Comments by "Posthumanist" (@thornelderfin) on "Fireship" channel.

Or exactly the opposite. AI scammers will be stuck in millions of hours talking to granny anti-scam AI bots just wasting their time. Kitboga9000. What a time to be alive!
237
Don't trust any benchmarks for any LLM. Even if done honestly (which most are not), they represent just a narrow line of questioning. It's best to test every model yourself with your own real-life problems and requests.
55
@suspicioussand Surprisingly, maybe live musicians will not be replaced by an AI (we already have flawless sound reproduction systems, yet people still go to live concerts). Also athletes. Robots can do every sport better than humans, yet people still go watch humans do it. ... But who will pay for anything in an economy where nobody has any job?
19
@Messmers_flame I've been telling this to my friends exactly. The $200 tier is fair. 128k context (compared to 8k for free and 32k with the $20 tier), plus unlimited Advanced Voice Mode (compared to none at all for free and 30 minutes per day with the $20 tier). Everything is unlimited with the $200. I cannot afford it, so I will stay on the $20, but I fully understand it will sell well to the people who need it. I just don't understand all the people being angry and outraged at $200 per month when it's obviously not aimed at them. Do you also outrage and get angry at the existence of a Business Class and First Class seats in an airplane? Or expensive cars that are obviously not meant for you and me?
11
It's not about ability of AI to do certain things (I am sure AI will be able to do everything that we can, and even better, one day), but it's all about costs. AI running costs are going to go down, but building a physical robot is like building a car - it's still very expensive and it will go down in price slowly. If you have dirt-cheap human labor that can compete with the robots, then it could take many years until robots replace all manual workers. But in the end, we will all lose jobs - which is comforting, because that means there will have to be a global solution (if people have no no money, who will buy the products and services? nobody... so there is no economy, there is no civilization ... which means people will just take physically whatever they want and there are no billionaires and millionaires anymore ... the only way is to give people at least some money (perhaps UBI, I don't know) to prevent them from revoling and to keep them buying products and servies so the economy can exist)
8
I use AI every day for my work and for my personal life. It's already tremendously useful. What "AI crash" are you talking about?
8
I remember GPT-3, when it could create a primitive landing page from a drawing on a napkin. We were all shocked. It's been just 3 years since then and it's doing complete refactoring of small projects and other amazing small things. Now imagine another 3 years... and another, and another, and another. It's not too far in the future.
7
@Kamilllly The fundamental problem with fake frames is that to calculate 3 fake frames, you need to have previous frame and the next frame, and so you are on frame 5, and you just start calculating frame 2 with AI, which still takes some time. This impacts your INPUT LAG severely. Essentially you have input lag 1 frame worse than what you would have without any frame generation (not only it does not help, it makes it worse). In games with little movemet this does not matter, but in first person games with lots of movement this causes various degrees of nausena. And then you have VR, and in VR this is absolutely and completely unuseable - you get nauseated even by slightest input lag. I love fake frames in movies, but in games it creates a lot of nausea.
7
@TheVertical92 Alright. Of course I agree with that. AI can be useful in many many products and services, but only where it makes sense and actually improves it. Also I've seen all the AI slop where they only "fake implement" AI, that's not really even an AI. But as a core technology, AI is as useful as the Internet / web or smartphones. It is not going away.
7
The world is changing and you just refuse to admit it. You are set in your ways and aggressively fight against having to admit that the world is changing. AI is accelerating every year and it will overtake absolutely every aspect of our lives. By the time you are old nothing will be even recognizable here. But you cling to mistakes of these AI prototypes we have today and say things like "AI will NEVER be able to do XY", so that you can avoid these negative feelings about the future. Accepting reality is the first step son.
4
I don't understand why so many people in the comments are angry at the $200 / month offer. It's obviously aimed at small businesses and not at you. Plus you can't even say they are paywalling something from you behind $200 / month, because that offer has the same models, but in unlimited quantity. If you pay $20 / month, you have 80 prompts every 3 hours. This is enough for any non-professional and hobbyist (I use it from morning till evening and it's enough for me). If you need more than that, then perhaps you are a business and should pay (NVIDIAs H100s are extremely expensive and these models are very demanding to run). On the one hand you laugh at OpenAI being -5 billion in loss every year, and on the other hand you think they charge too much (it is possible the $20 tier is subsidized that's why they are operating at a loss, but even this won't be enough for a lot of people... gmail and social media have taught us that we are entitled to expensive IT services completely for free, so paying for something that costs a lot to produce is suddenly outrageous for us).
2
@Wolfrich666 For a few years. But there will come a time when AI can do absolutely anything and everything a human can and it can do it better, faster, cheaper. There is no law of physics standing in the way of this. It's just a matter of time. We will all lose all our jobs, everyone everywhere.
2
@acoolcat. It's not like that. We are in prototyping phase of AI. The AI that you see today is a rushed unpolished raw prototype that has not been optimized for efficiency at all. In the past 2 years there have been some minimal efforts to for example train smaller models with better quantization on synthetic data (knowledge transfer) and it has yielded almost the same results for 1% of the inference costs. Every technology has to be rushed and prototyped first without any regard to efficiency, because you have to be first to market and get as much market share as possible, you need to constantly iterate and push progress as much as possible forward. Once the progress slows down (if ever), you will start optimizing for efficiency and safety and all the other things. Every technology works like this.
2
@niranjan-81 AI is constantly absorbing new data. Datasets now have trillions and trillions of tokens. Also more and more reinforcement learning is being introduced both to the pre-training and post-training processes. It will absolutely adapt to everything, it's just a matter of time. Most certainly better than you and me. Our brains are pathetic 1.4 kg and we have absolutely no way to upgrade them. You cannot double your brain capacity, you cannot 10x it. AI can. Easily. This is why it will always outpace you and me - because we cannot change or upgrade our hardware, computer clusters have no limits, they can 1000x. We are an outdated platform.
2
@niranjan-81 That's not how Transformers work, or any Machine Learning for that matter. Nothing is "instructed". What you refer to is procedural programming. That is not happening at all in Machine Larning. You start with blank randomized weights and biases on all heads in all layers of the Transformer model and as you cycle through trillions of tokens of training data, ALL the structures will form by themsleves due to backpropagation. Very similar to how neurons learn in your brain. I've seen many people having complete misconceptions how Transformer model works - some people thinks it's a search engine (it is nothing like a search engine), others think it has rigid procedural programming (it doesn't, everything it knows forms naturally during training). Do not use "philosophy" to try to understand it. Philosophy is not science. Just go and study the technical details and fundamentals of Machine Learning and Transformer model. You will understand how incredibly similar it is to human brain and how many emergent properties it has. It's a miracle.
2
@nietpraten020 I've been working with it since August 2022 every single day from morning to evening. The impact of the progress has definitely not slowed down. It still requires a lot of skill to operate (this will go away in time as the models get smarter and more agentic), but you can create miracles with it. Most people underestimate what AI can do because they just try some extremely basic prompt that is less than 1 page long and then ofcourse the result will not be impressive. But if you build massive meta-prompts and entire system and chain of prompts you can get shockingly good results, even in creative areas, art, programming. We will all lose our jobs, and it will come in a matter of years, not decades.
1
Why would you want anything without AI?
1
@alekseyburrovets4747 I respectfully disagree. If you use it only occasionaly (few times a day), then it's not worth it, but then why would you even need the highest tier. You would be perfectly ok with the $20 tier. However, if you are a small business or a freelancer, you get absolutely unlimited access to o1, 4o, and non-stop unlimited Advanced Voice Mode. It's like having a PhD. level advisor that you can talk to non-stop at your disposal in unlimited quantity for $7 per day. For any small business or a freelancer that can actually use such a system for that type of work, this price is a no-brainer bargain. $7 per day for an assistant. So it depends what your needs are.
1
@alekseyburrovets4747 Also, I frogot. In free tier, your GPT-4o has only 8k tokens of context (this is too little to do anything serious with the AI). In the $20 tier, you have 32k tokens of context (much better and enough for most people). In the pro tier you have 128k of context (this is a game changer for certain use cases where you have very long conversations or you need to feed the AI with a long documentation). Also being limited with certain number of messages per hour is changing your behavior, becuase you don't want to hit the limit and then be unable to work. Unlimited means your mind is free of such concerns and you can just use it whenever you need it without a second thought. I still think OpenAI is sponsoring / subsidizing the $20 tier (and definitely the 200 million free tier members they are bankrolling from their own pocket). It is unlike me to defend a billion dollar corporations, but they have real hard costs. It's not like NVIDIA is giving them the HW for free (take it up with NIVDIA and their insane margins) or Microsoft their datacenter HW and salaries of workers working there. I spent more than a decade of my career in datacenter business (and related). There are so many costs to running a server cluster, it's crazy. Most people have no idea how incredibly expensive it is. Then you have people working on those servers day and night (3:00 AM rush to the datacenter to fix a critical issue is common, so that people can enjoy uninterrupted service). Those people want to eat and pay their bills. Yet 200 000 000 OpenAI customers get all this for free. Somebody has to pay for it.
1
You know there is no logical limit to its intelligence or capabilities in the future. So it's just a matter of time. Maybe few years, maybe more, but it is inevitable. That being said, every single person on this planet will lose his or her job. We are not alone.
1
@maelstrom2313 I 100% agree with you. I didn't mean to say that all AI will replace all jobs the instant it becomes possible, I only wanted to denote that in the end, however long that may take, it will take all the jobs. I can see how slow things get implemented in my country and I have the same thoughts as you. BUT I think we need to factor in that this time no one will be asking the employees their opinion. AI companies will come to the company owners and offer them solution that will replace their entire company as a whole. Everyone will be fired, offices closed. Then it won't matter what is the cultural intertia and how fast can people addapt new technology... because there will be no people. And you know investors and owners will sign this without a second of hesitation. For them every company is a black box that just makes money - and now they will be offered exactly this.
1
@DonaldJevgenjevichPanasonic In USA? No. In developing countries? Yes, absolutely, even for less. Depends on what the capabilities of such robot are (humanoid robots are very expensive and limited, and non-humanoid robots are just machines we have in factories already - and they displaced a lot of workers, but still you have endless millions of people working there). Don't get me wrong, I am techno-optimist and I can't wait for the future. But it will take some time to beat the economy of human body which evolved to be self-replicating, self-repairing, self-maintaining, doesn't need any rare elements for construction or as fuel, only the common ones, and has tremendous flexiblity compared to machines of same size due to nature of muscles. Eventually, robots will replace everyone, but it won't be as soon as it might seem. Most of the world lives on far less than $1 per hour (and we need to count maintenance, replacement parts and energy costs for the robots, because all this is already included in the human salaries). But yes, high paying jobs will be displaced first.
1
@bromibromi Most early stage companies (like OpenAI) don't need to be profitable at all. They are just scaling and growing, focusing on R&D as fast as possible and customer acquistion. Profitability comes only much later. Many many companies did it like this and now are in the 500 most successful (and profitable) tech companies. This is standard operating procedure. Also, you have to understand 200 dollar per month (which is nothing for businesses and freelancers) is losing money because of o1-pro model. That's not a finished product, it's a rushed prototype that has not been optimized for performance at all (think of when GPT-4 came out first vs much more optimized GPT-4 Turbo later, which was one order of magnitude lest costly to run with only sligtly degraded performance).
1
@fufafac2270 In the 1980s I worked with SMEP computer that was the size of 2 rooms. I had 8-bit personal computer at home with 48kb RAM. Today my PC has 2 million times more memory. In the 1990s we all heard that Internet is a fad, that it's a bubble, that it's too expensive and too slow. Yes, my modem had speed 9600 bits per second. Now my Internet is 112 000 times faster for 18 € per month. First cellphone I had in 1990s was a huge brick (Siemens) and it didn't work half the time and most of the country had no coverage. Look at smartphone today, with 8 different cameras, 4k OLED display and a supercomputer in super thin package. AI will be no different. My point is that the AI you see today is just an embryo of the AI that is to come.
1
@fufafac2270 In the 1980s I worked with SMEP computer that was the size of 2 rooms. I had 8-bit personal computer at home with 48kb RAM. Today my PC has 2 million times more memory. In the 1990s we all heard that Internet is a fad, that it's a bubble, that it's too expensive and too slow. Yes, my modem had speed 9600 bits per second. Now my Internet is 112 000 times faster for 18 € per month. First cellphone I had in 1990s was a huge brick (Siemens) and it didn't work half the time and most of the country had no coverage. Look at smartphone today, with 8 different cameras, 4k OLED display and a supercomputer in super thin package. AI will be no different. My point is that the AI you see today is just an embryo of the AI that is to come.
1