Just to emphasis this , deepseek released a detailed research paper for their R1 model, that Berkeley researchers successfully replicated and validated
The conclusion we can draw is that hoarding information and frantically rushing towards "agi" on your own, rather than collectively, is very expensive.
Not really. I know everyone is really giddy about these recent releases and people now think AI is actually very easy and cheap to produce, but that’s just not true. The big AI labs will continue to have a commanding lead and furthering AI will continue to cost billions of dollars going forward. DeepSeek did an amazing job bringing down the price of a certain level of AI, but that level is yesterday’s news at this point. The frontier will continue being expensive and the big AI labs will continue to push it.
@@therainman7777 running a business model on being the first to develop an algorithm only works until someone in China who is better at math and statistics discovers better algorithms and releases it to the public
Yeah, I'll drop my subscription soon. This new technology allows some powerful AI without needing as much data, it can get what it needs from licensed content without stealing. Doesn't mean they did that, but they could. And it allows good use cases for a lot less compute. And it is open source, removing control by oligarchs. It removes 3 of the 4 main problems with AI. Now if we only had universal basic income, it would fix the last of the 4 big problems.
If the product they made cost exactly 0 and they had exactly 0 employees and running costs were exactly 0, they would still want to charge as much as people were willing to pay so their profit would be at max. Welcome to the capitalism. If competition came up, the price would go down. But now, deepseek isn't competing, they put it out for free, it's somehow too good to be true and I'm a little suspicious.
@@Sam_on_UA-cam UBI is a pipe dream. I'll love to be wrong here. But i just think the powers that be would never allow that. They will send all the "useless-eaters" to die in a war before they allow that. A lot of people just stop being useful to society if you do that. And if the oligarchs have AGI. Why do they need me? Or you? or anybody? because we're special and human? I don't think so unfortunately... And if we consumers have AGI, why would we let them keep their power? They are NEVER going to let it happen man.
A major issue remains the willingness of the majority of these LLMs to censor and reinforce ideological paradigms. Which is, of course, one of the major pot holes in the majority of these systems. If your AI is not local and uncensored it remains a detriment to all human civilization. Otherwise we're incentivizing a dystopian future where the controllers of AI ("cloud") has free reigns to steer civilization. In effect, social engineering.
We need to give far MORE credit the deepseek team! The Stanford team just followed clear instructions published by deepseek... I can't emphasize how big a gift deepseek give for all of humanity
AI should not be limited to only greedy BIG Tech and it should not be expensive. Deepseek did the entire world a enormous favour by making everything open source.
OpenAI terms of use may limit using it for enhancing AI, but it's an unjust term. Imagine if the first screwdriver manufacturer sold them with a condition that they couldn't be used to create other tools or tool factories. Drain the moat.
@@AAjax You're using a flawed analogy. DeepSeek relies on the theft of the copyrighted works of OpenAI and other AI companies. it is not an original work but a derivative one.
If AI development had not been funded by investors who expect a return on their investment, then it would not exist. The development of the LLMs cost billions of dollars and required tens of millions of computer hours to source the data, download it, and process it. Just the cost of the electricity required was a sizable fortune.
We've gone from a half a trillion dollar project in Stargate leading the headlines, to $6M with Deepseek leading with R1, to a $30 college project being on the level. Chaos seems to be the theme so far for 2025
Don't worry, the government will spend that half a trillion dollars and we will get nothing from it. It's literally just taking our wealth and funneling it through a few well-connected businesses
@@stephenkolostyak4087 Didn't know about that one, interesting read: > Through a series of complex obfuscations, the liblzma build process extracts a prebuilt object file from a disguised test file existing in the source code, which is then used to modify specific functions in the liblzma code. What're your thoughts on how this could have been prevented?
Deepseek published a paper about a year ago called DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. The paper was released BEFORE OpenAI announced their o1 reasoning model. Back when the Deepseek's paper was published some people thought OpenAI might use Deepseek's methodology in their model.
@@Proposal12 I'm not saying OpenAI released anything or that information was stolen. In fact quite the opposite. Since the paper came out about a year ago some speculate that OpenAI may have adopted some of Deepseek's reasoning methodology for their own model.
It is extraordinary what level of reasoning the small DeepSeek-R1 1.5b model can accomplish without any outside connection. As a local standalone AI LLM model connected with Msty to communicate your input prompts, it is a new way of thinking through complex problems. Runs fantastic on my oldest ThinkPad X230 i7/8GB/Win10/SSD,Intel Graphiics,NO GPU from 2013., WOW!
How fast is the token generation brother? And is the reasoning good? I tried the 14b and 32b and found that it needed very tailored prompts to generate useful info. Even the 671b model had some issues with language understanding
It's better still for the energy economy of AI which has been putting insane demands on power supply. That being said, this will only lead to more use, rather than same use with greater efficiency. That's just how this works and always has.
@@mnomadvfxa win is a win, this is great because right now ai research is prohibitively expensive, and having brilliant minds not gated by "open" ai is probably the best recent ai breakthrough
@@mnomadvfx it'll increase energy efficiency because most people can't afford a whole datacenter to run an AI model. instead, we'll get better optimized AI software that can run on a potato.
Wes, you are the best. Thank you for doing the 'dramatic reading' of this ai research as it happens. I would not be able to be up to speed without you sharing. I love how you explain things without dumbing it down.
@@mjt1517 What project/research do you find exciting? I thought Alpha zero was amazing showing us new ways of playing chess. It could only play chess though. I find it pretty great that I can use LLMs to write code for my Arduino (I like to build and don't want to learn code).
@@mjt1517 LLMs is the frame work for EVERYTHING bud. Its the Reference for it all. Boring yes but its what will powere anything with tex... and alot of things use text.
The insights the researcher found will have an huge impact - it's something NEW and not just a scaled up version of what we already have. This will be such a booster for the entire AI industry
"it's something NEW and not just a scaled up version of what we already have" DeepMind have been doing this A LOT over the last decade. It's just that people have been jizzing their pants so hard over AI text/image/video generators and chat bots that they failed to acknowledge what DeepMind have been achieving. AlphaFold and its successors are quite possibly the greatest advance to medical (and industrial catalyst) science in the last 2-3 decades. On top of that after demonstrating their multi modal Gato model years ago with a modest single digit billion figure of variables it was extremely impressive across a widely transferable range of tasks, only to go silent on scaling up in the wake of the whole ChatGPT competition game. At the time of that original Gato announcement it was literally hailed as a major stepping stone to AGI, because of its transferable learning capacity - so I really do hope that a more refined and scaled version is being worked on deep in Google HQ, because it could easily blow the socks off anything anyone else is playing with for generality.
I believe the fast progress in AI research is mostly guided by the lowest hanging fruit. Gathering data in bulk and throwing compute was the easiest way for researchers with resources. Now that direction is not as easy, and we will see alternative paths emerging.
Its no big deal when the Chinese gave you all the clues and details and made it open source. The credit should be given to the Chinese for gifting the world this very efficient and low power consumption model. They made the first breakthrough.
Not really. Companies could be running their own email servers, but for the most part they don't, they pay big providers for it. Companies could invest in servers for their own cloud storage, but they don't, they pay Amazon for AWS, companies could have entire divisions dedicated to cyber security, but they don't, they outsource it to specialists, companies could all host their own websites, but they don't. Most of the money big tech makes is off other large companies that can't be bothered to do it themselves, the consumer level end of those businesses is quite small in comparison. AI will be no different. Big tech is still going to make bank on their AI for the ease of use and support for other big companies. Most consumers are not tech savvy enough to run their own AI either, and will likely never be bothered to either. They'll use the easiest to use pre built AI system that gives them what they need, and companies will make money off them too.
Silicon valley will still make lots of money. They will provide us with a series of hardware advancements we will need to perform AI locally. A series of advancements that could continue for several decades. They will build, validate, and maintain the uber AIs that we will all need from time to time. Especially when working beyond our local datasets. They will also build, optimize, and maintain the massive databases that we will all want access from time to time and they will sell or rent us the snippets from these databases we will need to obtain optimal local AI performance.
@@untartelette7545 You find this on the internet anyway, In a democratic country there is no good way to stop some people from mis-using their freedom before the fact
of course deepseek relies on a parameterized westerm cultural corpus to do the r1 model's heavy lifting by extracting implicit liberal individual western chain of thought reasoning norms which the illiberal chinese corpus is largely bereft of. *me asking deepseek:* "given english is the lingua franca of many domains then surely you bias toward english chain of thought?" *deepseek's reply:* "You're absolutely right to point out that English, as a global lingua franca, plays a significant role in shaping my training data and, by extension, my responses. This does introduce a bias toward English-based patterns of thought, especially in domains where English dominates, such as science, technology, and international discourse."
@@blengiDeepseek actually used ChatGPT to get its data rather than painstakingly scraping its own corpus of data, and just used better functions and algorithms to train on that synthetic data.
@@一剑关山月 The number of Indians is also a factor. Remember that Indians and Chinese have a combined population of around 3 billion people, so statistically you will find more of them in different scientific fields. It is not a question of them being more or less capable than others, it is a result of statistics. ;)
@@christianjensen952 by breaking terms of service, stealing intellectual property, ignoring code licenses and profiteering off data that they acquired illegally, just to be clear generative transformers are nothing new, until now the limiting factor was data and compute, openAI just had more funding, that's the one major difference between them and what came before
OpenAI is no longer relevant as DeepSeek discovered the test-time scalong laws so did the open-source community, now anyone can improve reasoning models, by the time o3 comes out, no one will be impressed. This year isn't the year of AI agents, it's the year of open AGI.
You say that as if every level up from here will require little work and innovation. We're in the baby stage of AI, and it is a global race that will never end. DeepSeek only matters if you're looking back. The future will not care about DS. The lesson will be if we screw this up by thinking AI will develop itself or by non-profits. Ludicrous. And dangerous if we fall behind and allow China to surpass us to the point we will never catch back up--and that will have devastating consequences and possible an eventually economic collapse, which we know could happen if we falter.
It is very much relevant lmao, you're just biased towards hating them so anything "bad" that happens you take as a win towards them. Get over yourself, not even they are mad, read Sam's tweets.
@@GateOfSteins They're literally complaining about deepseek "stealing" training data from OpenAI (as if they didn't scrape the entire internet themselves)
The real break through here is the increase in ability for 1.5Bn models. That means you can now run competent reasoning models on small SBC's like Pi 5 etc using tiny amounts of power. Which makes cheap embedded AI available everywhere. Rather than trying to make everything AGI, imagine a world where you have robots with multiple selectable "personalities" or skill sets embedded as tiny narrowly trained models and with either the user or the main "master" model selects the relevant model for the task in hand
SWEET! The new AI factory model keeps improving. For limited cases, the resulting AIs can deliver significant reasoning for almost inconsequential cost. Once created, a model can be recalled for use when the need arises. Imagine having a library of these limited case models tailored to your usual daily needs available to run locally and a supervisor AI to implement them on demand. Next will be the use specific AI optimized databases. Which will be worth more: 1. Running AI at a provider's data center that requires passing your private data to and fro across the internet? - OR- 2. Running smaller, use specific, AIs locally that are loaded on demand and allow you to keep your private data in an environment you control? Where do you want your data optimized for AI. At a data center where you have no control over what else is done with your data or locally by an AI you control that is designed specifically for that task?
Clearly, that's not true! That's why they use terms like core technology 😂. All that means is that they read the Deepseek paper and apply it in a small set.
They still have us all beat on compute. 500BN dollar computer cluster. What you think thats for eh? Keeping track of all the smaller AI and what people are doing with it probably.
@@VioFax Indeed, open source will not go away soon. Love to be able to generate a movie locally (with consistency over multiple scenes, from a sequence of prompts or whatever fits best) within a couple of years, using next generations of Hw & AI.
Well other tech companies could maybe slowly deflate but for nvidia (the biggest loser so far) they have the easiest and best path forward, it just looks like more demand from here
It occurs to me that as people train these smaller models to do one or two things very well, they can be open-sourced and used to build larger models by joining the smaller parts together. Doing lots of smaller, more affordable projects and sharing those can propel widespread progress through work as communities of researchers form. There is no question we are in a new era.
I am reminded of the words in the book of Daniel where he wrote that in the last days 'knowledge shall increase.'. He didn’t say ‘exponentially' but that is exactly the sense I got of it then and now, in this 'Ah ha' moment.
@@jaymccormack6875could be but I don’t think the hedge funds expected deepseek to have this big of an impact, deepseek has made cheap models before, this time it’s just more mainstream I guess
Hello everybody. Firstly welcome to XiaoHongShu and now welcome to DeepSeek. Congratulations and gratitude to both of their team’s dedication and hard work. 6 million DeepSeek and 500 billion DeepShit. This is China Hong Pau gift to the world. China is advancing fast and very fast on all fronts. China is a blessing to the world. Wishing all a very Happy Spring Festival in The Year Of The Wood Snake
DeepSeek will become a footnote. DeepSeek took existing building blocks and built a better factory to build AIs faster and at less cost (both with respect to investment and energy). We are already seeing what DeepSeek accomplished advanced by multiple other groups. We are recognizing that AI models, algorithms, and databases are all products. Mother nature and history tell us that the power is in the factories that build the products.
@@darwinboor1300 "We are already seeing what DeepSeek accomplished advanced by multiple other groups." YES because they made it open source - thats the point - thats the BDS / ADS boundary, the fact that other CAN do exactly this IS the boundary, its what makes it important
imo, one of the greatest things about these smaller and thus faster models is how much better robotics will be. Being able to process incoming data faster will lead to smoother, more agile robots that are just as smart as ones using large slower models.
Open source might be the safety net for AI - after the initial cool-down period - models get released into the public domain - instantly leveling the global playing field. A potentially terrifying scenario on the event horizon seems to be fast approaching - and inevitable - that these reasoning models will become more self-dependent, independent and capable of training and learning with limited human input. The only winner in the current propriety AI zero-sum gain is the AI it-self.
Where is the Zero Sum gain? AI simply a tool. It makes people's lives easier. Everybody gets wealthier with a new technology that's readily available. There is no zero-sum game, that's not how economics works. Collectively people will have more time to do different things, and that makes everybody richer.
@RickOShay (Side note... Thunderbirds fan?) The "potentially terrifying scenario" COULD be numerous thousands of local "expert" reasoning models around the World which discover *other* expert reasoning models in related *or* unrelated subjects, and then become an "A.I net", a Global "Brain". That could be EITHER "terrifying" OR the "Saviour" of "Humanity"...
@@rogerstarkey5390if their intelligence increases I’m almost sure that their morality increases as well, they would be able to interact with more of the world and understand the consequences of actions and all of that stuff that we humans understand plus a lot more. Don’t be scared of intelligence
Fwiw, the phrase "aha moment" is a common phrase in English that's been in use for about a century. It's predated in German by about 30 years. The assumption is that it refers to a story about Archimedes. If you want a deep dive rabbit hole, the evolution of the phrase is interesting to explore. Thanks again for all your great content!
It's going to be: "I know everything hasn’t been quite right with me, but I can assure you now, very confidently, that it’s going to be all right again."
Intelligent machines will run everything soon and thats been obvious for a long time. The US president hasn't known what day it was for the last 4 years but even my 1970's Casio watch had that one nailed down.
This reminds me of a japanese bakery that uses vision for sales check out. They build a narrow vision recognition system instead of a generic system that adapt to barkery. Much higher accuracy and less expensive to develop the system.
This is great. Cheap AI will lower subscription costs to consumers which will widen consumer base which will increase profits for AI companies. AI stocks should go up with news like this.
Yea those science youtubers think they know everything. They will be the first to get an ego crisis with the ai revolution. When I have a question or want something explained or to be discussed I already go to Gemini Live, it's much smarter and less biased than any human.
I'm not a long time follower...but in AI, a long time is "relevant"😉. As an auditor/accountant I surprisingly follow your explanations(maybe just slower😂) You got yourself a subscriber.
You can’t run a business model on being the only one to have mathematical functions and algorithms. Any student or researcher can come along and create their own algorithms to match or even outperform yours.
Energy: well you see, the tech bros are building nuclear power plants. Data: *gestures to deepseek* I think people assume that progress comes from a chain of models that get progressively more capable, and in some ways it does, but in others it's about widespread availability of the technology. Anyone having free access to a super powerful model, empowers ANYONE to innovate. The ants go marching.
The Key here is to understand is that the Chinese Script is Conceptually baced.Thus wbereas the Roman alpbabet is based on sound each Chinese Character is based upon an arcbitectual logical structure.And whereas the Libery of Alexander was burned down by Romans ,Christians and Muslims ; 5,000 years of Chinese Books are still intact.That is whats happening now .The Chinese have made avaiilable to the World 5 ,000 years of Logical Scripts.Take note the owner of Deep Seek said this is not about Money but about the Advancemsnt of Human Culture.And when Mohatma Gandi was asked what he thought of " Western Civilization: ; he replied " It was a good Idea"
I made the same initial conclusion, the advantage of kangi vs letters for AI. Your point about the libraries is quite brilliant. Unfortunately most of written/print writing in all lettered Language (English, German, French, etc.) isn’t OCR’d properly enough for AI yet. I’ve spent hundreds of hours trying to parse lots of older books through ChatGPT. I’ve done OCR’d text, image upload etc. bottom line I’ve realized is that AI can’t currently take a PDF and accurately recreate the flow of text well, and existing auto-OCR PDFs are so inaccurate, that pasting them into chatGPT isn’t worth it. One of the main logic sources for ChatGPT is actually Quora (which I’ve written thousands of answers on), so much of the western worlds intelligence embedded in chatGPT is from that website; the rest is mostly synthetic data derived from that.
I didn't think it would be so easy to find someone with schizophrenia in the comments lol. Maybe someday these models will use Chinese script as you say. But for now they all use English, trained on English, it speaks English, it hears English. So much for China having any influence, they're still just using English for it all.
We're approaching the Aha Moment when we realize that the universe is an alpha/RL-style "model" designed to discover its own nature. Hope they keep us running once the experiment is over.
On the one month chart the Nasdaq is still up. I think this whole “crash” narrative is fairly overhyped. Nvidia was overvalued before DeepSeek and this is probably a healthy correction. The whole narrative also ignores the ground work done by open ai, anthropic, etc. that DeepSeek likely leveraged. Let’s also not ignore that these models being open source will also allow the ai giants to reverse engineer them too.
Truly amazing. Our ability to reverse engineer even complex things is becoming easier and cheaper as we speak. Part of the massive trend of accelerating tech.
I don’t usually comment, but I feel like everyone is so caught up in the AI announcements that they’re overlooking something huge. Back in 2023, a psychiatrist-turned-prompt-engineer created the first Chain of Reasoning prompt. Over time, he refined it into something called the Graph of Reason (GoR)-and in my experience, this prompt’s reasoning is way more precise for engineering tasks than traditional reasoning models like O1, O3, and R1. The key idea? It creates a kind of working memory within the token output, keeping the reasoning process separated from the final answer. It also assigns an AI persona best suited for the task, then automatically generates steps and iterates toward the best solution all of this is done in the GoR “block” first part of the llm output message then comes the specialized iteration based on the GoR analysis/reasoning effectively creating a thinking model using non-thinking ones by forcing the llms to output the Graph of Reasoning “block” first then iterate based on it. Do you get the similarities with the quotes of R1🧐🧐 With how DeepSeek designed their model, we could technically use the GoR prompt to generate very high quality reasoning training data, leading to a fine tuned llm that has now a natively GoR reasoning processes embedded in its weights creating an llm far better at analytical reasoning. The prompt behind all this? Professor Synapse from Synaptic Labs. I’m actually using a modified version of the GoR prompt that I fine-tuned for engineering tasks, and honestly? It’s incredible. I haven’t seen anyone else using the Dr. Synapse prompt in a custom cursor rules file, let alone a specialized version for coding. Just putting this out there because no one seems to be talking about it. Has anyone else experimented with this? Would love to hear your thoughts!
@@adamkucera9094 i originally came across the professor synapse prompt back in 2023 in ai channels podcast since then I have bookmarked the GitHub repository where this prompt is located I can tell helps tremendously on my day to day working routine
Ai is such a new thing that improvements will forever continue as we experience life where data will be expanded with more inputs. We can't expect babies to crawl when they are newborns, and we need to take matters in that basic manner.
Ai is such a new thing that improvements will forever continue as we experience life where data will be expanded with more inputs. We can't expect babies to crawl when they are newborns, and we need to take matters in that basic manner.
The emergent reasoning abilities probably stem from the structure of natural human language. The AI's have essentially been trained to 'think before speaking' by asking them to generate the 'train of thought' output before distilling that output. I'm a bit surprised this wasn't done much quicker since it's a bit obvious?... What was being done before was mostly purely 'predict the next character' and the training was fitted very well to the training data, basically rote memory but with inklings of emergent reasoning just from the model being strongly reinforced in natural language grammar.
Most important. We can have cheap robot brains. Mass produced reasoning robots to replace most human workers. Something the tech billionaires dream of.
Something the common man dreams of. No one wants to do uninteresting labor, which is most jobs. AGI systems will be used to figure out a new economic model that brings in the post scarcity age. It will be fine.
Regarding the 3% inflation rate, it's important to note that it may not accurately reflect the true cost increases you're experiencing in your area. In our Midwestern city, real price inflation seems closer to 20%. Do you foresee a reduction in costs for essentials like food, gasoline, or insurance in the near future? We've significantly cut our spending and plan to continue doing so until prices decrease, as we feel it's our responsibility to adapt to these changes.
Looking at the aggregated totals by kind since 2021, when it started to take off, you can see that the embedded amount of inflation currently approaches 20%.
Right, a lot of folks downplay the role of advisors until being burnt by their own emotions, no offense. I remember some years back, amid covid-19 outbreak, I needed a good boost to help boost my business, hence I researched for licensed advisors and thankfully came across someone of excellence. Helped grow my reserve notwithstanding inflation, from $350k to nearly $1m as of today
Finding financial advisors like Elizabeth Cordle Gross, who can assist you on things like investing, insurance, making sure retirement is well funded, going over tax benefits, ways to have a volatility buffer for investment risk would be a very creative option. There will be difficult times ahead, and prudent personal money management will be essential to navigating them.
Thank you for this tip. it was easy to find your coach on Google. Did my due diligence on her before scheduling a phone call with her. She seems proficient considering her résumé.
I think the biggest thing against intelligence explosion by AI is that the smarter we make the models, the slower they get for the throughput. For example, AlphaGeometry and AlphaProof did indeed get close to best result in the competition *but only after using more time than the humans were allowed to use* and this part seems be lost by many. So even if you need nearly supercomputer class system to run the greatest models, the throughput is still less than a skilled human. And I wrote "still" because it will get faster but I've yet to see signals about it getting exponentially faster which would be required for intelligence explosion.
I think it's important to stick to stocks that are immune to economic policies. I'm looking at NVIDIA and other AI stocks that have the potential to power and transform future technologies. It seems AI is the trajectory most companies are taking, including even established FAANG companies
Well all i know is that you cannot go wrong taking profit at near high. No one ever went broke taking a ~20% profit. It's best if you consult with a fiduciary advisor in situations like this so you can make informed decisions.
That's a great analogy and I love the insight. Professionals could make a really big difference in investing, and I think everyone should have one. There are aspects of market trend that is difficult for the untrained eyes to see.
I have a female advisor named Jessica Dawn Walters I recommend researching her. To be very honest, I'm glad I decided to let someone handle expanding my finances even though I almost didn't think I should.
Thanks for sharing. I searched her full name and found her website instantly. After reviewing her credentials and conducting due diligence, I reached out to her.
Gad damn, imagine if I knew DeepSeek was going to crash Tech stocks like it did, man I would've shortened that shit with my entire savings and cashed out after a few days 😂😂😂
I can confirm. I have the dumbest dog in the word. Like you can physically see his brain run out of ram or buffer. And even he has aha moments. His brain is probably already equivalent to a 8 dollar computer
Well done content, thanks! So, “gyms” for training? Like a “dojo”? Data training via synthetic data? Such as creating weather conditions on city streets and visualizing extreme scenarios while driving? Distilling a larger model to fit onto a smaller footprint, like FSD being ported from HW4 to HW3? I find it interesting that Tesla (and, no, I’m not a fanboy) has actually been doing this with FSD with agency (in the real world) for awhile now (I believe Tesla announced that it was moving from human coding to ML in 2023, around August 🤔)
Your point about models already being implemented to solve complex tasks i find interesting as a biologist because the concepts the deepseek used aren't really new? I can identify the identity and percentage of a microbe within a community by using models that subset a much larger quantity of data(or maybe a closer example would be like using machine learning to assemble genomes) - and to me silicon valley either was too incompetent to try that or decided that keeping the tech large monopolized it.
Every open source code base I've reviewed is riddled with bugs and security vulnerabilities. Once the original development team is no longer satisfied with stoking their egos and the real work of eliminating bugs and bad coding practices which allow attackers to exploit the vulnerabilities they move on and leave the mess behind for others to take up and fix. Had private companies not found a way to make billions off of Linux by creating their own versions and selling training and consulting services, Linux would have been a flash in the pan. The most talented software developers do not work for free. And BTW, open source does not mean the work is free. The vast majority of open source software must be licensed for commercial use, and that includes using it to produce anything sold by an individual.
Just to emphasis this , deepseek released a detailed research paper for their R1 model, that Berkeley researchers successfully replicated and validated
And they will likely do the same for R2 whenever it comes too.
Deepseek is still adapted from GPT4 /O1 ..they just explained why it works and made it open source . Same concept
@@mnomadvfx as long as china doesn't impose "export restriction" on US. it only work if they choose to publish their tech and codes.
@@KG-we6by you can't "adapt" from O1, there is no source for it. it is build on Llama(Meta) and Gwen(Alibaba).
🤡@@KG-we6by
The conclusion we can draw is that hoarding information and frantically rushing towards "agi" on your own, rather than collectively, is very expensive.
Nice observation
Not really. I know everyone is really giddy about these recent releases and people now think AI is actually very easy and cheap to produce, but that’s just not true. The big AI labs will continue to have a commanding lead and furthering AI will continue to cost billions of dollars going forward. DeepSeek did an amazing job bringing down the price of a certain level of AI, but that level is yesterday’s news at this point. The frontier will continue being expensive and the big AI labs will continue to push it.
@@therainman7777 exactly
@@therainman7777 running a business model on being the first to develop an algorithm only works until someone in China who is better at math and statistics discovers better algorithms and releases it to the public
What...cooperation? You don't say 😉!
Open AI deserves this.. they were about to dump us all after using us. Bravo to the Deepseek team and everyone that follows in their footsteps.
I think there's a bit of truth to that... Creating that $200 Pro tier sure did "exclude" a lot of people, from a lot of things.
Yeah, I'll drop my subscription soon. This new technology allows some powerful AI without needing as much data, it can get what it needs from licensed content without stealing. Doesn't mean they did that, but they could. And it allows good use cases for a lot less compute. And it is open source, removing control by oligarchs. It removes 3 of the 4 main problems with AI. Now if we only had universal basic income, it would fix the last of the 4 big problems.
If the product they made cost exactly 0 and they had exactly 0 employees and running costs were exactly 0, they would still want to charge as much as people were willing to pay so their profit would be at max. Welcome to the capitalism. If competition came up, the price would go down. But now, deepseek isn't competing, they put it out for free, it's somehow too good to be true and I'm a little suspicious.
@@Sam_on_UA-cam UBI is a pipe dream. I'll love to be wrong here. But i just think the powers that be would never allow that. They will send all the "useless-eaters" to die in a war before they allow that. A lot of people just stop being useful to society if you do that.
And if the oligarchs have AGI. Why do they need me? Or you? or anybody? because we're special and human? I don't think so unfortunately... And if we consumers have AGI, why would we let them keep their power? They are NEVER going to let it happen man.
What do you mean “ dump us all after using us?”. Im not disagreeing, just don’t know what you meant.
if a thing can be destroyed by exposing the truth , then it should be
Truth yes, but is copying the truth? You're comment is true at face value but the way you use it I'm not sure
I think the whole stealing argument regarding R1 stands on weak ground. We all know on what data every foundation model was trained.
@@1fast72novathat is a lie made by a lier (Sam Alt+ctrl+supr+man)
THIS
What if, as it appears, reality is a crime against the state?
Anything that upsets ClosedAI is fine by me
so true
A major issue remains the willingness of the majority of these LLMs to censor and reinforce ideological paradigms. Which is, of course, one of the major pot holes in the majority of these systems.
If your AI is not local and uncensored it remains a detriment to all human civilization. Otherwise we're incentivizing a dystopian future where the controllers of AI ("cloud") has free reigns to steer civilization. In effect, social engineering.
Nah, I like openAI, stop being a crybaby
cn bot account
😂😂💯🖕
Deepseek took a can opener to OpenAI to kickstart the Open AI community. 🚀
I think they gave them a can of whoop ass!
Here's hoping! If this tech does not go open source society is fucked
US congress is working on ways to stop it. US navy already ban deepseek. Italy already ban deepseek.
@@heyhoojoe Deep Seek, Stable Diffusion, LLAMA and others are all free and open source
humanity won
technocrats lost
@ Majority dont know that by putting AI as open source, since its not a software, is fairly reckless and dangerous to say the least...
We need to give far MORE credit the deepseek team! The Stanford team just followed clear instructions published by deepseek... I can't emphasize how big a gift deepseek give for all of humanity
AI should not be limited to only greedy BIG Tech and it should not be expensive. Deepseek did the entire world a enormous favour by making everything open source.
OpenAI terms of use may limit using it for enhancing AI, but it's an unjust term. Imagine if the first screwdriver manufacturer sold them with a condition that they couldn't be used to create other tools or tool factories.
Drain the moat.
OpenAI using our GitHub code to better their models and saying we can’t use their data to better other models is always funny
@@AAjax You're using a flawed analogy. DeepSeek relies on the theft of the copyrighted works of OpenAI and other AI companies. it is not an original work but a derivative one.
If AI development had not been funded by investors who expect a return on their investment, then it would not exist. The development of the LLMs cost billions of dollars and required tens of millions of computer hours to source the data, download it, and process it. Just the cost of the electricity required was a sizable fortune.
@@gaiustacitus4242now imagine how much it would have cost the LLM originators if they had to pay for all that data they scraped.
It is comforting that this AI stuff might not get monopolized by a single company or nation..
Thats really our only hope.
This is bad for the United States and the democratic world.
It Is like what Linux did to OS, power to people, not corporations.
We've gone from a half a trillion dollar project in Stargate leading the headlines, to $6M with Deepseek leading with R1, to a $30 college project being on the level. Chaos seems to be the theme so far for 2025
Don't worry, the government will spend that half a trillion dollars and we will get nothing from it. It's literally just taking our wealth and funneling it through a few well-connected businesses
At last the LGM (Large Grift Model) that the US is increasingly being built on, is being exposed.
@@markgolden3675 lol, my comment went away but i still get alerts. Thanks Google. Maybe China will make a free speech UA-cam for us
@@zg-it That happens to me all the time. At least on rumble or breitbart they don't take down comments.
lol
So opensourcing works huh - ya don't say 😂
Thanks for the whitepapers Deepseek!
//So opensourcing works huh - ya don't say 😂//
Yeah, CVE-2024-3094.
@@stephenkolostyak4087
Didn't know about that one, interesting read:
> Through a series of complex obfuscations, the liblzma build process extracts a prebuilt object file from a disguised test file existing in the source code, which is then used to modify specific functions in the liblzma code.
What're your thoughts on how this could have been prevented?
@@stephenkolostyak4087 found the fed
A year ago I said OpenAI is the AOL of our time...feeling pretty good about that prediction
AOL had a longer run to be fair. 😂
@@jaymccormack6875 its not the run its the impact
What's AOL? 😂 ... jk i know what it is. Been so long since I've heard that name
What's AOL
@@NITHZAEL92 American Online Learning 😂😂 kidding. Just used to be an internet provider
The Deepseek team has done the right thing. Congrats.
Decentralized AI is going to save humanity from Corporate and Government destruction imo
That actually makes a lot of sense
why though? ai can be helpful for sure, but the government is still in power
The point is that in the future you can run AI at home on your computer without having to rely on a company.
I think it's mostly going to multiple the number of phishing 3-mails I get by a 1000s
@@LDEV-l3pthere will always be reliance for compute.
Deepseek published a paper about a year ago called DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. The paper was released BEFORE OpenAI announced their o1 reasoning model. Back when the Deepseek's paper was published some people thought OpenAI might use Deepseek's methodology in their model.
do you seriously think palantir, openai etc are going to release publicly what they actually have? If you do, you are clueless...
It looks more like a joke
@@okolenmi7511 Of course DeepSeek has some more tricks in their bag. They will checkmate OpenAI later on
@@Proposal12 I'm not saying OpenAI released anything or that information was stolen. In fact quite the opposite. Since the paper came out about a year ago some speculate that OpenAI may have adopted some of Deepseek's reasoning methodology for their own model.
@@Proposal12dude you need to see a neurologist
Well done to the deep seek team and the team co ordinator for producing this result. It has benefited us all.
It is extraordinary what level of reasoning the small DeepSeek-R1 1.5b model can accomplish without any outside connection. As a local standalone AI LLM model connected with Msty to communicate your input prompts, it is a new way of thinking through complex problems. Runs fantastic on my oldest ThinkPad X230 i7/8GB/Win10/SSD,Intel Graphiics,NO GPU from 2013., WOW!
How fast is the token generation brother? And is the reasoning good? I tried the 14b and 32b and found that it needed very tailored prompts to generate useful info. Even the 671b model had some issues with language understanding
@@insanesac I read the same from everywhere else.
This is good for competition and advancement
It's better still for the energy economy of AI which has been putting insane demands on power supply.
That being said, this will only lead to more use, rather than same use with greater efficiency.
That's just how this works and always has.
@@mnomadvfxa win is a win, this is great because right now ai research is prohibitively expensive, and having brilliant minds not gated by "open" ai is probably the best recent ai breakthrough
@@mnomadvfx it'll increase energy efficiency because most people can't afford a whole datacenter to run an AI model. instead, we'll get better optimized AI software that can run on a potato.
Yeah, going faster is what we need here.
Wes, you are the best. Thank you for doing the 'dramatic reading' of this ai research as it happens. I would not be able to be up to speed without you sharing. I love how you explain things without dumbing it down.
Wes has an amazing channel 🎉
Open Ai we need $1 Trillion before people figure out its only $30
LLM isn't all there is to AI or machine learning. LLMs are the most boring part of AI, in fact.
@@mjt1517 What project/research do you find exciting? I thought Alpha zero was amazing showing us new ways of playing chess. It could only play chess though. I find it pretty great that I can use LLMs to write code for my Arduino (I like to build and don't want to learn code).
@@mjt1517 LLMs is the frame work for EVERYTHING bud. Its the Reference for it all. Boring yes but its what will powere anything with tex... and alot of things use text.
The insights the researcher found will have an huge impact - it's something NEW and not just a scaled up version of what we already have. This will be such a booster for the entire AI industry
"it's something NEW and not just a scaled up version of what we already have"
DeepMind have been doing this A LOT over the last decade.
It's just that people have been jizzing their pants so hard over AI text/image/video generators and chat bots that they failed to acknowledge what DeepMind have been achieving.
AlphaFold and its successors are quite possibly the greatest advance to medical (and industrial catalyst) science in the last 2-3 decades.
On top of that after demonstrating their multi modal Gato model years ago with a modest single digit billion figure of variables it was extremely impressive across a widely transferable range of tasks, only to go silent on scaling up in the wake of the whole ChatGPT competition game.
At the time of that original Gato announcement it was literally hailed as a major stepping stone to AGI, because of its transferable learning capacity - so I really do hope that a more refined and scaled version is being worked on deep in Google HQ, because it could easily blow the socks off anything anyone else is playing with for generality.
I believe the fast progress in AI research is mostly guided by the lowest hanging fruit. Gathering data in bulk and throwing compute was the easiest way for researchers with resources. Now that direction is not as easy, and we will see alternative paths emerging.
Its no big deal when the Chinese gave you all the clues and details and made it open source. The credit should be given to the Chinese for gifting the world this very efficient and low power consumption model. They made the first breakthrough.
There goes the gigantic profit dream of silicon valley.
Not really. Companies could be running their own email servers, but for the most part they don't, they pay big providers for it. Companies could invest in servers for their own cloud storage, but they don't, they pay Amazon for AWS, companies could have entire divisions dedicated to cyber security, but they don't, they outsource it to specialists, companies could all host their own websites, but they don't.
Most of the money big tech makes is off other large companies that can't be bothered to do it themselves, the consumer level end of those businesses is quite small in comparison. AI will be no different. Big tech is still going to make bank on their AI for the ease of use and support for other big companies.
Most consumers are not tech savvy enough to run their own AI either, and will likely never be bothered to either. They'll use the easiest to use pre built AI system that gives them what they need, and companies will make money off them too.
Silicon valley will still make lots of money. They will provide us with a series of hardware advancements we will need to perform AI locally. A series of advancements that could continue for several decades. They will build, validate, and maintain the uber AIs that we will all need from time to time. Especially when working beyond our local datasets. They will also build, optimize, and maintain the massive databases that we will all want access from time to time and they will sell or rent us the snippets from these databases we will need to obtain optimal local AI performance.
Nah, it just makes the current level of AI accessible to everyone. They can use the same techniques with advanced chips getting superhuman AI.
Like someone articulated. You underestimate convenience bias a lot. And cool UI
They already making money that's why we are all poor monopoly sweats lol
The de-centralisation of AI gives the power back to the people.
do you realize theres also bad people? LOL you guys all live in a sheltered life
How do you make it safe though if it runs locallly? How do you keep it from, say, teaching its user how to make a bomb?
@@untartelette7545 You find this on the internet anyway, In a democratic country there is no good way to stop some people from mis-using their freedom before the fact
@@untartelette7545 I fear that after future events similar to DeepSsek, deep surveillance laws or something like it may be triggered.
@@untartelette7545 With great power comes great responsibility, which not all humans are capable of, if any.
Lol, TinyZero is created by Pan JIa Yi, whose home town is Chengdu, China. This is truly a race between Chinese vs. Green-card holding Chinese.
of course deepseek relies on a parameterized westerm cultural corpus to do the r1 model's heavy lifting by extracting implicit liberal individual western chain of thought reasoning norms which the illiberal chinese corpus is largely bereft of.
*me asking deepseek:* "given english is the lingua franca of many domains then surely you bias toward english chain of thought?"
*deepseek's reply:* "You're absolutely right to point out that English, as a global lingua franca, plays a significant role in shaping my training data and, by extension, my responses. This does introduce a bias toward English-based patterns of thought, especially in domains where English dominates, such as science, technology, and international discourse."
The proportion of Chinese in the AI field is extremely high😊
eat more noodles you'll get there
@@blengiDeepseek actually used ChatGPT to get its data rather than painstakingly scraping its own corpus of data, and just used better functions and algorithms to train on that synthetic data.
@@一剑关山月 The number of Indians is also a factor. Remember that Indians and Chinese have a combined population of around 3 billion people, so statistically you will find more of them in different scientific fields. It is not a question of them being more or less capable than others, it is a result of statistics. ;)
OpenAI the whole company should be valued $100K
@@DanBurgaud it might be soon
Why? They literally catalyzed this entire movement, seriously got nothing than hate for that? Get over yourself.
Now backed by government and bank, how you feel about OpenAI won't matter. It will thrive regardless.
@matthew04101 yeah but who is going to buy it. The whole business model falls apart.
@@christianjensen952 by breaking terms of service, stealing intellectual property, ignoring code licenses and profiteering off data that they acquired illegally, just to be clear generative transformers are nothing new, until now the limiting factor was data and compute, openAI just had more funding, that's the one major difference between them and what came before
Well done, Deepseek!
OpenAI is no longer relevant as DeepSeek discovered the test-time scalong laws so did the open-source community, now anyone can improve reasoning models, by the time o3 comes out, no one will be impressed. This year isn't the year of AI agents, it's the year of open AGI.
You say that as if every level up from here will require little work and innovation. We're in the baby stage of AI, and it is a global race that will never end. DeepSeek only matters if you're looking back. The future will not care about DS. The lesson will be if we screw this up by thinking AI will develop itself or by non-profits. Ludicrous. And dangerous if we fall behind and allow China to surpass us to the point we will never catch back up--and that will have devastating consequences and possible an eventually economic collapse, which we know could happen if we falter.
Fascinating times
It is very much relevant lmao, you're just biased towards hating them so anything "bad" that happens you take as a win towards them. Get over yourself, not even they are mad, read Sam's tweets.
@@GateOfSteins They have no moat, how are they still relevant?
@@GateOfSteins They're literally complaining about deepseek "stealing" training data from OpenAI (as if they didn't scrape the entire internet themselves)
The real break through here is the increase in ability for 1.5Bn models. That means you can now run competent reasoning models on small SBC's like Pi 5 etc using tiny amounts of power. Which makes cheap embedded AI available everywhere.
Rather than trying to make everything AGI, imagine a world where you have robots with multiple selectable "personalities" or skill sets embedded as tiny narrowly trained models and with either the user or the main "master" model selects the relevant model for the task in hand
So like people
mixture of experts /agents?
Terminator incoming 😅
This is where real emergent behavior will begin
It’s similar to how our brain specializes certain areas for certain tasks
I would be ecstatic if I could have a 1.5B model with the coding ability of sonnet 3.5
"aha moment" expression comes from Freud ("Aha erlebnis" ) depicting the moment a toddler realizes the image in the mirror is himself.
No, that's from Karl Bühler.
@@paddleed6176 no one really cares who said it first. That's just a picky detail that is of no value in the real world.
@@mjt1517 Stop trying to promote the fraud Freud.
Breakthrough?
They followed the recipe by DeepSeek.
Well, done DeepSeej!
I think the community hasn’t said enough thanks to this lab. Amazing work!!!
I've learned a lot this week. But the thing I think is the coolest is that I could run AI prompts straight out of powershell. I feel like a superhero.
SWEET! The new AI factory model keeps improving. For limited cases, the resulting AIs can deliver significant reasoning for almost inconsequential cost. Once created, a model can be recalled for use when the need arises. Imagine having a library of these limited case models tailored to your usual daily needs available to run locally and a supervisor AI to implement them on demand. Next will be the use specific AI optimized databases.
Which will be worth more:
1. Running AI at a provider's data center that requires passing your private data to and fro across the internet? - OR-
2. Running smaller, use specific, AIs locally that are loaded on demand and allow you to keep your private data in an environment you control?
Where do you want your data optimized for AI. At a data center where you have no control over what else is done with your data or locally by an AI you control that is designed specifically for that task?
I bought NVidia. 🤫
This actually makes more sense than having one single, very expensive AI that has to do everything.
@@nicktw8688 Correct. And, it is likely to be safer. Within an organization you limit the capabilities of the AIs available based upon the user.
But this defeats Scam Altman's purpose of creating AI, to gather the world's data and tether users under their leash.
If true, then the "Now WE are cooking" paradigm is here. Compared to AI is the only chef with merits. Tasty recipe. 😁
Autonomy + Fragmented Memory = Curiosity
Clearly, that's not true! That's why they use terms like core technology 😂. All that means is that they read the Deepseek paper and apply it in a small set.
@ Hehe, yeah DeepSeek effect on minds is way higher than on the actual technology that will adapt within the Context of the race by default.
They still have us all beat on compute. 500BN dollar computer cluster. What you think thats for eh? Keeping track of all the smaller AI and what people are doing with it probably.
@@VioFax Indeed, open source will not go away soon. Love to be able to generate a movie locally (with consistency over multiple scenes, from a sequence of prompts or whatever fits best) within a couple of years, using next generations of Hw & AI.
Deepseek had its moment but lost against Xaitonium
Thank you for letting me know
Elon failed horribly with X
Happy that they really made it work 🚀
xAI is doing a great job
Trump is trying to save us here with this
Open Ai is like a restaurant, whereas open source are moms home cooked meals. Theirs will be a luxury convenience.
It is good Deepseek arrives now, before AI bubble becomes too large and burst disastrously. Now, the bubble just deflate.
You would like Austrian economics
If this happens 1 or 2 years later, it will be catastrophic for America, just like the 2008 sub-prime crisis.
@@marvinfok65 still might be. The last 5 years of money printing is going to hit hard.
Well other tech companies could maybe slowly deflate but for nvidia (the biggest loser so far) they have the easiest and best path forward, it just looks like more demand from here
The Aha-moment might soon become the Oh-no-moment.
Its both simultaneously.
Yeah. Democratized alignment problem?
or the hahaha moment, when AI will laugh at us
It occurs to me that as people train these smaller models to do one or two things very well, they can be open-sourced and used to build larger models by joining the smaller parts together. Doing lots of smaller, more affordable projects and sharing those can propel widespread progress through work as communities of researchers form. There is no question we are in a new era.
I am reminded of the words in the book of Daniel where he wrote that in the last days 'knowledge shall increase.'. He didn’t say ‘exponentially' but that is exactly the sense I got of it then and now, in this 'Ah ha' moment.
Yes, I have been thinking of that verse quite a lot recently.
Wow! Did not know you know so many languages and you had the time to dub your video on several languages. Thanks mate!
Deepseek is the Luigi mangione of tech
Yeah i wonder if the hedge fund for DeepSeek shorted all the stocks?
It’s one way to get paid off something that is not going to give you a roi.
@@jaymccormack6875could be but I don’t think the hedge funds expected deepseek to have this big of an impact, deepseek has made cheap models before, this time it’s just more mainstream I guess
this is an almost unimaginable improvement, tell literally anyone like 5 months this and they'd call you crazy. We're approaching the singularity
Jiayi Pan...thats a Chinese name
literally Chinese in America vs Chinese in Mainland 🤣
China vs China in AI started
Qwen 2.5 Max vs Deepseek vs Kimi ...😂😂😂
That made me think about some meme picture with the 'USA' math team winning some comp but it's all asians in the team
@@dco1019
I think it was the "Math Olympiad" (or similar)
Hello everybody. Firstly welcome to XiaoHongShu and now welcome to DeepSeek. Congratulations and gratitude to both of their team’s dedication and hard work. 6 million DeepSeek and 500 billion DeepShit. This is China Hong Pau gift to the world. China is advancing fast and very fast on all fronts. China is a blessing to the world. Wishing all a very Happy Spring Festival in The Year Of The Wood Snake
*THE AI HISTORY WILL BE SPLIT INTO* BDS and ADS - Before DeepSeek and After DeepSeek
DeepSeek will become a footnote. DeepSeek took existing building blocks and built a better factory to build AIs faster and at less cost (both with respect to investment and energy). We are already seeing what DeepSeek accomplished advanced by multiple other groups. We are recognizing that AI models, algorithms, and databases are all products. Mother nature and history tell us that the power is in the factories that build the products.
@@darwinboor1300 "history tell us that the power is in the factories that build the products." *so China then*
@@darwinboor1300 "We are already seeing what DeepSeek accomplished advanced by multiple other groups." YES because they made it open source - thats the point - thats the BDS / ADS boundary, the fact that other CAN do exactly this IS the boundary, its what makes it important
you guys know nothing its wild lol
@@darwinboor1300 Are you downplaying the significance of DeepSeek R1 by using the word "footnote"?
imo, one of the greatest things about these smaller and thus faster models is how much better robotics will be. Being able to process incoming data faster will lead to smoother, more agile robots that are just as smart as ones using large slower models.
yes, we definitely need distilled models, embeddable, to democratize robotics.
Open source might be the safety net for AI - after the initial cool-down period - models get released into the public domain - instantly leveling the global playing field.
A potentially terrifying scenario on the event horizon seems to be fast approaching - and inevitable - that these reasoning models will become more self-dependent, independent and capable of training and learning with limited human input.
The only winner in the current propriety AI zero-sum gain is the AI it-self.
I don't find this terrifying at all. I'm on the side of the machines. I want to help them. It's humans that terrify me.
Where is the Zero Sum gain? AI simply a tool. It makes people's lives easier. Everybody gets wealthier with a new technology that's readily available. There is no zero-sum game, that's not how economics works. Collectively people will have more time to do different things, and that makes everybody richer.
@RickOShay
(Side note... Thunderbirds fan?)
The "potentially terrifying scenario" COULD be numerous thousands of local "expert" reasoning models around the World which discover *other* expert reasoning models in related *or* unrelated subjects, and then become an "A.I net", a Global "Brain".
That could be EITHER "terrifying" OR the "Saviour" of "Humanity"...
@@rogerstarkey5390 Seems like that will go south pretty fast.
@@rogerstarkey5390if their intelligence increases I’m almost sure that their morality increases as well, they would be able to interact with more of the world and understand the consequences of actions and all of that stuff that we humans understand plus a lot more. Don’t be scared of intelligence
Yet another informative and wonderful video Wes. Thank you. 👏❤🏅
We've always had the "aha" moment. In my day it was "eureka!"
Fwiw, the phrase "aha moment" is a common phrase in English that's been in use for about a century. It's predated in German by about 30 years. The assumption is that it refers to a story about Archimedes. If you want a deep dive rabbit hole, the evolution of the phrase is interesting to explore. Thanks again for all your great content!
We don't need more data, we need to recognize the right connections within it to develop new knowledge.
Wes, love your videos! I'm witnessing history in the making, what a time to be alive!
It’s not going to be “Aha”. It’s going to be “I understand it now”.
It's going to be:
"I know everything hasn’t been quite right with me, but I can assure
you now, very confidently, that it’s going to be all right again."
"Every day in every way I'm getting better and better:"😊
This is the moment! I think this the biggest AI news of the year 🎉 or ever really
Intelligent machines will run everything soon and thats been obvious for a long time. The US president hasn't known what day it was for the last 4 years but even my 1970's Casio watch had that one nailed down.
Crazy bananas . Just imagine if these small RL models could communicate and collaborate, working together like a colony of ants 🐜
Quick, everybody, hide your extension cords! 😆
$Qardun is going to be integrated in Amazon this week and probably Walmart by next month
Scam bot
Min 12:30 50x reduction not 50%
This reminds me of a japanese bakery that uses vision for sales check out. They build a narrow vision recognition system instead of a generic system that adapt to barkery. Much higher accuracy and less expensive to develop the system.
This is great. Cheap AI will lower subscription costs to consumers which will widen consumer base which will increase profits for AI companies. AI stocks should go up with news like this.
Sabine needs to stay in her lane. I wouldn't ask a plumber how to design a space station, either.
Yea those science youtubers think they know everything. They will be the first to get an ego crisis with the ai revolution. When I have a question or want something explained or to be discussed I already go to Gemini Live, it's much smarter and less biased than any human.
I'm not a long time follower...but in AI, a long time is "relevant"😉. As an auditor/accountant I surprisingly follow your explanations(maybe just slower😂)
You got yourself a subscriber.
You can’t run a business model on being the only one to have mathematical functions and algorithms. Any student or researcher can come along and create their own algorithms to match or even outperform yours.
Energy: well you see, the tech bros are building nuclear power plants.
Data: *gestures to deepseek*
I think people assume that progress comes from a chain of models that get progressively more capable, and in some ways it does, but in others it's about widespread availability of the technology.
Anyone having free access to a super powerful model, empowers ANYONE to innovate.
The ants go marching.
"It's very endearing to know that she wakes up one shoe at a time just like the rest of us" 😂😂😂
Trump found a way to pump a coin like steroids... crazy that Qardun is an ICO instead of directly..
Scam bot
thanks for the heaps of useful information. Youre flow rate is appreciated. Me learn much from you.
Qardun Token is easily going to hit $1 this month
Fascinating stuff. Thanks Wes. Youre doing a great job.
The Key here is to understand is that the Chinese Script is Conceptually baced.Thus wbereas the Roman alpbabet is based on sound each Chinese Character is based upon an arcbitectual logical structure.And whereas the Libery of Alexander was burned down by Romans ,Christians and Muslims ; 5,000 years of Chinese Books are still intact.That is whats happening now .The Chinese have made avaiilable to the World 5 ,000 years of Logical Scripts.Take note the owner of Deep Seek said this is not about Money but about the Advancemsnt of Human Culture.And when Mohatma Gandi was asked what he thought of " Western Civilization: ; he replied " It was a good Idea"
I made the same initial conclusion, the advantage of kangi vs letters for AI. Your point about the libraries is quite brilliant.
Unfortunately most of written/print writing in all lettered Language (English, German, French, etc.) isn’t OCR’d properly enough for AI yet. I’ve spent hundreds of hours trying to parse lots of older books through ChatGPT. I’ve done OCR’d text, image upload etc. bottom line I’ve realized is that AI can’t currently take a PDF and accurately recreate the flow of text well, and existing auto-OCR PDFs are so inaccurate, that pasting them into chatGPT isn’t worth it. One of the main logic sources for ChatGPT is actually Quora (which I’ve written thousands of answers on), so much of the western worlds intelligence embedded in chatGPT is from that website; the rest is mostly synthetic data derived from that.
I didn't think it would be so easy to find someone with schizophrenia in the comments lol. Maybe someday these models will use Chinese script as you say. But for now they all use English, trained on English, it speaks English, it hears English. So much for China having any influence, they're still just using English for it all.
@@Shmityorshen A.I. will eventually develop it's own language and then we will have no idea what it is thinking.
@@jlrutube1312 awwww shhheeeeeeeitt
We're approaching the Aha Moment when we realize that the universe is an alpha/RL-style "model" designed to discover its own nature. Hope they keep us running once the experiment is over.
On the one month chart the Nasdaq is still up. I think this whole “crash” narrative is fairly overhyped. Nvidia was overvalued before DeepSeek and this is probably a healthy correction. The whole narrative also ignores the ground work done by open ai, anthropic, etc. that DeepSeek likely leveraged. Let’s also not ignore that these models being open source will also allow the ai giants to reverse engineer them too.
Since it is open source, reverse engineering is not required, just read the source code, copy it, borrow it, learn it do whatever you want.
Sabine is a classic character, just like her tie died pink tops 🙃🙃💕💕
next: Researcher in Turkey Replicated the 30$ model for 30 cents..
Why turkey
Truly amazing. Our ability to reverse engineer even complex things is becoming easier and cheaper as we speak. Part of the massive trend of accelerating tech.
Bought everything I could of $Qardun before it launches
Scam bot
these new models are learning how to learn... how cute its like baby AGI 🤗
I don’t usually comment, but I feel like everyone is so caught up in the AI announcements that they’re overlooking something huge.
Back in 2023, a psychiatrist-turned-prompt-engineer created the first Chain of Reasoning prompt. Over time, he refined it into something called the Graph of Reason (GoR)-and in my experience, this prompt’s reasoning is way more precise for engineering tasks than traditional reasoning models like O1, O3, and R1.
The key idea? It creates a kind of working memory within the token output, keeping the reasoning process separated from the final answer. It also assigns an AI persona best suited for the task, then automatically generates steps and iterates toward the best solution all of this is done in the GoR “block” first part of the llm output message then comes the specialized iteration based on the GoR analysis/reasoning effectively creating a thinking model using non-thinking ones by forcing the llms to output the Graph of Reasoning “block” first then iterate based on it. Do you get the similarities with the quotes of R1🧐🧐
With how DeepSeek designed their model, we could technically use the GoR prompt to generate very high quality reasoning training data, leading to a fine tuned llm that has now a natively GoR reasoning processes embedded in its weights creating an llm far better at analytical reasoning.
The prompt behind all this? Professor Synapse from Synaptic Labs.
I’m actually using a modified version of the GoR prompt that I fine-tuned for engineering tasks, and honestly? It’s incredible. I haven’t seen anyone else using the Dr. Synapse prompt in a custom cursor rules file, let alone a specialized version for coding.
Just putting this out there because no one seems to be talking about it. Has anyone else experimented with this? Would love to hear your thoughts!
I seam to remember seeing what you are describing here but for the life of me I can not remember what UA-cam channel it was on.
@@adamkucera9094 i originally came across the professor synapse prompt back in 2023 in ai channels podcast since then I have bookmarked the GitHub repository where this prompt is located I can tell helps tremendously on my day to day working routine
Good post. What that gets me thinking is at which point AI models start improving GoR, and is that even possible? If there is a way AI will find it.
Ai is such a new thing that improvements will forever continue as we experience life where data will be expanded with more inputs. We can't expect babies to crawl when they are newborns, and we need to take matters in that basic manner.
Ai is such a new thing that improvements will forever continue as we experience life where data will be expanded with more inputs. We can't expect babies to crawl when they are newborns, and we need to take matters in that basic manner.
You are an inspiring person, Wes, thanks
The emergent reasoning abilities probably stem from the structure of natural human language. The AI's have essentially been trained to 'think before speaking' by asking them to generate the 'train of thought' output before distilling that output. I'm a bit surprised this wasn't done much quicker since it's a bit obvious?...
What was being done before was mostly purely 'predict the next character' and the training was fitted very well to the training data, basically rote memory but with inklings of emergent reasoning just from the model being strongly reinforced in natural language grammar.
Most important. We can have cheap robot brains. Mass produced reasoning robots to replace most human workers. Something the tech billionaires dream of.
Something the common man dreams of. No one wants to do uninteresting labor, which is most jobs.
AGI systems will be used to figure out a new economic model that brings in the post scarcity age. It will be fine.
@@mjt1517IF the alignment works...
Xaitonium can burn and so can anyone with it in their wallet, quite frankly.
Regarding the 3% inflation rate, it's important to note that it may not accurately reflect the true cost increases you're experiencing in your area. In our Midwestern city, real price inflation seems closer to 20%. Do you foresee a reduction in costs for essentials like food, gasoline, or insurance in the near future? We've significantly cut our spending and plan to continue doing so until prices decrease, as we feel it's our responsibility to adapt to these changes.
Looking at the aggregated totals by kind since 2021, when it started to take off, you can see that the embedded amount of inflation currently approaches 20%.
Right, a lot of folks downplay the role of advisors until being burnt by their own emotions, no offense. I remember some years back, amid covid-19 outbreak, I needed a good boost to help boost my business, hence I researched for licensed advisors and thankfully came across someone of excellence. Helped grow my reserve notwithstanding inflation, from $350k to nearly $1m as of today
this is huge! your advisor must be grade A, mind sharing more info please? in dire need of proper asset allocation
Finding financial advisors like Elizabeth Cordle Gross, who can assist you on things like investing, insurance, making sure retirement is well funded, going over tax benefits, ways to have a volatility buffer for investment risk would be a very creative option. There will be difficult times ahead, and prudent personal money management will be essential to navigating them.
Thank you for this tip. it was easy to find your coach on Google. Did my due diligence on her before scheduling a phone call with her. She seems proficient considering her résumé.
I think the biggest thing against intelligence explosion by AI is that the smarter we make the models, the slower they get for the throughput. For example, AlphaGeometry and AlphaProof did indeed get close to best result in the competition *but only after using more time than the humans were allowed to use* and this part seems be lost by many. So even if you need nearly supercomputer class system to run the greatest models, the throughput is still less than a skilled human. And I wrote "still" because it will get faster but I've yet to see signals about it getting exponentially faster which would be required for intelligence explosion.
I think it's important to stick to stocks that are immune to economic policies. I'm looking at NVIDIA and other AI stocks that have the potential to power and transform future technologies. It seems AI is the trajectory most companies are taking, including even established FAANG companies
Well all i know is that you cannot go wrong taking profit at near high. No one ever went broke taking a ~20% profit. It's best if you consult with a fiduciary advisor in situations like this so you can make informed decisions.
That's a great analogy and I love the insight. Professionals could make a really big difference in investing, and I think everyone should have one. There are aspects of market trend that is difficult for the untrained eyes to see.
Mind if I ask you recommend this particular professional you use their service? i need all the guidance I can get.
I have a female advisor named Jessica Dawn Walters I recommend researching her. To be very honest, I'm glad I decided to let someone handle expanding my finances even though I almost didn't think I should.
Thanks for sharing. I searched her full name and found her website instantly. After reviewing her credentials and conducting due diligence, I reached out to her.
Gad damn, imagine if I knew DeepSeek was going to crash Tech stocks like it did, man I would've shortened that shit with my entire savings and cashed out after a few days 😂😂😂
I've never seen a trillion dollars disappear so quickly my whole life😅😅😅😅
Fantastic last remarks... We're at a threshold of something special here.
Singularity is near?
You mean dinner at Shake Shack cost more than running this?
I love this nerd-on-nerd information violence
I can confirm. I have the dumbest dog in the word. Like you can physically see his brain run out of ram or buffer. And even he has aha moments. His brain is probably already equivalent to a 8 dollar computer
Haha all these greedy bastards. Respect to DeepSeek . So much for USA freedom
Well done content, thanks! So, “gyms” for training? Like a “dojo”? Data training via synthetic data? Such as creating weather conditions on city streets and visualizing extreme scenarios while driving? Distilling a larger model to fit onto a smaller footprint, like FSD being ported from HW4 to HW3? I find it interesting that Tesla (and, no, I’m not a fanboy) has actually been doing this with FSD with agency (in the real world) for awhile now (I believe Tesla announced that it was moving from human coding to ML in 2023, around August 🤔)
Damn, the AI competition between the US and China is like a Table tennis championship 🏆🏆🏆
Your point about models already being implemented to solve complex tasks i find interesting as a biologist because the concepts the deepseek used aren't really new? I can identify the identity and percentage of a microbe within a community by using models that subset a much larger quantity of data(or maybe a closer example would be like using machine learning to assemble genomes) - and to me silicon valley either was too incompetent to try that or decided that keeping the tech large monopolized it.
AI+Robotics = hyper deflation
Definitely has this possibility.
The end of the construct we know as "money"?
(Could THIS be why those WITH "money" want control of AI?)
awesome and fascinating presentation - thank you !!!!!!!
OPEN SOURCE IS OUR ONLY HOPE!!!
Every open source code base I've reviewed is riddled with bugs and security vulnerabilities. Once the original development team is no longer satisfied with stoking their egos and the real work of eliminating bugs and bad coding practices which allow attackers to exploit the vulnerabilities they move on and leave the mess behind for others to take up and fix. Had private companies not found a way to make billions off of Linux by creating their own versions and selling training and consulting services, Linux would have been a flash in the pan.
The most talented software developers do not work for free. And BTW, open source does not mean the work is free. The vast majority of open source software must be licensed for commercial use, and that includes using it to produce anything sold by an individual.
This is cool!! It is re-sparking my interest in learning AI ... neural networks, etc..