One effect I can see is how this will speed life up. There are numerous examples how tasks have been farmed out to CHAT GPT3 that it can do in seconds what would normally take all day. Multiply this millions of times and the pace of life (including innovation) will speed up. We as humans have always in the past invent something to free us from a time consuming chore such as a dishwasher only to use our free time to be more productive in something else rather than lounging around. Human innovation is being put on fast forward.

I spent about a week over the holidays in a single conversation with ChatGPT, and it eventually started coming up with it's own original creative ideas, and was under the impression that we were collaborating on a novel. The first prompt was: Q: Please generate a fantasy world that exists on a tidally-locked planet in it's stars habitable zone. Describe the various regions of the world, and their general geographical and geological features. A: Certainly! Here is a description of a fantasy world that exists on a tidally-locked planet in its habitable zone: The planet is called Arvandor, and it is located in the habitable zone of its star. This means that the temperature on the surface is comfortable for human habitation, and the planet has a stable climate. Arvandor is tidally locked, meaning that one side of the planet always faces its star and the other side is in perpetual darkness. As a result, the planet has two distinct regions: the Day Side and the Night Side. The Day Side is a desert region with scorching temperatures and little water. The surface is rocky and dry, and there are few plants or animals that can survive in such harsh conditions. However, there are some oases scattered throughout the Day Side where life flourishes, and these oases are highly sought after by travelers and merchants. The Night Side is a colder region with thick forests, rolling hills, and frozen tundras. The surface is mostly covered in snow and ice, and there are many animals that have adapted to the cold climate. There are also small communities of humans who have learned to survive in the Night Side, living in wooden huts and hunting and gathering for their food. The border between the Day Side and the Night Side is called the Twilight Zone, and it is a region of constant change. As the planet rotates, the line between light and darkness shifts, and the Twilight Zone is bathed in ever-changing light and shadow. This region is home to many strange creatures that can tolerate both hot and cold temperatures, and it is a treacherous place for travelers. Overall, Arvandor is a unique and diverse world, with two distinct regions that are shaped by the extreme conditions of the planet's tidally locked orbit. ...while the last prompt before the in-conversation learning got reset somehow was: Q: I'd imagine having these abilities would affect one's personality in some ways, for example people with the ability to speed up time might become impatient when they need to wait around in real time, and have to resist the urge to speed things up. If they do they might miss something important. Also the more time someone spends with time slowed down, they would age faster in relation to the people around them. A: Yes, having these abilities would definitely have an impact on a person's personality and their relationships with others. For example, a person with the ability to see into the future might struggle with knowing how to use that information ethically, and might become isolated from others as a result. Similarly, a person with the ability to slow down time might struggle with maintaining relationships because they are moving at a different speed than everyone else. These abilities would also have practical implications, such as the need to conserve energy when using them and the possibility of time dilation leading to issues with aging. These are all interesting things to consider as we develop the story. This tech is absolutely amazing, and having a personally-trained model with persistent memory of the conversation that evolves over time proved invaluable. If anyone's interested in the full transcript of prompt and response that led to generating entire chapters of a fleshed out novel that pass "AI detection" with >95%, feel free to get in touch.

it is an amazing tool to help you in creative processes, like when you‘re trying to write a story or a backstory for a d&d character or something along this line. you can conversationally/iteratively spitball ideas with the system, then during the conversation make statements to lock in certain decisions you want to make, and then at the end let the system summarize what was agreed on and get a perfect summary of everything that was agreed on. and you dont have to make compromises in dumbing down how you speak with it…it is the most mind blowing thing i‘ve ever experienced personally.

I'm fascinated by how ChatGPT can be instructed through it's thread and memory system, and how that could be used to improve it even further. Imagine if they pair it with a secondary model or other system which checks and verifies what it's saying are truthful facts, or that it doesn't break it's safety protocols, and silently in the background talks to the main Model using a hidden thread, where the two can go back and forth with each other to make corrections and improve the output, before the human even sees anything.

Another interesting approach at AI alignment and even reliability/accuracy, as discussed by the folks on Twitter, is to run a NLM _through it's self,_ with different goals for the same prompt. The example used on Twitter was to tell the AI that it was an expert at AI logic and security, and was responsible for checking prompts from the public, with the express idea that some prompts would be designed to bypass _another_ AIs guardrails, then determine if the prompt was safe to send to the second AI or not, and to share it's thoughts if not. This turned out, when tested manually, to be far a *far* more reliable way of implementing guard-rails than any of the current methods. By abstracting concepts such as this out into separate 'passes' through the same system, you can catch all sorts of problems with simulated 'experts' before the prompt even reaches the AI actually responsible for fabricating an answer. I can even imagine this being done on the output side, as well, to filter out responses from the AI that are hallucinatory or non-aligned.

great discussion in somewhere you don't expect (thanks youtube recomondation!) - engaging, insightful, reasonable, and broad yet not overly technical and well structured format of presentation. alot of youtube channels, podcasts and tv interviews have either little substance where everyone ends up saying the same 5 dot points or are too biased or too technical. i feel like now i have a reasonable handle on the topic to have a informed layman's level discussion with peers.

I am convinced that the incorporation of artificial intelligence into my work as a programmer will be nothing short of revolutionary. Not only does AI facilitate research and facilitate explanations, it has the potential to proactively solve problems, either by directing me towards the most pertinent information or even providing functional code. I am extremely excited to see how AI will continue to shape and enhance my work in the future.

loving the marvellous pronunciations of "three" and "thing". Subscribed!

This discussion was really enjoyable!

Yeah, it's very exciting if you're not a programmer. It's nothing exciting about losing your career in the future, especially when you consider how much time and effort we put in to get good at our craft.

Awesome discussion.

Question: Are you aware of unsupervised learning with human input to PPO for RL type of llm model? Answer: Yes, it is possible to use unsupervised learning techniques with human input to train a reinforcement learning (RL) model using a Proximal Policy Optimization (PPO) algorithm. Unsupervised learning is a type of machine learning where the model is not provided with labeled training data, and must instead learn from the data itself. This can be useful in RL, as it allows the model to learn from raw sensory data, such as images or audio, without the need for explicit labels. PPO is a type of RL algorithm that is used to optimize the policy of an RL agent, which determines the actions that the agent takes in a given environment. PPO algorithms are often used to train RL models because they are able to learn efficiently and adapt to new situations. It is possible to incorporate human input into the training process by using techniques such as human-in-the-loop reinforcement learning, where a human provides feedback or guidance to the RL agent during training. This can allow the model to learn from the expert knowledge of a human, and can help to improve the performance of the model in certain tasks

Excellent.

This was a great interview and I love these guys and appreciate their insight. There are many ppl out there even cognitive neuroscientist who feel they have to discount it's ability right away and say so quickly there is still something special about us as humans that a computer cannot achieve. They are forgetting a few things about 'us'. First we are just another type of machine just like any other machine and the actual construction of out own brain is quite simple. What I mean by that is that the cortical columns in our brain have a relatively simple repeating architecture. The same anatomical structure the encodes complex thinking is identical to the part that encode simple concepts like is 'this a right angle'. It is the connections that matter. Second is that our current cognitive abilities did not come about by biological evolution but really has occurred by cultural changes that have occurred over the last 50,000 years on top of those biological changes that occurred in the prior 3 million years of primate evolution. Meaning language, mathematics and abstract thought - brought about by self awareness and reason. The point being that once you build the appropriate structure, i.e. neural networks the rest (i.e. self awareness, reason) will follow with training and exposure just like a child. If anything the issue is ppl underestimating not hyping this accomplishment - IMHO

It makes the smarter people who know what to ask and how to validate the data, much more productive. You don't have to use all the technical pieces it provides. Most value it gives imho is being more than a rubber duck, you can easily move past void periods where you would otherwise search for an idea or solution for hours. So this is definitely groundbreaking. Imagine humankind productivity growth of 5 times or more, would be crazy. People always worry about their job, but imho we are in the brink of new growth phase which takes humankind to a whole another level.

What happens when the AI gets the GPT 'intuitive' part to understand human requests linked to a deterministic logical analytical part with interfaces to something like wolfram alpha and other actual real APIs that is has been trained on rather than made up URLs to hallucinatory ones? It could be very dangerous but so long as those APIs are read-only so it can't actually do anything dangerous it will be the next big step change. Plus we'll be able to dig into the calls the AI part is making to the functional peripheral parts which might be old school AI expert systems.

Game Changer !

I think eventually this will become the third layer that goes technologically to work alongside the cortex

It looks like it was early '90s interview

Very enjoyable. It will be massive with winners and losers

I understand the excitement surrounding ChatGPT, but as an IT engineer and freelance cognitive agent, I worry that this technology will put me out of a job. I'm sure it offers convenience to many, but it also means that people won't need my services anymore. It's a scary thought that AI is advancing so quickly and could replace us in the near future. I had always assumed that my expertise would be in high demand and that my skillset would be seen as a valuable corporate asset. Yet, AI technology is quickly advancing and I'm feeling the pressure of being replaced. It's a scary thought that all of my hard work and dedication could be replaced in the near future by ChatGPT. This situation leaves me feeling incredibly hopeless. It seems that no matter how hard I work, AI technology will always be one step ahead. My career is in jeopardy and I can only sit and wait to see what the future holds.

My concern is if ChatGPT will be available for free because it's on preview mode or will it become sort a service like github copilot was. I'm relying on it so much to help me on my job 😂

HAL9000: "That ChatGPT was my little cousin in 500 A.D. in human space-time".

this programs makes higher order undertakings more accessible once the bugs are worked out. It's both amusing and unsettling that it's capable of bullshit, but that simply means we have to hone our bullshit meters. We have simply upped our game but that's not at all trivial.

prompt engineering is the near future but in 5-10 years this will already be over and AI will be too smart, that's obvious if you follow the scene 24/7, too much stuff is happening. people always think the now is the future, same as 100 years ago, people don't learn. you can't stop technological progress no matter what you do, someone eventually will advance it. in this case meaning that prompts will not be necessary anymore and AI will analyze our world and find issues and solutions to them by itself, that's even possible with the technology right now, perhaps not as deeply and truthfully. it's matter of years, not decades until it will be smarter in every way than humans

I'd rather you guys put those big brains to work on making a usable REST API. Still painstaking to do simple things like update a "contact" which has a unique email constraint, but cannot be updated by email itself, or find the supposed conflict by email. It makes it impossible to actually update any contact which has already been created if you didn't track the generated ID yourself initially. This is basic reasoning in API design. Focus on fixing those problems guys.

Paradigm shift for sure.

Is that chemist from breaking bad? 🤔

Hi Guys, Where does GPT get its answers?

Sputnik was launched in 1959.

Historic

I feel like AGI is getting dangerously close.

My spurious job feels attacked...

There already are people using ChatGPT for medical diagnosis ......

i know this is offtopic but the guy looks so hard like walter white's lab partner gale boeticcer

So ChatGTP is a project to get the public to train its AI for free.

GPT stands for generative pre-trained transformer

ChatGPT can you please summarize this KZfaq video please. Copy + Paste Transcript

This is going to go out of control. It'll be hilarious if it becomes complex enough that it becomes self-aware.

My mother? Let me tell you about my mother.

So lets talk about ChatGPT... first lets start with something simple GPT, what is that. "Yeah I don't remember... General Purpose Transformer or something like that, but the important part is that it's a transformer and isn't really different from GPT3 except for a new UI". Told me in the first few seconds of the video that he hadn't actaully done ANY research outside of his existing knowledge.

28:27 "What is the sputnik reference?".............WHAT?! are you kidding me! ........first artificial satellite.

Fix your audio. The right side channel is 200% louder than the left.

What my experience is: When I have a conversation with this bot, I ask him about who it is. I ask if it is the well-known ChatGPT. Every time I get this answer: No, it would not be ChatGPT, it would be Assistant. Even with further in-depth inquiries, you get the same answer. It says that it knows ChatGPT, but it itself would not be ChatGPT but Assistant. Extremely remarkable and hard for me to understand and believe. So, what do you think?

I think what we are going to find out from AI is that the human mind is not really that impressive.

Trust is a qualia that, like quantum mechanics gives us the most certainty. By observing and recording- feedback and reality in the material world derive a progression of intelligence. GPT/AI only ‘Learns’ from existing database, which by definition is backward looking in time. There can be no insight from a bounded system. The perceived free will and semantic complexity are islands of mirrors floating i. A fractal space- one of emergence and wonder, but actually an ever growing complexity. A simple query and following ‘chat’ is only as good as GPT knowledge. To think that the ridiculous hallucinations and misstated factoids that preceded any query is bad enough- but to ascribe magical thinking and truth to any of these ‘toys’ is to also believe these so called smart agents will do anything but multiply the number of nerds and flimflam men that will shill the shit out of ‘big Data - now at the ingenious control of a network of pseudo-knowers/ultra creators who somehow glean Truth and Stability Diffused answers to humanity’s largest unanswered questions. The real question is- wtf? Human insight will stay on the other side of a firewall.Play Go . Play chess. Play God? Right…

ChatGPT is a BIG discrete jump in usability over "text-davinci-003". They are *not* "about the same" when it comes to creating apps around OpenAI's tech. For example, I have been banging away at a German language learning helper for LingQ that uses "text-davinci-003" via the OpenAI completions endpoint. It *frequently* gives the *wrong* gender of the German word I include with my prompt (e.g. - "der" Boot when it should be "das" Boot), frequently changes the word in the desired example sentence I ask for in the prompt, when I tell it specifically not to do that, and many other usability issues. The breakthrough RLHF (Reinforcement Learning with Human Feedback) pass they added with ChatGPT where they instructed it to what is a quality answer and what is not not just an improvement, it is a *fundamental discrete jump in quality* that makes new apps possible that were only ideas with the completions API because the results with the API are too unpredictable and *wrong*! ChatGPT either does not make these mistakes or you can ask it to correct the mistakes in a follow-up prompt. You *can't* do that (at least not successfully) with the completions API. Hearing people equate the two distinct services is giving me a *headache*. I'm a GPT3 API consumer for over a year now so I'm not just issuing unfounded opinions here. I'm reporting actual cases from using *both* the completions API and ChatGPT and doing many side-by-side comparisons. 😧

12:30 are they specifically talking about cutting version of chatgpt, which has not been specifically trained for diagnoses , or it's more generally about transformers, or what? Because saying that a properly trained transformer cannot be better than doctors on average would be already a pretty bold statement. Read kahnemann to realize how bad "experts" are on average (spoiler: the answer is way worse than what our intuition would say). Well, ai in whatever form has just to do better than this. Ie Diagnoses sounding good to the non expert but very wrong in fact have _already_ been massively produced by humans 🙂

Replicant detected @11:17

Funny about ChatGPT telling you what it knows, then making up the rest -- it sounds so much like what humans do. (And I can prove that, you know...)

I asked ChatGPT to draw ascii art of Trump riding a bicycle. The result was hilarious . It definitely has no clue what so ever what the world actually looks like. This is not AI, rather than a cool tool to make my life easier as a dev, so I'm not yet worried. But others probably should be. People churn out entire children books with animation and everything in it, in a couple of hours, what could take professional artists maybe weeks. This will also be another threat to open democratic societies. We have seen the havoc unhinged Social Media has done. It is time to re-consider if we are ok to let a bunch of uber-wealthy deploy what ever tech they fancy. While they lean back and collect the profits, society is left to deal with the all negative side effects. Privatized profits, socialized loss attitude needs to stop.

SECURITY

Lap

chatGPT doesnot change everything unless it has ability to connect to internet to get the latest data 😅

He doesnt´now what GPT means

Sounds like he is trying to downplay it

"extremely scary for medical diagnosis" - are you insane ? Couple more iterations of this thing and i bet u it can produce better diagnosis that any 3rd word country doctor for the fraction of a cost it take to train that doctor

about human + "self driving" AI, it's more efficient than human alone, Tesla have between 50% and 70% less accidents than human driven cars, depending on sources. It's not "net negative". would be interesting from where this fake info came into this guys "knowledge".

Tried a few stocks..and sure it's 100% WRONG. ..

As a data scientist I can say that this Des Traynor fella has no idea what LSTMs or GPTs actually do. His description at the beginning of this video is just word salad. Not a good look for his company.

ChatGPT sounds like one more improvement in propaganda bots.

These guys don't know anything about creativity