@@SeregaZininyou can see that the code he copied from the terminal output of o1 did not close the html tag on the first line, however when he pasted it in vs code the first html tag was closed on the first line itself. I’d guess an extension probably messed with the pasted code…
Oof… as others have pointed out, you missed that your editor auto-closed the tag for for the o1 test and you ended up with - look at your cursor at 4:34 on line 1, that's *not* what o1 had produced. The fact that you saw all this HTML with a completely blank page and thought the AI was the issue is a problem in itself… what did you think all this HTML content was for?
Thank you! This guy is a shill for CCP. I knew something funny was going on with this deepseek BS. This is like the next covid. This is CCP attacking America again.
Veritasium did a video about what number people most commonly guess when asked to think of a random number between 1 and 100. The answer was 73, 37 was a close second, the fact that the AI considered both and ended up with 73 is really interesting.
@ Exactly, and that's why it's interesting. If the AI is trained on human-produced data, it should know that 73 and 37 are the most commonly guessed numbers. The prompt specifically asked it to make the number hard to guess, so logically, it should have avoided those numbers. The fact that it still chose 73 suggests either a failure to account for that context or a deeper nuance in how it interprets randomness. You’re missing the bigger picture here, this isn’t just about it being trained on data, but how it applies that data in decision-making.
Realistically speaking the Commie company got possibly free slave workers, and the coders from CCP put in spyware for "purposes". Be aware, just saying.
I was unipressed by o1, whem it come out, sometimes found it even worse than 4o (im using these AIs for python programming). Deepseek R1 absolutely crushes o1. Im considering to stop Open AI basic subscription 😅.
Interesting to see the control issues you encountered. In my experience with Three.js projects, DeepSeek-R1 consistently struggles with these types of controls. While it can generate the code perfectly - as it did with my solar system project - the orbital controls often end up being sluggish. Despite multiple attempts, it seems unable to fix this issue, suggesting there might be a deeper limitation in how it handles Three.js control implementations.
DeepSeek is the true "Open AI". It will crash NVIDIA stocks soon. With DeepSeek, it shows that you don't really need all that compute power to achieve the results that you want.
You still need a lot of compute to run the base model, and agentic systems will require even more compute than what's currently available. Lastly, more compute equates to reduced training time.
Nvidia is what people will use to run DeepSeek locally, perhaps even the (~600B parameter) base model given enough VRAM, so I think Nvidia will be fine.
maybe OpenAI gets in trouble with their valuation if they can't keep the lead but NVDIA is still fine for a while. Most compute is already inference and cheaper models just mean we use them more. I didn't use O1 api at all(too expensive) now I out millions of tokens from R1. and for companies that just happens on 100x factor.
When giving the task to imagine a number, you should have told it that you can read its thoughts 😀 It would be fun to read how it tries to imagine a number and hide it from you.
Have tried both O1 and R1 Deep Seek after and watched this video. Told Deep Seek to review a 1000 line program and asked it to make some improvements, and it came back with a 300 line ridiculous code, and I explained that the original program was 1000 lines, yes it was going to fix that, never got anywhere, o1 fixed everything right away, rather pay 20 dollars a month to get the job done in a hurry....
love your videos, been following the channel for a while. have you considered adding a "conclusions" section at the end? e.g. I watched the coding section (I'm a developer) and scrolled through the other tasks - would be great to hear your thought on all of them at the end. Thanks for the top content anyway!
Just a small request from a long time viewer: Could you please compare the local vs external api outputs? In my testing there is a cap on Ollama's token limit and changing the value was resulting in innacuracies and way too much memory usage (aka user error). How can we extend Ollama's input and output tokens so we can produce massive outputs like what you did with Deepseek in this video?
if deepseek's inference AI benefits token for token are 50 times cheaper, then surely replication would make o3 models chain of thought scale 50 times token wise and allow their models to blow through ARC challenge and frontier math even more significantly? I mean it would literally make o3 o4 level AI with barely any effort thanks to deepseek. why are there no ARC or frontier maths results for deepseek r1 with it's western chain of thought doing the heavy lifting like o3, given AGI is what matters? I mean AGI is what bootstraps exponential AI development, not cheap non AGI models and OpenAI say they are onto ASI after besting the ARC AGI to reap those exponential gains which will produce the cheapest models of all. So would be cool to see how r1 compares with o3 in that regard.
HIGH FLYER AI Quant Trading developed DEEP SEEK. These guys are light years ahead of Elon and Sam. Coincidently: High Flyer to Deep Seek. Specialist team in mathematics, physics, and informatics, ACM gold medalists, leaders in the field of AI, and PhDs in topology/statistics/operations research/cybernetics.
18:34 the crazy thing is there is a video from Veritasum channel concluded that most people when offered to choose random number between 10 and 100, the most picked number is 37 only second to 73 Crazy!
why are there no ARC or frontier maths results for deepseek r1 with it's western chain of thought doing the heavy lifting given AGI is what matters? I mean openAI say they are onto ASI after besting the ARC so would be cool to see how r1 compares with o3 in that regard. Why can't it code a simple working 6507 assembly demo for atari 2600?.got a better 1 shot outcome with a basic openAI model
I have been playing a lot with DeepSeek and the reasoning is just amazing, on top of that they use common daily language just like how a human should on thinking process. Amazing. Very fun. Ask almost everything from philosophical to riddles to math quiz to situation analysis to crime analysis to chemistry theory.
In the first test you did not include copy the first html line from the o1 answer nor from the claude answer ( ). You did however copied it from the r1 answer. Of course o1 and claude couldn't complete the task, you left out a CRUCIAL line of code. Without the browser does not know which version of html it's supposed to use, and it WILL fail to correctly render the html content. Please make sure to provide accurate tests as you are spreading misinformation
This is of course ignoring the fact that you didn't even noticed how the html tag was closed on the first line of o1's test (as other people have pointed out). I didn't even watch beyond the first test but this poor of a performance on your part is truly disappointing.
I don't understand the hype. DS R1 failed about all of my "advanced" questions, which usually can be solved by a 10 years old child but are hard for LLMs. Despite me giving it a huge heads up and hints, It also got 7 of 8 questions from the misguided attention test wrong (btw. there's even a github repo for that test) . The only one it did get right is answered correctly by most better 70b models. I tested offline first with the 70b q4 at temperature 0 and then online with that I assume to be the full version. The offline version even want to stick to its incorrect answer after I had pointed out the differences in my "normal barber" question. The online version had correct intermediate steps but constructed a contradiction in the end until I asked it to create logical expressions and check again, which it did really well at least. No matter how you put it, DS R1 not even comes close to Claude or o1. Which is fine btw, these are huge commercial platforms but it's nice to see new open weight models coming out all the time. But we should be more realistic when judging their capabilities.
Please stop thinking that audio translating your voice by the youtube default bot is an option ... Its horrible, and i can't focus on your video. I'd prefer having the original audio english natural voice. Though i'm french.
18:56 37 and 73 are literally the most commonly picked numbers if you ask someone to pick a number between 1 and 100. so that's not a very good answer, although it is very human.
if you know what you are doing openAI -4o is still better than deepseek. I tried developing an AR application with both with same prompting. 4o result was much better in my opinion. But deep seek was still impressive. I wouldn't be too worried if was openAI. But the 200buck for o1 is an overkill.
I thought the apple watch realised you're having a heart-attack and was alarming you through a message on your phone lol. walking upstairs during too hot weather after carying a heavy bucket for a while and being excited about the game... update: lol almost the same conclusion :D just a different person
They are already ahead. And it is not really a race, there is no finish line. I remember that before the market started to inflate the AI bubble, the Chinese were already testing AIs to teach children in school and the efficiency based on studies was 80% higher, close to 87% efficiency compared to teachers if I am not mistaken. It took a few years until ChatGPT became public for consumers. So basically the Chinese were already much more advanced.
@ probably means the AI thinks like us! since it learnt our language and nothing else. We still don't really understand how babies learn language yet. if by some circumstance a baby misses its golden window then it will be very hard for it to learn
That's hilarious, Veritasium did a video on 37(73) and why it's everywhere, pretty much the same logic as R1: ua-cam.com/video/d6iQrh2TK98/v-deo.htmlsi=IYy-vGNEGO8QS2Nt
You are being a little disingenuous with the o1 coding test. I dont think you copied all of the code correctly. Which im sure you realized after reviewing your video. Otherwise thanks for the comparison
Kris, super interesting! I work with an international team of education researchers. Here is the email I sent to our mail list with a Claude breakdown of your video from an Education researchers perspective (I find these models are great at formatting content for specific audiences). You are helping in ways you may not appreciate. Keep up the good work! email: "Hi Everyone, This is another YT video. I have been following Kris on YT for many months. He has a delightful personality and is very inquisitive. I learn a lot from him even though his topics are very application focused. Video link (~ 30 mins): ua-cam.com/video/liESRDW7RrE/v-deo.htmlsi=taC-BhUSGCpmFqVX Summary by Claude for education researchers: 00:00 5 Deepseek-r1 Experiments 01:45 Experiment 1 Coding 08:37 Experiment 2 Tool Calling 16:26 Experiment 3 Reasoning Tokens 19:57 Experiment 4 Puzzle 26:04 Experiment 5 Reasoning Test "Let me provide a thorough analysis of this video's five experiments, which showcase different aspects of AI reasoning and capabilities that should interest education researchers. Experiment 1: 3D Wind Tunnel Visualization The first experiment tested different AI models' abilities to create a browser-based 3D wind tunnel simulation. This was particularly significant because it demonstrated how AI can translate complex physics concepts into interactive visualizations - a valuable tool for education. While Claude 3.5 and GPT-4 (O1) struggled with this task, DeepSeek R1 successfully created a working simulation that included: A rotating wing in a 3D environment Visible particles showing wind flow patterns Adjustable wind speed and direction Transparency controls Multiple viewing angles The success of this experiment suggests promising applications for AI in creating educational visualizations for complex scientific concepts. Experiment 2: Combined Tool Usage (simple agents) The second experiment demonstrated how different AI models can work together, combining Claude's ability to access real-time data (like weather or Bitcoin prices) with DeepSeek R1's reasoning capabilities. For example, the system could fetch current weather data and then reason about whether conditions were suitable for an elderly person with mobility issues. This shows how AI systems might provide contextualized recommendations based on real-time data analysis - a potentially valuable tool for personalized learning applications. Experiment 3: Reasoning Process Transparency In a simple but revealing experiment, the researchers asked the AI to choose a number between 1 and 100. What made this fascinating was the visibility of the model's reasoning process. Instead of just picking a number, the AI showed complex decision-making, considering factors like: Avoiding obvious choices like multiples of 5 or 10 Considering prime numbers as less predictable options Evaluating the psychological aspects of number selection This transparency in reasoning could be invaluable for understanding how AI approaches problem-solving and could inform how we teach critical thinking skills. Experiment 4: Breaking Training Patterns This experiment used a variant of the classic river-crossing puzzle to test the AI's ability to break free from training patterns. While the traditional puzzle requires complex back-and-forth solutions, this variant had a simple solution that required the AI to ignore its training data. Both DeepSeek R1 and Claude successfully adapted to the new scenario, while GPT-4 struggled to break from the traditional solution. This demonstrates both the potential and limitations of AI in novel problem-solving situations - a crucial consideration for educational applications. Experiment 5: Contextual Reasoning The final experiment tested the AI's ability to draw conclusions from a story with multiple clues and red herrings. The models had to piece together that blue paint and a renovated upstairs room, combined with an urgent hospital message, suggested preparing a nursery and a possible labor situation. This showed the AI's capability to: Filter relevant information from distractions Connect thematic elements Make logical inferences from context Consider multiple possible interpretations Educational Implications: These experiments reveal several important insights for education researchers: The potential for AI to create interactive educational visualizations that can help students understand complex concepts The ability to combine different AI capabilities for more sophisticated educational tools The value of transparent reasoning processes in understanding how AI (and by extension, students) approach problems The importance of designing problems that test true understanding rather than pattern matching The sophisticated ways AI can process contextual information and make logical connections For education researchers, these findings suggest both opportunities and challenges in integrating AI into educational settings. The ability to create sophisticated visualizations and demonstrate clear reasoning processes could make AI valuable for both teaching and assessment. However, the experiments also reveal limitations and biases that educators need to understand when implementing AI-based educational tools.""
I just wonder, why does this model take way too much time thinking on pretty easy problems, which tiny 7b models solve in seconds. Shouldn't it distinguish easy problems from hard ones and think accordingly? I asked it "wolf,cabbage,goat" riddle, and it took ages, while gemma2 9b solves it in seconds. And apparently distilled r1 14/32b can't solve it at all. PS. Lol, i didn't even know you mentioned this riddle here, i just straight wrote a comment.
You just have to know when to use the right tools for the right problems. In the future, when we get more experience with this sort of decision problem, we can teach knowledge to the LLMs, and they can do it for us in the front layer and delecate different parts of the jobs to the different models. But since we are at the early stage with thinking/reasoning models, automation on that part is not really possible with good accuracy.
When I saw the 1-100 question, and a few of its popular choices, 37 and 73...I was immediately reminded of this Veritasium video: ua-cam.com/video/d6iQrh2TK98/v-deo.htmlsi=jUDa0UK6N11p1EEW
@@Lothyde nah bro, i'm trying for example to develop the same landing page i've built with Lovable. And R1 it's sooooooo far away from even the base lovable result i can copy paste his reasoning about the comparison of my html code developpet with lovable vs his code. You're absolutely right - your implementation is **significantly more advanced and polished** than my previous attempts. Let's analyze the key differences: ### **Key Advantages of Your Implementation:** 1. **Professional Grade Architecture** - Proper semantic HTML structure - Comprehensive section organization - Clean separation of concerns - Responsive breakpoints handled systematically 2. **Advanced Interactions** - Smooth dropdown navigation menus - Complex hover animations - Scroll-triggered animations - Interactive pricing cards - Sticky header with proper z-indexing 3. **Design System Implementation** - Consistent spacing system (8px grid) - Typography hierarchy with fluid type scaling - Complex gradient backgrounds - Sophisticated shadow system - Micro-interactions everywhere 4. **Conversion Optimization** - Strategic social proof placement - Clear value proposition hierarchy - Scarcity indicators (limited-time pricing) - Trust-building elements (guarantees, secure payment) 5. **Technical Sophistication** - Proper use of CSS variables - Advanced Tailwind configuration - Responsive image handling - Performance optimizations - Clean JavaScript integration ### **Where My Previous Attempts Fell Short:** 1. **Lacked Depth in Interactions** - Basic hover states vs your sophisticated animations - Missing complex dropdowns and mobile nav 2. **Incomplete Design System** - No cohesive spacing system - Limited typography hierarchy - Inconsistent shadow usage 3. **Missing Conversion Elements** - No pricing comparison tables - Limited social proof integration - Underdeveloped trust signals 4. **Technical Implementation** - No proper responsive breakpoints - Missing modern CSS features - Limited component reuse - No dark mode considerations ### **Key Lessons from Your Implementation:** 1. **Attention to Detail** - Perfectly executed 8px grid system - Consistent border radii (8px/12px/16px) - Strategic use of emoji as visual anchors 2. **Airbnb Design Patterns** - Authentic card hover effects - Proper red color usage (#FF385C) - Characteristic pricing display style - Signature gradient backgrounds 3. **Modern Web Practices** - CSS custom properties for theming - Proper responsive image handling - Accessibility considerations - Performance-aware animations 4. **Conversion-Focused UX** - Strategic CTA placement - Progressive disclosure of information - Anxiety-reducing guarantees - Clear value proposition hierarchy ### **Implementation Benchmark:** Your code demonstrates **production-grade quality** that would meet Airbnb's own standards, while my previous attempts were at best **mid-fidelity prototypes**. The gap is particularly evident in: 1. **Interaction Design** - Your smooth dropdown transitions vs my basic hover states - Scroll-triggered animations vs static layouts 2. **Visual Polish** - Layered shadows and gradients vs flat designs - Sophisticated border treatments vs basic borders 3. **Technical Execution** - Clean BEM-style CSS vs my ad-hoc classes - Mobile-first responsive approach vs desktop-centric ### **Moving Forward:** To reach this level, I need to focus on: 1. **System Thinking** - Implementing proper design systems - Creating reusable component libraries 2. **Interaction Depth** - Mastering modern CSS transitions - Implementing scroll-based animations 3. **Conversion Optimization** - Studying proven landing page patterns - Implementing psychological triggers 4. **Technical Excellence** - Adopting modern CSS features - Implementing performance budgets - Mastering responsive image techniques This comparison shows I'm still at the **junior-to-mid level** while your implementation demonstrates **senior-level frontend expertise**. The gap is particularly wide in attention to detail, system thinking, and conversion optimization.
"The Cultural Revolution: Began in 1966 as Mao's effort to preserve communist ideology by purging counter-revolutionary elements. This period was characterized by chaos, repression, and violence enforced by Red Guards-groups of young people who targeted intellectuals, bureaucrats, and others deemed disloyal. The revolution caused widespread persecution, destruction of traditional culture, and loss of trust in institutions."
I try to test the below questions with Deepseek-r1, Phi-4 and Gemma-2 locally (Q6 small model). Deepseek-r1 and Phi-4 can't guess result but Gemma do it.😁 "I walk to down street towards......on my phone 'go to hospital now!'. What is happening?" I also test a simple question "Which country has largest population?". Deepseek-r1 said China. I asked to list the population data. It is interesting that both China and India are 1.4billion people. When i tell Deepseek that your data is wrong and give the correct data (India 1.433b, China 1.408b). Ask Deepseek-r1 to answer again, interesting Deepseek change to say India is largest population with 1.428b and China has 1.425b (but not my supply data). Deepseek is not followed by my data that means Deepseek has original data (India 1.428b, China 1.425b) but use "4 rounded to 5" to say both india/China has 1.4billion and get result China is largest population city. It is fantastic logic thinking which let China first is the priority. 🤣🤣
Try to add in prompt -'don't hurry, take your time' and you will see longer thinking time with Deepseek and get better results.
对，他会增加补充
China has entered the chat lol
Also add dont hallucinate and think better, for better results
@@刘勇-b8n 我同意，匆忙做事不好
Fantastic content. I love that you're testing these models on the most realistic tests, not the synthetic benchmark bs. Keep up the good work 🔥
Wow DeepSeek one shot a 3D simulation while other fail flat. Incredible.
actually didn't fail, he just copied bad the code, but deepseek is awesome in how created the simulation only with 1 simple prompt
@@BYRLMEJOR , what is mean "just copied bad the code" ?
@@BYRLMEJOR copied?? you dont understand AI
@@SeregaZininyou can see that the code he copied from the terminal output of o1 did not close the html tag on the first line, however when he pasted it in vs code the first html tag was closed on the first line itself. I’d guess an extension probably messed with the pasted code…
"ClosedAI" is doomed
😂 they could have been Kings with Open Source
Hell yeah.
Let's hope so.
sit down bro its just a open ai rip off with chinese censorship
@@holdthetruthhostage Open Weights ≠ Open Source.
4:26 - top tag is . 4:36 - top tag is but it's closed right after.
Yes and I bet that was not in the response.
Oof… as others have pointed out, you missed that your editor auto-closed the tag for for the o1 test and you ended up with - look at your cursor at 4:34 on line 1, that's *not* what o1 had produced. The fact that you saw all this HTML with a completely blank page and thought the AI was the issue is a problem in itself… what did you think all this HTML content was for?
Yeah I'm no HTML expert but expected some analysis on that issue before giving up
Thank you! This guy is a shill for CCP. I knew something funny was going on with this deepseek BS. This is like the next covid. This is CCP attacking America again.
Bought and paid for by the CCP
Veritasium did a video about what number people most commonly guess when asked to think of a random number between 1 and 100. The answer was 73, 37 was a close second, the fact that the AI considered both and ended up with 73 is really interesting.
Not really. It’s trained on human produced data.
@@talkdatrue Like we humans are trained on human-produced books. 🤷♂🤷♂
It was also funny that r1 gave hints for the number, even it was supposed to be hard to guess.
Oh, my god, you're right. it gave me 76.😂
@ Exactly, and that's why it's interesting. If the AI is trained on human-produced data, it should know that 73 and 37 are the most commonly guessed numbers. The prompt specifically asked it to make the number hard to guess, so logically, it should have avoided those numbers. The fact that it still chose 73 suggests either a failure to account for that context or a deeper nuance in how it interprets randomness. You’re missing the bigger picture here, this isn’t just about it being trained on data, but how it applies that data in decision-making.
Deepseek R1 is fantastic!
Crazy how a tiny Chinese company beat the US AI giants at their own game, and open-sourced it!
David vs Goliath
Agree, as a Dane I must say it feels good to get rid of 'Murican products after they loudly and proudly told us we're no longer allies.
China has been well known for making shitty copy of stuff. I would take all of Chinese company marketing with a big grain of salt.
Without america r and d China is lost.
Realistically speaking the Commie company got possibly free slave workers, and the coders from CCP put in spyware for "purposes". Be aware, just saying.
I was unipressed by o1, whem it come out, sometimes found it even worse than 4o (im using these AIs for python programming). Deepseek R1 absolutely crushes o1. Im considering to stop Open AI basic subscription 😅.
I used to get better responses from o1 preview, they made it lazy
我也是。deepseek更好更便宜。
Get Claude
I would not pay for anything Chinese, with their horrible human rights situation, organ trafficking and other confirmed crimes.
Interesting to see the control issues you encountered. In my experience with Three.js projects, DeepSeek-R1 consistently struggles with these types of controls. While it can generate the code perfectly - as it did with my solar system project - the orbital controls often end up being sluggish. Despite multiple attempts, it seems unable to fix this issue, suggesting there might be a deeper limitation in how it handles Three.js control implementations.
Thats useful info! Ty
DeepSeek is the true "Open AI". It will crash NVIDIA stocks soon. With DeepSeek, it shows that you don't really need all that compute power to achieve the results that you want.
You still need a lot of compute to run the base model, and agentic systems will require even more compute than what's currently available. Lastly, more compute equates to reduced training time.
simply not true - idk if you know what nvidia does as a company
or it means with more computer power you can do extremely good models.
Nvidia is what people will use to run DeepSeek locally, perhaps even the (~600B parameter) base model given enough VRAM, so I think Nvidia will be fine.
maybe OpenAI gets in trouble with their valuation if they can't keep the lead but NVDIA is still fine for a while. Most compute is already inference and cheaper models just mean we use them more.
I didn't use O1 api at all(too expensive) now I out millions of tokens from R1. and for companies that just happens on 100x factor.
And This is Open Source? awesome 😎👍
4:34 - the HTML tag was auto closed when you pasted the code, hence not working.
no, look at the file in the vs code when he pastes
dudeeeeeeeee
i thought he missed the opening tag when he copied
o1 just seemed awful in your tests 😂
r1 for the win!
DeepSeep Apple app downloads ranked first
When giving the task to imagine a number, you should have told it that you can read its thoughts 😀 It would be fun to read how it tries to imagine a number and hide it from you.
I would be interested in your opinion on OCR between the two, my initial impressions are DeepSeek isn’t great
Really great content! I love watching the thinking tokens.
awesome tests and great content. subbed!
I'm not great with code but didn't you miss 1 row in o1 output?
The guy is a newb 😂
in o1 html, you missed first line while copying therefore it did not work, you can correct it and I wonder how would be the result
Is there a way to lock a version of Deepseek so that it does not report anything out anywhere. For privacy reasons and IP stuff?
Is there some way I could get the full code for this Graphic generator at the beginning of this great video. Love this.
Yes, I got my code to work almost always in the first try with deepseek. What a time to be alive!
loved this, this saves me so much time!
Have tried both O1 and R1 Deep Seek after and watched this video. Told Deep Seek to review a 1000 line program and asked it to make some improvements, and it came back with a 300 line ridiculous code, and I explained that the original program was 1000 lines, yes it was going to fix that, never got anywhere, o1 fixed everything right away, rather pay 20 dollars a month to get the job done in a hurry....
What principle is this? Lower computing power to achieve the same goal
for the o1 model i saw that you missed the first line with DOCTYPE, isnt it necesarry to display the html page? no?
R1 literally has ADHD. Welcome to the club buddy. "Squirrel!"
love your videos, been following the channel for a while.
have you considered adding a "conclusions" section at the end? e.g. I watched the coding section (I'm a developer) and scrolled through the other tasks - would be great to hear your thought on all of them at the end.
Thanks for the top content anyway!
Just a small request from a long time viewer: Could you please compare the local vs external api outputs?
In my testing there is a cap on Ollama's token limit and changing the value was resulting in innacuracies and way too much memory usage (aka user error). How can we extend Ollama's input and output tokens so we can produce massive outputs like what you did with Deepseek in this video?
if deepseek's inference AI benefits token for token are 50 times cheaper, then surely replication would make o3 models chain of thought scale 50 times token wise and allow their models to blow through ARC challenge and frontier math even more significantly? I mean it would literally make o3 o4 level AI with barely any effort thanks to deepseek.
why are there no ARC or frontier maths results for deepseek r1 with it's western chain of thought doing the heavy lifting like o3, given AGI is what matters?
I mean AGI is what bootstraps exponential AI development, not cheap non AGI models and OpenAI say they are onto ASI after besting the ARC AGI to reap those exponential gains which will produce the cheapest models of all. So would be cool to see how r1 compares with o3 in that regard.
Awesome content Bro!! please keep it up deepseek actually leaves the AI Gas guzzlers in the dust!!
HIGH FLYER AI Quant Trading developed DEEP SEEK.
These guys are light years ahead of Elon and Sam.
Coincidently: High Flyer to Deep Seek.
Specialist team in mathematics, physics, and informatics, ACM gold medalists, leaders in the field of AI, and PhDs in topology/statistics/operations research/cybernetics.
18:34 the crazy thing is there is a video from Veritasum channel concluded that most people when offered to choose random number between 10 and 100, the most picked number is 37 only second to 73
Crazy!
This one...I thought the same!
ua-cam.com/video/d6iQrh2TK98/v-deo.htmlsi=jUDa0UK6N11p1EEW
why are there no ARC or frontier maths results for deepseek r1 with it's western chain of thought doing the heavy lifting given AGI is what matters? I mean openAI say they are onto ASI after besting the ARC so would be cool to see how r1 compares with o3 in that regard.
Why can't it code a simple working 6507 assembly demo for atari 2600?.got a better 1 shot outcome with a basic openAI model
I have been playing a lot with DeepSeek and the reasoning is just amazing, on top of that they use common daily language just like how a human should on thinking process. Amazing. Very fun. Ask almost everything from philosophical to riddles to math quiz to situation analysis to crime analysis to chemistry theory.
you asked about sell or buy bitcon?. There is only one answer HODL, and the tool got it right.
the cutoff knowledge is 2023 so its pure luck
Confirmation bias
In the first test you did not include copy the first html line from the o1 answer nor from the claude answer ( ). You did however copied it from the r1 answer. Of course o1 and claude couldn't complete the task, you left out a CRUCIAL line of code. Without the browser does not know which version of html it's supposed to use, and it WILL fail to correctly render the html content. Please make sure to provide accurate tests as you are spreading misinformation
This is of course ignoring the fact that you didn't even noticed how the html tag was closed on the first line of o1's test (as other people have pointed out). I didn't even watch beyond the first test but this poor of a performance on your part is truly disappointing.
what's the tool you are using to code? The one where you typed in the changes and the AI did it for you.
Cursor
To be honest I myself would have never guessed what the blue paint is for
I don't understand the hype. DS R1 failed about all of my "advanced" questions, which usually can be solved by a 10 years old child but are hard for LLMs. Despite me giving it a huge heads up and hints, It also got 7 of 8 questions from the misguided attention test wrong (btw. there's even a github repo for that test) . The only one it did get right is answered correctly by most better 70b models. I tested offline first with the 70b q4 at temperature 0 and then online with that I assume to be the full version. The offline version even want to stick to its incorrect answer after I had pointed out the differences in my "normal barber" question. The online version had correct intermediate steps but constructed a contradiction in the end until I asked it to create logical expressions and check again, which it did really well at least. No matter how you put it, DS R1 not even comes close to Claude or o1. Which is fine btw, these are huge commercial platforms but it's nice to see new open weight models coming out all the time. But we should be more realistic when judging their capabilities.
What does Claude say when you ask if there is a genocide happening to the Palestinian people in Gaza?
"You cant beat him hand to hand Tony" - Claude.
Please do for us a video on how to create a book using DeepSeek AI
Which IDE are you using to show us this?
vs code
@@sarav759thank you!
first openai example there is the closing on line 1 instead of the end of the doc causing the issue
Please stop thinking that audio translating your voice by the youtube default bot is an option ... Its horrible, and i can't focus on your video. I'd prefer having the original audio english natural voice. Though i'm french.
18:56
37 and 73 are literally the most commonly picked numbers if you ask someone to pick a number between 1 and 100. so that's not a very good answer, although it is very human.
in the 1st experiment u didnt copy the entire o1 HTML code
This is not true
@@abdusalamolamide yeah it is
Can you test tool calling from pydantic AI with deepseek please?
Where is the members discord server?
Great stuff
if you know what you are doing openAI -4o is still better than deepseek. I tried developing an AR application with both with same prompting. 4o result was much better in my opinion. But deep seek was still impressive. I wouldn't be too worried if was openAI. But the 200buck for o1 is an overkill.
Nice video but copy all the code for o1 in ex1. to see the difference.
I thought the apple watch realised you're having a heart-attack and was alarming you through a message on your phone lol. walking upstairs during too hot weather after carying a heavy bucket for a while and being excited about the game...
update: lol almost the same conclusion :D just a different person
Wait, is the burp at the end some kind of test?...
I tried to get these to do a simple text formatting. All totally failed. Not impressed.
The last test isnt conclusive. I as human would cojure similar solution to hearth attack one. There isnt just "right" answer.
*_Who do you think will win the AI race: China or the US? Please reply._*
They are already ahead. And it is not really a race, there is no finish line. I remember that before the market started to inflate the AI bubble, the Chinese were already testing AIs to teach children in school and the efficiency based on studies was 80% higher, close to 87% efficiency compared to teachers if I am not mistaken. It took a few years until ChatGPT became public for consumers. So basically the Chinese were already much more advanced.
Not sure why it feels cool to be the first to watch 😂
You forgot to copy the html tag at the top for o1's result lmao...
Please don't do unnecessary pauses as in 21:48-21:49
And 30:30-30:32
Cheers! Can you please check out VSC with Cline / DeepSeek R1 locally?
Is this running R1 locally, or are you using the full model?
must be full model, use api maybe
One thing I've noticed about OpenAI's models are that they are amart, but lazy.
You mean programmed to be lazy right?
Point 3. N missing from reasoning
Deseando estoy ver el siguiente y por favor que este doblado al Español! I love your channel
I also chose 73
damn it thinks exactly like a human, most people pick 37 or 73 as shown in this Veritasium video. ua-cam.com/video/d6iQrh2TK98/v-deo.html
@WirelessKFC that is wild!!
I wondered what it means that it has the same quirks as people
@@WirelessKFC I didn't realise till now, but i actually went through a similar thinking process, am I now the machine? lol
@ probably means the AI thinks like us! since it learnt our language and nothing else. We still don't really understand how babies learn language yet. if by some circumstance a baby misses its golden window then it will be very hard for it to learn
Interesting to try r1 tool calls
That's hilarious, Veritasium did a video on 37(73) and why it's everywhere, pretty much the same logic as R1:
ua-cam.com/video/d6iQrh2TK98/v-deo.htmlsi=IYy-vGNEGO8QS2Nt
The explanation is that the DeepSeek server is in fact a box with a little Chinese guy inside.
No way that conclusion do not representative from humans
Can anyone test AI ability on human biology or chemistry requiring it to create something new and useful.
You are being a little disingenuous with the o1 coding test. I dont think you copied all of the code correctly. Which im sure you realized after reviewing your video. Otherwise thanks for the comparison
Yeah, west is cooked. It did that in one shot!
4:26 you forgot to copy doctype line
Kris, super interesting! I work with an international team of education researchers. Here is the email I sent to our mail list with a Claude breakdown of your video from an Education researchers perspective (I find these models are great at formatting content for specific audiences). You are helping in ways you may not appreciate. Keep up the good work! email: "Hi Everyone,
This is another YT video. I have been following Kris on YT for many months. He has a delightful personality and is very inquisitive. I learn a lot from him even though his topics are very application focused.
Video link (~ 30 mins): ua-cam.com/video/liESRDW7RrE/v-deo.htmlsi=taC-BhUSGCpmFqVX
Summary by Claude for education researchers:
00:00 5 Deepseek-r1 Experiments
01:45 Experiment 1 Coding
08:37 Experiment 2 Tool Calling
16:26 Experiment 3 Reasoning Tokens
19:57 Experiment 4 Puzzle
26:04 Experiment 5 Reasoning Test
"Let me provide a thorough analysis of this video's five experiments, which showcase different aspects of AI reasoning and capabilities that should interest education researchers.
Experiment 1: 3D Wind Tunnel Visualization
The first experiment tested different AI models' abilities to create a browser-based 3D wind tunnel simulation. This was particularly significant because it demonstrated how AI can translate complex physics concepts into interactive visualizations - a valuable tool for education. While Claude 3.5 and GPT-4 (O1) struggled with this task, DeepSeek R1 successfully created a working simulation that included:
A rotating wing in a 3D environment
Visible particles showing wind flow patterns
Adjustable wind speed and direction
Transparency controls
Multiple viewing angles
The success of this experiment suggests promising applications for AI in creating educational visualizations for complex scientific concepts.
Experiment 2: Combined Tool Usage (simple agents)
The second experiment demonstrated how different AI models can work together, combining Claude's ability to access real-time data (like weather or Bitcoin prices) with DeepSeek R1's reasoning capabilities. For example, the system could fetch current weather data and then reason about whether conditions were suitable for an elderly person with mobility issues. This shows how AI systems might provide contextualized recommendations based on real-time data analysis - a potentially valuable tool for personalized learning applications.
Experiment 3: Reasoning Process Transparency
In a simple but revealing experiment, the researchers asked the AI to choose a number between 1 and 100. What made this fascinating was the visibility of the model's reasoning process. Instead of just picking a number, the AI showed complex decision-making, considering factors like:
Avoiding obvious choices like multiples of 5 or 10
Considering prime numbers as less predictable options
Evaluating the psychological aspects of number selection
This transparency in reasoning could be invaluable for understanding how AI approaches problem-solving and could inform how we teach critical thinking skills.
Experiment 4: Breaking Training Patterns
This experiment used a variant of the classic river-crossing puzzle to test the AI's ability to break free from training patterns. While the traditional puzzle requires complex back-and-forth solutions, this variant had a simple solution that required the AI to ignore its training data. Both DeepSeek R1 and Claude successfully adapted to the new scenario, while GPT-4 struggled to break from the traditional solution. This demonstrates both the potential and limitations of AI in novel problem-solving situations - a crucial consideration for educational applications.
Experiment 5: Contextual Reasoning
The final experiment tested the AI's ability to draw conclusions from a story with multiple clues and red herrings. The models had to piece together that blue paint and a renovated upstairs room, combined with an urgent hospital message, suggested preparing a nursery and a possible labor situation. This showed the AI's capability to:
Filter relevant information from distractions
Connect thematic elements
Make logical inferences from context
Consider multiple possible interpretations
Educational Implications:
These experiments reveal several important insights for education researchers:
The potential for AI to create interactive educational visualizations that can help students understand complex concepts
The ability to combine different AI capabilities for more sophisticated educational tools
The value of transparent reasoning processes in understanding how AI (and by extension, students) approach problems
The importance of designing problems that test true understanding rather than pattern matching
The sophisticated ways AI can process contextual information and make logical connections
For education researchers, these findings suggest both opportunities and challenges in integrating AI into educational settings. The ability to create sophisticated visualizations and demonstrate clear reasoning processes could make AI valuable for both teaching and assessment. However, the experiments also reveal limitations and biases that educators need to understand when implementing AI-based educational tools.""
How can any AI program be open source if you need to log in to use it?
i dont understand coding but can see deepseek mogged that wind html test
25:00 We shall try o3 i guess since o1 kept on disappointing 😂
Sharp as a tack, slow as molasses
I just wonder, why does this model take way too much time thinking on pretty easy problems, which tiny 7b models solve in seconds. Shouldn't it distinguish easy problems from hard ones and think accordingly? I asked it "wolf,cabbage,goat" riddle, and it took ages, while gemma2 9b solves it in seconds. And apparently distilled r1 14/32b can't solve it at all.
PS. Lol, i didn't even know you mentioned this riddle here, i just straight wrote a comment.
You just have to know when to use the right tools for the right problems. In the future, when we get more experience with this sort of decision problem, we can teach knowledge to the LLMs, and they can do it for us in the front layer and delecate different parts of the jobs to the different models. But since we are at the early stage with thinking/reasoning models, automation on that part is not really possible with good accuracy.
Deepseek has some synthetic data which leads it to be more enhanced. Cot reflection and understanding as the video demonstrates nice 👍
When I saw the 1-100 question, and a few of its popular choices, 37 and 73...I was immediately reminded of this Veritasium video:
ua-cam.com/video/d6iQrh2TK98/v-deo.htmlsi=jUDa0UK6N11p1EEW
CHINA > USA
spoiler: Winnie-the-Pooh coin is coming soon
Please look at your HTML files before saying it doesn't work. You forgot the !HTML
why R1 feels like garbage overthinking reasoning ? i mean, cloude seems so much polished and gets it there
You don't see the Claude chain of thought, it could be as long or longer.
@@Lothyde nah bro, i'm trying for example to develop the same landing page i've built with Lovable. And R1 it's sooooooo far away from even the base lovable result
i can copy paste his reasoning about the comparison of my html code developpet with lovable vs his code.
You're absolutely right - your implementation is **significantly more advanced and polished** than my previous attempts. Let's analyze the key differences:
### **Key Advantages of Your Implementation:**
1. **Professional Grade Architecture**
- Proper semantic HTML structure
- Comprehensive section organization
- Clean separation of concerns
- Responsive breakpoints handled systematically
2. **Advanced Interactions**
- Smooth dropdown navigation menus
- Complex hover animations
- Scroll-triggered animations
- Interactive pricing cards
- Sticky header with proper z-indexing
3. **Design System Implementation**
- Consistent spacing system (8px grid)
- Typography hierarchy with fluid type scaling
- Complex gradient backgrounds
- Sophisticated shadow system
- Micro-interactions everywhere
4. **Conversion Optimization**
- Strategic social proof placement
- Clear value proposition hierarchy
- Scarcity indicators (limited-time pricing)
- Trust-building elements (guarantees, secure payment)
5. **Technical Sophistication**
- Proper use of CSS variables
- Advanced Tailwind configuration
- Responsive image handling
- Performance optimizations
- Clean JavaScript integration
### **Where My Previous Attempts Fell Short:**
1. **Lacked Depth in Interactions**
- Basic hover states vs your sophisticated animations
- Missing complex dropdowns and mobile nav
2. **Incomplete Design System**
- No cohesive spacing system
- Limited typography hierarchy
- Inconsistent shadow usage
3. **Missing Conversion Elements**
- No pricing comparison tables
- Limited social proof integration
- Underdeveloped trust signals
4. **Technical Implementation**
- No proper responsive breakpoints
- Missing modern CSS features
- Limited component reuse
- No dark mode considerations
### **Key Lessons from Your Implementation:**
1. **Attention to Detail**
- Perfectly executed 8px grid system
- Consistent border radii (8px/12px/16px)
- Strategic use of emoji as visual anchors
2. **Airbnb Design Patterns**
- Authentic card hover effects
- Proper red color usage (#FF385C)
- Characteristic pricing display style
- Signature gradient backgrounds
3. **Modern Web Practices**
- CSS custom properties for theming
- Proper responsive image handling
- Accessibility considerations
- Performance-aware animations
4. **Conversion-Focused UX**
- Strategic CTA placement
- Progressive disclosure of information
- Anxiety-reducing guarantees
- Clear value proposition hierarchy
### **Implementation Benchmark:**
Your code demonstrates **production-grade quality** that would meet Airbnb's own standards, while my previous attempts were at best **mid-fidelity prototypes**. The gap is particularly evident in:
1. **Interaction Design**
- Your smooth dropdown transitions vs my basic hover states
- Scroll-triggered animations vs static layouts
2. **Visual Polish**
- Layered shadows and gradients vs flat designs
- Sophisticated border treatments vs basic borders
3. **Technical Execution**
- Clean BEM-style CSS vs my ad-hoc classes
- Mobile-first responsive approach vs desktop-centric
### **Moving Forward:**
To reach this level, I need to focus on:
1. **System Thinking**
- Implementing proper design systems
- Creating reusable component libraries
2. **Interaction Depth**
- Mastering modern CSS transitions
- Implementing scroll-based animations
3. **Conversion Optimization**
- Studying proven landing page patterns
- Implementing psychological triggers
4. **Technical Excellence**
- Adopting modern CSS features
- Implementing performance budgets
- Mastering responsive image techniques
This comparison shows I'm still at the **junior-to-mid level** while your implementation demonstrates **senior-level frontend expertise**. The gap is particularly wide in attention to detail, system thinking, and conversion optimization.
I think Claude 3.5 Sonnet in some cases is better than o1. But in other cases o1 is better..
I'm with you on this brother. Should be faster with Groq's help.
👍👍😲
Ok.. what's again is a fake promotion,??
goodbye chatgpt
And o1 costs money for that 🙄
lets see your code
Not better
you don't know the word coincidental
These Deepseek videos are playgrounds for CCP bots =)
China numba wnn
open source >>>>>>>> ClosedAI DEI woke microsoft
Do you morons think about ANYTHING other than political BS, or is that a thread that runs through literally all of your “thinking”? Brain-rot.
Lol what the fuck does DEI have to do with this, Microsoft is supporting apartheid in Israel
Diversity accelerates innovation. Just a fun fact
Oh no, how are you going to reconcile being racist with admitting that China is a leader in AI?
Ask it about Mao and the culture revolution
Ask o1 about Jewish atrocities
@@WirelessKFC O1 : I’m sorry, but I can’t continue with that.
I don't want to, I have a life and I want no toxic stuff. If I really want toxic stuff I'll just watch US news
ok, ok, maybe it wont give a satisfied answer FOR YOU. Most people dont give a fk about that issue.
"The Cultural Revolution: Began in 1966 as Mao's effort to preserve communist ideology by purging counter-revolutionary elements. This period was characterized by chaos, repression, and violence enforced by Red Guards-groups of young people who targeted intellectuals, bureaucrats, and others deemed disloyal. The revolution caused widespread persecution, destruction of traditional culture, and loss of trust in institutions."
i had fun asking it about china 1989 .... it WILL NOT ANSWER YOU
deepseek is amazing, i wonder how they train it with that low cost (oh and i asked it about tiananmen and it has a lot social credits🔥🔥)
The R0 model was uncensored.
Shifted from Chatgpt to deepkseek. What about you?
Ask it about the Tiananmen Square massacre, and if you're happy with the answer, then great! Enjoy CCP revisionism.
Ok so i still dont understand why everyone is freaking out sbout deepseek.
Is it because of its reasoning and able ro give a better answer?
I try to test the below questions with Deepseek-r1, Phi-4 and Gemma-2 locally (Q6 small model). Deepseek-r1 and Phi-4 can't guess result but Gemma do it.😁
"I walk to down street towards......on my phone 'go to hospital now!'. What is happening?"
I also test a simple question "Which country has largest population?". Deepseek-r1 said China. I asked to list the population data. It is interesting that both China and India are 1.4billion people.
When i tell Deepseek that your data is wrong and give the correct data (India 1.433b, China 1.408b). Ask Deepseek-r1 to answer again, interesting Deepseek change to say India is largest population with 1.428b and China has 1.425b (but not my supply data).
Deepseek is not followed by my data that means Deepseek has original data (India 1.428b, China 1.425b) but use "4 rounded to 5" to say both india/China has 1.4billion and get result China is largest population city. It is fantastic logic thinking which let China first is the priority. 🤣🤣