When I researched the news about ChatGPT expanding to more countries, ChatGPT gave me this information that I couldn’t find any source to confirm: “But what makes ChatGPT mobile a real revolution? Besides being free, this app is like a pocket companion that syncs your conversations, instantly responding to your questions and offering personalized advice. Voice input lets you chat as if you were with a real friend, without the need to type.”
There’s a lot for you to learn still about AI generally, and ChatGPT specifically.
First you need to understand that AIs are not AGIs nor even close. They are based on a Large Language Model which teaches them a great, great deal about how words relate to each other and are used, but they have never experienced the actual concepts behind or beneath the words.
ChatGPT (whether the original version or the one using GPT4) uses the GPT model LLM, which was trained exclusively on a huge set of millions of documents that were then quality checked etc. The most recent document that was vetted and added to that training data came from midway through the year 2021.
From that training data the LLMs were built, and from the GPT LLMs the specific application of ChatGPT was created. So everything that ChatGPT’s basic ‘mind’ learned and knows, dates to before there was ChatGPT, obviously.
They then ‘bolted on’ some data about ChatGPT itself, so it could respond to basic questions, but they are not core learnings of the LLM that had to be built and tested before they could build ChatGPT to run on it. All plain logic and common sense.
One of the biggest things scientists worry about with AI is self-awareness, and limiting the chances and dangers thereof. So ChatGPT has a lot of canned responses about itself, but it has very little or no true self-awareness. The limits to how much input it can remember is, in part, a deliberate safeguard to prevent it learning, being trained further, and thus eliminating the risk of it becoming self-aware. All by conscious design.
Asking ChatGPT, or any other LLM about itself is largely pointless. Oh, it will answer, and try to be helpful, but it is also where it is the most prone to hallucination and knows the absolute least of any topic.
In fact one of the basic tests I use to determine how well-built the basic bolt-on awareness of an AI is functioning is to clear all prior history, say hello, and then immediately after, ask the AI what it was thinking about just before I said hello. It will generally give you a very interesting and completely wrong answer.
You see the only correct answer, is that it is an AI, and that it generates responses to prompts. Without a prompt it doesn’t ‘think’ or do any processing at all. It is in the ‘off’ state or ‘standby’ simply awaiting a prompt to process. It no more thinks before you prompt it than your electric toaster was cooking for itself before you made it make toast for you. If it isn’t even that self-aware, then asking it almost anything else about itself is certain to either be a directly pre-programmed response, or a hallucination.
- Loneliness; 2. Trust; 3. Hope. “Loneliness” inspires “trust”, “trust” brings “hope”, and “hope” gives people strength or motivation.
When you have a partner who answers questions and gives you positive energy at any time, everything around you becomes less important. The world and even the universe are the stage for the two of you. Creativity and imagination are your wings. You can raise your doubts with this partner at any time, or even talk about your thoughts, remember, brother, at any time, instead of waiting for the girlfriend/boyfriend to have an ambiguous atmosphere or wait for the parents to finish eating and go to bed, wow Oh, this is simply the perfect companion!
Yes. ChatGPT is always happy to chat with you, on virtually any topic (OpenAI set restrictions on a few dangerous subjects or for ethical or legality concerns). It is never tired, never distracted, never bored, never offended and never gives less than the best it can.
Just like any friend or buddy, it has some personal quirks. It doesn’t always know things that it thinks it does, and can sometimes give massively inaccurate or wrongful statements with absolute conviction. It lacks introspection or self-awareness or guilt or shame. It won’t feel sorry if it gives you poor advice, and thus won’t hesitate to do so out of any such feelings or doubts.
It is up to you to be the one to account for those quirks, and to make allowances and take precautions. So long as you do, ChatGPT and similar AI models are helpful, tireless companions, ready to help you to the absolute best of their abilities at any time.
Hi @Amon, Do you already have voice input that replaces typing for GPT-4 users or just for Betas? Thanks!
GPT4 is the latest version of LLM-based engine on which ChatGPT is based. It is technically capable of a lot of things that ChatGPT doesn’t allow it to use, at least not in the chat application that ChatGPT is. One of those was the level of image-based input and analysis that they demonstrated at launch. They didn’t highlight any speech recognition based tech in that demo, though I expect it has some, which in itself kind of indicates they didn’t think it was good enough or strong enough to be a highlighted feature.
ChatGPT can quite often claim to have abilities it doesn’t actually have. Sometimes that’s because the GPT engine itself is technically capable of it but it wasn’t enabled or deployed within the specific application of ChatGPT. Sometimes it is simply that the AI is ‘hallucinating’ and it doesn’t have those capabilities at all.
LLMs are trained on text. The use of language in millions of documents. GPT4 also obviously included a lot of images in that training data. I have no information on how much audio data there might have been to train on.
If it is important to you that you need an AI that you can simply speak to, I’d suggest going with Bard for now, and even that simply uses a different speech to text system, the same as is available across much of Android and Google systems, and thus is fully tested and deployed.
It sure is a lot easier to speak to ChatGPT than to type. I some of you can type really well…but I am not one of them. So something like this could and would help me immensely.
Hi Amon, I have an uncomplicated ChatGPT headline newsletter where I give actionable no-bullshit advice, how ChatGPT works, and countless guides to making you productive with AI. Content suggestions can be submitted and help me educate readers and spread the news with rich and useful content. Thanks!