Vollständiges Transkript anzeigen (2.976 Wörter)
Shh. >> [sighs and gasps] >> Artificial intelligence updates every Friday at 2:00 p.m. Eastern today, Friday, June 19th. And as you see some good updates. Today is actually federal holiday. And let's switch to the next. So, this is our leaderboard from 3 days ago. Well, the usual stuff on the coding and chatting. So, on the coding on the top as usual, Claude. Claude Fable 5 was discontinued and now being negotiated whether it will be opened or not. Again, we have GLM 5.1 and GLM 5.2. There is a lot of talking about GLM 5.2 specifically for coding. It looks like very good. Okay. Everything else seems as before. So, let's go to the next. Okay, so this is update about Fable 5. Claude, I mean Anthropic released it and then 4 days later removed it. And now you see Anthropic and Trump administration negotiating conditions under which access can be partially or fully restored. Not clear what exactly happened. Apparently, Amazon CEO Andy Jassy alerted Trump administration that the model can be hacked, can be jailbroken, and then used, well, against USA. So, administration asked Anthropic to remove it. They said no because it's stupid. This can be done with every model. But then they released the [sighs] export control directive ordering Anthropic to block foreign access to Fable and Mythus. And at this point they decided to disable the model and disable it globally, not for only not for foreign but also for US internally. The early users of Anthropic models, which is government and some big corporations, they still have access despite this US order. And we will see what happening. They're negotiating. Okay, Anthropic delays Claude Agent SDK billing change. This is actually very good news for us because we were building application which uses Anthropic SDK. And we can use subscription, monthly subscription, to use it. So you have $20, $100, and $200. And $100 a month it's pretty good. Uh Um but they promised that on June 15th, they will remove SDK usage from subscription. But luckily for us, they kept it at least for now. But in the meantime, we decided to try to use other models like DeepSeek and whatever. And again, luckily for us, DeepSeek was not added to blacklist. Uh well, American side has a blacklist of companies which government not supposed to use. And for example, they put Alibaba company in there and and some other companies. So which because Alibaba models are very good. Anyway, Kimi K 2.7. So the previous one was 2.6. Uh so it's open weight coding focused. So they kind of follow Anthropic to make their models mostly for coding. Uh 1 trillion parameters, 32 billion active. Works with long file, multi-file repositories. A tool and agent workflows is the focus. Very affordable, as you can see. And you can run it via Ollama, for example. So, these are examples how to launch it. Uh Hugging Face and so on. So, it's open weight model. Uh Kimi work uh is a moonshot local desktop AI [snorts] employee. It's kind of like a Claude co-work and many other agents which we have seen recently. And it can coordinate multiple agents, automate knowledge work. It can Okay. Uh powered by Kimi K 2.6. So, they haven't switched to 2.7 when I was writing it. Maybe for now they already switched. Okay. GLM 5.2 by ZAI Zippo AI. Uh so, previous version was 5.1. 1 million token window open source MIT license, which is very permissive. Uh relatively small uh s- s- s- 744 billion parameters mixture of expert index share architecture accelerates sparse attention via cross-layer index reuse, significantly fewer flops per token. So, it makes it very, very affordable. Uh plans light, pro, max, team. Roughly matching or slightly beating top closed models on coding and reasoning benchmarks. There were a lot of videos where people compare GLM 5.2 with GPT or with Claude. Strong long context integrity up to hundreds of thousands of tokens. So, this is good model. And again, you can launch it uh like this, for example. Uh anyway, uh Open router fusion. This is big thing. I I this is really, really big. So, what's the idea? Uh you know OpenRouter, it gives you access to multiple models. They usually uh charge extra, sometimes charge much more. Like, for example, we compared for DeepSeek, they charge four times more. >> [laughter] >> But still, because Chinese models are so cheap, uh uh even with OpenRouter extra, it's still cheap. Like, here for example, they used the Gemini 3 Flash, which is a cheap model from Google, and the two Chinese models, Kim 2.6 and DeepSeek V4 Pro. And what Fusion does, so OpenRouter provides it allows you to create a like ensemble model, a model which consists of several models. And when you give a query, uh it sends it to all these models, in this case three models, and then it aggregates their responses to make one good response, synthetic response. What this thing does, it removes hallucinations, it increases accuracy, it makes the model much smarter, and because you're using cheap models, you actually end up cheaper than let's say Claude. So, here what do we have? Here we have Claude Fable 5, which currently is not available, and then you see below, you you see this combination of Gemini, Kim, and DeepSeek, and it's basically the same uh on this diagram. And above, uh you see uh Claude model combined with other. Again, it's Fusion, but including Claude, and you see that you can improve the model. But this is absolutely amazing. Uh >> [snorts] >> so, what we have similar to this is Perplexity computer. Uh Perplexity computer uh if you pay $200 a month you have access to Perplexity computer and one of its features it can use multiple models up to 19 models and combine their outputs. So it uses Claude Opus as a judge and which increases the quality of the response. So agentic task-oriented orchestration. We can also do local fusion. In fact this would be a very interesting project to use let's say a llama models or cheap models >> [snorts and laughter] >> and write a small controller script in Python which will expose this small models as one model. Okay. This is a plug for my channel. It's not growing really fast. Same number of subscribers and videos. But I ask you to subscribe. The slides under the video there are links for GitHub and for Google Drive with the slides and summaries. And please pause the video and answer a pin question which usually lately is if you use agents and how you use them and which agents you use. Okay. Perplexity brain memory. So this is the project which currently only starting rolling out to Max and Pro subscribers and the idea is similar to Hermes which is Perplexity remembers >> [snorts] >> the what you were doing with it. It has a memory and it uses this memory to self-improve. So it continuously reviews activity, writes structured knowledge into memory allowing system to improve over time without manual input. Enables tasks to start with irrelevant prior knowledge. Well, this is by the way one of the features I why I like Perplexity because I ask a short question and it has knowledge of my previous conversations and it gives me meaningful answer based on all this context. This is actually very very important. Okay, SpaceX IPO it went up then it went a little bit down but still the goals which Elon Musk wanted to achieve he achieved. So company received a lot of money. He became a first trillionaire. And many people blame Elon Musk that he's chasing the money but Peter Diamandis for example wrote a very good post where he explains well he knew Elon before he became so rich when he for example tried to build the first rockets and everything was exploding and he invested all his money into this and he lost all his money and only the fourth last rocket by some miracle like it was was a success and after that he got support from NASA and so on. So yes, so SpaceX went public. It's now on Nasdaq and ticker SPCX. You can just Google it. Okay. Oh, this this is interesting abbreviation, right? Fang is familiar. Facebook, Amazon, Apple, Netflix, Google. Fang. And now we have the new one Mangos. Meta, Anthropic, [clears throat] Nvidia, Google, OpenAI and SpaceX. Okay, Tata Consultancy Services. You know what? I used to work there for about half a year. This is a big big Indian company with 400,000 consultants. And what they're doing now, they make a major partnership with Claude, with Anthropic, to use Claude, and they're rolling it out to 50,000 their employees. Uh from AGI to ASI. So, this is a paper by Google DeepMind. So, this is the beginning of this paper. And they're trying to think what will happen after AGI is achieved. So, imagine millions of fast coordinated AGI agents that share knowledge instantly outperform top human teams. Progress could come through scaling current methods, new paradigms for recursive self-improvement, collective AI systems. However, limits remain. Data scarcity, cost, energy, increasing research difficulty. A key unknown is whether AI can generate truly novel ideas rather than mixing existing knowledge. Okay, SpaceX acquires Cursor for $60 billion. So, Cursor is a famous uh editor for coding, uh AI first. It was originally fork of uh uh Visual Studio Code. And uh even before the IPO of SpaceX, they made a partnership agreement that after the IPO, they will acquire it. And now they did it. They started the process acquiring uh Cursor for $60 billion in in stock. In stock in stock options of SpaceX. It's a huge deal. This is Michael Truel. He's one of the co- I think it was four co-founders of Cursor. Okay, loop engineering. This is term coined and becoming very popular recently. So, their first chats, like GPT-3.5, you were sending your request to the chatbot and it would respond. It was just LLM and it was basically predicting the next token. This is not how those chats work now. Now it is a loop inside. It gets your request, it understands it, it creates a step-by-step plan and it start answering it and improving it, checking if the response is correct and eventually answering it. So, it is self-prompting in a way and this is what loop is. So, a loop is automated workflow where agent manages its own tasks based on schedule such as heartbeats, specific times, or events by pursuing defined goal until completion. So, today's agents and tools, they all have some sort of agentic loop inside. Okay. And Vibe Coding by Matthew Berman. This is a really, really good video where he explains what Vibe Coding is and provides examples and key practices. And files like agents.md and so on. Benefits of cloud agents for isolated parallel processing, free online loop library. So, this is the link. This is his stuff. signals.forward-future.ai/loop-library. Where he collects different loops. Okay. Agent engineering, again, talking about loop engineering, emerging paradigm where developers design autonomous event loops that guide agents rather than manually prompting models. And this is agent engineering life cycle on the right. So, role, goal, and intent, that's your starting point. Define a persona, core objective, scope, authority, architecture and logic, planning, control loop design, state management, infrastructure and capabilities, front engineering, tooling and harness, memory systems, and quality and operations. Verification and evaluation, very important. The agent should verify that what he did it is right using LLM as a judge. Safety and guardrails and observability. People want to actually see what is happening and why. Okay. From Databricks to Agent Bricks. So, [clears throat] you know a famous company, Databricks. It came from Apache Spark project. And first they were managing Spark on clouds and distributed file systems. Heavily used of parquet and columnar data. Just a reminder, the Hadoop, the original distributed system, was all built on files operations. But Spark started not saving to files, but just pass data via network. And because of that, it's like 30 to 100 times faster. And guys who created Apache Spark and then put it in open source, that's why it's called Apache, Apache Spark, they created company, Databricks, which is now 10,000 people. Very famous. And now they use AI more and more in their system. So, this is kind of how it went from year to year. So, late 2010s, introduction of Delta Lake, that means adding transactions. Um Then launch Delta engine in Databricks SQL. So, basically what you're getting in Databricks, it behaves like a regular database with transactions, with everything. Uh expansion into broader data plus AI platform with MLflow, Unity Catalog, governance. Uh >> [clears throat] >> last year introduction of Lakehouse and OLTP Postgres compatible engine and launch of Agent Bricks and articulation of LTAP Lakehouse and Lakehouse. So, both for analytics and transactions as a unified governance substrate and persistent memory layer. Uh company continues to grow. Their technologies are absolutely incredible. If you have enough data to justify using because this is usually for big data streams and real-time processing and like big analytics, this is definitely very good platform. And you can use [clears throat] Python. Okay, DeepSeek. DeepSeek raises 7.4 billion at 50 billion valuations. So, becomes China's most valuable AI startup. And uh thanks God that it was not blacklisted uh at least now. Uh its founder Liang Wen Feng invested around 3 billion in the fund raise. So, out of this 7.4 billion, uh 3 billion came from one person. And before that, he was working at a hedge fund or managing hedge fund. He previously held about 90% of the company, but now it's changes. Okay, a government-backed fund invested very small amount actually, 150 million. DeepSeek plans to use the new capital to advance the research and expand the computing infrastructure. Okay, um from rack to memory systems, building stateful AI architecture. So, basic rack is just retrieval and cannot provide a real continuity, a real structure. Well, we spoke about it before. This is a common vibe. So, a production agent need a typed memory system and memory manager rather than better vector store. So, here in this article, the author defines five distinct memory types: policy, preference, fact, episodic, and trace. Each with its own schema, life cycle, and retrieval pattern. The article warns against common anti-pattern of dumping everything into single vector index. Durable memory must live in real database. Long conversation, central memory manager. So, this is a very, very good article. And thank you, Sandra, for sending me the link. It's It's really a treasure. I highly recommend to read this article. Okay, DeepSeek plus Claude code. What you get as a result is 5-10 times cheaper workhorse to use. Uh So, this is a video in Russian, but you can ask to convert into English. So, the presenter demonstrates how to set a free Claude code or DeepSeek terminal UI, which are tools that allow you to run DeepSeek models directly from VS Code. So, you can use in the same session Claude Opus as orchestrator and DeepSeek for simple coding tasks. So, this is advanced, maybe, but really, really good way to go. Agentic resource discovery, ARD. So, this is a joint effort. You see many famous company: Google, Microsoft, Salesforce, Snowflake, ServiceNow, Databricks, Hugging Face. And they created this ARD. So, agentic resourcediscovery.org is the main website. You see all these players together. It's a open Apache license specification. Let's publishers expose their AI tools, skills, MCB servers, and so on. And let's clients discover those capabilities. Now, what's interesting, OpenAI and Anthropic are not in the initial group. Their focus remains proprietary. Okay, Amazon is preparing to sell their training chips. So, we know that Claude is running on Amazon, and that means it's running not on Nvidia. It's running on those training chips. And now, Amazon challenging Nvidia. So, it plans to sell their training chips to other data centers, not Amazon data centers. So, it will be a direct competitor to Nvidia. Okay, Lean open-source small vector database. So, this this is really good. So, it's a very very small database. There is a corresponding module in Python, which [snorts] you you import. And you see this for the same data from 200 GB down to 6 GB. Very small, very lean, very effective vector database. And the first use, of course, is for local use. If you have a system, some sort of a rack system which you want to use locally on your laptop, well, try this first. Okay, Midjourney medical ultrasonic scan. So, it's a full body ultrasound scanner, and this is images it produces. It is 10 times cheaper than MRI and 60 times faster. So, they plan to deploy them in multiple scan offices, they Midjourney spa wellness center. So, they plan to have tens of thousands of those centers where people just can come in and do the scan. The resolution is sub millimeter. Well, at least that's what they promise. So, and they are of course much healthier than doing MRI. So, yeah, this is this is a big thing. So, and it uses AI for image processing. That's why I put it like AI update. Although major part of the system of course is just regular scan. Okay, this are updates about layoffs, but you see in June there are not many layoffs. And this is me as usual, and thank you.
Video
Video
Video
Video
Video
Video
