00:00Are you going to lose your job in 2026? I'm going to give you 3 predictions for 2026 that will tell you what's going to happen.
00:08You have 3 signals to consider. Should you sell Nvidia now? Will the job market collapse in 2026?
00:14OpenAI has already started to model the doctor's profession, the doctor's profession completely, with ChatGPT.
00:21I'm not talking about the launch of ChatGPT Health, I'm talking about the system where doctors trained ChatGPT to be a medical assistant.
00:29And today, the only obstacle is whether your job is one of the 1350 tasks that have been modeled by OpenAI?
00:36OpenAI has just entered into a $10 billion partnership with Cerebras.
00:41Here's what this changes and why Nvidia is likely to be the stock you should sell off your hands.
00:46Andrew Feldman gave a lecture. Here's what he explained.
00:49Let's see how it goes.
00:50Now that it's over, applaud enthusiastically for at least 30 seconds for the experience I'm about to share.
00:59And there you have it, thank you.
01:01It is not simply a specific model, but an entire range of inferential and non-inferential models.
01:10Larger, smaller models, before Cerebrus had an output of 200 tokens per second.
01:17Thanks to our intervention, each model has been considerably optimized in terms of order of magnitude.
01:25At the very moment I am speaking to you, you do not know it, but ChatGPT is now capable, with instructions, of reproducing the behaviors of a doctor.
01:33But also that of a pharmacist, a bank employee, a real estate agent.
01:38All of this is written within the framework of ChatGPT's new capabilities.
01:43The upcoming version 6 will completely revolutionize the world of work.
01:47The only thing currently slowing down the speed of deployment is the computing speed.
01:51And that's where the partnership came in. The time it takes for the model to process the information to analyze chest X-rays will be about 10 minutes, patient.
02:00Very well, the entire GPU community, there are between 212 and 200 tokens generated every second, according to the analysis.
02:09Exactly. Where are we at the moment in terms of approach and progress?
02:12In truth, we were so far to the right that people had difficulty noticing us.
02:16That's why the orange arrow had to be added for clarification. Okay?
02:20Imagine that, starting from the same image, the result appears in just a few seconds.
02:25ChatGPT is capable of diagnosing, analyzing, reporting to me, and producing content for the medical field.
02:32That's exactly what we can already do with ChatGPT.
02:34But this operation will take me between 5 and 25 minutes.
02:37Imagine that now, we can reduce this time to the order of a second. That's what happened.
02:41I will now present Mistral, located on the far right, Anthropik, running on GPU in the central position and OpenAI, using the GPU on the left.
02:53Now, the human Mistral is much weaker.
02:56Their model is significantly smaller, but its reasoning ability is significantly accelerated thanks to Cerebrus.
03:02This is how you can compete with bigger and richer people by doing it faster.
03:13Honestly, I can't wait any longer to let you explore all of ChatGPT. Take about 50 seconds.
03:19First sign, Nvidia, should we sell Nvidia?
03:22Nvidia was the benchmark, the Oper H100 chip, the H200 chip, the benchmark on all DLL server systems.
03:31But today, this paradigm has a limit.
03:34The limit, the speed at which the models respond.
03:36That's when a new partner appeared, Cerebrus.
03:39What are we doing? We are working tirelessly on bottom-up engineering,
03:44fearlessly solving extremely difficult problems.
03:48Our approach began with the creation of the most imposing processor in the industry.
03:53To give you an idea, that little gas on the right is the B200.
03:57These are 208 billion insignificant transistors.
04:01The one on our left is of course our third-generation platen scale engine.
04:05We put this chip, this chip at the level of the wafer in the system
04:10and we deliver it to you, our customers, whether on your premises or via our cloud offering.
04:17And you need to understand that the industry and the world of work are going to be impacted right now.
04:21Everything is being done to ensure that AI agents take our jobs.
04:25Not only will you fill these facilities with your work,
04:29but also the new partnership with Meta that we announced a week ago
04:35where we will make their new API service available to the world's largest open weight community.
04:42A few days later, we announced that we would also be offering baby services.
04:56Beyond the IB, Madam, the cloud is reaching the other end of the corporate demographic spectrum.
05:03the 100 or 150 largest organizations in the world served by IBM.
05:07Over the course of this year, our customer base has grown by leaps and bounds.
05:12because large companies like our friends at GlaxoSmithKline...
05:17We just need to realize that today, all the largest energy processing companies,
05:23health organizations are in the process of signing partnerships with Cerebras.
05:27What will change is the power. What you've seen is just the beginning.
05:30Pharmaceutical companies that are 300 years old, as well as startups with 4 people.
05:37joined our cloud and benefited from the advantages of extremely fast reasoning.
05:44So, should Nvidia be sold? For me, there are signs that India will no longer have a monopoly.
05:49Today, it is very clear, there is a new market and that market is that of AI agents.
05:54Whether it's AI agents or even robotics, our jobs are directly impacted.
05:58The question is, do you know how to convert ChatGPT into AI agents with your $20 per month subscription today?
06:05The instructions today are very different.
06:07Agentic systems are based on reasoning systems that will consume tokens.
06:12These tokens are precisely what we call inference time.
06:15And that's what Cerebras' chip will solve.
06:18Isn't that right? We're running out of data.
06:20Therefore, it is clear that pre-trained extension, while important,
06:24is not entirely sufficient to reach the next level of intelligence.
06:28We still need more calculation, but not just in training.
06:32We have now entered a new era focused on computationally oriented towards optimized inference time.
06:37And, unsurprisingly, this has fascinating implications.
06:40It also obeys its own scaling laws in the reasoning process.
06:43The more tokens used, the higher the intelligence.
06:46The processing power will increase to a rate of over 2600 tokens per second.
06:52where today we are limited to a ceiling of 200 tokens per second.
06:55The inference model is too slow.
06:57The O3 model, which is known as the inference model, is a quote and is extremely slow.
07:02Computing power will mean that time, production capacity, and work,
07:06You are capable of writing system instructions, and will only be limited by the power of the processors.
07:11And today, that processing power is instantaneous.
07:14I don't have to wait any longer.
07:16If I need to assign a task instantly, I can simply ask the AI to do it instantly.
07:20Imagine how many tasks per minute can be executed.
07:25That's because we were limited by the reasoning factor of the model.
07:28Time has led most people to not mention the models of reason.
07:31But reasoning models are only one element of the AI power loop.
07:35The real power lies in the instructions.
07:37That is why these models are called inference models,
07:40because they can think and reason reactively before responding.
07:45Today, every cutting-edge model is an inference model.
07:48This is the intelligence index of all popular models today.
07:52Purple represents reasoning, blue represents non-reasoning.
07:54As you can see, all the best models are inference models.
07:58In fact, it is interesting to note that even the worst inference model
08:01is better than the best standard model of non-inference.
08:04Reasoning is the only way to achieve the most advanced intelligence.
08:08My second prediction is that 2026 will be the year with the biggest impact
08:12on the work and figures of the shadow market.
08:15Today, we've all been told, learn to use artificial intelligence,
08:19But what we did was we didn't learn how to use artificial intelligence.
08:21We learned how to chat with chatbots.
08:23The era of chatbots is dead.
08:25Today, systems such as OpenAI, Anthropie, and Cérébras show us one thing.
08:29Today, we are architects of IAT structures.
08:32What we see here is a direct consequence of the law of inferential time scaling that we just saw.
08:38The more tokens used, the higher the intelligence.
08:41More tokens equals higher intelligence.
08:43But we also know that from a practical point of view,
08:45Faster reasoning equates to more tokens, given a fixed time budget.
08:49It's almost like demonstrating a mathematical proof where, thanks to the nature of transitivity,
08:54We can conclude that faster reasoning processes
08:58ultimately lead to a higher level of intelligence.
09:01We must imagine the system as a probabilistic system
09:04which will determine, based on your instructions, the actions to be taken.
09:08So, how was this work carried out by the teams?
09:10First, skills were provided.
09:12That is to say, humans, doctors, lawyers, real estate agents
09:16spent time, hours, programming the AI to behave,
09:20the processes that are used in the areas of business tasks.
09:23What is changing, what we need to imagine today,
09:25that's what models like OpenAI with CHGPT or Anthropie are for.
09:29They are currently the only two models that have these capabilities.
09:31What we do in terms of training,
09:32It's about giving them the ability to work.
09:34And you, in fact, need to know how to use the learning blocks.
09:38and coordinate them with each other.
09:39In this way, you will be able to build workflows.
09:42The model is already to create PowerPoint presentations.
09:44The model is already about searching for information.
09:46The model already outlines what to do in the fields of medicine.
09:49finance, management, accounting.
09:51But it's up to you to organize it in what way and in what order.
09:55In fact, criteria, you will use the links of these different functions.
09:58Imagine all of this as bricks, and you're going to assemble them.
10:01You go from a job where you execute,
10:03where you become, in fact, an architect who executes
10:05by a fleet of artificial intelligence that you will configure.
10:09That's my biggest prediction for 2020.
10:11If you want to learn how to use artificial intelligence
10:13in the medical field,
10:15Office automation, the best of AI,
10:1765 hours of courses on artificial intelligence,
10:20update included for all those who are intended
10:23to create agentic systems
10:25where AI takes control of business tasks.
10:28Prompt Engineering Excellence,
10:30This is the elite, today, of the I reasoning system.
10:33The information is in the description.
10:34Those who own Nvidia shares today, in my opinion.
10:36should ask themselves the question.
10:38The shock is real.
10:39For me, Nvidia is no longer the leader today.
10:41He was not a leader.
10:42It was simply erased by the arrival of CR.
10:45So, of course, the market will need
10:47to always have multiprocessors
10:49with several suppliers.
10:50But today, there is very clearly a gap
10:52who will put on the market
10:54those who have an immediate answer
10:55and those who will wait more than half an hour
10:57to get an answer with heavenly intelligence.
10:59If, until now,
11:00We were limited by computing power.
11:02Today, it becomes instantaneous.
11:05The servers are already deployed.
11:06currently in use.
11:08The question is,
11:09Is that what you learned
11:10in art-sky intelligence
11:12will allow you today to really
11:13to use this computing power?
11:14And if you don't do it,
11:16Your colleague will take your place.
11:17the one who excels.
11:18It's obvious, the public authorities
11:20do not have fire speed
11:21to monitor the evolution of strategies
11:23of the use of tools.
11:25Europe is considerably behind.
11:27We realize that we are still,
11:28at the time, influencers.
11:29Let me explain: press a button
11:30to arrive at an answer
11:32of artificial intelligence.
11:33But today,
11:34true competence,
11:35it's about understanding
11:36how do they work
11:37the systems architectures
11:38and to be able
11:39to organize workflows.
11:40Those who already use systems
11:42such as ChatGPT,
11:43Agents,
11:43with systems like Claude's,
11:45Coworks are already able
11:46to perform tasks
11:47for more than an hour
11:48with Vinière.
11:49We delegate business tasks.
Commentaires