Vollständiges Transkript anzeigen (4.239 Wörter)
Artificial intelligence updates every Friday at 2:00 p.m. Eastern Time. Today, Friday, June 12th, a lot of updates as usual. The epigraph for today's presentation, if your agent has no memory, it has no career. And let's start with the leaderboard. We have Claude Fable 5 at the top. Which is great. So, Fable is a new model. You see, before Claude was version 4. So, you see Claude Opus 4.8, 4.6, 4.7, 4.5. Now, we have five. And it's Fable and it's Mythus, and we'll talk about it. The rest is looks approximately the same. And here, by the way, I provide information about sizes of the models. Sometimes we know the size in billions of parameters, sometimes it's approximate. So, for example, Claude Sonnet is approximated 1 trillion parameters. Opus 4.8 is 5 trillion parameters. And Fable is 6 trillion parameters. So, these are really huge models. And this kind of explains why they're also more expensive. Okay. So, this information about Claude Fable and Mythus. So, [snorts] they both are version 5. So, it's Anthropic Claude version 5. And it's Fable or Mythus. It's actually under the hood absolutely the same model, same parameters, same weight. The only difference is that Fable is made for general public. So, it is made secure. If you ask a potentially dangerous question like how to create a virus or a weapon, it will redirect to the previous model, which is Opus 4.8. But other than that, if you use it, for example, for coding, you will see that Mythus Fable and Mythus Preview, I mean it's basically the same. Well, here they actually put Mythus and Fable in one column because it is the same model. And you see the red background means it's the winner. It's the champion. And Claude Mythus Fable is champion is in almost all categories. So, this is and this is Mythus Preview. >> [laughter] >> So, comparing to other Claude and GPT and Gemini, so so this is amazing model. Now, this is about the sizing. So, you see Fable is 6 trillion, Opus 5 trillion, Sonnet 1 trillion, and Haiku this is small model is only 20 billion. So, it's much much much smaller. That's why it's stupid. Now, if you go in Claude code and you don't see a Fable yet, what what you can do, you can actually edit by just running this command, explicitly give the command model name and it will edit and then you'll be able to use it. Now, unfortunately, it's very expensive. So, the token price you see it's 10 dollars 50 dollars per million tokens input output tokens. This is twice more than Opus. Opus was 5 and 25, here we have 10 and 50. For now, it is included Mythus and not sorry, Mythus Fable is included in your subscription. If you have Pro mark subscription, you can use it, but it will only last until June 22nd. After that, they remove this model from subscription and move it to a credit plan where you basically pay via the price of API tokens. So, the more you use, the more you pay, and it's very expensive. Okay, uh this is Anthropic and CLI, which is the way to work with let's say Claude uh from the shell. So, this you can brew on Mac, brew install Anthropic tab and and then you see you can log in, and you can send messages. So, here for example, and messages create, you say which model, blah blah blah, and what So, this is the way to use Claude from shell, from Unix shell. >> [snorts] >> Uh Anthropic stops testing new mythus model after Yeah, there was a a Shauna's leak. This is already old news, but this just shows but nobody can guarantee security. Oh, this is a good milestone. Internet now has more information from bots than humans. >> [snorts] >> Uh so, this is according to Cloudflare CEO Matthew Prince. Now, Cloudflare, you know, it's a huge company. It's a content delivery network. A lot of traffic goes through it, so they have statistics, and yeah, this is This is really important milestone. Well, maybe in several years we'll have uh ratio of bot traffic to humans like 100 to 1 or 1,000 to 1. So, basically, there will be no internet. There will be conversations between bots. Uh Claude pricing. Yes, this what I want to tell you. So, right now we have three plans. We have pro, max, and max five and max 20. So, pro plan it's a monthly subscription. Uh token usage is 44,000 tokens per five-hour window. So, if you used all the tokens for this window, you have to wait until reset every five hours it resets. Right, max gives and these two max plans gives you of course more tokens. And additionally start starting June 15th, starting June 15th what will happen we will also get a monthly agent SDK credits. So for example, if you have subscription of $100 you will also get credits. And if something is not included in this subscription you [snorts] can start using credit and after it expire you can just use regular API tokens and credit will be also used at price of API tokens. So anyway, it is expensive. And what when the credits will be applied? Well, let's say when the credits will not be applied. When you have subscription and you're using Anthropic Claude native tools, like for example, you use Claude co-work, Claude desktop application, Claude CLI in the terminal, right? Claude code in the terminal, native Claude code, then it is covered by subscription. But if you use Claude code inside Visual Studio Code then it's not covered. If you're using it instead inside Z editor or any other IDE or you wrote some Python script or whatever, it will not covered because it recognizes that you're using under the hood you're using agent SDK and that means you have to pay per token, which is very very expensive. Okay, the next starting June 22nd. So oops, sorry. I'm I'm talking about this. Another bait and switch happens. Claude fable 5 will be removed from subscriptions completely. And you only can use it using credit pricing and API pricing, which is very, very expensive. Okay? Um another thing about Claude Ultra code. So, when you encode code, you can give {slash} effort Ultra code. Ultra code is a combination of two things. It's a high reasoning effort, extra high reasoning effort, and dynamic workflows. Dynamic workflows is a new thing which allows you allows Claude to spawn multiple sub agents, hundreds of them, maybe up to a thousand. Uh when you do Ultra code, it becomes very powerful, but also very expensive. Just think about hundreds of agents spending your tokens and at high reasoning effort, right? Really, really. Now, if you have Max plan, then dynamic workflows, which is multiple agents, is on by default. Uh okay, it requires the recent model Opus 480 or 47. And uh yes, okay, next. Anthropic employees about AI versus human. So, these are just some interesting quotes. Uh on days where everything works well, I can't help but think nothing I I do matters. Everything is automated and better and faster than I ever will be. But then there are days when everything breaks and I don't understand why and I realize I have no idea what I've been up to anymore. >> [laughter] >> Anyway, >> [clears throat] >> Claude partner hub. So, Claude follows Microsoft. You know, Microsoft has these certifications for people, for companies. So, uh people uh spend money on exams, on certifications, and then they become partners of Microsoft and uh Microsoft can give them some consulting gigs. And many people and companies actually make very good living on this model. So, now Anthropic doing something similar. So, they pre-created a partner hub. So, you see partnerhub.anthropic.com. And you can sign and you can pass the test and get education. They have courses. And you see select preferred global partner. These are for companies. So, if you have a company with 10 people who passed Anthropic training, so you can become like a select partner. And this like global premier, this is a huge consulting company, I guess. And they will get a lot of customers from Claude. So, this is a deep So, Anthropic creates an army of companies and developers to use their technologies. Okay. This is plug for my channel. So, on YouTube, I haven't updated these numbers, but whatever. Name of the channel Left Selector, which is my name. I talk about AI every week. I provide slides on GitHub and on Google Drive. Links under the video. And usually I pose a question under the video. Please stop the video and answer the question. Uh yes, SpaceX IPO. Uh they evaluated SpaceX at 170 or 75 billion No, I'm sorry, 1.75 trillion dollars. Whatever. Um but some other independent estimators say it's trillion dollar more than it actually worth. And actually it may be only 780 billion dollars, something like that. But regardless, it's starting trading. And it is already oversubscribed. So, there are more people who want to buy stocks and the price of the stock will go down go up. That is absolutely 100% will happen. And uh it means that Elon Musk will become first in history trillionaire. Even more than that, self-made trillionaire. Right. Uh there was a event, Apple event. Uh so this is uh uh world developer conference and uh So there is uh several things. One is of course that uh Tim Cook is leaving his position. It is his last conference. They showed the redesigned Siri and it's now powered by Google Gemini. So they were selecting between different companies, decided to go with Google. And now you have Google, so there's new operating system, new version. And it's 27, so it's 27 on phone and it's 27 on Mac. Uh well, it was not released yet. I don't see it in my base, but coming soon. And on all other platforms, you see 27 everywhere. So 27 is this new step. Uh Pentagon adds Alibaba, Baidu, UID and so on to military blacklist. 200 Chinese companies, some of the top technological companies, like for example robotics, whatever, added to this list. What it means is that military cannot do US military cannot do contracts with them. And next year they will not even be able to buy any products from these companies. I'm not sure why uh Pentagon is doing this uh making separation instead of cooperation because it already backfired when government was banning uh uh sales of Nvidia chips to China. And uh China just created their own chips and now this ban is removed, but China doesn't doesn't buy Nvidia anymore because they have their own chips and now they can run their models and train their models much faster. Anyway, Google diffusion Gemma. So Gemma is open source model and it comes in different sizes and different variants. And this particular one is experimental 26 billion parameters mixture of experts using text diffusion. Diffusion usually associated with speed. It's a faster model. So on H100 Nvidia, it can run at 1,000 tokens per second. With the four-bit floating point precision, of course, it's uh like decreased in size, but still it's very, very impressive. Uh Eli Lilly. So Lilly may become the largest company in the world. Well, right now it's already a trillion market cap, but so this is a company which makes pharmaceutical and they have done a lot of research, they have a lot of data and now they partnered with AI companies to do more like drug discovery research and they are growing very fast and uh yeah, so I think we need to see what will happen uh next. Oh, choose an agent. So before we spoke about the models and every time I show you a leaderboard about the models, but now you can also there are leaderboards about agents and you can think which agent you want and there are open source agents and closed agents. Like for example, open source, of course, you've heard about open claw Oops, sorry. And uh Hermes, for example, and whatever. You see you see a list that actually maybe much bigger. And then closed some Abacus deep agent. I spoke about it multiple times. And oh, here we have of course a Microsoft, Google, Perplexity. Perplexity computer people say it's really really good. Salesforce, IBM. Like lots of agents to choose from. Some of them are free, some of them are uh you have to pay. Uh but anyway, uh or maybe you want to create your own agent. An agent means you need to choose a harness. So agent is a horse, but you need to use some sort of harness to make this horse useful. And again, you have open harnesses and you have closed harnesses. So we use Claude code which is a closed harness. It's not uh we don't have to pay for it, it's free, but uh it is uh closed. Claude code. Uh but we have open harnesses as well. And this here is some list of open ones. Uh the benefit of open is that you have more control. And uh uh but for example, we used Claude for for a long time I used Claude, with Claude. But then we switched to Claude code and we found it's much better. Although Claude code is not open, but it's simply better tuned to work with Claude models. It's just amazingly better. So we never looked back. But if for some reason you need to make a custom agent where you need more control of what this agent is actually doing, then you may go with the open source harness. Um harness is probably more important than the model itself. Like for example, with Claude code you don't have to use Claude model, you can also use other models and they will work. A harness is what important. Okay, uh this is interesting. This is how you can make your own harness. Well, we didn't go that far. I mean me and my partner, but it is possible to create our own harness. And this describes what are the moving parts of the harness. Okay, fraud detection and AI. Now you see this AI in fraud detection, but frankly, most fraud detection today runs on classical machine learning. Isolation forest, XGBoost, plus some good feature engineering and rules. So, if you have transactions, let's say credit card transactions, there are millions and millions and millions of those transactions, you cannot really analyze them using like regular large language models, it will be very very slow. So, [snorts] you use regular scripts, like regular machine learning models, which are fast. You can still use AI for, for example, document analysis, for forgery detection, attack detection, conversational fraud probing. So, there areas where AI can be used. But usually when people say uh fraud detection in finance, this means transactional like bank operations. And it's it's still done the same way. So, you have checks, like for example, suddenly you see transaction which is coming from a different country where the person was never buying anything. So, this is a red flag. Or maybe amount is bigger than usual. Or maybe a time is unusual. So, these are all very very simple deterministic handwritten rules. And then the model can do some statistical analysis, some adaptive ML. Like I mentioned isolation forest. AWS has its version of isolation forest for a big production like enterprise deployment, which is called a random cut forest, which can process streaming of huge amount of data and adapt as it goes without stopping. So this this is really great. And many many other models can be used, but this is all kind of classical ML. Uh Stateful swarms, how persistent memory beats traditional agent architectures. And this is good article and the whole idea is you have to have some sort of persistent memory throughout your architecture. And it just call it blackboard, blackboard memory. So all agents as they work, they add like append something to this memory. This is a very very simple architecture, but it's amazingly how efficient effective it is. So you need to have some sort of memory. And that's why the way I put this slogan at the very top that the agent without memory has no future. Okay, agent arena. We used arena leaderboard for models, but now there are multiple arenas also for agents. And here's example arena AI leaderboard agent, right? It's live. It already has a lot of agents and a lot of data. So here are links on some other leaderboards. You can compare agents when you select them. Okay, Google Dream Beans app. This is maybe a small thing. What what it does overnight, it creates 10 14 beans for you. Bean is a small piece of information like curated personalized story card based on your data and photos and so on. So this is nice to have, right? This is very good conversation with Linus Torvalds about AI and how it is used in development of Linux kernel and his personal experience. And he is generally positive. He likes the technology. He likes to play with it. He likes how it may be effective, especially in debugging. It finds bugs. But what happened before, if somebody would find a bug, you can keep this bug a secret until you fix it. But now, because many people using AI, they all find this bug at the same time and they all send a message about it. So, the mailing lists are getting flooded with redundant bug reports. So, it kind [snorts] of changes the dynamics of how you work with the code and how you debugging of the code. Okay, AI in Snowflake and the next will be Databricks. So, Snowflake is SQL database which keeps data in Amazon S3 buckets, right? And it's generally standard SQL data warehouse. But of course, for the marketing nowadays, all people say that they are AI first. So, they put their Cortex code which is for coding inside. They have co-work. Remember Claude co-work? Similar. They have MCP servers. So, you have data and in front of your data tables, you can put MCP layer. They have Cortex code CLI, the command line interface like Claude code CLI or Gemini CLI. They have traditional machine learning models there. So, you have a table, you have like let's say you have two columns, X and Y, and you can do, I don't know, linear regression and put it in column Z, something like that. What's they claim which is good that it follows all the governance, privacy rules, and it's all running inside Snowflake governance perimeter, which is important. Um I don't believe uh Snowflake should be first choice to run uh AI. Uh it's a database. Database can always uh sit separately. Now, Databricks on the other side is much better suited for that. And uh just to remind you the story about uh Databricks uh is founded in 2013 by researchers in Berkeley. So, some young guys uh decided, well, Hadoop is slow. Hadoop uh allows you to create clusters of servers and uh uh do some distributed uh work on data. The idea actually came from uh Google Bigtable, where you have a query in the browser. You're searching for something, and this query mapped onto millions of servers, and these servers Each server uh contains information about some websites, and it tells, yeah, I found it on this website, and they rank it. And then all this information come to the center and uh kind of reduced in one prioritized list. So, it's a map reduce operation. You map to outside servers, and then you reduce it back to the center and show it in the browser, in the search. So, this is kind of a reproduction of the same map reduce idea, but the Hadoop implementation was in Java. It was working on CSV files. It was very, very slow. And um this team from uh Berkeley, they decided, uh why we saving files all the time? We saving files, we reading files. Let's do it on network without saving mostly. And they achieved 30 to 100 times um speed up. And they called their system Spark, and then and put it in the open in Apache Foundation, so this is Apache Spark. Uh And this this is a great great uh software, and then they formed a company uh 4 years after that, which is called Databricks. Uh today Databricks has I don't know, maybe 10,000 people. Yeah. 10,000 10,000 employees. It's a big company. It's growing very fast. So, what it allows you to do, it allows you to uh process data which is distributed between many many many files, and you can use CSV files, you can use parquet files, and recently they created something which is called delta files. And the idea is you want to support uh transactions. So, you want to do something and then either commit it or roll back. And return. So, it's kind of like having control Z with your data. Uh very good technology, and it's very fast. It may run uh on huge clusters in the distributed, and it's not expensive, frankly. So, they're very smart people, and uh you can run scripts, you can run notebooks like kind of Jupyter notebooks. Uh you can have data frames which similar to pandas data frames, but under the hood they are distributed and on many many different files and many many different servers. So, very very good technology. Company is growing very fast. Uh you can work with data using SQL. So, it's basically like a SQL database. It has SQL, it has transactional support, it understands all the permissioning, and uh you see multi-layer security with full R back and fine-grained visibility. This is a role-based access control, grant revoke, and so on. So, it is enterprise compliant. It's great thing. So, you see here Snowflake versus Databricks. Snowflake is you mostly running a SQL to make reports or whatever. So, you're just doing some business intelligence. Whereas, Databricks, if you need some serious machine learning, a large language model, agents, pipelines, and so on, Databricks, of course, is is better choice. But, you don't have to to use either one of them. You can just run Python scripts. Okay, build agent workflow pipeline. So, if you want to have a workflow pipeline, there are multiple sources, open source and commercially available. Some of them, well, for example, Microsoft, Zapier, very very famous, Make, LangChain, n8n, people love it, AutoGen. You see, there's a big list to choose from. So, if you want to automate make some sort of workflow with nodes, yeah, multiple things to choose from. Difficult to choose. Okay. Oh, this is interesting idea. You know how how Microsoft, uh, they usually not very good at creating new stuff, but they're very good at buying or somehow consuming technology created by other companies, other people, making their own, and and then using it, making money. So, this is the ultimate example of this strategy. So, these are three brothers. They all are self-made billionaires, and they're working together. Uh, you see, copycats. And what their model is very simple. Find a hot US startup, copy it fast, launch it in Europe, and then sell back to the original company. So, in >> [laughter] >> like 27 years ago, they cloned eBay, called it Alando, and sold it back to eBay just 100 days later for $43 million. So, then they did a similar thing for a mobile content platform, Verizon, sold it to Verizon, and so on, so on. You see Groupon, many, many if if you go to Wikipedia page, you will see like dozens of companies which they start. It's not really engineering and it's not AI, it's more like a business strategy. And it's very stable because they don't take risk. They take something which they know already works, and they just apply it to a different market. That's all they do. Very easy to copy. Okay, jobs by demand. Here on the left, you see list of jobs uh for AI. And you see that there are only two jobs in this list which are not technical. One is AI product manager, and another is chief AI officer. Everything else is basically engineer. There is a lot of demand for uh engineer plus AI. You see AI engineer, applied machine learning engineer, ML ops, AI solutions architect, data AI focused AI. Same thing. So, the combination of AI and software is in high demand, and generally it's also paid more. So, there are multiple sources. So, you can expect 10, 15, 20% more simply if you add the word AI to your resume. Okay? Uh next, this is statistics about jobs. June have started, not many layoffs yet. Okay, this is me as usual, and thank you.
Video
