Transcription
Hi guys, my name is Sahil Khanna and let's start today's tech news. The first news is about RAM prices. Allegations are being made that RAM prices have been artificially increased by the big companies. Basically, there are only three leaders who control the entire market. All three have doubled down on AI and left the common public behind. In that process, RAM prices have been seen increasing. The allegation is that Samsung, Hynix, and Micron have deliberately stopped supply so that prices can increase. Prices have been seen increasing by 700% in the last 4 years. Due to this, prices have been seen increasing in the US. If we talk about market share, Samsung holds 41% market share. If we talk about the memory business, Hynix has 29%, Micron has 20%, and all others in the market are above 10%. And this is not happening for the first time. Before this, in 2000, Samsung and Hynix were fined for doing similar things. In 2016-17, a class-action lawsuit was filed against them, which was dismissed. Where prices were seen to almost double. After that, in September-November, the prices of DDR5 RAM suddenly increased. Almost 60% of the 64GB kits, which used to cost $15, were seen going up to $78. And recently, a case has been filed against them regarding this matter.
Now, if I talk about business sense, in my opinion, this is pure business. Companies have realized what the common public will pay. Suppose you have four items. You can make one thing by combining four items, and that item will sell for ₹1000, versus you can make the same item and sell it for ₹100. In which will your margins be higher? So, as a business, if someone wants to make a profit, and they are doubling down on the ₹1000 category, then in my opinion, there is no problem with that. In my opinion, it will not make any difference. RAM prices will be seen increasing. Unless, because recently reports were coming out that Chinese companies are providing DRAM, storage, etc., cheaply. If that thing scales, then who knows.
Before moving on in the video, I would like to remind you that you haven't pressed the like button. Press the like button. Subscribe to the channel. I often bring such tech news for you.
Besides this, five big things have happened in the AI space in the last week. Some of them we have already covered. Like the news about Sonet, we talked about the new Sonet model that has come out. It is better than the old Sonet model. It's not as good as Opus, but it's comparable. It's fast. It has come out with a 1 million context size. And I would say that since its release, I have been using it a lot. I would like to share my personal experience here. If we talk about browser usage, if you are getting something done agentically, like go to the browser and do this, do this, do this, then the cost is very low, the results are very good, and it is also very fast. Almost, I have the big plan. Almost only 2-4% of my usage is consumed, which earlier used to be 20-25-25-30-30%. Although I used to use Opus before, Sonet 5 is doing very good for me. Benchmarks themselves have shown that Opus is much better. But my workflow says to use Fable first. Use Fable for all structured documentation and all those things, and then tell it to cover everything within that documentation. Then get it made like Son. This is my workflow, which I have been trying to crack for the last week and a half, and I can see it is working for me.
Apart from this, allegations are being made about the new GPT model that has come out, regarding the tampering done in its benchmarks, that it is actually not as good as it is being shown. Meta has also announced a model called Watermark, but has not released it. It is being trained 10 times more than the previous model and is giving GPT 5.5 level results. I feel like whoever comes along, they just brush aside GPT 5.5 and move on. That's how it's going. First, we saw Chinese models like GLM, etc. They said they are better than GPT 5.5, better factors, better, better. But now these are coming. LongCat is another model that has come out, which is open source. Meaning you can run it for yourself. And it is giving GPT 5.5 level results. Apart from this, Mistral has released Langstream, which is open source. It is a niche model. It has been released to solve problems like math. And Google has released NanoBanana 2 Lite, which is faster, gives comparable results, and is also cheaper. But if we talk about some highlights, the highlight was that another new model from China has come out, called Mute. I might not be pronouncing it correctly. It is a 1.6 trillion parameter model. They have not used any Nvidia chips. So, in itself, it is a victory for the Chinese. And the game of the Chinese is completely different from that of the US. I said this in the last business news that if you understand the strategy of the Chinese, they are trying to undercut the US by releasing US-level things in open source. They want to end the AI race itself. They are trying to bring some democratization in AI. They themselves are not, but they are trying to do it, which is a great thing in itself.
If we talk about Generative AI, GLM has released its $18 subscription for its light version, while the max version costs $10. Where the big plans of Cursor, Cloud, and Gateb are all $200 plans. So, where others are charging $200, these people are charging less and giving comparable features. So, there is nothing better than that, right? For LongCat, which they have released, just look at the benchmarks to see what level of results are coming out. Even in terms of comparison, it appears comparable in GPQA to other big models. And in SWEE Bench Pro, it is absolutely devastating. It was better than KMI 2.6. It is better than GPT 5.5. Now, the Chinese have recently released many models that are open source or open weight. For example, if we talk about LongCat, it is open weight, with 48 billion active parameters. It comes with an MIT license. DeepSeek's V4 Pro is also open weight. It is from China. KMI 2.6 from Moonshot. This is also open weight, and all of these are seen to be winning against GPT 5.5 in many aspects, which is a closed-source, trillion-dollar company.
On the other hand, if we talk about the US, the US keeps getting into fights with its own people. Recently, they even banned Anthropic for rejecting their model. Earlier, this deal was seen to be happening. For those who don't know, let me tell you the story. Anthropic got a tender to become an AI supplier for the US government. After that, Anthropic said that they cannot do this deal because they want full control. It can be misused, and they don't want that. They backed out. The US banned them. After that, GPT got this tender. After the news came out, GPT's market share was seen to go down, and Anthropic's market share was seen to increase. Now, Anthropic went to court. After going to court, they got relaxation. Due to this relaxation, they are now doing everything. Now the US has created another new clash: when Fable came out, or some new powerful model was coming out, it said that you cannot release it in the global market. It is too much. US security will be at risk because of this. Many people thought that this was done because of the ongoing clash with Anthropic. But then OpenAI's new model announcement happened. There, the same thing was seen: the good premium models that will come out will first be available to the US, the government, limited companies, and not to the general public.
So, what is the noise now? It's about AI sovereignty. Meaning, if I am creating my own product, the model it runs on, because the future of companies will run on AI. We are seeing that many companies are already deploying AI agents. What does AI agents mean? All the digital work of a company is being done by AI. For example, if I need to do some work on my browser, check something, do anything, I tell the AI, and it does that work for me. For example, this PPT that has been made is a WordPress MCB that I have created on the Laps Voice website. From there, it checks how many news items have come. Out of all of them, which ones are clearly visible to it. It merges all of them. It creates a story. It does research for me. I have made some rules behind it, and it gives me PPTs like this. So, this work was supposed to be done by a human. Today, it is doing this work beautifully for me, and if you look at it, it is taking someone's job. Now, suppose my workflow is dependent on AI. Although I have a team, we can do it manually as well. But let's assume all of this is done by AI. And not just me, but 500 people work in my company, and all 500 are AI. Whatever AI models I run in the background, whether it's Claude or OpenAI's, if the US government shuts it down tomorrow, my entire business will stop overnight, and this is the future of business. Here, agents will run, and all work will be done.
So, the noise that is being made here is about this very thing. Sovereign AI, meaning we will use some AI over which we have complete control. And here, I only see Chinese models if I think about it. Because Chinese models are giving us US-level work, which they are giving at an open-source level. Meaning, I can run that model on my GPU anywhere. Whether I rent it, because there are many companies where I can run my AI model, and I have to pay per instance. Meaning, I have to pay according to the load I put. So, what will happen here? Tomorrow, no matter what happens, my business will not be seen to stop. And it's not like Chinese models are very backward. They are models from three years ago. They are comparable. They are only one or two months behind. In fact, ChatGPT hasn't released its latest model. It hasn't been publicly released. So, Chinese models have surpassed them. This is a humiliating thing for ChatGPT. It's a big humiliation, and because of this, their user base is seen to be drastically decreasing.
Now, a good thing about this is that the cost of AI compute, of doing any work, is seen to be decreasing over time. So, a very good story is emerging for us. OpenAI recently said they have found a way to reduce their inference cost by 50%. Meaning, the cost of their models will come down by 50%. I have already told you from my experience that Sonet 5, which has come out, is very good. It is giving good results, and it is also fast and cheap. Some time ago, a report from Amazon came out saying that Amazon is developing its own model using Claude, which is part of their deal. Because of this, the model is costing them less internally. It is costing up to 75% less. Meta is playing a different game altogether. If they have spare compute, meaning they have spare GPUs, they are renting them out to others. So, if there are open-source models, Chinese models, etc., that you want to run, then compute is also available in the market where you can run them. So, if you look at it, the story is shaping up where things will be seen to become cheaper.
So, here I ask you a question: in your opinion, what is the right option for you? To use subscriptions like Claude, ChatGPT, or to switch towards your own systems? Perhaps for many people, it might not be possible right now. But in the future, this is a possibility. Earlier, I thought that all computing would be done on local computers. Someone would buy a system with 256GB or 512GB RAM, it would be a one-time expense, and then their computer would keep running because Chinese models are coming. There is a model from Queen now. It is open source. You can download it. It is a distilled version of Fable 5. So, it has the thinking properties of Fable 5, and it is Queen's model. It is 32 billion parameters, in my opinion. And it's doing good. So, my theory was that such things would come out that would run on local computers. But the problem is that in the last some time, RAM prices and everything have been seen to increase. So, personal computing has also become very expensive. What seemed affordable some time ago, like the Mac Mini, cost me around ₹48,000 after all taxes and discounts, and today it is going up to a lakh. So, computing is becoming expensive personally, but it is becoming cheaper at the data center level. So, if you look at it, using Chinese models is not bad. You can use it by hosting it somewhere else at the API level. Suppose you don't trust China, but there must be some GPU provider in India where you can get DeepSeek's model or any Chinese model through an API. There must be some companies in the US that are providing it, so you can get it from there.
Otherwise, I think if they were doing something wrong, things would have opened up in the market. If we talk about the CEO of Planter, he talked about no points, which talks about sovereignty. That the future is sovereignty. Your data is your business. If you have given your data to some AI, then your edge is seen to be ending. He spoke against token maximization, saying that the best model for the best thing is not always necessary; for many tasks, it's not. He talked about controlling weights, saying that higher weight doesn't always mean everything is good. For example, if your use case is small, then perhaps you can work with 2 billion, 5 billion, 10 billion, 15 billion, 22, 30 billion. You might not even need a trillion-parameter model. The biggest thing said is that technology should not be politicized, which is happening. The whole story I told you, because of politics, the US will be ruined. No doubt their models are working well, but because of the politics of not providing chips to the Chinese, not giving them new GPUs, not giving them X-rays of this, they have started making their own chips. Now they have made their own AI model without Nvidia, so this is proving detrimental to the US. Now the Chinese will build their data centers cheaply all over the world, harming Nvidia's business. I am not saying this will happen, but it could, right?
Apart from this, expertise has been discussed, that expertise is the most important thing. I see many people doing anything on AI, but we shouldn't become journalists, we should become specialists. If you are a specialist in something in AI, you can do a lot by harnessing the power of AI.
Besides this, if we talk about Moore's Law, it is seen to be beaten. Moore's Law states that the transistors in any chip are seen to double every two years. And this was valid for the last few years, but it is no longer happening. For a long time, there was a lot of noise about it not being possible to go below 1 nanometer. But IBM has achieved this. They have achieved their new chip at 0.7 nanometers. Now, what has been done here? Stacking has been done. Normally, what used to happen was: there was a single chip, and all transistors were placed on it. Now, what they are doing is: stacking transistors one on top of another, on top of another, on top of another. And the reason for this is: the demand is increasing very rapidly. Every year, talking about AI chips, by 2026, there will be a demand of almost 1 million wafers per year, which was 370 in 2024.
Now, to explain transistors, I will have to get a bit technical. For this, I have also made an infographic to explain it to you. In engineering, perhaps someone has studied gates and weights. If you haven't studied it, or studied it and don't remember, then you will remember it a bit now. Basically, what is a transistor? A transistor can be considered a wire. You put current into it from here. Current came out here or not. If current came, it became Yes. If it didn't come, it became No. Yes means 1, No means 0. So, you must have heard that computers operate in 0s and 1s. So, computers operate in 0s and 1s through transistors. Now, if a transistor's wire is broken. For example, look at the first one, the output one, the wire is not connected there. So, data came in, and the answer came out as 0. When the wire got connected, the answer came out as 1. Now, understand that there are multiple transistors in a chip. Current enters from one place. Yes, No, Yes, No, Yes, Yes, Yes, No, Yes, No, Yes, No, Yes, No, Yes, No. It has made its own path, and according to that, a final output is coming out. 0, 1, 0, 1, 0, 1, 1, 0. Meaning, what answer is coming in the last output nodes. The computer understands that answer. So, the earlier theory was that all these gates should be very...
We cannot pack closely. Again, let's think of this in the current way, that if two wires come close to each other, a short circuit happens. So, similarly, the logic was that if transistors come too close, they will interfere with each other's output, due to which the output result will appear to be spoiled. But a new solution was created: stack them on top of each other, and eliminate the tension.
Apart from this, a report has come out which reveals that if you use AI, your performance initially improves, but later you become foolish. Basically, people who are using AI for all their tasks are gradually losing their ability to think. Initially, because their minds are working well, they are able to get proper output. But gradually, they don't use their brains for many things because they have become reliant on AI, thinking, "Why should I bother when it will do it?"
Here, I am reminded of my biggest personal example. I was good at math. Good at math means engineering was entirely math. I did Chemical Engineering. Many people think Chemical Engineering means Chemistry, but Chemical Engineering means math. Perhaps, in my opinion, I studied math for eight semesters. I probably studied math for seven semesters. So, my math was decent. When I was preparing for CAT, my math was very solid at that time. My calculations were like finger-counting, it would happen automatically. Since childhood, I had seen my father calculating like this. I also did it myself, and I had my own systems built.
Then I went to do my MBA. Calculators were used extensively in MBA. They were also used in engineering. But in MBA, it was common. Now, what is happening there? I am using a calculator even to calculate 72 * 2. I am using a calculator even for 72 - 7 + 3 * this, that, and that. Which I used to do quickly before. So, one day, I was giving an exam, and I wrote 2 + 2 = 4. I thought, it is 4, right? I took 5 seconds to think about it. Then it hit me. I thought, "What am I doing?" So, AI is making you dependent in the same way. And this is a bad thing.
And what is the solution for this? I don't know either, because I am also becoming very dependent over a period of time. When I read this report, I also felt a little, I thought, "Brother, I need to control it a bit because I have also started relying on AI for many things." Yes, it will do it, why should I do it when it can do it? Some things I used to prefer to learn. Now I don't prefer to learn at all. I think, "AI will do it itself, why should I learn?" I tell it, "Learn yourself, plan yourself, and give it to me."
So, this experiment was conducted on 2681 people for 30 months, where it was observed that initially, there was an 18% boost in homework scores, and this was an illusion. Later, this number decreased from 18% to 24% within 2 years. And the sad story is that along with this, jobs are also seen to be decreasing in the market for fresh graduates. So, many people are selling courses saying, "AI will not take your job. The person who is using AI will take your job." But if a person has been using AI for a long time, their intellectual power has already diminished.
Earlier, if you look at the market, demand is low. Jobs are seen to be decreasing. Admin roles are down by 90%. IT programming roles are down by 80%. If we talk about other graduate vacancies, they are down by 61%. This is data from Hong Kong. If we talk about data for freshers, then it is not good news. So, what is the future? Time will tell.
Regarding Apple's new earbuds that were supposed to have a camera, visual intelligence at a very dangerous level would have come to Apple, and they stopped it. Reports came out just a short while ago that the prototype was ready and was going to be scaled up to a mass level. It was just about to be seen, and I thought, "This is destruction." People are adopting glasses. Now this will also come. What all things are going to come. I was very excited. But they have dropped it. I will explain the reason for dropping it.
What is it? Apple's intelligence is not ready. Visual intelligence is definitely not ready. You talk about visual intelligence. It is constantly on and is watching and understanding everything about me. So, we have not reached that world yet. Hardware prices have been seen to be increasing for some time. iPhone prices have been seen to be increasing. So, making this type of device was going to be expensive for them in the first place, due to the increase in RAM prices. Plus, it is a small device, it has a camera, everything. So, the battery will also not be at that level. So, the battery backup will not be at that level. So, the product became unviable. That's why it was probably folded, according to the reports.
And where are they putting their resources? In the new iPhone, which is going to be foldable, which will be their state-of-the-art new iPhone. I don't know, they have made a lot of speculations. So, they are thinking of 10 million orders. They are planning to make 10 million devices. Earlier reports were saying 7 to 8 million. If we talk about the total lineup of Z Fold, it is at 5 to 6 million, and it is a very good phone. Z Fold has proven itself in iterations that their device works well. They have created a product with a good experience. It is smooth, powerful, has a good camera, and a good screen. They have also made it thinner. They are also improving the aspect ratio. So, despite being a superior phone, Flip is at this number, and I am talking about the total. And they are expecting 10 million. Where their prices have already been seen to increase. It will come around $2500-$3000. So, how many people will adopt it? No idea. Will it be something of that level of value? Because Apple's intelligence is not ready at that level yet, right?
Samsung is not calling itself intelligent, but it is Gemini-backed and is doing many things because of Gemini. Samsung has already proven many things. Apple does not have anything like that. Yes, in iOS 27, they are introducing a Scam Shield that will protect from scammers, which is a good thing. Scams in the US will be prevented. Will scams in India be prevented? Let it come, then we will see.
Apart from this, if we talk about Japan, they are going to take robotics to the next level. They have set a target of 10 million robots by 2040. Meaning, they will install robots in their factories. 10 million means 1 crore. And I will give you an idea of where the market situation is right now. China is the market leader at this time. It has already deployed 2 million robots. There are 4.66 million robots worldwide, out of which China has more than 2 million. And this is data from 2024. This number must have increased by now, because now robots are being seen in roads and offices there. Japan has set a target of 10 million by 2040, which is good. But I think by then, China will probably have moved further ahead. We are not even talking about India. We mean, things are being made in front of us. We don't think anything about it. Nothing happens, and we remain where we are. There is a big difference, isn't there?
I was reading about Anthropic. They have launched a paid fellowship program. They are giving 70-80 thousand to people, saying, "Come and learn AI, and in return, we will give you money." Think about it, this is happening in the US. It has not come to India. But there are many companies in India like Sarvam, etc., that can do this. Yes. Okay. They cannot give $0000. They can also give it because they have launched a fund of 150 million dollars, so you can launch a fund of 10 million dollars where you can help many people, provide compute, give your resources, saying, "Come, apply to us for startup funds, and we will give you something for free." If we want to become the next superpower, we have to do these things. The US has become ahead because of this. We are not even talking about robotics yet, and in China, 2 million were deployed in 2040. Do you understand? Dark factories have existed in China since when? Dark factories mean where lights are not on. Everything is happening in automation. Since when did this happen? But here, we are not even thinking about robotics. We are a top player in cyber attacks. Meaning, we are the ones being cyber-attacked, and in cyber attacks, it is mainly ransomware. 45% of India's cyber incidents have fallen under the ransomware category. This is a jump of 165%. Recently, 630 GB of data from Tata Electronics is circulating on the dark web, where a lot of data from Apple's suppliers is present. Look at the rise. In 2025, it was at 100, and in 26, it reached 265, and it's only been half a year. Half a year is still left.
India is doing one good thing, which is that now India is able to manufacture iPhones and sell them worldwide. Many other things are going in the opposite direction for India. And as AI is transitioning, these cyber attacks and all these things will be seen to increase, and we are not ready for these things either.
So, with this, we end our tech news for today. I hope you enjoyed it. Please like, share. Comment and tell me how you liked the video. Bye, good night, shubhkamnaayein, good morning, afternoon, whenever you are watching the video. Jai Hind, Vande Mataram. Remember me in your prayers.