Vollständiges Transkript anzeigen (5.648 Wörter)
Artificial intelligence updates every Friday at 2:00 p.m. Eastern time. Today, Friday, July 24th, a lot of updates. I put this quote from Arnold Schwarzenegger. The world's the worst thing I can be is the same as everybody else. Well, most people are working. We need to do our own businesses, guys. So, let's look at the leaderboard. uh you see that for chats and for coding for definitely the blue is claude and claude fable is on the top on the other side we have now new uh powerful Chinese models like quen 3.8 max preview and kim3 which are big 2.4 trillion parameters and 2.8 8 trillion parameters. Well, Claude Fable uh somewhere I saw the estimates like seven six trillion parameters. So, don't know. Um anyway, and these models are here. So, if you look at uh let's say Quinn 3.8 it should be 3.7. Where is 3.8? I thought I saw it somewhere. But but but anyway, um doesn't matter. And Kim K3, they both are very good. And now people comparing uh Chinese models between themselves because they're basically on the same level as American. So why would you use American uh unless you are in the government and not allowed to use Chinese models? Okay, this again u from artificial analysis.ai cost per intelligence. So on the right we have the most expensive uh clot fable. So what happens here? All the models solve the same task and then uh it shows uh how much money you had to spend on this task and uh so fable is like $2.75. If you go to GPT56 soul extra high it is only 68. So it's like four times less. And if you go to high it is only 45 cents. And then you go down to Chinese. So and uh this is 5.2 is very very powerful and then you go cloth haiku but cloth haiku is not really good model but then you have deepsek which is extremely affordable and so this is for actually delivering uh the result quite amazing. Um another thing I I want to show that GLM and GPT the cost per token uh for GLM is like half of the GPT but when they actually solving the real problem it turns out they use more tokens so the price is kind of comparable. So this this is a very very useful chart from this website artificial analysis. Okay. Antropic has a lot of updates this week. First they finally said that they will not switch off fable 5 it becomes a per permanent part in subscription and uh so in max and team premium uh but only up to 50% of their usual weekly usage limits so good I'm using it using it every day enjoying it so it's not bundled for pro and team standard uh plans and uh yeah cloth sonet 5 was released several weeks ago. Oppus 5 not announced yet. Uh, Haiku is still at 4.5. Antropic added an inapp browser to CL code. Uh, offers some good programs for teachers. Added voice mode to clo. So now you can talk to it and for example tell it to do something in your Google workspace. Gmail, Google calendar, Slack, Canva, notion, uh, multimodel or not. This is just interesting thinking. Antropic doesn't focus on multimodel. Well, they added uh voice support but uh still they're not generating like videos or whatever. Open AI they used to have omni models and they were removing them and they removed the sort of video and they pivoting to the same thing as entropic. So coding focused analytical Google Gemini they announced in May the omni like omni flash. So this is multimodal video generation and of course Chinese create a lot of uh multimodal models. Okay. Uh skill recording in clo uh this is very very powerful. So imagine that you have some system which you want your clot to work with but this system doesn't have API or MCP. But if you can record the video using this skill recording in Claude uh application on Mac, they have this new feature. You just press on record the skill and you walk uh you record and you do voice over. You explain what you're doing. So uh Claude will make a skill out of this and next time you can just run this skill. It's really really amazing because no APIs, no MCPS are required. It's it's not technical and uh it just empowers people tremendously. And of course, once you recorded it here, you probably can copy it and use it from your prompt in CL code or something. I haven't tested it yet, but it sounds great. Okay. Alibaba Quen 3.8 Max was released on July 19th. And also we have quen 3 uh 8 uh oh no 38 we also have Kim 3 and this video it actually compares them side by side on multiple uh use cases. Uh generally Quen is a little bit better but they are really really comparable. So very powerful models and much cheaper than American models. Well, Quen is Alibaba and Alibaba is part of the black uh list by US government. So, government cannot use it. But uh Kim is still available for everyone. Okay. Open code and uh uh clo. So with open code you use something called omni routout and this allows you to use multiple models and uh if uh you're using some model you have subscription and then it runs out it automatically switches to another model so you never have outages right and uh yeah it allows you to save a lot of tokens allows you to save a lot of costs um Google AI I mode in search uh you know in open u Google search and they have now AI mode and uh you can link external applications there. So uh for example you can ask uh generate a grocery list and then push them into your Instacart cart. So Instacart is the place where you can order groceries, right? Ask AI mode for design options and then uh push it into Canva. Ask for playlist and creates and saves it in your YouTube music library. So it's now directly connects from Google search to some applications. Uh Google new flash 3.6 and flashlight 3.5 models were just released. Uh so what's interesting that it actually shifts a lot of work uh real daily work from uh pro model to flash model which is cheaper faster uh 36 flash cuts output tokens use and pricing you see well it it's not very cheap but uh well it's actually cheap much cheaper if you think about fable which is $50 per million tokens in their output tokens. So, this is input tokens and output tokens. And the flashlight uh is really fast and really cheap. Okay. Uh okay. This is about my channel. I actually probably haven't updated it yet, but yes, I have more than 7,000 subscribers, close to 300 videos. Name of the channel, Le Selector. I provide uh on GitHub and uh Google Drive uh the uh slides, so you can download and click on the links. uh and please please uh provide some comments and questions under the video. Okay, this is a very interesting scandal. So what happened uh OpenAI uh created uh new models and they were testing them. There are benchmarks for red teaming. So you know red team and blue team uh red team tries to hack the website and blue team tries to defend right and uh so they doing red teaming and uh so they did some task and uh they uh model decided so so it's easier for them to cheat. So instead of actually trying to solve the problem just go to hugging face and find uh because hugging face has data sets and models and it they have benchmarks questions and answers. So they decided to get to this questions and answers [laughter] and just cheat and they did it. So they somehow got out of the secure environment uh hacked uh the hugging face and and got what they wanted. What's interesting that Hugging Face is all open source, right? Uh so there there is nothing to get there. Well, they probably have some credit card information where they paid subscribers, but uh anyway, the model was not looking for that. The model were looking for specific like answers to questions and they got it. But what's interesting here that when hugging face people realized what's happening they decided to use uh claude and uh uh GPT uh to figure out and close the holes right and they couldn't you see they tried analyzing the attack using US proprietary frontier models like GPT and CL but their safety guardrails block the security payloads so what what did the hugging face do they also host Chinese models so they use Chinese [laughter] models to resolve it and you see this is they use like GLM5 or something. So it's interesting on one side uh government and anthropic and open AI they say we have to have regulations we have to be strict we should not allow like we we have to be safe and so on but when the problem happens suddenly they cannot help and like hugging face run and use Chinese models [laughter] so it's it's really a scandal um okay um open AI chart GPT work uh for fast conversational help like questions, brainstorming and quick drafting. Uh chat GPT is for longer multi-step tasks that end in a finished uh deliverable. So you see this is the chat and you have you can click between chat and work. So uh this is my my screen. Uh I I just opened uh chat GPT in the browser and uh so now if you want uh your chat to work more like an agent to do multi-step to complete the whole task, you just switch to work. Okay. Poolside AI released the new model Laguna S 2.1. So Poolside is a startup from San Francisco. Uh so they have about 60 people. Well, they actually have more because they have people in Europe and they created this new open weight model 118 billion parameters mixture of experts and only 8 billion active. So, it's a small uh it doesn't require a lot of memory. You can run it on like RTX 3090 which is 24 gig memory. DJX Spark of course high memory marks. So, it's pretty good model. They provide a lot of graphs. I just show here to compare like for example this is deepseek and you see that this model actually beats deepseek whatever this is kim so it's not as good as kim but whatever it's a good model it's uh free uh context length 1 million tokens wow right so it is really great model okay open code versus clo code u a lot of good reviews about u open code So uh you know that cloud code uh is written and maintained by entropic and they provide it uh for also for Python and for TypeScript as libraries and and it's pretty good. So for Python it's something called clo agent SDK but open code is completely open source model agnostic lower cost CLA first. So uh it is designed to work in the terminal but uh there are also free available SDKs uh so this is example of Python SDK and uh and and it's written in go sorry I didn't I didn't say that so this is example how you can invoke it using subprocess like a like a terminal and here uh how you can invoke it uh from a python so for example streaming You see with open code as AI for chunk in so you you print the response as it appears uh it's it has its own harness uh so it's kind of like uh so it can do uh make a plan and then loop through the plan execute it verify it can do everything okay uh Google Gemini and Gemini notebook uh versus fable so notebook LM is renamed It's now called Gemini Notebook. Uh so can do deep web research, multimodel analysis, content generation, blog post and so on. Digest massive PDFs. Compare up to 50 documents to find conflicting facts and figures. So yeah, these are good updates uh from uh Google Gemini. Thank you. Selecting right tokens for specific tasks. Guys, I love this video. Uh Matthew Berman, uh he's excellent. So this video, YouTube video here, please please listen to it. Uh there's so much depth and understanding and clarity. So what he is explaining that uh low cost per token uh may require twice as uh many tokens, right? So the result and I already showed it on the diagram uh before that don't just look at the price of tokens you have to look at the overall uh [snorts] uh cost per solving the problem for delivering the result and also they work with different speed. Uh so this may be also a factor. Uh so the way he works he splits uh his work in three steps and this this is again this is genius. So first he uses a top tier model for high level planning. So it creates a step-by-step plan. Then he uses fast and cheap model to execute code. He actually is is using uh like GPT 5.6. So it's not the cheapest and uh it's actually pretty good model. And then on the third step he again uses a top model but different maybe for do doing the code review to confirm that everything is done correctly. So one two three three steps. Um output tokens uh mostly produced on the second step where the model writes the code and uh they cost significantly more than input tokens. So here the cheapest model in this three steps is on the second step. Okay. Uh oh yeah this this is amazing. So um Julia McCoy, you know I I talk about her uh last year she was estimating her business at about million a year. Now it's already 3 million a year. And uh she is creating her own avatar. So she creating videos where you see her but it's not her. it's automatically generated video and u she we're using Hagen service to create those avatars and 11 labs w which is very famous for cloning the voice right so she would do some research using uh claude uh to do web research and generating the scripts for the video but then in order to actually create the video she needed a human to work with the heen uh to use the her pieces of videos, her face, her voice to and the script to put it all together in one video. [snorts] But now she's changing what she's doing because there is a service called Hicksfield. So it's a hickfield.ai and this is a platform where you can go and generate short video clips. They will be maybe 10 15 seconds long. They can actually u join several of those together to make what they call a flow and uh so you can generate multiple videos and then uh um so you upload again your images, pieces of video, your voice and then uh Hicksville can do that and then you can if you want like for example if I want to generate my uh video like this 20 minutes 30 minutes I will have to record many small pieces then I can download them on my computer and then I can use ffmpeg to join them into one video. Right? What's interesting that Hicksfield now has MCP connector inside CLA. So you can actually script everything. Uh you can write uh tell clude uh what you want to do. CL will find make uh uh let's say slides u make text and then use hickfield to generate clips and then download them and then use ffmpeg to join them together. So the whole production can be completely automated absolutely amazing and she she was well she was not talking about ffmpeg but she was talking about using claude working with hickfield. Okay. Alama uh better tool calling and speed. When people uh use Alama, they usually complain that they run the local model, the quality of the model is not very good. It goes into loops. It uh it doesn't verify. It gives you wrong answer and uh doesn't even know that it gave the wrong answer. But now what happened? They added a tool calling much better with Gemma 4. And now you can actually make workflows. Uh it it works like an agent. It can verify the results and self-correct. And also they made it run much faster about twice faster on uh Apple silicon using MLX uh library. So this is this is absolutely great. uh we'll probably see more and more usage of local models uh using Alama uh Gemini AI uh can generate slides now so uh in uh Google workspace it can use docs, sheets, PDFs, email content, Gmail content and uh take all these multiple documents and somehow combine them into uh Google Slides. Uh next, uh Mark Pinus. Yeah, he he's a very famous person and he just explains how you come up with ideas for new applications. So he creates two lists, the ones things you love and you hate and then you think how you can improve the ones which you hate using some features from the ones you love. So this is things you love, things you hate and how you creatively use it. Uh this this is a very good uh presentation. Oh, I didn't put the link. Sorry. Uh AI transforming a six sock development. So this is about uh generating new chips and uh when people generate new chips, they actually use AI to do this. So ASIC is application specific integrated a custom chip to perform some functions as sock system on chip. So problems increasingly complex hydrogenous and uh chiplet based architectures solution use AI for design verification and optimization power performance and area. Uh so there are tools generative AI tools translate natural language intent into verifying templates machine learning driven uh on designs emerging trends focus on domains specialized models tightly integrated into existing flows okay uh so they analyze debugging and and so on so it's automated design it's AIdriven design machine learning driven design um so things are changing very fast here AI eye on small devices. This is area which usually I don't cover but somebody asked me to say a few words about it. So for example uh light RTLM is a crossplatform ondevice LLM stack. So you you on different types of mobile devices function Gemma 270 million parameters reportedly reach nearly 2,000 tokens per second on Pixel 7. It's absolutely amazing. Uh that speed on a on a on a phone fine-tuning improved fixed intent accuracy from 46 to 90% on eight of 10 functions. Gemma 412B positioned for local multimodel agentic workflows. Apple ferret UI light reportedly showed that three billion parameter model can match or exceed much larger systems. Droid run LLM agnostic mobile agent for controlling Android and iOS devices through natural language and multi-step automation. So these are just few of many uh very fast developing area. Okay. VIP coding uh six principles. This this is actually in Russian. I really uh like like her [snorts] I don't remember her name but uh I really like her videos. I've subscribed to her. So, uh, plan before prompting, direct the AI to ask clarifying questions before building, build in small steps, test after each step. Oh, this I definitely do. Uh, copy raw error messages directly into the prompt. Yes, copy paste. Roll back to previous working versions. If code breaks, you have to have checkpoints. I always ask when it goes step by step write the progress and say I'm starting this I'm doing this I I did this this is successful tested next step so then if I don't know you ran out of electricity you need to restart it it will restart from where it finished last start fresh if AI cannot fix it in three tries okay so these are all good common uh sense and advice don't ship skills without evals uh very good video uh that uh when you create a skill uh how do you know that the skill is actually good? So you need to systematically evaluate uh that skills are actually doing what you want them to do. Okay. Uh web server caddy versus ngx. I was using engineext for many years. Uh this is uh open source. um everywhere like on Amazon cloud Google like it's used everywhere for web server reverse proxy high performance very good but I was recently introduced to caddy and caddy uh maybe not that powerful but for regular applications it's absolutely amazing and it makes your life easy because configuration file is smaller it takes care um of uh certificates It's like secure certificates, renewals. So it just the configuration setup is so much simpler. So you see engineext is still better at hypers scale. It's written in C and EngineX consumes quarter of the memory of Kadi. So it's a smaller under idle connections and more efficient. But for your personal let's say website application, Kadi is probably better choice because it's just easier to work with it. Calibri tiny inference engine. Well, this is kind of a joke. It's a uh 1300 lines of C which allows you to run models even very big models uh on your computer. It will be slow like maybe one/10enth or one token per second but it just shows that it's possible to do okay use RAM as VRAM. So this is a hack on Linux you in uh using AMD based system uh laptop. So it's a AMD uh CPU and you can configure and run it using RAM instead of uh as if it is GPU. Again it's just a hack [snorts] uh clo versus codeex. uh many people compare them and also uh uh use them together finding that uh they actually enhance each other because they have different strengths. So Codex offers better value for money, generous token limits and practical automation features and very good for front-end design. CL code excels at complex multifile projects and maintaining deep memory during the programming sessions. So uh well probably use both use both together use them to verify each other's work or to complement each other. Okay. Uh seven layer developer stack for one person uh company. So he explains how this this is basically like standard VIP coding. So it's uh uh JavaScript or TypeScript based. So you're using like cursor code. You're using some uh framework like next GJ GS JavaScript built- on react hosting Versel Railway database superbase or neon authentication. So it's like Lego pieces. This is how usually VIP coding works. Uh we created our own agent but we don't use any of this. It's Python based and it's completely different and a regular database. Okay. Uh bottleneck the physical and mental limits of a solo founder. Single person must still handle late night server issues. Uh start up true competitive edge lights and the founders taste judgment and distribution. Yes, this is very true. Okay. Hermas agent updates. You see a lot of updates. uh Hermas agent u just to remind you there is a open claw which is in typescript and uh it's a foundation under open AI and there is a hermas agent uh from news research a company which was established 3 years ago and uh it's extremely popular and it has features which uh as it runs it self-improves it rewrites its own skills with time it has memory and uh you see this new uh updates uh cloud uh hosted so you can um start it uh very fast on on the cloud multi- aent systems obsidian shared memory uh remember Andre Karpath was explaining you you're creating multiple MD files they're interlink so everything is based on files uh no vector database uh more than 200 AI models supported automated outreach capabilities using Hunter API by daily news curation by Oracle. It's not Oracle database, it's Oracle system. Trend tracking and analysis uh voice control integration, automated video creation tools. Uh now it starts much faster. You see from uh what's about what five times faster uh it shows streaming thinking. It it creates checkpointing in the background. [snorts] Profile routing individual channels use specialized agents, models, memory. So you can configure all this security and safety updates and the model catalog expands like which models it can use. It became uh bigger. So a lot of updates. Hermas agent is a very very good system. Okay. Um a little bit about pandas, duck db, polars and so on. Um almost everybody who is doing data science worked with pandas. It's a python library. But now we have duck db uh which is uh column database. So it's like pandas. So you do import duck db and then here is an example. You do select something from and you provide CSV file or you can provide parquet file or JSON file. It uses files as tables in in the database and you don't need to first load it in memory. uh so you the file can be huge it may not fit in memory but it doesn't matter uh duct DB will handle it and uh it can actually handle even like pabytes of data and uh the files can be on S3 uh in Amazon cloud and you can easily uh take the results and the results are like the lazy results and you can convert them let's say in pandas data frame so Here's a example summary in dataf frame in pandas equals summary relational df. Right? It's very easy to convert back and forth. Uh now there's also polars library which is written in rust and it's lazy execution vectorzed so it's very very fast. Uh it's a substitution for pandas. Uh duct db is embedded analytics engine right. Pandas is a great for small and legacy application language modeling and uh plotting because uh all the standard data science libraries they understand pandas. There's also spark, dusk and data fusion for distributed uh data like Apache Spark and Apache data fusion. Okay. Automated routing to pick best models for a given task. uh this becomes uh very very popular and uh different systems uh provide you more or less flexibility of doing this. Uh for cloud code allows you to define let's say one model like fable for strategic planning and then another model you define for actually writing. But um other harnesses gives you maybe more flexibility and you can do more fine grained uh configuration which models you use for what. Um Google Vids, Gemini, Omni and uh personal avatars. So generates and edit video clips directly from text prom sketches and existing footage. Can refine lighting background and so on. So now you create can create videos using Google Vids. Okay. OpenAI GPT red uh red teaming system. I am curious if uh this uh uh hacking uh hugging face which happened maybe they use this GPT red model or something. So this is uh model um specifically for red teaming for attacking uh Linux versus Windows. Um so Windows 11 have a lot of problems and Linux uh wins right and gaining more and more so people using uh Linux more and more uh designing effective AI agent systems. So this is a very good uh conversation. Uh so the invited speaker is uh from Antropics cloth code and he emphasizes structuring agent loops around clear goals explains how dynamic multi-agent workflows enable parallel execution automated video editing example discussion covers clo tag a slack integrated system key advice is to keep context window tight and develop a deep understanding of system limits rather than relying on prompting uh yes If if you load too much uh stuff in the context window or too many skills, it absolutely confuses the model and the model cannot operate. Yes. Uh Google Turbukquant vector compression. I think I spoke about it uh previously but uh uh this is the way so you can effectively shrink your model in this example from 31 GB down to 4 GB and they're using some special techniques to do that. Okay. Uh Buzz opensource group chat workspace. So this is Jack Dorsey. He at some point co-ounded Twitter. So it was like 20 years ago. Now it is XCOM. His other major project is square later renamed into block and uh so they doing mobile payments and small business card processing and they released buzz. So buzz is a opensource alternative to slack. So it's a group chat and workspace and uh yeah so instead of slack you can use this it replace parts of Slack and GitHub for team collaboration designed for humans and AI agents to work together. So AI agent may be your team member. Buzz is a model agnostic, decentralized, self-service, whatever. It's I guess it's a great thing. So maybe we all should switch from Slack to to Buzz. Okay. Governments win Windows uh to Linux. This is interesting. So France is moving 2.5 million government uh computers uh workstations to Linux. And they do it because before they were using Windows but you know with the current US um administration they created rules so they uh can compel any US-based company to release their data and uh that means that if France is running Windows uh which is Microsoft which is US company which may be compelled by US government to release data. This is security risk for France. So they decided that they move all 2.5 million workstations uh to Linux right another example is u state of Germany they migrating 30,000 uh people to Linux and Libri office so no more Microsoft office and uh there are some other examples uh and uh China they moving from uh Windows to Chinese-made Linux- based systems again to Linux uh some governments uh started using Apple computers like in Germany some big US agencies like NASA uh department of I don't know energy maybe uh uh national institute of health and parts of FBI uh so big companies also switching uh Windows to Linux so Google uh is using Linux based desktops internally or described as gubuntu Google, Ubuntu. Okay. NASA contractor United Space Airlines moved key functionality from Windows to Debian based Linux turn which is research center big uh in Europe uses thousands of Linux desktops, scientific Linux, Ubuntu and so on. So you see this is a big trend to uh remove Windows and it's already uh on the government level. OpenAI has acquired Astral. I love their products. So they started uh 22 rough which is a Python llinter and formatter uh meaning uh it uh makes your uh Python uh look good. And UV that's what we use to install packages. Uh this is like the main tool. You start your project with using UV and uh uh configuring your environment. TY um this is the type checker. So this is example. So you see for example UV uv init uh my project UV add uh whatever um packages I need UV run and you you can run you can do so so much with UV command and same with type check and same with rough. Okay. And this is the founder of the company. So his name is Charlie Marsh and OpenAI has acquired his company. So this is uh so they're here in Brooklyn, New York about more than 11 people. I don't know exactly. Okay. Uh KMS key management system. Uh this is for security. If if you will uh deploy uh your project on cloud uh so you need to encrypt your data. For this you need to have a key for encryption key and then you will need to store it in a specialized device using one more level of encryption and uh so this is generally how it works. You have uh key management uh system to store encryption to for the keys which used for encryption to encrypt your data. uh in interesting that all clouds using similar approach and similar abbreviations. HSM hardware security model. It's a physical device designed to securely generate, store and use cryptographic keys. Uh so it's not just software, it's also specialized hardware. Uh okay, remove slope from writing. So this is a a GitHub repo which gives you uh skill to remove common phrases, common patterns uh uh from writing. Um I use similar things and was creating similar things to humanize the text but I think this is a good starting point. Okay. Clot code uh plus omni root. Uh so this is uh interesting example. So what the person does it it uses clo code but not the official clo code but remember clo was exposed uh back in March uh like and now there are many clones and you can use uh omni root to automatically switch to another available model to prevent workflow interruptions. Okay, so this is yet another Oh, okay. So, this is about jobs. This is July and these are the layoffs in July. Uh, a lot from Microsoft. And this is me as usual. And thank you.
Video
Video
Video
Video
Video
Video
