun story, I was at a Buddhist meditation session once, and while sitting there I discovered that I can turn my thoughts off almost completely. Which includes the processing of sensory information, like leg going numb or the guy next to me farting. These continued to exist but lost any meaning or relevance to react upon. It was a very weird experience, turning my mind completely inwards, experiencing an intense but pleasant void of complete nothingness while being conscious. It took about an hour for me to get back to normal afterwards. I asked the Buddhist guys later if that is an expected experience with this type of meditation, and they had never heard of such a thing.

    Odd how different people’s brains can be wired internally. And yet, I am able to communicate with all those fancy “internal monologue” people as if I was one of them.




  • “Microsoft changed course” (about shoving Slop into every nook and cranny of their products)

    I very much doubt that. When I have to open Office365 (Copilot… Sloppity… WhateverItIsCalledThIsWeek) in a browser they manage to hide the actual apps like Word, Excel etc. a little more behind the giant empty chat interface. Last time I checked my pinned Apps were gone and now there is a tiny button containing some of the apps. To make it worse, this button changed the location on screen almost every time I opened this website in the past months (I’m not using it very often).