Vollständiges Transkript anzeigen (4.935 Wörter)
[gasps] >> Artificial intelligence updates every Friday at 2:00 p.m. Eastern. Today is Friday, May 29th, and as you see, there are a lot of updates. [snorts] Let me go to our leaderboard arena. So, what you see here is as usual Claude Opus is on the first place. We have this Chinese model GLM 5.1, which is right after Claude, before Gemini, before Muse, before everything else. So, this is new. Okay, next Claude Opus 4.8. It was released yesterday, and it is a huge, huge thing. So, first of all, it is better than Opus 4.7 in many benchmarks, in many like parameters. Also, [snorts] they have same price, which and it became new default. So, if you go in any tool where you used to use Opus 4.7, now it's Opus 4.8. Same price. Now, they have couple features which will increase your usage. One of them, they have so-called fast mode. So, it gives you 2.5 faster speed, but at twice price. So, instead of 5 million input, you will $5 per million tokens and 25 for output tokens, you will pay 10 and 50. But, the speed is 2 and 1/2 times faster. Another way you can increase your usage by using multiple sub-agents. And they have this new thing called dynamic workflows. It's also called Ultra code. So, Ultra code setting. By default, it's on. Uh so, it can spawn up to like thousand parallel sub agents to do the task. So, what Claude now can do, it can orchestrate given the task, it will create a workflow where multiple agents will work in parallel to execute the task. And they have a very important features like they, for example, have a judges to verify that the task completed successfully. Which is very, very important. So, what it means that when it does the task, um you can trust it much more. So, it's becoming more honest, like less hallucinating, less making mistakes. So, this is actually huge. You can also have control or about effort, and they have multiple levels levels of control, and so on. Uh there are so many videos, and I put this link. Thank you, Sandra, for sending it. Uh this is very good article which shows how to actually use this features. So, how to enable massively parallel execution, adversarial verification. So, when agents are in adversarial relationship, one agent does something, another uh criticizes it and objects or whatever. Scalable test time computer usable automation. So, this is this is great release. Um Cursor Composer 2.5 model. Model. So, a composer you think it's UI? No, no, it's it's a model. And it's built on top of Kimi K 2.5. So, this is Chinese model, and it has trillion parameters. And usually in America, people don't like to use Chinese models. But here, what they have done, they took this model which is open source, and they trained it. And when I say they, it is Cursor. Cursor is a company which provides AI enabled editor for coding and they have a lot of data which for programming, for coding. And they made a partnership with Elon Musk's companies. In fact, Elon Musk companies will acquire cursor, but because they're in the process of IPO, they made a contract so that they partner, but they haven't acquired yet. They will do it after the IPO. But anyway, the cursor got access to huge data centers which Elon Musk's companies have and they use these data centers to train the model which was already good LLM model to train it specifically for coding and they achieved very good results and they provide this model inside the cursor editor. So, it is not available by itself anywhere else. It only available inside cursor. And it is 10 times cheaper than other models on this platform. So, what they're trying to do, I guess, is to promote cursor. So, it's called cursor composer. The model. Now, interesting about token costs, enterprises start using agents more and more and agents consume a lot of tokens and suddenly you have costs of tokens as high as human salaries. So, if you're spending, let's say, $700 running your agent cloud code or whatever to do the development, multiply it by 22 days in a month, so you have $15,000. So, it's a human salary. Right? At the same time, what's interesting, like in China, Xiaomi decreased token price like 50 to 100 times. So, you're talking about, you see, 0.0036 per million tokens and they really somehow achieved very very low prices. And pretty fast, 57 tokens per second. This is this is good enough. Deep Seek when they released they gave gave discount like 75% discount and now they decided to keep it to make it permanent. So it looks like token prices go down but at the same time token usage goes up and it goes up faster than >> [laughter] >> prices of tokens going down. Okay. So here companies are shifting towards a hybrid approach using frontier models for heavy architectural planning and routing to a cheaper models to do like simple work. Same with Google, SpaceX and so on. Okay, Cohere command A plus was open sourced. So Cohere is a company which produces models and systems for enterprise and they also took their model and it's a 218 billion parameter model and they made it open source. It runs on just two H100 GPUs. Coding today is mostly done by AI. Well, I didn't touch code. I'm coding every day but I didn't touch code with my hands for probably half a year. Google says 75% new code is AI generated, Anthropic 90%, Meta now maintain data 75% for most of their code. Amazon rolled out Cloud Code for every corporate employee. Boris Cherny thesis is that the engineer in the future is a builder, someone who decides what to make and directs AI to make it rather than writing the code themselves. So Deep Seek discount is permanent. I already said that. Oh, this I really like. If you just news.google.com/foryou if you put it in your browser and uh make it a favorite, this is now my homepage in in Chrome. Because I usually interested in AI and it gives me all the news about AI and technology. Just very good way to stay informed. KPMG integrates Claude across its core business. So, KPMG is this big company dealing with accounting, audits, and so on. And you see 276,000 employees worldwide now will use Claude. And named Anthropic's preferred partner for private equity consulting. So, this is also a big thing. Okay, users are upset about Gemini 3.5 Flash. So, well, you know, Google this month they had a Google I/O and they many announcements and one of them is Gemini 3.5 Flash. And they say, well, intelligence and speed is awesome, amazing, incredible. But it's been trade to Max Evils for benchmarks and it doesn't really useful in when you work with it like iteratively going back and forth. Claude actually is very good at that. But apparently Gemini is not. The model learned that doing extra stuff correlates with higher scores, so it generates that pattern anyway. So, what it says that it does more than you ask ask it to do. So, it's not exactly what you're asking it. Okay, this is very interesting piece of news. Trump approved Nvidia chip for sale in China, but Beijing doesn't want it. Nvidia says it has largely conceded China's AI chip market to Huawei. So, first there were restrictions and the Chinese found their own way to do it and now they have their own chips and they can do training cheaper, and they don't want American chips anymore. Okay, MCP tunnels. This is a very very good idea. Uh so, Claude um released uh managed agents. Uh so, these are agents which you can configure on Anthropic uh cloud on uh and then suppose that this agent it's running on the cloud, but it needs access to data which is uh locally in your company behind your firewall. Like usually uh how people were solving it, you would expose this data through some server, and then this agent uh or in Anthropic cloud has to log in into the server using I don't know, username, password, some keys, somehow. And that means that you expose your data to outside. The idea of MCP tunnel is that you do it differently. You uh dial out from your internal uh system, and you connect to uh to cloud uh this way. So, you keep the keys inside your company. You don't give it to external agent. So, who initiates this connection? And in this case, it's not agent from outside, it is you establish this encrypted tunnel uh so that then uh you can use it. And uh so, both Anthropic and OpenAI now have this uh capability. So, uh what you do, you install uh Cloudflare, you see blue install Cloudflare, and then you create a tunnel, and then from inside your code, you tell it to use this tunnel. It's really really uh simple. So, here you see actual examples of code. Uh server.py runs inside your private network. Uh Cloudflare uh tunnel agent creates outbound only encrypted connection. And yeah, that's it. That's it. Uh okay, this is uh about my channel. So, we have Well, I haven't updated these numbers, but uh we have uh close to 7,000 subscribers and close to 300 videos. Uh name of the channel is selector. I provide uh slides links for the slides on Google Drive and on GitHub under the video. And I usually ask the question. And recently the question uh is like how which agency you using how you use them and what's your experience. But anyway, please stop the video and answer the question. Okay, use cheaper sub agent models in Claude code. Uh in Claude code uh you can use multiple models uh in the same session. So, you can for example use Claude Opus for planning and architecture. And you can use Claude Haiku for actual execution on simple tasks. So, Claude code creates a task list and then you give it to simpler model to execute. And it doesn't have to be Anthropic Claude model. It may be even local model or DeepSeek or whatever. You can configure uh multiple models. And uh here is an example how to do it. Uh you do pip install uh light LLM proxy and then you create the YAML file where you define these models. Uh this is how you do uh environmental variables. Uh this is uh you create definitions of your different agents uh for different models. So, this is planner. This is executor. And you see it says the model. Here it's Claude Opus, let's say 4.6. And here Claude Haiku 4.5. And here it's Qwen 3 Code and next. So, it's a Chinese model, whatever. And then in your Claude MD, you explain which models to use where, and that's it. And then when you start, you start with the planning model with the most smart model, and then it will do all this orchestration internally. So, it's not difficult to do. Uh and you can decrease your cost this way. Uh harness engineering a system around the model. Okay, this is generated by probably Gemini. Okay, tools constrain feedback loops interfaces instead of observing over the perfect prompt for obsessing over the perfect prompt, focus on how the agent operates over time, what's allowed to do, how to check its work, and how failures are caught and corrected. Okay, this is the link on GitHub. Practical harness usually includes So, it's kind of how to make your own harness. There are a lot of now publications and videos on how you can build your own harness or maybe build it on top of several harnesses. Uh start small. Each time agent fails, don't just tweak the prompt, add or refine the harness rule, tool, and so on. Okay. Uh Yeah, here is example. Hermes agent runs Codex, which is OpenAI, and Claude in one agent. So, you you're basically using two harnesses. Uh demonstrates a collaborative AI workflow by combining three top tools, Hermes agent, Codex, and Claude. ECC, everything Claude code. Very good. It's on GitHub and has 192 thousand stars. Probably more because this was a week ago when I looked it up. Uh built on top of Anthropic Claude code coding agent. Uh you see everything Claude code extends Claude code with structured skills, reusable workflows, custom slash commands, and and so on so on. Repository, you see it's 182, it's actually 192. Now it's even more. And yeah, okay, Cognition. Uh Cognition is a famous company. They uh created Devin, which is a coding assistant. Uh they never were open source. They were like for money, for payment. But, it's it's good. It's a high quality. And they recently raised 1 billion at 26 billion valuation. So, they grew very fast. The revenue went from 37 million to almost 500 million in just 1 year. And it writes 89% of its own code. So, this is Devin. Okay, Hermes and OpenClaw. Uh I just provide the recent updates. Every week you see a lot of progress, a lot of new features, and number of stars grows. So, it's 172,000 stars for Hermes and 375,000 for OpenClaw. Okay, and uh yeah, anyway. Uh next, Google. Why Google invested 40 billion in Anthropic. So, what's what's happening? Uh in April, right? Anthropic released data that their uh subscription rate increased from 1 billion per year to 9 billion per year in just 1 year. So, it's a very fast growth. On April 20th, Amazon uh they didn't actually wire them the money. They they made a contract on compute credits. So, for 5 billion with a 20 billion option on top. As you know, um uh Claude runs on um uh Amazon Cloud on a Bedrock. And uh yeah, so they made a commitment. And then, 4 days later, Google actually invested real money, 10 billion dollars to Anthropic with the option to another 30 billion um in in the future. So, while why Google uh did did that? Because Google has its own models, right? It has Gemini models. But at the end of the day, Google makes money when models run on their infrastructure, right? Payment for tokens. So, with running uh Claude, they maybe make a smaller portion, but Claude now dominates the market. They have 54% of corporate markets right now. Um OpenAI used to have something like 70%. Now they're down to 20%. So, Claude is this big whale, and both Amazon and Google they're trying to be vendors to Claude to get piece of this action. Right? China Deep Seek drops the V4 model running on Huawei chips, completely bypassing Nvidia. Yeah, and by the way, both Amazon and Google they don't use Nvidia. They have their own chips. Amazon have their own chips. Google have these TPUs they're famous for. Uh right. Now, Anthropic just closed the the funding. So, I I I said last month that they closed No, they didn't close last last week, sorry. They only closed this week. And they were supposed to raise 30 billion. They actually um uh got 65 billion funding at 90 100 billion valuation. Right? So, this is the largest private AI fund raise in history. And these are the sizes of some of the biggest companies. You see Nvidia 5 trillion, Alphabet almost 5, Apple 4, Microsoft 3, Meta 1.5, SpaceX 1.5. Well, they're targeting IPO right now. And OpenAI is less than a trillion. So, Anthropic is now bigger than OpenAI, definitely. Uh next uh Oh, yeah, this is from Star Trek. Compassion, that's the one thing no machine ever had. Maybe it's the one thing that keeps men ahead of them. This is your famous quote from Star Trek. I don't know why I included it. Okay, Claude managed agents on Cloudflare. So, Cloudflare, I use Cloudflare a lot. This is a very good service. This is a content management content delivery network, CDN. And you can host there your websites and there's a lot of services. I I I highly recommend. So, now they have this Claude managed agents hosted. Uh the integration uses Cloudflare global network to provide secure isolated sandboxes. Anthropic describes architecture as decoupling the brain, the agent, from the hands, the execution environment running on Cloudflare. Um anyway, uh Anthropic, Claude, and Cloudflare uh my two most favorite tools. So, I highly recommend to look into this. Okay, um Google AI co-scientist and AutoResearch Claude. So, this is um about science. So, Google published AI co-scientist research in Nature. So, this is a paper in Nature introducing hypothesis generation, a Gemini-powered tool that uses competing AI agents in idea tournaments to generate and rank scientific hypothesis. Inspired by AlphaGo, agents propose, critique, and refine ideas. In one Stanford test, co-scientist drug lead reduced uh liver fibrosis scarring by 91%. So, it's a very powerful way uh >> [snorts] >> to do science. And uh AutoResearch Claude, self-reinforcing autonomous research with human AI collaboration. So, it's an open source framework. Uh Mellon, Google, Stanford, UC Berkeley. Okay. Auto Research CLAW hosted on GitHub and uses open CLAW and CLAW for science. Yeah, this is very, very interesting. DeepSeek: Thinking with visual primitives. So, most of the models we're talking about are large language models. They're thinking in text. But, DeepSeek actually can think in visual primitives. Instead of describing images using text, the method allows AI to think using visuals, seamlessly handle tasks like counting, topological reasoning. Incredibly, the system uses 90% fewer visual tokens than top commercial models. Okay. OpenAI models solved 80-year-old math problem. So, it is disproved a 1946 conjecture by Paul Erdős about planar unit Okay. This problem had students solve for nearly 80 years. The breakthrough came by connecting discrete geometry with algebraic number theory, two fields which are rarely bridged. Experts say AI didn't use exotic math, but excels at cross-disciplinary synthesis, finding a path that human specialists could have discovered if they like looked at these two fields at the same time. Okay. OpenSpec spec-driven development. Well, you know what we're doing now. It used to be that you write the code and then later you create documentation. Now, it's completely different. You first create documentation, you create specs in a very, very detailed and thoughtful manner. You talk to AI back and [snorts] forth, back and forth creating the specs. And once the specs are created, then AI just writes the code very, very fast. Right? So, OpenSpec is an open-source spec-driven development framework for AI coding assistants like Claude Code and Cursor. OpenSpec solves the core problem by of preserving the intent and requirements. It solves the core problem with AI assistant coding. Intent and requirements get lost across scattered chat sessions. Uh so, it creates structured specification files directly in your repository and and so on. So, these are the directories on OpenSpec dev, OpenSpec pro, and GitHub. Okay, everything is in plain markdown, so it's a good thing. Um Harnessing [snorts] LLM agents with skill programs, HASP. H A S P. So, this is paper by N NYU and Salesforce. Wraps a base LLM in an external Python control harness, converts passive textual skills into executable program functions. Uh operating like physics Maxwell demon, uh harness monitors agent state and then is failure-prone state is detected, harness actively intercepts it, either overriding the action or injecting corrective context. Uh test shows this inference-time loop intervention dramatically boosts accuracy. So, it monitors what's happening and if necessary intervenes like a Maxwell demon. Okay, AI compute cost exceeded employee salary. Well, I already spoke about it, but actually corporations uh are taking action about it. Like, for example, Microsoft recently uh Microsoft has multiple contracts and licensing agreements with Claude, but they started canceling most of their direct Claude Code licenses and redirecting engineers to use GitHub Copilot CLI. Now, GitHub Copilot CLI also has Claude as one of the models inside it. So, they're not really uh like forbidding using Claude, but they're just forcing to use GitHub. Claude models remain available. Okay, NVIDIA confirmed a similar reality stating that compute cost on his team is exceeded employee salaries. Similar with Uber. So, paradox, while individual token prices are going down, total bills rise because AI agents consume more tokens per task. Goldman Sachs projects a 24-fold surge in token consumption in the next years. Gartner warns that cheaper tokens won't necessarily translate into lower enterprise costs. Okay, Microsoft and Anthropic. Microsoft and Anthropic, you see they forged a partnership in November last year, which is about half a year ago. And Anthropic committed purchasing billions of Azure compute capacity from Microsoft. NVIDIA joined the deal additional 100 10 billion investment. On product side, Claude models were integrated into GitHub Copilot Microsoft 360 sifter. So, they they continue to cooperate. It's just that Microsoft trying to save some money. Okay, they also entered talks into AI chip deal involving Microsoft Maya server chips. Okay. Next, quantum-informed AI for chaotic processes. So, this is interesting. This is a paper and the idea is so this University College London that in certain situations you you have data like a quantum data which is chaotic in nature, but model can make some judgment about this data with a better accuracy if it uses quantum as part of the pipeline. So, you see the quantum informed AI can predict highly chaotic and complex systems such as weather, turbulence, and disease spread with about 20% greater accuracy than classical AI. So, interesting. So, this is one of the first demonstrations that quantum calculations can actually help. Top view agent version 2 with C dance 20 model from Bytedance. So, you know Bytedance is a huge Chinese company. And this model is for generating long-form videos. So, it comes with different prices from 0 16 44 50 dollars per month. And it does like everything scripting, voices, images, clips, editing all together in one interface. You just talk to it and it does things for you. Yes. Next, multiverse computing LLM plus quantum blocks. So, this is parallel to what I was just talking about. But this is metal llama model and a small set of quantum blocks like really small number of parameters running a hybrid system on IBM qubit quantum processor reduced the model perplexity. Improve may maybe modest, but it had meaningful real world effect. The quantum enhanced model correctly answered astronomy and biological questions that original llama model got wrong. So, again in same direction. AI will deliver wisdom. This is interesting. This is from Peter Diamandis. I mentioned him many times. I recommend you to subscribe to him. So, wisdom is a probabilistic pattern recognition across vast human experience. Future slim a new benchmark replaces real world events and ask AI to forecast 90 days ahead. It shows that Frontier models like GPT-55 already outperforming crowd prediction markets. AI can simulate billions of interventions to identify optimal outcomes. This artificial wisdom may matter more than AGI since intelligence solves problems while wisdom chooses which problem to solve. So this is very good. And group ring 2.6 1 trillion parameter model. So this is Chinese. Designed to execute complete multi-step business workflows without requiring manual interventions. 262K token window and these are all the links and it's available oops. Through open router. Hold on a second. Why I cannot do it on the whole screen. Oh gosh, what's going on? Okay, this okay, robots Boston Dynamics upgraded the Atlas humanoid robot so it can now carry 100 pound fridge as you can see. And Hyundai plans must produce and deploy Atlas units as you know Atlas was acquired by Hyundai. And Unitree introduced the real-time voice-driven motion generated for its G1 humanoid robots. And so on. Okay, quantum is breaking encryption. So this is paper and YouTube video. Securing elliptic curve cryptocurrencies against quantum vulnerabilities. A team of Google researchers demonstrated that short algorithm can break 256 bit elliptic curve discrete logarithm problem. Using fewer than 1200 logical qubits and 90 million to 40 gates. Yes, so this work on quantum encryption or how to break it >> [laughter] >> making people paranoid and switching to better security in the systems. You see, it urges all vulnerable cryptocurrency communities to migrate to post-quantum cryptography without delay. Okay, Obsidian for memory, rag, and MCP. So, Obsidian stores data as a directory tree of markdown files. It's called vault, but it's just a directory with a subdirectories markdown files plus small amount of app metadata. It's basically a personal wiki with some indexing. Each node is a plain MD file, markdown file. You can mix images, PDFs, and other arbitrary files in the same tree. Obsidian adds .obsidian folder at the root. No database no proprietary format. You can use grep, ripgrep, and so on directly over the vault to search for stuff. Obsidian uses wiki-style links between the files. A vault is a folder and subfolders. Many people use folders like docs, projects, attachments. When you paste or drag and drop image into a node, Obsidian saves it in the file and inserts a markdown image link. And so on. So, it's Obsidian is a very very simple and useful paradigm to use for your memory, for your knowledge base. Full-text search in the app uses metadata cache. If an external agent modifies markdown files, Obsidian rescans and updates the cache. Wiki engine, you can have multiple vaults, multiple directories. You can use Google Drive to sync folders across devices. These are community plugins to do that. So, this is a very good thing to know and to use uh Obsidian. Andre Karpathy like famously was recommending Obsidian. Okay, medv.org. Uh, so this is uh Uh, okay, so Matthew Galgker, a 41-year-old self-taught coder from LA launched a GLP weight loss uh uh health platform. So, this is medv and this is the website. And why I'm talking about it? Because they're generating a lot of money. First year, 400 million in revenue and 65 in net profit. And second year, 1.8 billion projected and almost 300 in net profit. So, and this company is basically run by two people. And uh they're using AI for everything, for customer service, for sales, for for promotions, for advertising. Uh, and yeah, what they're doing is basically a marketing company. They're selling product which is a weight loss product which is in high demand. They also have some other products. Uh, and they're getting something like 15 or 20% on top for their sales effort. And yeah, it's highly highly profitable. Okay, build business with AI. Yet another video, actually two videos, how you build your business using AI. And I will not go through this because we're already over time. But here for example, uh start with docs, rules, skills for agents. So, you create your own employees, they're simply agents. And agents are defined by MD files. And then he has a claw chief which is the main agent uh which runs cron jobs like every 15 minutes and then and it does something. So, every 15 minutes it wakes up to do something. Every night the AI does lead generation via fire crawl, writes CRM Google sheets, and so on and so on. Okay, uh about jobs. Jack Dorsey, so his company Block, so he is flattening the org chart, going from five layers to maybe two, three layers. He wants to redefine managers as coaches overseeing hundreds of people rather than directive supervisors of small teams. He already requires all engineering managers to contribute code directly. Uh from catastrophic predictions of white-collar job loss toward viewing AI as productivity multiplier. So, on one side we see that some um tasks are now outsourced to AI. On the other side, we see a lot of requirements for people who can do AI and orchestrate AI. Uh so, interesting. Bolt CEO, Ryan Breslow, has fired entire HR team uh for creating problems that didn't exist. So, no HR. Problems disappeared when I let them go. This This is him. In first quarter of this year, tech companies laid off 80,000 workers, approximately like half of them directly attributed to AI. Goldman Sachs report, like whatever. Klarna cut 700 customer service jobs with AI. Shopify, whatever, many, many examples. Fastest vanishing roles, mid-level managers who deal with scheduling, status updates, whatever, because it's easy to automate with AI. Junior analysts, which is cheap junior labor, and one-trick specialists, people who do just one application or one platform. So, these professions, these are roles which can be easily outsourced to AI. So, recommended steps, list repetitive tasks in your job which can be outsourced to AI, automate one task to build a skill, so become a AI specialist yourself. Position yourself to deploy AI as a hybrid worker. So, you can do the work, but you can also outsource it to AI and use AI. Okay, these are stats and you see in May we have more layoffs than last year. Okay, and this is me as usual and thank you.
Video
Video
Video
Video
Video
Video
