swampash wrote: ↑2 months ago
For speed I tend to use a hyphen, especially if I'm using my iPhone or iPad. On the Mac it's pretty easy to get an em dash by simultaneously pressing the option and hyphen keys:
hyphen: -
em dash: –
Interesting to hear what you had to say about AI's limitations re. the unknown unknowns. Isn't that the fundamental difference between AI and the human brain? AI can trawl through the universe of knowns and deliver a probability based answer but cannot speculate about what it doesn't know, although the spikey knowledge boundary argument might permit that it might be able to infer?
The human brain though is able to speculate about the unknown unknowns. Is this a unique capacity that fundamentally differentiates AI from the human brain?
On another tack, I recently had to take issue with an academic journal over an article it published. I'll pm you with the details if you're remotely interested (?) rather than post them here. AI is being increasingly used to generate fake academic papers – there are even online agencies offering to generate them, for a fee, too. There's a real problem heading down the line for academic research...
There may well be true AI experts (and i am not one) who are shouting at their screens over what i'm writing here, Swamps. I sort of feel it goes deeper than "what it doesn't know" in that, as i said before, an LLM doesn't really "know" anything. It's been trained on an absolute shitload of information but it isn't an entity that fundamentally "knows" stuff in the way humans do... or at least, that's the way it feels to me. I'm definitely down with the article's writer in that AI has no real world interface with which to experience things and no actual way to process or conceptualise "experience" and, thus, is very much limited to what it is fed or "trained" on.
For instance... I'm assuming you've now had a little bit of a tinker with one or more LLMs in their standard chatbot form (if not, please actually do so). Anyway, any feeling that you're having real-world two-way conversation is being supplied mostly by smoke and mirrors. In the end, every single time you send a new prompt, the LLM is starting from scratch. It has no intrinsic "knowledge" or "memory" of the preceding prompts (conversation) you have been sending. The way it gives the illusion of remembering the thread is that the interface is continually re-sending (in the background) as much of the preceding conversation as possible so the LLM can essentially analyse the full conversation - including its own side of the conversation - and pick up from there. In LLM terms, this is known technically (and unsurprisingly) as "context" and context space is very much limited by how much memory and processing power is behind the system and the LLM being used. Due to this, things like ChatGPT use a sort of sliding window system to send as much context as possible and slowly but surely dropping off the oldest parts. Trying to send everything all of the time will eventually consume all available context space and the system runs out of (hardware) memory.
Obviously, OpenAI, Google, Anthropic, etc. have absolutely vast buildings full of unfathomable amounts of processors and memory but it's actually surprisingly easy to roll your own LLM chatbot with limited consumer-grade hardware. I know this because i've done it and have a fairly workable chatbot running on an albeit quite recent and expensive laptop - but it's not too difficult to do. The most difficult part is making it seem like it has a memory of the conversation.
Regarding em dashes. What you posted above is a hyphen and an en dash. I'm sure you are aware but the "en dash" is so called because it is the same typographic length as the "n" character. The "em dash" is longer, being the same length as the "m" character. You need Shift + Option + hypen on an Apple computer to type one and it's even more of a faff on a PC. I wasn't making a big deal out of this to have a pop at you, though. The simple fact of the matter is that no fucker would bother to specifically type an em dash unless they had very good reason to and i'd argue that a massive amount of the population are blissfully ignorant that such a set of different dashes even exists... yet ChatGPT and it's friends use them liberally and almost exclusively. It's a dead giveaway for AI generated copy.
With regard to your beef with a journal - why don't you just post the details here in this thread as I'm sure i'm not the only person here who might be interested.
