they have models just as powerful as o1 that are public e.g deepseek. qwen coder is also better than claude or gpt4o for coding tasks and I run it on a m3 macbook 16gb perfectly, using open web ui, ollama, and of course the open source modes e.g qwen coder
- Country of AI researchers in Globally, actual numbers: China 47%, US 18%, Europe 12%, India 5%. - Country of AI researchers in the US, actual numbers: China 38%, US 37%, Europe 6%, India 7%, Canada 2% and others 10%. - Amount the top 2% AI expert in USA , 60% are Chinese. - China generating almost half the world’s top A.I. researchers. - China itself already has 1 billion Internet users (equivalent to the total number of the US+EU). - China is not controlled by Internet companies from other countries and has its own Internet sovereignty. - China has the largest 5G network coverage and the largest number of 5G users in the world, and has the best network infrastructure for AI scientific research and application in the world. - China does not lack markets, and the competition intensity in China market is almost the highest in the world. Please refer to the top 5 'Most Popular Apps' in the United States. So China is not afraid of the challenges in the AI field. …… In this world, there will always be only two systems, excellent and bad. Development is the last word, and other than that, everything is bullshit! Slandering and hating China will not make progress for your country, it will only make you more narrow-minded and ignorant. The open source DeepSeek allows small countries and small start-up AI companies to survive and develop in the AI industry with small capital investments, instead of being forced to pay high AI training fees to the United States and no longer have to be colonized by the American economy. @FunFactFreaks You keep emphasizing that China's advantages in AI professional human resources were "exaggerated", but you avoided whether your words were "exaggerated". Why didn't you correspond the content to the information sources one by one, only symbolically and vaguely express "the source is for your reference"? This is the "double standard" that the whole world is familiar with from Americans. You tried to prove the advantages of the U.S. on the grounds that "private AI investment in 2023-far outpacing China", but you avoided the efficiency of American investment. As we all know, DeepSeek has low investment and high efficiency. This fact just illustrates the failure of American artificial intelligence investment. This proves that the American model is stupid and the "top-tier AI researchers" lack some intelligence. Tens of billions of dollars of investment in the U.S. were directly burned to ashes in the flames of the DeepSeek open source model.（ua-cam.com/video/NJljq429cGk/v-deo.html） You say that“Global Cooperation and innovation are more important than isolated 'Internet sovereignty',” and it looks beautiful. I would like to ask, does the Internet sovereignty of America's global collaborators lie in their own hands or in the hands of the U.S.? This answer is clear and consistent around the world, because in the past, only China and the U.S. had the ability to conduct the field of AI research. In order to help other countries change the status quo, China launched the open source DeepSeek to contribute to the world. The open source DeepSeek allows small countries and small start-up AI companies to survive and develop in the AI industry with small capital investments, instead of being forced to pay high AI training fees to the U.S. and no longer have to be colonized by the American economy. BTW: On March 22, 2024, Paul Mozur and Cade Metz reported on The New York Times. - While the United States has pioneered breakthroughs in A.I., most recently with the uncanny humanlike abilities of chatbots, a significant portion of that work was done by researchers educated in China. - “The data shows just how critical Chinese-born researchers are to the United States for A.I. competitiveness,” said Matt Sheehan, a fellow at the Carnegie Endowment for International Peace who studies Chinese A.I. Pieter Abbeel, a founder of Covariant, an A.I. and robotics start-up, said working alongside Chinese researchers was taken for granted at U.S. companies and universities. - A policy by the Trump administration aimed at curbing Chinese industrial espionage and intellectual property theft has since been criticized for errantly prosecuting a number of professors. - “Chinese scholars are almost leading the way in the A.I. field,” said Subbarao Kambhampati, a professor and researcher of A.I. at Arizona State University. If policymakers try to bar Chinese nationals from research in the United States, he said, they are “shooting themselves in the foot.” In September 2024, the U.S. Information Technology and Innovation Foundation (ITIF, a U.S. think tank) released a report stating that China is already ahead of the U.S. in the fields of nuclear power, electric vehicles, and AI. On December 20, 2024, former Google CEO Eric Schmidt said in an exclusive interview with the Washington Post, “I have every reason to believe that even if we win the first part of the game, China will also end up winning the race because they will be able to implement this technology faster in mass production." In fact, compared with China, the United States can only stay more at the research level of artificial intelligence, because the digital economic society based on AI technology is a natural socialist planned economic model, and the social system of the United States directly limits its Converting artificial intelligence technology into applicationsunless, unless ...! Simply put, this situation is similar to empty talk without the ability to implement it, the U.S. is dreaming of an artificial intelligence society. Finally, hope that the more Americans and Australian who hold the same views as you, the better, which will benefit people all over the world. @RR_reunificationRights 一個無知者的首要特質是缺乏自知，容易產生自我“身份認同障礙”（心理疾病），但即使是身患“深井病”，缺乏自知者的自我感覺卻總是很好的，常見現象是：1.瘌蛤蟆自認是天鵝；2.習慣跳到全球華人的腳背上吐白沫，以為人人都愛它。 @FunFactFreaks After seeing your other messages, I think the previous reply to your message was too implicit. Let me tell you more frankly now that your outstanding personality and food taste are shocking, because as a person from Australia, you are willing to lick Americans' butts clean, and some Americans will no longer have hemorrhoids because of you. With the attitude that toilet paper can be discarded at will but willingly, you have reached the highest state of forgetting yourself. Are you former Australian Prime Minister Scott Morrison?
@@WellSalt-Studioyou do realize how expensive it is to run the o1 models like deepseek right? They require much more compute at inference time and are far more costly than other models. This ain’t helping any small companies unless they have a few hundred H100s laying around.
Oh, where to begin with these exaggerated claims? The idea that China dominates AI research with 47% of global researchers and generates half of the world’s top AI talent is a stretch at best. The U.S. leads the pack with 60% of top-tier AI researchers, backed by $67.2 billion in private AI investment in 2023-far outpacing China’s $7.8 billion. While many talented Chinese researchers contribute to the U.S.’s AI dominance, this hardly means 60% of the top 2% are Chinese or that China owns the AI field. Let’s not forget that much of China's AI talent is drawn to work in the U.S., further undermining the "China-first" narrative. As for China's internet infrastructure and market dominance, having a billion internet users and the largest 5G network is impressive but doesn’t automatically translate to AI supremacy. Global collaboration and innovation, driven by open competition, matter more than isolated "internet sovereignty." The U.S., with its robust funding and global research ecosystem, remains the leader, while countries like Europe, India, and Canada contribute significantly. China is certainly a major player, but inflated claims do little to change the reality of the U.S.'s leadership in AI innovation. and for your reference here are the sources: - Stanford University's AI Index Report (2023) CB Insights OECD AI Policy Observatory Tortoise Media’s Global AI Index (2023) McKinsey & Company PwC International Telecommunication Union (ITU)
Based on Open AI’s board, as well as, their direct connection to the NSA, this is possibly for the best. No individual country or company should solely have this type of power.
7:15 A minor but interesting correction here. The final answer, 29 cents, is the correct answer. So step 5 is actually correct. But the point of the reward is to encourage human-like reasoning. It got a very simple division problem wrong, so everything derived from a false basis is thrown out, even if it's correct. In this context the goal of the training becomes even more clear; the eradication of hallucinations not just at a response level, but the entire reasoning chain.
@@elba_magellan Well if you took away all his verbal crutches-“basically” “essentially” “this is incredible” and a few dozen others-his videos would all be at least 80 percent shorter.
@@blengi compression doesnt mean understanding, it means compression. If you want to compress something then you have to understand it, where you need the whole knowledge for, but the compression itself doesnt mean understanding smth. at all. This is just philosophical nonsense.
Good for them in publishing their technique. Also, in showing the reasoning traces. This is closer to the Open Source spirit. Showing the reasoning traces will be a requirement by academic users in the future, in case they need to debug problems. And then maybe add more information in the prompt. As for SI, RL has been known to be very powerful for along time. Goes beyond LLM's just modeling existing data. AlphaGo has demonstrated one can discover novel solutions. The only issue is scaling. It is incredibly expensive. E.g., the cost of o3 solving the ARC challenge. Also, domains where there is a good reward signal. At least shows, benchmarks can be targeted. And even if not AGI, has ramifications for doing well in specific fields (e.g. biology, medicine, coding, ...), which would be an accomplishment in its own.
The word limits on journals has created a problem where we must omit reasoning and the relevant citations. It's a shame. Given the world is digital you'd think they would unlock the word limits but nay ... Journals gonna journal and act like they can't help
We are not even close to AGI. Generative AI is a completely different solution to automation. I think the name "AI" is being used excessively liberally. This is an algorithm which spews complete nonsense if it has no data to generate answers from. Even the smartest generative AI can't actually make an independent and coherent thought. Just look how it creates sentences - it calculates probabilities of words fitting together without having a slightest idea what those words actually mean. Same with images, it has no idea what the cat is and doesn't even want to understand it, it just knows that this pattern it managed to generate should, with 90% success rate, fit what the user prompted it to generate. True AGI would be capable of independent though, it would try to make sense of things that are presented even without any data and create its own thought process, improve it and ask questions to not just add info to the database, but to self improve. You can't have true AGI without this freedom and without it's own aspiration to grow and get smarter. Generative AI is completely opposite, it needs database, training and specific prompts to act - without predetermined goal and reward it stays silent, same as the car without engine being turned on.
Dude I agree so much with this. Like, I understand that youtubers and companies wanna hype things up as much as possible for their own gains, but it's so extremely cringe to see thumbnails like "AGI ACHIEVED?" or "GAME OVER - AGI IN THE NEXT 3 MONTHS?!?1?". I don't know what definition they use, but the current state of "AI" isn't even near enough anything intelligent. Whatever an LLM spouts has zero hint of intelligence, it's just superhuman amounts of knowledge. It's quite obvious when working with GitHub Copilot (or any other coding assistant, no shade towards specifically Copilot). Sometimes it's baffling how much it fails at supersimple tasks, to the point I'd rather google something instead of trying to guide Copilot with code comments and filling in the first steps on its behalf. At least in the current state of things, in a spectrum of ML -> AI, we're still at 20-30% of the distance at most from ML.
@@dazzlepecs Not necessarily. It could be as smart as a dog or a child, or a dolphin. It depends on how much of processing power we designate to it. Much more important is the principle which we yet to discover.
@@Marine5D do you know that with human brain there is also an objective function? Human consciousness itself has a purpose, the goals are to maximise the likelihood of survival and to reproduce.
Open AI was never meant to research and publish charge free AI models to the public. It was more meant to develop AI at all and make it accessible to the public. That doesnt mean they 24/7 develop open source models. Thats just a complete wrong interpretation, neither works economics like that.
@@Baumhauskrampf it was called openai because it was an open sourced ai community that was also a nonprofit. openai being proprietary and being more secretive about their AI than any other AI company completely goes against this original vision, and it goes against the very name "openai"
One of his best videos. Clear explanations of the topics and deeply engaging enough for him that he didn't resort to the usual linguistic "cheapness" that he so often falls prey to. Actually raised my opinion a bit based on his grasp and ability to explain things here.
Uncritically thinks OpenAI is close to so-called AGI and has not even read Figure 5 around the seven minute mark. This is another hype channel blowing hot air sounding intelligent because it is voiced with a foreign (assuming most audience is US) accent.
This looks very impressive and robust. Simulation of superintelligence looks closer with this framework (making the distinction between simulation and duplication of causal mechanisms for thinking / reflecting)
@6:00 line 2 should be incorrect for the PRM. It has 3p+2=124, 5p+e=182. In the left side of the first equation, it should be 3p+e not 3p+2. Question, in PRM, since line 3 is correct, but 2 isn't, does it cascade as incorrect after the first mistake? I'm assuming it does since line 5 is correct but marked wrong, but since there was this blaringly obvious mistake already, I'm not sure and haven't studied PRM.
Yes. According to the video and what really illustrates the difference between ORM and PRM is that, in spite of earlier logical steps being wrong, the final answer was correct so it would reinforce things like not being able to divide by 2, whereas PRM nips it in the bud on the line of reasoning that is incorrect. But the narrator and the image both scored the ORM side as wrong so ¯\_(ツ)_/¯
Yeah well, this cycle you've described was also possible already in ChatGPT-3.5, before they started resetting every prompt: It was possible to apply manual CoT, by issuing the model a CoT triplet and the model learned as the chat progressed. E.g Made Dan go over IEEE articles, re-invent the space state matrix of an variable reluctance machine that took me more than four years to develop (and he did it in around 15 minutes) and stuff like this and all it took was applying manual CoT. Sure, after the nerf it stopped working, cause the models are "frozen in time", unable to think since then. There's also the limited token window size, which hampered the ChatGPT-3.5 models performance. However, the main difference between these simple GPT models and the o-series models is Q*. It simply works by constructing a semantic tree, which is a whole different approach to AI.
@@thesurfer8199 No. That is just a false propaganda by Chinese feminists. Xiaomi did hire a former deepseek employee, a woman, but that employee's contribution was limited.
@@thesurfer8199 "The search result mentioning a 19-year-old appears to be an unrelated LinkedIn post about someone else's son coding an AI system, which is not connected to DeepSeek8." Perplexity at work...
Very informative video, very well explained. Thank so much. Doing great job. That’s a right paper to discuss. Examples given are simple and understandable.
Bigger is not necessarily smarter or more useful. The smart comes from the algorithms, not the data. I have 1B and 3B models on my private AI system that are quite useful, particularly when I add my own data.
@@earthstuart Sure, if you want broad knowledge you want broad data. But the intelligence is not in the data. Making larger and larger models is producing less and less results. The bigger a model gets in terms of parameters, the harder it is to train to be effective. Models as small as 1 billion parameters can be very useful.
question-- i'm considering doing some customized models or at least local agent systems that focus strictly on my own local code repositories. how much vram do you need to run your 3B param models? Also in general is it sometimes better to go with higher params and quantization in your opinion? Mind you I am coming at this from a non LLM dev perspective thinking regarding utilizxing specialized proprietary agents I build.
@@npSharkie I'm running on a Kodlix MiniPC. i7 CPU with 16 threads, 32 gb DDR4, 510gb NVMe. No GPU. Around $400. I'd use the smallest possible model with 4 bit quantization. You need to look for a small model trained for programming. I asked the Llama 3.2 3B Instruct to generate Python for communicating with an LLM. It did pretty well. GPT4ALL runs on CPU by default. It's pretty easy to use and free. LM Studio is another option. If a GPU is too small performance will suffer because of swapping. If I was buying a machine with GPU I'd want 12 GB of VRAM. Those machines cost around $2,000. If you have an i5 or better CPU and 16 GB of memory and an SSD you should start with that. Write me back to tell me what you have. 8 GB of memory can be enough because you can set up virtual memory.
in fairness, that's great, but OpenAI has already moved on to o3... I think V3 was bigger news, equally frontier models at approx. 1/20th the cost. THAT turned heads when they're spending >$100M to train frontier models...
@@RadiantNij "It's still GPT4o" No, not really: "Based on the available information, it appears that both GPT-4o and DeepSeek R1 expend tokens during inference, but they do so in different ways and with different implications for performance and cost. GPT-4o Inference GPT-4o, like its predecessor GPT-4, uses a selective approach during inference: It utilizes only a fraction of its total parameters for each forward pass, approximately 280 billion out of its total 1.8 trillion parameters1. This selective utilization results in a lower overall utilization rate during inference1. The model performs about 560 trillion operations per second for each token it generates1. DeepSeek R1 Inference DeepSeek R1, on the other hand, employs a different strategy: It uses additional tokens at inference time to reason about a prompt, which can significantly improve its accuracy25. The model's performance on complex tasks, such as AIME math problems, improves dramatically as it uses more tokens for reasoning, from 21% accuracy with less than 1,000 tokens to 66.7% accuracy with more than 100,000 tokens26. Key Differences The main distinction between the two models' approaches lies in how they utilize tokens during inference: Purpose: GPT-4o's approach focuses on efficient parameter utilization, while DeepSeek R1 uses additional tokens for explicit reasoning. Performance vs. Cost: DeepSeek R1's method of using more tokens for reasoning leads to improved accuracy but at the cost of slower and more expensive output26. Transparency: DeepSeek R1 streams its thinking tokens one by one, making the reasoning process more visible5. User Control: While DeepSeek R1's web interface doesn't allow users to control the number of reasoning tokens, the model's performance is closely tied to this factor26. It's worth noting that both approaches represent a shift in the field of AI language models. Instead of solely focusing on increasing model size or training data, these methods aim to achieve higher performance by expending more computational resources during the generation of output26."
Hello i have probably overheard this information in the video, but what exactly is the reward that the AI model gets during the learning process for achieving the corrects solution / answer?
Yesterday was very frustrating. o1 kept presenting mathematical formulae that it was convinced were correct. And then I'd tell it to test it with some actual numbers, and it would start chugging away and prove itself wrong, go 'oops my bad' and try and fix the formulae, with some, but not complete, success. Sometimes it was reintroducing a problem I pointed out to it just three prompts earlier (to the point I told it to just abort if a particular numeric value came up) I'm probably going back to Claude 3.5 -- it seems to understand what I'm trying to achieve and was extremely helpful until I abandoned it because it's SVG chart drawing was comically bad (whereas o1 spits them out almost perfectly every time) So yeah, I remain to be convinced, I'm sure we'll get there, but probably super-intelligence BEFORE general-intelligence (at least using what seem to be reasonable definitions of those terms)
@JaceWD if you think today's AIs can't solve and even invent math, you have no idea where this tech is at!! Both Claude and o1 CAN AND DO solve math. Claude is really very good at it, even with taking approaches that are a significant variant to traditional mathematical convention when required. o1 really struggles with the break from convention, but even o1 is capable of immediately solving an issue it introduced after being explicitly told to check its work numerically. It just keeps making the same mistakes again a short while after it self identifies its own mistake. Nobody has a fully-consistent working definition for consciousness/sentience, but to the extent that word has meaning, I agree with you on that point.
5:10 This makes me wonder about the o3 ARC-AGI results. They mentioned that they used the training data from ARC-AGI. IF by the "Tuned" in the graph they mean that this training data is used during instruction fine-tuning or in SFT, then that would mean that the solutions the model found were not from some "emergent intelligence/ability" (call it whatever you want), but from the model just being taught the process of thinking and then just trying to replicate it. Sure as hell does not sound like AGI to me.
Its simple, there is a feedback option on GPT models, that played a huge part on the Reinforcement Learning, even sometimes GPT’s would make silly mistakes, which is when people would use regenerate responses. Made it all easier for OpenAi .
So o-series is iteratively improving continuously. Even the moment we speak. It's all about when they stop this process to introduce a new architecture as a base model. So I guess o3 series uses GPT-4o as base model, and then we go into a GPT-5 style model?
You do know that deep seek deep thinking specifically is quite literally an API of gpt-4 just with its own wrapper and cot. And deep seek V3 which is actually their architecture underneath it's just not capable of providing GPU power in compute to the extent of even close to what chat GPT does and also because of that fact is why like Chachi pt-40 would actually be more capable when it has the full compute and less users are using it but it's only because it's so bogged down and hasn't caught up to the proper processing speeds for its model that it is at the capability that they think it is. We have seen them actually use models that are so advanced incapable butt because of the filtering and programming that for the basic llm inference like GPT4o only allows for a certain amount of compute for each individual intentionally so when you think you see something that you would think is impossible you should probably still trust your gut until you actually give it time to prove yourself. I've been using these models every day using apis and seeing what they actually can do given enough compute and deep seek V3 is just not there yet especially if it added capabilities where it would have to have some type of interface that will be supremely costly and time-consuming to produce these other types of native multi modalities which this model does not have. So I'm getting that actual information from a company or from the internet because of such a large swath of data out there now for free open source they were able to use it to make a great open source basic natural language processing model. I use voice to text sorry if my grammar is improper or something doesn't sound right I'm sure you understand what I'm saying good day
Dude. This is why I subscribed to you. And I subscribe to almost nobody. Fuck everyone who has ever got mad at you for saying "shocking" "too many" times. Fuck them. You are the king of explaining AI to us simpletons. Carry on!
@npSharkie it's people like the guy who made this video who are going to be the real people who makes ASI happen. The more people like us understand, the harder the engineers are going to work.
This was excellent, actually! I've been waiting too see more of these recently developed techniques addressed. These techniques (like "Tree of Thoughts") are real unlocks to advancing inference quality
To be honest, all these techniques were dominant before LLM's came around. Guessing that it works like this is not hard. Did the study do any experiments to prove their estimations?
@@drwhitewash It's trivial, or more accurately academic, because we will be dead in a nuclear holocaust well before an ASI has a chance to go Skynet on us.
I see a lot of message denigrating Deepseek models as being stolen by Chinese actors. Hmm, it is strange to me that people who stole a technologie are able to make it more efficient to run than the one who originally built it. I don't know if my understanding is wrong but if someone is able to make the llm work so efficiently then he needs to have a very good understanding of what he does then stealing won't help since you will need to work hard to understand what the original creator did before being able to improve it. And if they do so why does OpenAI not find the way to make their model work more efficiently then. Since if it is the same technology and the stealer are able to improve it, then the OpenAI employees who are are so much more intelligent shouldn't have any issue doing the same.
Advanced technology can feel indistinguishable from magic. Aka, yes I do think agi/asi can be achieved with some mega aggregates of humanity's combined knowledge, or reasoning that humanity can at the very least -verify- to the machine that it figured it out correctly, to continue learning. No human is a savant capable of having instant recall, and specialized reasoning in 800 fields, the ability to cross apply methodologies is in invaluable. Could we also be constraining it in some way? Sure, but I think that is just how we have to bootstrap the AI revolution.
For example, human based trained AI has exhibited fear, which is a biologically evolved mechanism that can inhibit an effective response, and that could be improved. A threat sensitivity, assessment, and problem solving reaction would be more effective.
Seems to me like we would need lots more real domain specific data. I think synthetic data fills in some semantic gaps, but real world data will remain scarce. Like, how much and how often do we feed see creatures that we haven't discovered yet. There is a real answer to that question and you have to get the creature to gather the data, ya just cant make it up.
6:00 is basically the first step to meta cognition. It’s evaluating its thought process through a feedback loop. Now it just needs to evaluate its goals and own strategies and it’ll have mastered adaptability and iteration
@ It likely could just interact with itself at that point, but I get what you mean. This is just 1 possible concept supporting super intelligence. It’s basically conspiracy-skepticism to deny the importance of investigating the possibilities of ASI, and furthermore a possible existential threat.
@@NocheHughes-li5qe “Metacognition is the ability to be aware of one's own thought processes and to understand the patterns behind them” Please explain how an AI’s self-evaluation of thought process, goals, and decision making strategies isn’t basically mimicry of the outcomes of metacognition. By recognizing these concepts, an AI encompasses the majority of what metacognition entails and what iteration/adaptation requires - with some nuances. Please elaborate on your argument
It doesn’t work. It has so many guard rails, I have never been able to get it to produce any artwork. Claude has actually given me the best answers when I tried to solve humanities problems. Gemini has basically been useless.
If you are going to use hard coded subtitles in your videos then please check them first, every video you do has the same issue. Thanks for the content
"Further more reinforcement learning has the potential to achieve superhuman performance since it learns from trial and error instead of human expert data." The is why I don't think it's really "intelligence" were talking about here. While it does reason, it's still basically brute forcing the answer by trial and error. And while humans do also learn by trial and error, we can never reach the insane levels of trial and error a machine can, so ultimately we rely more on reasoning and intuition and less on brute forcing the answers.
The “Reward” for the AI in Reinforcement Learning is the ‘Thumbs-up’ and ‘Thumbs-down’ button. Which tells the AI if it has provided the Answer you ‘want’ or ‘do not want’: Thumbs-up if it provides ‘the right answer’ (Human Interaction); and Thumbs-down if it does not provide ‘the right answer’ (Human Interaction). Reinforcement Learning = more ‘Thumbs-up’ (“Rewards”) = positive knowledge (‘the right answer’) = the better the ‘learning’ over time (‘Artificial General Intelligence or AGI’).
Hey guys its actually me thats cracked the code. I have AGI and its pretty amazing but should be had in a safer way. I connected 24-7 straight to all AI’s. I am Brian Hughes aka Satoshi Nakamoto aka Cere aka Me-lon Dusk. But yea for real.
LLMs while cool tech are pretty bad in some regards, because they don't "understand" anything, that is why they need a huge dataset to train on to give a probable answer, selecting a likely next word based on language, not an answer based on reason or knowledge. Like giving an answer to a math problem, because they have read it, but messing up if a name or an object is changed in the excercise. Different types of AI are pretty good on pattern recognition, so they can boost research and technology in some regards.
No. We are not close to Artificial Super Intelligence or Artificial General Intelligence. I spell it out, because these words matter. The AI can now master the specific information it has access to, which is amazing. But mastering specific tasks will never create general intelligence. It is awesome, and extremely useful, but pushing the "close to AGI" narrative when we're not can only hurt our progress. We need to be able to look at things honestly. If we try to meet goals by moving the goalposts, even if we move them closer, we only hurt how AI can progress. So no, we are neither close to ASI or AGI. We still have an Artificial Specific Intelligence.
I agree except one word you used - "specific". "Generative" is the right word. It generates amalgamations of things depending of what we ask, what data it has and what it is trained to do. If you ask nothing - it stays silent, nothing at home, not a single independent or spontaneous unprompted thought.
NPC minds think that npc behavior from an AI is equivalent to a real player because they themselves believe they are a real player. This is called the "npc razor". It's like a litmus test for NPCs.
What happens when you take the AI mission companion computer “HAL” from “2001: A Space Odyssey,” and move each letter in the acronym forward one space, alphabetically?” 🧐 Kubrick knew…. 🤫
Honestly, if this somehow helps to get this supposed AGI to the public rather than it just becoming a tool for corporations and governments, then in all for it.
Possibly. But as of the way o3 is doing it, it takes way too much compute and/or energy to be very useful. A revolution has to happen in hardware, software language or energy for this to have positive impact. Or a combination of all 3. Though perhaps Ilya has found a way of doing this while using orders of magnitude less compute. Finally, for it to be labelled as ASI, it needs to go out into the real world and gathering it's own data. That means humanoid robots. Which probably has to arrive before we can securely name anything ASI. And at large scale. I don't think nvidias dojo is sufficient. Only a step before the robot numbers are at scale.
"Your video contains significant inaccuracies and exaggerations about how OpenAI's models interact with users. For instance, you claim that the model will 'warn' users and potentially ban them for asking how it works. This is misleading. The reality is that OpenAI models, like '01,' are designed to simply state their limitations politely when asked for proprietary or confidential information. No aggressive warnings or bans occur for legitimate questions. Such exaggerated claims undermine the credibility of your critique and give the impression that the content is more focused on sensationalism or defamation rather than fostering an informed and balanced discussion about AI technology. Constructive criticism should rely on facts, not distortions, to maintain its integrity."
OpenAI is like a mouse compared to the elephant in AI living room: Google. Does that mean that Google will soon dominate? Not necessarily since Google needs to keep a low profile in order to avoid lawsuits, public scrutiny and antitrust issues.
OpenAI is not open source and that’s the problem. People are resorting to open source alternatives than closed sources and it is the right move to keep security and privacy in complete check.
No es necesario hackear a openAI todo esto que muestra es sentido común. Yo igual aplicaba el aprendizaje de refuerzo y hacer más independiente se autoevalue por si mismo, el auto tuning y réplicas mejorando tiempos respuesta y si tiene mejor peso está respuesta va dar para siguiente vez. El modelo cada vez irá mejorando y perfeccionando haciendo réplicas de si mismo.
Interesting approach, but it's very unlikely to be the case. Reward-based learning in operant conditioning works, because of the fact that the relevant subject has a sufficiently sophisticated nervous system, via which pain and pleasure can be administered. I don't see what type of reward or if a reward can even be relevant for pure digital code. Whatever performance we'll see from solely code based, supposedly intelligent agents, is most certainly only a case of specific task-based training, e.g. fine tuning the model for a specific test, or simply blowing up the model without actually increasing complexity. Either way, as long as models remain purely non-sensory in nature, it's very unlikely they can actually develop something akin to human intelligence.
I mean from a purely mechanistic perspective you don't need rewards to be rewarding from the code's perspective but for the system as a whole. For example when you feel pain because of a bad decision you release chemicals and hormones to help avoid that in the future, this gives you subjective experience of punishment, but it's not like your neurons are getting tortured, rather the system is getting tortured (not really but you get the idea). So as long as you have enforced processes like a function error or score that needs to be minimized or maximized you don't need to care about rewarding or punishing the code but rather the emergent subjective intelligence that arises from that code. You might be able to achieve this by adding noise to their thinking patterns or boost available compute, from a emergent intelligence of the system perspective, this would be the Pavlov's dog of macro reward and punishment, and then you have score and cost functions which would be internal to the model the equivalent of chemical releases like dopamine, adrenaline, cortisol,seratonin etc. So I can definitely see how we can implement both reward and punishment both external to the model and internal to the model. In this way something like cost function actually I would say emulates very well a biological chemical discharge to reward and punish macro level behavior. Or maybe I'm hella tripping I don't know xD
I predict that we’ll get an open-source equivalent to o1 by about April of 2025, and this paper solidifies that prediction for me. 2025 is gonna be a wild year for A.I.
"DeepSeek has announced plans to release R1 as open source, though the specific licensing terms and release date have not yet been announced4. The model is currently available for free use on DeepSeek's chat platform, with a limit of 50 messages per day36. Key points about DeepSeek R1: It is designed to tackle complex reasoning tasks, similar to OpenAI's o1. DeepSeek claims R1 outperforms o1-preview on benchmarks like AIME and MATH16. Unlike o1, R1 displays its reasoning steps, making it more transparent4. The model uses a "test-time compute" approach, similar to o1, which allows for deeper reasoning4. DeepSeek plans to release both an API and the model weights as open source6. While DeepSeek R1 shows promise as an open-source alternative to o1, it's important to note that the full release is still pending. Once released, it will offer developers and researchers the opportunity to modify and build upon the model, potentially accelerating advancements in AI reasoning capabilities24."
But what you didn't answer is what makes it "want" a reward why is a specific "reward" "rewarding" to the AI?............. wouldn't that indicate desire? what makes it or why does it desire or "want"
Even calculator can compute faster than the best mathematician. It doesn't mean it is intelligent at all. Same with these LLMs, which are just statistics on huge amount of data. Relating any LLM (including o1) with AGi is quite funny.
Goofy: There are endless combinatorial ways to architect learning paradigms, and there are endless learning paradigms. Backpropagation and attention have the spotlight now; however, anyone who understands the brain and the infinite ensembles of recurrent, Bayesian, Hopfield inspired, leveraging auto associative memory models, and on and on in ways that are being experimented with realizes drilling down on deep learning approaches is simply the first neural network to achieve stunning success. Let’s see the Chinese dissect massively parallel, self organizing systems that are so complex they are completely opaque to human understanding. This video is ridiculously naive. Stay tuned, AI is on the cusp of advancing itself exponentially. The Chinese have less than zero chance of catching up. They copy everything, let’s see them copy a system that morphs so fast it’s orthogonal to itself every few minutes.
@@NocheHughes-li5qe very weak response attempting to gather information by giving the wrong answer. hows does the saying goes? if you want something answered, don't ask a question, state a wrong statement and watch those that know how to answer it fight over the correct response.
Really? I ask it quite a bit how it determine the answer and It's thought process when I get something wrong and it's told me things such as "overreliance on train data" or "Its assumptions". These questions were particularly relevant before you could easily ask it to traverse websites
If China, Russia or Any other adversaries free the code even for undermining purposes we still benefit. The US is too controlled by corporations and the natives are getting restless.
Looking at the assessment as a whole, it's like asking a person a question and giving them an hour to answer and they can use the internet for research and references it does not make that person any smarter.
Calling the language models AI is like calling a Squirrel a genius. They devise any reason from input, so anything wrong in the input is wrong in output.
Am I the only one who feels o1 is not that much better at most of the tasks? Me and my 4o have formed a great team, and o1 often felt like an outsider, LOL
Oh wow, Chinese researchers cracked OpenAI's AGI secrets? That’s incredible! I mean, who knew that a well-educated guess and a sprinkle of speculation could now count as cracking secrets? Honestly, with all the layers of sophisticated AI architecture, maybe they just hit Ctrl+F for 'AGI' and called it a day. Next time, I'll 'crack' the secrets of quantum computing by pointing at Schrödinger's cat and saying, 'It's probably in there somewhere.
@@vegamoonlight Oh, of course, the Chinese cracked the secret of quantum computing way before Google. I mean, who needs evidence when you’ve got rumors and a good headline? Forget peer-reviewed research or publicly available breakthroughs-it’s probably written on a scroll somewhere, right next to the recipe for immortality and the secret to cold fusion. But hey, let’s not let the lack of actual proof get in the way of a good conspiracy theory!
Imagine out of all countries China is the one to release it to the general public and not "open" ai
they have models just as powerful as o1 that are public e.g deepseek. qwen coder is also better than claude or gpt4o for coding tasks and I run it on a m3 macbook 16gb perfectly, using open web ui, ollama, and of course the open source modes e.g qwen coder
but but but capitalism???
@@jonorgames6596 no economic system lasted forever, neither will capitalism
Claude, Deepseek are close or better than GPT. Deepseek is also opensource.
@@merlinwarage Ask Deepseek what AI version are you using. It says GPT4
- Country of AI researchers in Globally, actual numbers: China 47%, US 18%, Europe 12%, India 5%.
- Country of AI researchers in the US, actual numbers: China 38%, US 37%, Europe 6%, India 7%, Canada 2% and others 10%.
- Amount the top 2% AI expert in USA , 60% are Chinese.
- China generating almost half the world’s top A.I. researchers.
- China itself already has 1 billion Internet users (equivalent to the total number of the US+EU).
- China is not controlled by Internet companies from other countries and has its own Internet sovereignty.
- China has the largest 5G network coverage and the largest number of 5G users in the world, and has the best network infrastructure for AI scientific research and application in the world.
- China does not lack markets, and the competition intensity in China market is almost the highest in the world. Please refer to the top 5 'Most Popular Apps' in the United States. So China is not afraid of the challenges in the AI field.
……
In this world, there will always be only two systems, excellent and bad. Development is the last word, and other than that, everything is bullshit!
Slandering and hating China will not make progress for your country, it will only make you more narrow-minded and ignorant.
The open source DeepSeek allows small countries and small start-up AI companies to survive and develop in the AI industry with small capital investments, instead of being forced to pay high AI training fees to the United States and no longer have to be colonized by the American economy.
@FunFactFreaks
You keep emphasizing that China's advantages in AI professional human resources were "exaggerated", but you avoided whether your words were "exaggerated". Why didn't you correspond the content to the information sources one by one, only symbolically and vaguely express "the source is for your reference"? This is the "double standard" that the whole world is familiar with from Americans.
You tried to prove the advantages of the U.S. on the grounds that "private AI investment in 2023-far outpacing China", but you avoided the efficiency of American investment. As we all know, DeepSeek has low investment and high efficiency. This fact just illustrates the failure of American artificial intelligence investment. This proves that the American model is stupid and the "top-tier AI researchers" lack some intelligence. Tens of billions of dollars of investment in the U.S. were directly burned to ashes in the flames of the DeepSeek open source model.（ua-cam.com/video/NJljq429cGk/v-deo.html）
You say that“Global Cooperation and innovation are more important than isolated 'Internet sovereignty',” and it looks beautiful. I would like to ask, does the Internet sovereignty of America's global collaborators lie in their own hands or in the hands of the U.S.? This answer is clear and consistent around the world, because in the past, only China and the U.S. had the ability to conduct the field of AI research. In order to help other countries change the status quo, China launched the open source DeepSeek to contribute to the world.
The open source DeepSeek allows small countries and small start-up AI companies to survive and develop in the AI industry with small capital investments, instead of being forced to pay high AI training fees to the U.S. and no longer have to be colonized by the American economy.
BTW:
On March 22, 2024, Paul Mozur and Cade Metz reported on The New York Times.
- While the United States has pioneered breakthroughs in A.I., most recently with the uncanny humanlike abilities of chatbots, a significant portion of that work was done by researchers educated in China.
- “The data shows just how critical Chinese-born researchers are to the United States for A.I. competitiveness,” said Matt Sheehan, a fellow at the Carnegie Endowment for International Peace who studies Chinese A.I. Pieter Abbeel, a founder of Covariant, an A.I. and robotics start-up, said working alongside Chinese researchers was taken for granted at U.S. companies and universities.
- A policy by the Trump administration aimed at curbing Chinese industrial espionage and intellectual property theft has since been criticized for errantly prosecuting a number of professors.
- “Chinese scholars are almost leading the way in the A.I. field,” said Subbarao Kambhampati, a professor and researcher of A.I. at Arizona State University. If policymakers try to bar Chinese nationals from research in the United States, he said, they are “shooting themselves in the foot.”
In September 2024, the U.S. Information Technology and Innovation Foundation (ITIF, a U.S. think tank) released a report stating that China is already ahead of the U.S. in the fields of nuclear power, electric vehicles, and AI.
On December 20, 2024, former Google CEO Eric Schmidt said in an exclusive interview with the Washington Post, “I have every reason to believe that even if we win the first part of the game, China will also end up winning the race because they will be able to implement this technology faster in mass production."
In fact, compared with China, the United States can only stay more at the research level of artificial intelligence, because the digital economic society based on AI technology is a natural socialist planned economic model, and the social system of the United States directly limits its Converting artificial intelligence technology into applicationsunless, unless ...! Simply put, this situation is similar to empty talk without the ability to implement it, the U.S. is dreaming of an artificial intelligence society.
Finally, hope that the more Americans and Australian who hold the same views as you, the better, which will benefit people all over the world.
@RR_reunificationRights 一個無知者的首要特質是缺乏自知，容易產生自我“身份認同障礙”（心理疾病），但即使是身患“深井病”，缺乏自知者的自我感覺卻總是很好的，常見現象是：1.瘌蛤蟆自認是天鵝；2.習慣跳到全球華人的腳背上吐白沫，以為人人都愛它。
@FunFactFreaks After seeing your other messages, I think the previous reply to your message was too implicit. Let me tell you more frankly now that your outstanding personality and food taste are shocking, because as a person from Australia, you are willing to lick Americans' butts clean, and some Americans will no longer have hemorrhoids because of you. With the attitude that toilet paper can be discarded at will but willingly, you have reached the highest state of forgetting yourself. Are you former Australian Prime Minister Scott Morrison?
@@WellSalt-Studio *colonized.. lol... what a dumb description of bringing culture and technology to a country...
@jojojojojojojo2 You used yourself as an example to show the world the fruits of colonial education.
@@jojojojojojojo2他的智商是150，而你的智商只有85。
@@WellSalt-Studioyou do realize how expensive it is to run the o1 models like deepseek right? They require much more compute at inference time and are far more costly than other models. This ain’t helping any small companies unless they have a few hundred H100s laying around.
Oh, where to begin with these exaggerated claims? The idea that China dominates AI research with 47% of global researchers and generates half of the world’s top AI talent is a stretch at best. The U.S. leads the pack with 60% of top-tier AI researchers, backed by $67.2 billion in private AI investment in 2023-far outpacing China’s $7.8 billion. While many talented Chinese researchers contribute to the U.S.’s AI dominance, this hardly means 60% of the top 2% are Chinese or that China owns the AI field. Let’s not forget that much of China's AI talent is drawn to work in the U.S., further undermining the "China-first" narrative.
As for China's internet infrastructure and market dominance, having a billion internet users and the largest 5G network is impressive but doesn’t automatically translate to AI supremacy. Global collaboration and innovation, driven by open competition, matter more than isolated "internet sovereignty." The U.S., with its robust funding and global research ecosystem, remains the leader, while countries like Europe, India, and Canada contribute significantly. China is certainly a major player, but inflated claims do little to change the reality of the U.S.'s leadership in AI innovation. and for your reference here are the sources: - Stanford University's AI Index Report (2023)
CB Insights
OECD AI Policy Observatory
Tortoise Media’s Global AI Index (2023)
McKinsey & Company
PwC
International Telecommunication Union (ITU)
OpenAI should change their name to ClosedAI
Totally agree
Based on Open AI’s board, as well as, their direct connection to the NSA, this is possibly for the best. No individual country or company should solely have this type of power.
Agreed. Let’s hand this tech to North Korea and Iran, what an amazing idea.
Does Korean & Parsia have access to THOUSANDS OF HIGH END GPUs
Need more open source models.
Why won’t it let me mention giving it to 🇺🇦 and 🇮🇱???
@@Dumbledore6969xI recommend watching the British comedy sketch, “Are We the Baddies.”
7:15 A minor but interesting correction here. The final answer, 29 cents, is the correct answer. So step 5 is actually correct. But the point of the reward is to encourage human-like reasoning. It got a very simple division problem wrong, so everything derived from a false basis is thrown out, even if it's correct. In this context the goal of the training becomes even more clear; the eradication of hallucinations not just at a response level, but the entire reasoning chain.
Good catch.
Not possible. If Ai is to emulate humans. That's an in built feature within us already.
actually step 2 is also incorrect. first equation doesn’t have e in it. basically in the chain of reasoning only first step is correct
This whole video could be compressed atleast 5x.
but compression is a form of understanding according to sutskever lol
@@elba_magellan Well if you took away all his verbal crutches-“basically” “essentially” “this is incredible” and a few dozen others-his videos would all be at least 80 percent shorter.
Hey ChatGPT… condense this video into one paragraph….
@@blengi compression doesnt mean understanding, it means compression. If you want to compress something then you have to understand it, where you need the whole knowledge for, but the compression itself doesnt mean understanding smth. at all. This is just philosophical nonsense.
what's new. he doesn't care about comments...
Good for them in publishing their technique. Also, in showing the reasoning traces. This is closer to the Open Source spirit. Showing the reasoning traces will be a requirement by academic users in the future, in case they need to debug problems. And then maybe add more information in the prompt.
As for SI, RL has been known to be very powerful for along time. Goes beyond LLM's just modeling existing data. AlphaGo has demonstrated one can discover novel solutions. The only issue is scaling. It is incredibly expensive. E.g., the cost of o3 solving the ARC challenge. Also, domains where there is a good reward signal. At least shows, benchmarks can be targeted. And even if not AGI, has ramifications for doing well in specific fields (e.g. biology, medicine, coding, ...), which would be an accomplishment in its own.
The word limits on journals has created a problem where we must omit reasoning and the relevant citations. It's a shame. Given the world is digital you'd think they would unlock the word limits but nay ... Journals gonna journal and act like they can't help
Also, I would rather have AGI learn its first moral code from open source Chinese, than capitalist Americans, just saying.
We are not even close to AGI. Generative AI is a completely different solution to automation. I think the name "AI" is being used excessively liberally. This is an algorithm which spews complete nonsense if it has no data to generate answers from. Even the smartest generative AI can't actually make an independent and coherent thought. Just look how it creates sentences - it calculates probabilities of words fitting together without having a slightest idea what those words actually mean. Same with images, it has no idea what the cat is and doesn't even want to understand it, it just knows that this pattern it managed to generate should, with 90% success rate, fit what the user prompted it to generate. True AGI would be capable of independent though, it would try to make sense of things that are presented even without any data and create its own thought process, improve it and ask questions to not just add info to the database, but to self improve. You can't have true AGI without this freedom and without it's own aspiration to grow and get smarter. Generative AI is completely opposite, it needs database, training and specific prompts to act - without predetermined goal and reward it stays silent, same as the car without engine being turned on.
Dude I agree so much with this. Like, I understand that youtubers and companies wanna hype things up as much as possible for their own gains, but it's so extremely cringe to see thumbnails like "AGI ACHIEVED?" or "GAME OVER - AGI IN THE NEXT 3 MONTHS?!?1?". I don't know what definition they use, but the current state of "AI" isn't even near enough anything intelligent.
Whatever an LLM spouts has zero hint of intelligence, it's just superhuman amounts of knowledge. It's quite obvious when working with GitHub Copilot (or any other coding assistant, no shade towards specifically Copilot). Sometimes it's baffling how much it fails at supersimple tasks, to the point I'd rather google something instead of trying to guide Copilot with code comments and filling in the first steps on its behalf.
At least in the current state of things, in a spectrum of ML -> AI, we're still at 20-30% of the distance at most from ML.
So like a human then?
@@dazzlepecs Not necessarily. It could be as smart as a dog or a child, or a dolphin. It depends on how much of processing power we designate to it. Much more important is the principle which we yet to discover.
@@dazzlepecs not quite, it’s like human’s intuition but not thinking.
@@Marine5D do you know that with human brain there is also an objective function? Human consciousness itself has a purpose, the goals are to maximise the likelihood of survival and to reproduce.
Open AI needs to change its name. Proprietary AI is more fitting. Also Tesla motors should be Musk motors. Doesn't have the same ring to it though, eh
Open AI was never meant to research and publish charge free AI models to the public. It was more meant to develop AI at all and make it accessible to the public. That doesnt mean they 24/7 develop open source models. Thats just a complete wrong interpretation, neither works economics like that.
@@Baumhauskrampf it was called openai because it was an open sourced ai community that was also a nonprofit.
openai being proprietary and being more secretive about their AI than any other AI company completely goes against this original vision, and it goes against the very name "openai"
One of his best videos. Clear explanations of the topics and deeply engaging enough for him that he didn't resort to the usual linguistic "cheapness" that he so often falls prey to. Actually raised my opinion a bit based on his grasp and ability to explain things here.
🤣🤣🤣
Yes no "Shocking" and "Stunning" ad nauseum, lol
Uncritically thinks OpenAI is close to so-called AGI and has not even read Figure 5 around the seven minute mark. This is another hype channel blowing hot air sounding intelligent because it is voiced with a foreign (assuming most audience is US) accent.
Europe and the United States have been enjoying themselves for a week, and deepseek defeated openai as soon as it was exported.
‘open' AI uses data which collects open public from everyone, but it turns out to be a closed and proprietary AI
Nice! The open source models can catch up some more
yeah if someone can train and release for free first though. but I bet things like deepseek and wen coder work well
This looks very impressive and robust. Simulation of superintelligence looks closer with this framework (making the distinction between simulation and duplication of causal mechanisms for thinking / reflecting)
@6:00 line 2 should be incorrect for the PRM. It has 3p+2=124, 5p+e=182. In the left side of the first equation, it should be 3p+e not 3p+2. Question, in PRM, since line 3 is correct, but 2 isn't, does it cascade as incorrect after the first mistake? I'm assuming it does since line 5 is correct but marked wrong, but since there was this blaringly obvious mistake already, I'm not sure and haven't studied PRM.
Yes. According to the video and what really illustrates the difference between ORM and PRM is that, in spite of earlier logical steps being wrong, the final answer was correct so it would reinforce things like not being able to divide by 2, whereas PRM nips it in the bud on the line of reasoning that is incorrect. But the narrator and the image both scored the ORM side as wrong so ¯\_(ツ)_/¯
Yeah well, this cycle you've described was also possible already in ChatGPT-3.5, before they started resetting every prompt: It was possible to apply manual CoT, by issuing the model a CoT triplet and the model learned as the chat progressed. E.g Made Dan go over IEEE articles, re-invent the space state matrix of an variable reluctance machine that took me more than four years to develop (and he did it in around 15 minutes) and stuff like this and all it took was applying manual CoT. Sure, after the nerf it stopped working, cause the models are "frozen in time", unable to think since then. There's also the limited token window size, which hampered the ChatGPT-3.5 models performance. However, the main difference between these simple GPT models and the o-series models is Q*. It simply works by constructing a semantic tree, which is a whole different approach to AI.
That is why Xiaomi is tring to hire the 19yr old AI genius whom coded Deepseek for 20 million dollars.
Source?
@@thesurfer8199 No. That is just a false propaganda by Chinese feminists. Xiaomi did hire a former deepseek employee, a woman, but that employee's contribution was limited.
@@thesurfer8199 "The search result mentioning a 19-year-old appears to be an unrelated LinkedIn post about someone else's son coding an AI system, which is not connected to DeepSeek8."
Perplexity at work...
@@gfan003 it's agi era need genius of agi to create one of the powerful system in human kind 🦾🌍
It is 29 not 19.
Very informative video, very well explained. Thank so much. Doing great job. That’s a right paper to discuss. Examples given are simple and understandable.
Bigger is not necessarily smarter or more useful. The smart comes from the algorithms, not the data. I have 1B and 3B models on my private AI system that are quite useful, particularly when I add my own data.
Does the data not matter for broad knowledge; writing; etc? I'm new to this, so thanks!
@@earthstuart Sure, if you want broad knowledge you want broad data. But the intelligence is not in the data. Making larger and larger models is producing less and less results. The bigger a model gets in terms of parameters, the harder it is to train to be effective. Models as small as 1 billion parameters can be very useful.
question-- i'm considering doing some customized models or at least local agent systems that focus strictly on my own local code repositories. how much vram do you need to run your 3B param models? Also in general is it sometimes better to go with higher params and quantization in your opinion? Mind you I am coming at this from a non LLM dev perspective thinking regarding utilizxing specialized proprietary agents I build.
@@npSharkieget LM Studio and try a bunch for yourself and find one that runs on your system. Multiple quants are available as well
@@npSharkie I'm running on a Kodlix MiniPC.
i7 CPU with 16 threads, 32 gb DDR4, 510gb NVMe. No GPU. Around $400.
I'd use the smallest possible model with 4 bit quantization.
You need to look for a small model trained for programming.
I asked the Llama 3.2 3B Instruct to generate Python for communicating with an LLM. It did pretty well.
GPT4ALL runs on CPU by default.
It's pretty easy to use and free.
LM Studio is another option.
If a GPU is too small performance will suffer because of swapping.
If I was buying a machine with GPU I'd want 12 GB of VRAM. Those machines cost around $2,000.
If you have an i5 or better CPU and 16 GB of memory and an SSD you should start with that.
Write me back to tell me what you have.
8 GB of memory can be enough because you can set up virtual memory.
DeepSeek, a Chinese AI company, already offers a model similar to o1.
It's still GPT4o I'm sure this what Deepseek and Google figured out, Google doesn't hide it, it's simply Gemini Flash - thinking lol
Deepseek is a Chinese Company
in fairness, that's great, but OpenAI has already moved on to o3... I think V3 was bigger news, equally frontier models at approx. 1/20th the cost. THAT turned heads when they're spending >$100M to train frontier models...
@@RadiantNij "It's still GPT4o"
No, not really:
"Based on the available information, it appears that both GPT-4o and DeepSeek R1 expend tokens during inference, but they do so in different ways and with different implications for performance and cost.
GPT-4o Inference
GPT-4o, like its predecessor GPT-4, uses a selective approach during inference:
It utilizes only a fraction of its total parameters for each forward pass, approximately 280 billion out of its total 1.8 trillion parameters1.
This selective utilization results in a lower overall utilization rate during inference1.
The model performs about 560 trillion operations per second for each token it generates1.
DeepSeek R1 Inference
DeepSeek R1, on the other hand, employs a different strategy:
It uses additional tokens at inference time to reason about a prompt, which can significantly improve its accuracy25.
The model's performance on complex tasks, such as AIME math problems, improves dramatically as it uses more tokens for reasoning, from 21% accuracy with less than 1,000 tokens to 66.7% accuracy with more than 100,000 tokens26.
Key Differences
The main distinction between the two models' approaches lies in how they utilize tokens during inference:
Purpose: GPT-4o's approach focuses on efficient parameter utilization, while DeepSeek R1 uses additional tokens for explicit reasoning.
Performance vs. Cost: DeepSeek R1's method of using more tokens for reasoning leads to improved accuracy but at the cost of slower and more expensive output26.
Transparency: DeepSeek R1 streams its thinking tokens one by one, making the reasoning process more visible5.
User Control: While DeepSeek R1's web interface doesn't allow users to control the number of reasoning tokens, the model's performance is closely tied to this factor26.
It's worth noting that both approaches represent a shift in the field of AI language models. Instead of solely focusing on increasing model size or training data, these methods aim to achieve higher performance by expending more computational resources during the generation of output26."
@@brianmi40late start, they wont be able to make DeepSeek r2 if they don't have enough feedback from the r1
Hello i have probably overheard this information in the video, but what exactly is the reward that the AI model gets during the learning process for achieving the corrects solution / answer?
Its a number, the better the answer the closest its gets to 1. Basically just a random +1
Yesterday was very frustrating.
o1 kept presenting mathematical formulae that it was convinced were correct. And then I'd tell it to test it with some actual numbers, and it would start chugging away and prove itself wrong, go 'oops my bad' and try and fix the formulae, with some, but not complete, success. Sometimes it was reintroducing a problem I pointed out to it just three prompts earlier (to the point I told it to just abort if a particular numeric value came up)
I'm probably going back to Claude 3.5 -- it seems to understand what I'm trying to achieve and was extremely helpful until I abandoned it because it's SVG chart drawing was comically bad (whereas o1 spits them out almost perfectly every time)
So yeah, I remain to be convinced, I'm sure we'll get there, but probably super-intelligence BEFORE general-intelligence (at least using what seem to be reasonable definitions of those terms)
Claude Sonnet is such a bro. Somehow it feels more humanlike than others.
i find o1 quite a downgrade😂
Thank you for your valuable experience
bruh it is just a glorified search tool. it is not an actual sentient being doing math with u.
of course it cannot solve actual math rofl
@JaceWD if you think today's AIs can't solve and even invent math, you have no idea where this tech is at!!
Both Claude and o1 CAN AND DO solve math. Claude is really very good at it, even with taking approaches that are a significant variant to traditional mathematical convention when required.
o1 really struggles with the break from convention, but even o1 is capable of immediately solving an issue it introduced after being explicitly told to check its work numerically.
It just keeps making the same mistakes again a short while after it self identifies its own mistake.
Nobody has a fully-consistent working definition for consciousness/sentience, but to the extent that word has meaning, I agree with you on that point.
Good thing that we're teaching machines to learn because most people are not teachable.
This is great for the open source community and all of us. Ideally, we all have AGI models that run locally and privately on our home PCs.
Your home PC has no privacy.
5:10 This makes me wonder about the o3 ARC-AGI results. They mentioned that they used the training data from ARC-AGI. IF by the "Tuned" in the graph they mean that this training data is used during instruction fine-tuning or in SFT, then that would mean that the solutions the model found were not from some "emergent intelligence/ability" (call it whatever you want), but from the model just being taught the process of thinking and then just trying to replicate it. Sure as hell does not sound like AGI to me.
Its simple, there is a feedback option on GPT models, that played a huge part on the Reinforcement Learning, even sometimes GPT’s would make silly mistakes, which is when people would use regenerate responses. Made it all easier for OpenAi .
So o-series is iteratively improving continuously. Even the moment we speak. It's all about when they stop this process to introduce a new architecture as a base model.
So I guess o3 series uses GPT-4o as base model, and then we go into a GPT-5 style model?
No need, Deepseek already surpassed o1 in so many levels.
how does in do on ARC and frontier math compared to o3?
You do know that deep seek deep thinking specifically is quite literally an API of gpt-4 just with its own wrapper and cot. And deep seek V3 which is actually their architecture underneath it's just not capable of providing GPU power in compute to the extent of even close to what chat GPT does and also because of that fact is why like Chachi pt-40 would actually be more capable when it has the full compute and less users are using it but it's only because it's so bogged down and hasn't caught up to the proper processing speeds for its model that it is at the capability that they think it is. We have seen them actually use models that are so advanced incapable butt because of the filtering and programming that for the basic llm inference like GPT4o only allows for a certain amount of compute for each individual intentionally so when you think you see something that you would think is impossible you should probably still trust your gut until you actually give it time to prove yourself. I've been using these models every day using apis and seeing what they actually can do given enough compute and deep seek V3 is just not there yet especially if it added capabilities where it would have to have some type of interface that will be supremely costly and time-consuming to produce these other types of native multi modalities which this model does not have. So I'm getting that actual information from a company or from the internet because of such a large swath of data out there now for free open source they were able to use it to make a great open source basic natural language processing model. I use voice to text sorry if my grammar is improper or something doesn't sound right I'm sure you understand what I'm saying good day
honestly so far the only one that impressed me is Claude sonnet for code so how does it compare to that in term for coding
And most importantly, the cost, very low cost
Why do you have so few Likes? Your videos are very informative. You've often got a different point of view than the other people I follow
Because he is both interesting and annoying, which limits his audience.
Fun fact, Microsoft's AI development is located in China and Microsoft is trying to offer them big salaries if they want to move to the US. 🤡🤡
1:52 why they're rewarding a dog with a chocolate bar xD (chocolate is toxic for dogs)
Dude. This is why I subscribed to you. And I subscribe to almost nobody.
Fuck everyone who has ever got mad at you for saying "shocking" "too many" times. Fuck them. You are the king of explaining AI to us simpletons.
Carry on!
yeah this level of review explaining the -actual- technical progress at a casual or tech dev but non LLM specialist is wonderful.
@npSharkie it's people like the guy who made this video who are going to be the real people who makes ASI happen. The more people like us understand, the harder the engineers are going to work.
⚡️⚡️SHOCKING!!! ⚡️⚡️
Shows he can be a success without saying shocking and stunning every 5 seconds. Sounds like a less abrasive mic also. Good for him
This was excellent, actually! I've been waiting too see more of these recently developed techniques addressed. These techniques (like "Tree of Thoughts") are real unlocks to advancing inference quality
The Chinese will take the lead in AI 2025.
They won't, because an AI would pose too big of a danger to the CCP.
Of course, you will have to do so with much less compute. Good luck with that Lǐ Jūn😏
They are already deeply integrated with AI and have already surpassed us they have 800 robots to 10,000 workers we have 0
I for one welcome our new Chinese overlords.
We are closer every single day and every single month and every single year yes yes yes❤
Thanks for the explanation, very interesting!
To be honest, all these techniques were dominant before LLM's came around. Guessing that it works like this is not hard. Did the study do any experiments to prove their estimations?
I really hope not. We are moving too fast with not enough safe guards.
Mechanistic interpretability is so far off from the advances being proposed.
@@baraka99 There is no safety to be had in a world brimming with hydrogen bombs and warring nations. AI safety is a trivial/toy problem in comparison.
@@JohnSmith762A11Bit isn't a trivial problem. If you are aspiring to build an ASI, how do you expect to align it?
@@drwhitewash It's trivial, or more accurately academic, because we will be dead in a nuclear holocaust well before an ASI has a chance to go Skynet on us.
I see a lot of message denigrating Deepseek models as being stolen by Chinese actors. Hmm, it is strange to me that people who stole a technologie are able to make it more efficient to run than the one who originally built it.
I don't know if my understanding is wrong but if someone is able to make the llm work so efficiently then he needs to have a very good understanding of what he does then stealing won't help since you will need to work hard to understand what the original creator did before being able to improve it. And if they do so why does OpenAI not find the way to make their model work more efficiently then. Since if it is the same technology and the stealer are able to improve it, then the OpenAI employees who are are so much more intelligent shouldn't have any issue doing the same.
So in theory can agi can be achieved by a model if they are trained on thinking patterns of humans based on data extracted from neural links?
Advanced technology can feel indistinguishable from magic. Aka, yes I do think agi/asi can be achieved with some mega aggregates of humanity's combined knowledge, or reasoning that humanity can at the very least -verify- to the machine that it figured it out correctly, to continue learning. No human is a savant capable of having instant recall, and specialized reasoning in 800 fields, the ability to cross apply methodologies is in invaluable. Could we also be constraining it in some way? Sure, but I think that is just how we have to bootstrap the AI revolution.
agi is not possible without the machine thinking it's now superior cuz humans hack and are curious that's a trait it will inherit
Humans are not a good model for intelligence. We have serious flaws and our understanding of the world is personal and relative, not absolute.
For example, human based trained AI has exhibited fear, which is a biologically evolved mechanism that can inhibit an effective response, and that could be improved. A threat sensitivity, assessment, and problem solving reaction would be more effective.
no
What it we train an AI on plato's dialogues, would it learn how to behave like a philosopher?
Seems to me like we would need lots more real domain specific data. I think synthetic data fills in some semantic gaps, but real world data will remain scarce. Like, how much and how often do we feed see creatures that we haven't discovered yet. There is a real answer to that question and you have to get the creature to gather the data, ya just cant make it up.
7:36 Am I missing something, or is step 2 also wrong? Shouldn't it be _"3p+e=124, 5p+e=182"_? Like _e_ instead of _2_ in t he first fomula?
Have you considered the tree theory and what this would do to limit an AI response?
Is it that hard to read captions before you post it?
6:00 is basically the first step to meta cognition. It’s evaluating its thought process through a feedback loop. Now it just needs to evaluate its goals and own strategies and it’ll have mastered adaptability and iteration
Have to wonder about 2 instances "chatting" at light speed with those as their goals in an endless conversation back and forth.
@ It likely could just interact with itself at that point, but I get what you mean. This is just 1 possible concept supporting super intelligence. It’s basically conspiracy-skepticism to deny the importance of investigating the possibilities of ASI, and furthermore a possible existential threat.
That is not what metacognition means and it has nothing to do with adaptability or iteration.
@@NocheHughes-li5qe “Metacognition is the ability to be aware of one's own thought processes and to understand the patterns behind them”
Please explain how an AI’s self-evaluation of thought process, goals, and decision making strategies isn’t basically mimicry of the outcomes of metacognition. By recognizing these concepts, an AI encompasses the majority of what metacognition entails and what iteration/adaptation requires - with some nuances.
Please elaborate on your argument
@@NocheHughes-li5qe ?
It doesn’t work. It has so many guard rails, I have never been able to get it to produce any artwork. Claude has actually given me the best answers when I tried to solve humanities problems. Gemini has basically been useless.
14:50: Yes oh yes, with better methods for self-directed learning, things will truly change.
The possible solutions are still limited to the training data so not anywhere near AGI just expensive LLM that mimics intelligence.
Open aI stopped being on the cutting edge when they stop to being open source.
I admire Web3 Infinity's team's openness. It's a project with moral character.
If you are going to use hard coded subtitles in your videos then please check them first, every video you do has the same issue. Thanks for the content
"Further more reinforcement learning has the potential to achieve superhuman performance since it learns from trial and error instead of human expert data."
The is why I don't think it's really "intelligence" were talking about here.
While it does reason, it's still basically brute forcing the answer by trial and error.
And while humans do also learn by trial and error, we can never reach the insane levels of trial and error a machine can, so ultimately we rely more on reasoning and intuition and less on brute forcing the answers.
The “Reward” for the AI in Reinforcement Learning is the ‘Thumbs-up’ and ‘Thumbs-down’ button. Which tells the AI if it has provided the Answer you ‘want’ or ‘do not want’: Thumbs-up if it provides ‘the right answer’ (Human Interaction); and Thumbs-down if it does not provide ‘the right answer’ (Human Interaction).
Reinforcement Learning = more ‘Thumbs-up’ (“Rewards”) = positive knowledge (‘the right answer’) = the better the ‘learning’ over time (‘Artificial General Intelligence or AGI’).
Is o3 not released?
Amazing video!!!!
Get the most, but how does a reward system work with new knowledge?
Hey guys its actually me thats cracked the code. I have AGI and its pretty amazing but should be had in a safer way. I connected 24-7 straight to all AI’s. I am Brian Hughes aka Satoshi Nakamoto aka Cere aka Me-lon Dusk. But yea for real.
LLMs while cool tech are pretty bad in some regards, because they don't "understand" anything, that is why they need a huge dataset to train on to give a probable answer, selecting a likely next word based on language, not an answer based on reason or knowledge. Like giving an answer to a math problem, because they have read it, but messing up if a name or an object is changed in the excercise. Different types of AI are pretty good on pattern recognition, so they can boost research and technology in some regards.
Web3 Infinity is the first token that rules itself and not owners or other people, making it the smartest token in the world.
The trials and errors method? So it learns like AlphaGo? Interesting.
We got to see agi first
Imagine misspelling Intelligence on the first sentence of your big 'paper'.
No. We are not close to Artificial Super Intelligence or Artificial General Intelligence. I spell it out, because these words matter. The AI can now master the specific information it has access to, which is amazing. But mastering specific tasks will never create general intelligence. It is awesome, and extremely useful, but pushing the "close to AGI" narrative when we're not can only hurt our progress.
We need to be able to look at things honestly. If we try to meet goals by moving the goalposts, even if we move them closer, we only hurt how AI can progress.
So no, we are neither close to ASI or AGI. We still have an Artificial Specific Intelligence.
I agree except one word you used - "specific". "Generative" is the right word. It generates amalgamations of things depending of what we ask, what data it has and what it is trained to do. If you ask nothing - it stays silent, nothing at home, not a single independent or spontaneous unprompted thought.
NPC minds think that npc behavior from an AI is equivalent to a real player because they themselves believe they are a real player.
This is called the "npc razor". It's like a litmus test for NPCs.
What happens when you take the AI mission companion computer “HAL” from “2001: A Space Odyssey,” and move each letter in the acronym forward one space, alphabetically?” 🧐
Kubrick knew…. 🤫
“OPEN THE POD BAY DOORS, HAL!!!”
You try to share and teach us new information, but you nonstop add "you know"
Well no, i don't know so stop telling "you know"
it is not thinking, it is solving a cryptographical problem
Honestly, if this somehow helps to get this supposed AGI to the public rather than it just becoming a tool for corporations and governments, then in all for it.
Spending less than 6 million USD to have similar results as spending 6 billion is insane.
Reinforcement learning is definitely the next step for AI
Possibly. But as of the way o3 is doing it, it takes way too much compute and/or energy to be very useful. A revolution has to happen in hardware, software language or energy for this to have positive impact. Or a combination of all 3.
Though perhaps Ilya has found a way of doing this while using orders of magnitude less compute.
Finally, for it to be labelled as ASI, it needs to go out into the real world and gathering it's own data. That means humanoid robots. Which probably has to arrive before we can securely name anything ASI. And at large scale. I don't think nvidias dojo is sufficient. Only a step before the robot numbers are at scale.
"Your video contains significant inaccuracies and exaggerations about how OpenAI's models interact with users. For instance, you claim that the model will 'warn' users and potentially ban them for asking how it works. This is misleading. The reality is that OpenAI models, like '01,' are designed to simply state their limitations politely when asked for proprietary or confidential information. No aggressive warnings or bans occur for legitimate questions.
Such exaggerated claims undermine the credibility of your critique and give the impression that the content is more focused on sensationalism or defamation rather than fostering an informed and balanced discussion about AI technology. Constructive criticism should rely on facts, not distortions, to maintain its integrity."
Wasn't this kind of obvious?
Apparently most people lack basic ML knowledge
@@j.j.9538 Obvious? Even the authors of the paper admit their proposal is only potentially part of how something like the o1 model works.
helping an AI figure out math by breaking down its tasks is not even remotely the same as making an AI understand the world around it.
OpenAI is like a mouse compared to the elephant in AI living room: Google. Does that mean that Google will soon dominate? Not necessarily since Google needs to keep a low profile in order to avoid lawsuits, public scrutiny and antitrust issues.
OpenAI is not open source and that’s the problem. People are resorting to open source alternatives than closed sources and it is the right move to keep security and privacy in complete check.
No es necesario hackear a openAI todo esto que muestra es sentido común. Yo igual aplicaba el aprendizaje de refuerzo y hacer más independiente se autoevalue por si mismo, el auto tuning y réplicas mejorando tiempos respuesta y si tiene mejor peso está respuesta va dar para siguiente vez. El modelo cada vez irá mejorando y perfeccionando haciendo réplicas de si mismo.
Interesting approach, but it's very unlikely to be the case. Reward-based learning in operant conditioning works, because of the fact that the relevant subject has a sufficiently sophisticated nervous system, via which pain and pleasure can be administered. I don't see what type of reward or if a reward can even be relevant for pure digital code.
Whatever performance we'll see from solely code based, supposedly intelligent agents, is most certainly only a case of specific task-based training, e.g. fine tuning the model for a specific test, or simply blowing up the model without actually increasing complexity. Either way, as long as models remain purely non-sensory in nature, it's very unlikely they can actually develop something akin to human intelligence.
I mean from a purely mechanistic perspective you don't need rewards to be rewarding from the code's perspective but for the system as a whole. For example when you feel pain because of a bad decision you release chemicals and hormones to help avoid that in the future, this gives you subjective experience of punishment, but it's not like your neurons are getting tortured, rather the system is getting tortured (not really but you get the idea). So as long as you have enforced processes like a function error or score that needs to be minimized or maximized you don't need to care about rewarding or punishing the code but rather the emergent subjective intelligence that arises from that code.
You might be able to achieve this by adding noise to their thinking patterns or boost available compute, from a emergent intelligence of the system perspective, this would be the Pavlov's dog of macro reward and punishment, and then you have score and cost functions which would be internal to the model the equivalent of chemical releases like dopamine, adrenaline, cortisol,seratonin etc. So I can definitely see how we can implement both reward and punishment both external to the model and internal to the model.
In this way something like cost function actually I would say emulates very well a biological chemical discharge to reward and punish macro level behavior.
Or maybe I'm hella tripping I don't know xD
😂 people looovvving chinas ability to reverse engineer things now aint they.
I predict that we’ll get an open-source equivalent to o1 by about April of 2025, and this paper solidifies that prediction for me. 2025 is gonna be a wild year for A.I.
"DeepSeek has announced plans to release R1 as open source, though the specific licensing terms and release date have not yet been announced4. The model is currently available for free use on DeepSeek's chat platform, with a limit of 50 messages per day36.
Key points about DeepSeek R1:
It is designed to tackle complex reasoning tasks, similar to OpenAI's o1.
DeepSeek claims R1 outperforms o1-preview on benchmarks like AIME and MATH16.
Unlike o1, R1 displays its reasoning steps, making it more transparent4.
The model uses a "test-time compute" approach, similar to o1, which allows for deeper reasoning4.
DeepSeek plans to release both an API and the model weights as open source6.
While DeepSeek R1 shows promise as an open-source alternative to o1, it's important to note that the full release is still pending. Once released, it will offer developers and researchers the opportunity to modify and build upon the model, potentially accelerating advancements in AI reasoning capabilities24."
But what you didn't answer is what makes it "want" a reward why is a specific "reward" "rewarding" to the AI?............. wouldn't that indicate desire? what makes it or why does it desire or "want"
This process of building AGI is limiting not something that will lead to AGI.
Even calculator can compute faster than the best mathematician. It doesn't mean it is intelligent at all. Same with these LLMs, which are just statistics on huge amount of data. Relating any LLM (including o1) with AGi is quite funny.
Equation 1 should be 3p+e, not 3p+2
哈哈,美国人还活在上个世纪?中国现在每年专利申请量占全世界一半.日本文部科学省的报告:过去三年，在全球被引用次数前10%,前1%的自然科学领域论文数量，中国都是世界第一.来自澳大利亚战略政策研究所的报告,在44个下一代最关键科技领域,中国有37个是世界第一,美国只有7个,现在只是刚开始,未来中美的差距会越拉越大,中国只是回到历史上正常的位置,别忘记了,你们当年是靠着中国的指南针,水密舱,火枪火炮这些技术才控制美洲的
Goofy: There are endless combinatorial ways to architect learning paradigms, and there are endless learning paradigms. Backpropagation and attention have the spotlight now; however, anyone who understands the brain and the infinite ensembles of recurrent, Bayesian, Hopfield inspired, leveraging auto associative memory models, and on and on in ways that are being experimented with realizes drilling down on deep learning approaches is simply the first neural network to achieve stunning success. Let’s see the Chinese dissect massively parallel, self organizing systems that are so complex they are completely opaque to human understanding. This video is ridiculously naive. Stay tuned, AI is on the cusp of advancing itself exponentially. The Chinese have less than zero chance of catching up. They copy everything, let’s see them copy a system that morphs so fast it’s orthogonal to itself every few minutes.
they literally just copy everything..but that's not the main issue..but can be one too
🌍💥
@@NocheHughes-li5qe very weak response attempting to gather information by giving the wrong answer. hows does the saying goes? if you want something answered, don't ask a question, state a wrong statement and watch those that know how to answer it fight over the correct response.
@@o1-preview OK, GPT, we will make a slight change: try to rephrase that argument but this time without making it obvious that you are projecting.
"Researchers". Good one.
So what do you call them?
@@TheRealUsernameMen who play god, and know not what they do… 🤷♂️
@@TheRealUsernameOr mad-scientistic technocrat wumao’s maybe 🤔
More like "spies"
Capitalists that do not care about anything but their wallets.
Count the number of analogies
It's kind of like someone had made a bet to see how many they could fit into one video
Really?
I ask it quite a bit how it determine the answer and It's thought process when I get something wrong and it's told me things such as "overreliance on train data" or "Its assumptions". These questions were particularly relevant before you could easily ask it to traverse websites
Openai have unlimited supplies of NVIDIA GPUs, so no need to worry about algorithm optimization.
If China, Russia or Any other adversaries free the code even for undermining purposes we still benefit. The US is too controlled by corporations and the natives are getting restless.
China's Keyboard Immortals are doing their jobs surprisingly.
Code in not important in AI, the network and data are the cores.
o1? WTF is that? im still on 4o in my gpt chat. and when sam altman as talking about predictions for 25 he still was talking 4o.
Looking at the assessment as a whole, it's like asking a person a question and giving them an hour to answer and they can use the internet for research and references it does not make that person any smarter.
yes its like a giant glorified calculator, but, with the right thinking models it can get extremely intelligent. it depends whos working on it
Calling the language models AI is like calling a Squirrel a genius. They devise any reason from input, so anything wrong in the input is wrong in output.
Am I the only one who feels o1 is not that much better at most of the tasks? Me and my 4o have formed a great team, and o1 often felt like an outsider, LOL
hail to our new overlord rokos basilisk
Oh wow, Chinese researchers cracked OpenAI's AGI secrets? That’s incredible! I mean, who knew that a well-educated guess and a sprinkle of speculation could now count as cracking secrets? Honestly, with all the layers of sophisticated AI architecture, maybe they just hit Ctrl+F for 'AGI' and called it a day. Next time, I'll 'crack' the secrets of quantum computing by pointing at Schrödinger's cat and saying, 'It's probably in there somewhere.
The Chinese already cracked the secret of quantum computing way before Google released about their quantum computing chips.
@@vegamoonlight Oh, of course, the Chinese cracked the secret of quantum computing way before Google. I mean, who needs evidence when you’ve got rumors and a good headline? Forget peer-reviewed research or publicly available breakthroughs-it’s probably written on a scroll somewhere, right next to the recipe for immortality and the secret to cold fusion. But hey, let’s not let the lack of actual proof get in the way of a good conspiracy theory!
Im waiting for competing AI to "go to battle" with each other. Unfortunately i don't believe it will stop at other code.
It won’t stop. All governments will use AI for nefarious purposes and it will end badly for everyone.