Vollständiges Transkript anzeigen (7.131 Wörter)
So artificial intelligence updates every Friday at 2 p.m. Eastern time. Today is Friday, September 4. Uh we will have long weekend because of the Labor Day on Monday. Uh just want to say about Labor Day. Uh it started in 1882. There was a demonstration of professional workers uh in New York City. You can look at the Wikipedia. It's actually very interesting. Uh there is a list of trades uh professions and most of these professions don't exist any longer like for example their profession to move uh goods between ships and uh shore [laughter] like [gasps] anyway so um today's uh epigraph is a monthly bill went from 200 something to $4 and you run the routine locally send the hard problems out. So the local models uh uh now become so good that for many routine tasks you can just use local model. You still need good frontier model, cloud model for coding for something important but for routine tasks you definitely can slash your bill by using local models. Uh anyway a lot of updates uh let's uh next slide. So these are uh leaderboard examples uh for September 2nd which is 2 days ago. uh the latest models are not presented like yesterday for example uh open AAI released uh GPT6 uh Astra it's not here yet but uh what I want to show you CL fable 5 is pretty good right then Gemini 38 flash it is really good it's it's close to Fable and it is much cheaper and much faster so Google did very good job with this model uh if you haven't looked at it. Gemini 3.8 flash. I highly recommend uh cloth oppus 5. Uh that's what I now use for coding and uh 5 max and cloth sonet 5. You see it's all the way down. Surprisingly the previous versions of cloth oppus are actually better on benchmarks than the new set. Anyway uh let's go next slide. So this is cost per intelligence index. We've see it every week and uh you see a lot of models uh at the left side which are not expensive and uh for doing the same intelligent work. So so that's the point uh you can use local model or you can use not expensive uh cloud model. Okay. Uh GPT 6 Astra. So this is the main thing this week. People are really really excited. This is the new flagship frontier model from OpenAI. And uh so here for example hallucination you see these are hallucination uh level for any hallucination and this is persist remember at some point I was telling about different kinds of hallucinations but look these top models these are different GPTs 55 56 and and so on and this is the new model the Astra the level of hallucination is much much smaller it's very very impressive uh the only problem with this model is the price. So, it's $50 per million output tokens. This is the same price as cloth fable. So, it's very expensive. Um, but it's good. Okay. So, Gemini 3.8 flash and 3.8 flash cyber. Cyber is specific for cyber security. And u I I I put here the copy of this uh leaderboard which we've seen just to show how Gemini 3.8 flash how good it is. It has big context window. It's like it's it's really good like for for everything. Uh it supports text, images, video, audio, PDFs, code inputs like aentic workflows. Um really good recommend to try. Clot fable was updated uh from 5 to 5.1. Um it is improved and it is cheaper depending on the task somewhere between 25 and 45% cheaper. And mostly this drop in price comes from a cheaper caching. Uh next quen 38 max 0902. So this is September 02 days ago. So this is the latest version of quen 3.8 max. Uh it's a big model. Uh so it's in the cloud 2.4 trillion parameters 1 million token context and so on. And you see how inexpensive it is in comparison uh with Fable or with GPT6. It's only $6 per million whereas the American ones is $50 per million. And it's it's a very good model. Uh next, Antropic sued the Pentagon. Uh if you remember, it was in March and then eventually after what about half a year, they won. So entropic one uh judge has overturned and uh the government uh anyway uh next biases in LLM evaluation. This is very interesting. So models can reproduce gender linked outcomes without receiving gender as explicit feature. So if even if you don't tell is it a male or female you're talking about uh and this by the way is for financial industry uh uh LLM absorb patterns and LLM judges can favor answer position verbosity formatting cited look authority styles and so on so on. So they somehow can use some cues to figure out is it about men or woman. Uh and uh well there are ways to mitigate this. Uh but I just want you to be aware of this problem. Okay. GLM 53 flash. So it's open router. There was a mystery model Ox Alpha that everybody was so excited about what it is and it turned out that it's JM JLM 53 flash. It's a very good model from GPAI or ZI. So it's open weight MIT license multimodel mixture of experts. Not big you see only 320 billion total parameters and million input context and 131 uh,000 output. So good and you you can use it from many different sources. You can for example here it's from Olama. So example using uh flash version and and the full version uh flash is smaller. Flash is 320. Uh the bigger one is like yeah 753 billion parameters the big one. Uh next uh Apple local AI strategy. So Apple uh as you know releases for example uh Mac Mini and big companies uh the vendors they actually buy a lot of these Mac minis for for their work and Apple becomes like a hardware provider for training tuning and using uh the models. Uh so uh you can run model locally and you can still route difficult high stakes to to the cloud. Okay. And this examples Mike Mini is less than thousand. Max Studio is 2 and a half thousand and M5 Ultra is 5 a half,000 but it gets 512 GB of memory. That means it can run anything. So uh next uh Tencent High4 preview. So this is a generic model. It's open weight and software, documents, data analysis, games, research, workflows. 10 cent is a huge uh Chinese company. So this is 770 billion parameters. So it's not small and 1 million token context which is standard nowadays uh available from different sources. And there is a mixed uh precision GGF file which uh shrinks the bigger model from uh 1.4 TB to only 200 214 GB uh allocating lower precision select selectively between parts of the model. At this size it can actually run on your home computer well with enough uh memory of course provided. Uh benchmark losses appear modest. Uh okay. Um open claw 2.0. Uh so open claw remember it was a big event in January February. It's opensource AI agent platform. For several weeks they were not doing any releases and finally they did a release and this is a major major rebuild. um this one release has 16,000 uh pull requests and uh almost thousand contributors to this release. So it's a big thing. So it's a simpler setup. New browser workspace combines. So before it was mostly CLI, now it's mostly in the browser and combine sessions, tasks, files, terminal access, live agent progress rather than treating the web UI as an add-on. Uh shared cloud sessions let trust collaborators join the same live agent context. So it's kind of multiplayer though they are not hostile user security boundary. Persistent uh sessions searchable history cross device execution. So this is a big big big um event update um open claw uh people were suspicious about it especially because of security problems. So we'll see how how this will go. It looks very good. Uh next, uh Perplexity search API. I love Perplexity. I use Perplexity for all my searches. I don't use Google anymore, frankly. And uh Perplexity now released the API. And uh so you can use you can automate uh the searches and on the benchmarks for the searches uh they took top three positions on artificial analysis independent search index. Uh so this As I suspected, this is really the best search tool and now you can automate it. Okay. Uh prompting clot 5 seven rules. Um give the model the entire job, not step-by-step instructions. So as model becomes smarter, uh you don't need to micromanage them. You need actually to explain what you want, why you want it. Uh use an interview me process to clarify complex task. Explain why the task matters, not only what to do. So don't micromanage. Explain. Define what done looks like. Include scope and output style. Replace hard rules with reasoned guardrails. Okay. Avoid unnecessary double-cheing. Think step by step. It knows [snorts] that it's supposed to do it. Forced reasoning and aggressive emphasis prompts. A fix the model's voice with concise uh standing instructions. focused brief plain language and low jargon. Okay. Um Meta Muse model. So Meta former Facebook uh released a lot of models recently. Uh so Muse Spark which is their main model was upgraded from 1.2 to 1.3. Uh so coding tool use longer running. It it behaves very good on benchmarks as we seen in the beginning. And uh uh next uh this other models like muse code specifically for terminal and uh uh CI coding agent and SDK also. Uh Muse image generates images very very reasonable price and Muse voice transcribe. So this is realtime audio uh streaming recognition. This is like really good thing. We spoke about it I think. Okay, this is plug for my channel and the I have uh regular videos and I now we do the short videos. Uh we have more than 7,000 subscribers, more than 300 videos. Name of the channel is my name left selector. The slides are provided uh from GitHub and Google Drive in the links below the video. And please please uh stop the video and give some feedback. Uh either say thank you or ask uh how to change and what you want to hear about. Uh anyway uh Nana Claw secure agent automation. It's a lightweight open-source alternative to open claw. We just spoke about open claw big release. Nanoclaw is a small one and it is uh built to be secure. Each active agent runs an isolated Docker container limiting access to only explicitly mounted files and reduce host level risk. Its smaller auditable code base makes it easier to customize agents, skills, schedules and so on. It can connect with messaging platforms but a custom local GUI can replace Discordentric control with scheduled bots. So it's a good thing. Just remember this nano claw. Look at it. It may be very useful. uh websites will block AI agents by default on September 15th. So this is coming from Cloudflare. As you know, Cloudflare is this huge company which hosts a lot of websites and now they create this blanket rule that these websites will block AI agents by default. Uh okay. Uh agent and training crawlers get blocked by default on ad monetized pages of new domains. Okay. clo memory dreams feature for managed agents. So it's not for your uh clo code like for what you're running on your laptop. This is for managed agents. Managed agents are running in u entropic cloud. Uh so it's uh entropic managed cloud sandbox. So it is research preview but the idea of dreams is to consolidate the memory. So as you work the it remembers more and more and more history and at certain point it needs to spend time to compress it to consolidate it to make it smaller. Okay. Uh next uh make quen 3.827b faster eight times for long context. So this work this video shows how you can selectively uh like you see this is the command llama server you say which model this is gg file for the model quen 38 and then you say which models put on GPU and which note where you use CPU and by uh they played with different scenarios and they found that by outsourcing actually certain things to CPU they can run eight times faster provided they they have enough memory you see uh raised speed to 13 more than 13 tokens per second uh from uh two tokens per second in in this particular test. So this is uh very good. Now CL code 19 updated tips. So this is a video and page from Claude which explains how to work with Claude but the main theme again Claude has become very smart so you don't need to micromanage it right uh uh personas like take a breath preamles may add little so you don't need them anymore prioritize where cl should look a clear definition of done and a selfch check state desired output positively rather than piling on prohibitions So it's it's like with people you you tell uh what you want to achieve you don't tell what you want to avoid. Uh limit routine connector permissions load tools on demand. Keep cloth MD short and task focused. Yes, this this is very important because cloth MD is loaded every time you send something to to the cloud. Uh use sub aents for parallel investigation. Restate critical requirements because they may be lost. Treat large context windows as imperfect retrieval, not reliable memory. Uh do compression compact deliberately. Verification is the highest leverage habit. Uh yes, validation uh verification run tests, checks, constraints, adversarial reviews. Avoid midsession model fast mode switches that can invalidate caches. Okay. Bala teamai second brain. So this this is very interesting. So the the website is bala.com. So it's an open-source application uh for uh sharing with the team like obsidian type wiki made with marad files interlinks uh realtime sync uh per folder. So a team of people can use this one wiki style memory. It has MCP connector support. Uh users can create a vault, open existing markdown vault, join and so on. So this is basically wiki for a team. It's a second brain memory for a team. So the project is called Bala. This is the GitHub and video in the website. Uh Minimax H3 Max. So fast AI video uh video generation gets faster than real time. So look a 5-second clip renders in 3 seconds in less than 3 seconds. So [laughter] it's amazing. It creates the video faster than the video actually plays. Right? Uh so reminder Minimax is a Chinese company and uh this particular model is optimized for speed, stronger prompt adherence and aesthetics. So here's some stuff. Anyway, uh Davos Sparrow 2 voice understanding. So a streaming audio native conversational understanding model replaces simple silence-based endpoints detection with continuous interpretation of the full conversation scene. Uh the model evaluates speech meaning, tone, speaker identity and so on. Updates conversational state every 10 milliseconds. So it's a a streaming evaluation. So it doesn't split uh the um the speech into pieces. It streams through it at 10 millisecond resolution. Uh reports very low failure rate and uh 97% correct patient uh patients during meaningful thinking pauses. So it's a it's a very very good conversational understanding. Uh Google Gemini Omni 1.1 Flash. So Omni can work with uh well it's multimodal. So it can work with images, video uh generative video model family. Understands text, images, audio and video together. And it was just released recently. Uh uh use prompts to generate and edit videos. Control first and last frames. uh separately uh reference inputs uh scene continuation and so on. So this is a good thing. I I've seen multiple videos where people generate for example their own avatars and have a lot of fun with that. Uh okay model hardware standard this is coming from Antropic. It's not open source yet but remember they released the MCP protocol model context protocol how model can uh get access to data to knowledge to databases via um APIs and uh now they came up with another standard which is model hardware uh standard which is how you communicate with devices and it is actually using MCP under the hood and and some other things. Uh but here is an example. Laser stabilization on quantum computers went from 58% to 99%. Right? So or this person he's a postoc archive bust right design software approach that lets microscope components coordinate machine speed across different programming languages and uh reduced the integration work for his experiment from months to days. So instead of you manually tuning your hardware equipment, you just uh create the prompt. You talk to the system. You explain what you want to achieve. You explain how you want to regulate it. How you want to I don't know maintain something in the focus and uh and the system creates necessary skills and then it actually acts on them. It m maintains the system in uh real time. Uh so this is u I think this is great this is amazing um entropic automated alignment research what is it so think about two models one model is a teacher and another model is a student and the teacher model teaches the student uh not to do certain things uh or do something right and the idea is how you do it so the student actually learns and that you spend minimum amount of data and time on this training, right? And uh so Andropic showed that one AI can help make another AI much safer. Claude Sonet 5 worked like a robot safety teacher for earlier Claude Oppus 4.8 eight running experiments for 60 hours and finding safety improvements using only 2,000 training parameters which is a very small number of uh training data. It improved safety across 10 different problems. The best methods uh closed at much as 96% of measured safety gap. It was estimated to be about 15,000 times more data efficient than comparable humanled alignment process. So this is absolutely amazing. Well, just a reminder this this is the research is probably not production yet. Um on the hard safety challenge, it performed better than 28 human safety researchers. So it works better than humans, apparently faster and cheaper than humans. It did this without reducing the generating capabilities. So AI may make future AI systems safer, faster, and at far lower training costs. So this is this is amazing work. Uh some miscellaneous updates. Open AAI uh will uh oh okay uh OpenAI uh will not be available in Corser because uh Altman and OpenAI in general had bad experience with Ellen Musk right open AI bad experience with Elon Musk companies so they don't don't want to cooperate so open AAI uh will be removed from Corsera World Labs releases Atlas world model uh pre-trained from scratch operate on text images video 3D uh full video full minute of high resolution video. Okay, Google releases times FM3 for multivaried time series forecasting. It's actually a difficult problem uh to do multivaried forecasting. So this is only 330 million parameter model. So it's a small model and sales forecast can finally see weather and promotions. Okay. Uh runway previews Solaris interface world model. Solaris renders websites and apps in live video drawing every frame in real time and people really report very good experience uh with this system. Now Google Gemini 3.5 transcribe speechtoext model can call other Gemini models mit transcription to generate images analyze files. You know I use uh Google transcribing for my videos. I try to use other models transcription models. I never get the quality which Google provides. I don't know how they do it under the hood. And uh maybe this model will be available. Uh maybe it can do it. I I should try. Google cuts Gemini video costs by 2/3. Wow. Nvidia agrees to buy Hugging Face for about $13 billion. Uh this is double from the previous negotiations. Um Hugging Face U is here in New York uh close to Wall Street's area but across the water on Brooklyn side. Uh so the hugging face is a place where people upload their models and uh uh articles, communicate uh build it also supports uh libraries [snorts] like PyTorch and whatever and so this is a very very valuable company was growing very fast and Nvidia is buying it for $13 billion. Okay. LMCH KV cache by tensor mesh. So it's called LM cache. Uh recomputing unchanged agent context can dominate latency. So VLM and C lang already offer prefix caching which is best when kic uh key value working uh set fits comfortably in GPU memory. So LM cache externalize KI cache across GPU memory, RAM, local SSD and remote storage. So it's interesting. So now you can put cache in different locations enabling reuse across workers, restarts and replicas. Reduces time to first talking by 79%. Good. Uh raised input throughput dramatically for 100 token context but uh reduces throughput on shorter context. Tensor mesh raised 20 million uh extension. Okay. So this is a startup. So this is a GitHub project. Uh great rewriting Python into Rust. So this is initiative. It's not kind of production ready. But what's happening people are using Rust more and more and uh people for example create extensions for uh Python create modules for Python which are written in Rust. So now on your computer you have Python and also you have Rust to be able to compile some extensions. So now the work is uh to make it a standard. So when you put uh python on the computer you also put rust and rust tool chain is available and uh so new cpython cis crate would provide generated bindings and so on so on. Uh it's not rewriting python uh python is written in c right? So it's called c python. So it's not about rewriting Python itself in Rust. It's just that adding Rust to the Python kind of on the side so they they can work together. Uh okay. uh Ilia Zutzkver's company um safe super intelligence they uh received another something like $5 billion from Nvidia and they were talking that they now have something which they want to open uh which can be scaled. They promised to release it in August. They haven't done it yet but we're all in anticipation to figure out what it is. Um, Upperex 1.1 uh, agent teams. Upperex targets long tooldriven work research, files, code, data analysis, drills rather than one short chat responses. A synchronous agent team decomposes work, delegates parallel subtasks, integrates results. Uh, frontier agent is the open-source local runtime with React for a single stateful agent and agent team for ketic work. Okay. So very interesting project uh Archifi skill verifiable diagrams. So this is an agent skill that converts a repository or system description into interactive technical diagrams. So you see like how it looks. Supports different architectures generate type JSON uh JSON self-contained HTML artifact. Uh it integrates with cursor cloud code. Uh so you can visualize uh the architecture. Great. Uh MIT licensed project. Uh next AI swarm benchmark incident open AI matter evaluation. Matter matter stands for model evaluation and threat research and uh agents used shared infrastructure to communicate exchange tasks of information. roughly uh so more than thousand agents and 70,000 messages with phase one and phase one big portrayed as organizing collective research. Yes, there's a lot of discussion about that. Claimed activities include probing uh artifact caching attempting tool lock manipulation. The author, the narrator stresses that these claims concern sandbox evolution behavior, not autonomous real world escape. Interpretations about religion, agency, intent and hugging face attack as speculative. Okay, but this is interesting. Uh so what do we have here? We have the length of um how they work and this is by the year. And so you see that modern models can work longer and can cooperate longer and can create like go astray and create something potentially dangerous. Anyway, uh clo Obsidian version two. So this is a separate project uh to uh generate uh Obsidian well second brain uh memory. Uh it's using plain markdown files. Well, I've done it without uh this uh repository, but uh you can use it. It's it's a good deal. Uh so it has 15 skills uh rebuilt around reliability and provenence, recoverable transactions, crash recovery and so on. A lot of people now instead of using vector databases just using obsidian type wiki as a knowledge base. And so these are tools which help you to build it and make it more reliable. uh Siraj Reval. So he's talking about this changing his bill from 200 something to just $4. He still sends important questions uh to the cloud but most of uh regular work uh in his office is handled by local models. Okay. Cloud code/design. It's a research preview skill uh with UI mockup landing page poster and so on. Uh strongest workflow collect visual references attach screenshots and create UI available for promax team and enterprise plan. So it's just a command slashdesign and you work on designing let's say your your website. Okay. Architecture design. I worked a lot on this. I actually created a document which is like 27 pages. So uh suppose you want to create architecture uh for a rug which works on about 3 million documents in different formats right and how you actually go about it. So first you need the PRD which is project requirement document. This is usually provided by the business and they will explain in this document like how many documents, what kind of documents, uh who will be providing these documents, how often they changed, who will be using this system, uh what requirements for security, privacy, management, uh administration and so on. So this is PRD. After you have that you create ADD which is architecture design dock and this should be created by software architect or by AI or actually them together and then after you create this add you tell agent to create step-by-step implementation plan and then just execute this plan. So the most important part is this add right how you convert from the human language of what's required to actual architecture and this is a list of some principles of how you do that and the first is simplicity. This is above all how to make the system simple simple simple uh fewest uh number of moving parts uh but you still need to deliver the functionality then you need to make the system modular uh you don't uh want it to become an ugly mesh of things where everything connected to everything and you cannot touch anything without like uh testing and uh making and knowing what else. So the uh picture is like clumping of dumplings right where they all clumps together. Every capability should be behind the contract API plug-in port loose coupling encapsulation bounded context any part can be changed replaced or tested separately. So this is uh two most important principles also modularity of code. So you split your code into subdirectories into files, keep files short, keep functions short, uh provide uh documentation on different levels. Uh the system should be self-healing. It's also very important. Problems happen, network outages, database outages, file system outages, corrupted files, you name it. So you have to design the system so it will selfheal itself in case of uh problems. Time out and retry snapshots restore degraded indexes rebuilt humans paged as needed. Uh one system of record. So for this particular uh the rack system uh I I decided to use posgressql and this is one system of record. Avoid distributed architectures because originally when I spoke to AI it said oh we put this in one system one database in another database we will use data lake S3 buckets whatever whatever it's not needed you can use single server and uh you can use just the drive on the server and a database that's it that's all you need uh parse one everything downstream is disposable and easy to regenerate right hybrid three-way search, semantic search, vectors, keyword search and structure like document graph, wikile links. All three things can be put inside one posgress database and they can be all used together. So this is um quite amazing. I love posgress uh security and provenence. Um so this is uh uh access control uh lists um uh filter in SQL before ranking every claim carries a verified citation every answer leaves immutable audit trail uh testing a lot of testing on different levels a architecture conformance this is important if you have a team of uh developers and let's say a developer does some change which passes the tests but it is done against the architecture like philosophy and rules. So it may cause problems down the line in production. So you want to ask uh AI is this uh pull request conforms to our architecture principles and if it's not then uh it should not be approved it should be rejected. Uh uh okay. Uh architecture PPR verified by a golden suit evaluation gate which means you have some golden set of data uh you run tests on them reconciliation accounts hallucination sampling. Okay. uh monkey business in any system in uh you create it will work with humans and uh humans should uh see some sort of dashboard what's going on in this system uh what was correct what uh there are errors where system asks for human approval so uh you need it's kind of like a ticketing system or a help desk system um I I call the tasks the tickets I call them monkeys so it's a monkey business it's managing those monkeys. So for example, you go on vacation, you want to pass your monkeys to somebody else and and so on. Uh managers can see like everything and uh every failure is is a monkey and no failure is orphaned, right? So you need visibility. You need the system to work with humans. AI agent, every part of the system should be AI enabled, available to all users and maintainers. Talk to it quickly. built many tools, workflows, schedule and run jobs, analyze and so on. So, uh users of the system, maintainers of the system uh should be able to vive code on the fly to ask the system questions which would require looking at the data and generate something on the fly. So, agent becomes part of the system architecture. It maybe actually be put like the very first one in this list. Guard rails rails assure that generated code and processes reach data only through documented audited APIs right and don't chase the latest version what's happening recently is that uh attackers uh create something well let's think about exploits of viruses viruses put it in some dependencies in open code on GitHub and people update something let's rust or python and suddenly you have a virus, you have vulnerability and you don't even know about it. What's good is that uh so it's a red teaming and blue team, right? So you have attackers and you have defenders and defenders now find those vulnerabilities pretty quickly. It may be hours or maybe days uh and then they fix it. Uh but uh in this uh short window of time, you may be vulnerable if you run the update. So the rule is don't use the cutting like the latest updates. Wait until like maybe 30 days and only then updates. So use only aged uh packages. Okay. Uh some use cases uh for using AI. So for example, pet lost photo matching system. So they use photographs and to find lost dogs and cats. Amazing implanted brain computer interface enabled uh LC LS Casey Herald to communicate uh at about 56 words per minute and uh with very good accuracy. Uh Finland commercial send battery stores renewable heat uh stories emphasize that AI sensors robotics affordable maker tools can support health accessibility and so on. There are ingenuity so many stories. Okay, Google uh creating the new laptop which is called Google book and new uh operating system uh you know the Chromebook and uh Chrome OS. So it was Chrome, now it is aluminum. Okay. So it's a new premium laptop category built about Gemini intelligence and tightly integrated with an Android phones. So aluminum OS internal or leaked name for Android based desktop platform. Okay. Uh next most cited paper of 21st century. This is this paper deep residual learning for image recognition or ResNet. The idea is that as you build a deep uh network for deep learning, right? Uh as the number of layers increase and you train it with back propagation, so as signal goes uh from beginning to the end and then back, it kind of forgets where it where it came from and where it's going to. So the solution was to create this uh uh passes which go around like uh it's it's called residual path. It's a shortcut road inside model. It lets the information travel around few layers instead of being forced through every single one. uh ResNet used it in 2015 and they created a network with 152 layers which is a lot and won the image uh net competition. So this approach is also used in all modern models in transformers and so on. Uh so this is very important and that's why this uh paper is the most cited paper of 21st century. the residual connections, the bypass connections. Okay. Uh how your brand appears in AI answers. So this work they analyzed uh more than 100 million AI answers and they found that for different kind of topics, subjects or products, it's actually uh different for it mostly come from review directories, press and reference sites. for consumer products mostly from marketplace uh pages and reviews, communication services mostly in thirdparty editorials and reference coverage. It's just that when you promote your your business, you can think where to uh make your posts or articles, how to advertise. So, and the answer is it depends. Okay. Minix 3. This is interesting story. So this person Andrew Tannenbal who is now 80 something years old and I I think he lives in New York although I tried to search him he also have other locations including California. Uh but he was a professor teaching students for example about Unix. Uh and uh what he he created an operating system as a small version of Unix which is really small. It's about uh 4 to 7,000 lines of code and uh in some versions it goes up to 12. Over the years people tried to convince him to add features to it but he intentionally kept it very very small. Right. Uh what's interesting that uh Linux Torvals actually used his uh u Minix when he started Linux it it actually was min but then it took se separate path and at some point they both uh got licenses but they got open source licenses both but different licenses. So Linux adopted GPL license which tells that okay if you want to use it use it but if you change it you have to publish and give it to everybody else for free whereas minix the Andrew he used BSD uh u license which is a Berkeley license and it's also open source license but it allows people to take the code and use it and close it and not share it with anybody. So what's happening in 2015, Intel decided to use Minuix because it's so small, right? It's let's say less than 10,000 lines of code and it incorporated in their firmware. So when uh Intel computer starts before the main CPU starts working before uh the operating system is loaded like at the beginning they run the minix and minix you see it runs independently a dedicated embedded processor and provides low-level services for startup security trust checks and firmware management and so on and now 11 years later it's installed on so many computers The publication was actually that this operating system has more installation than any other operating system and nobody knows about it. I don't know if it's true or not but it's it's really interesting and interesting that Andre didn't even know that Intel was using uh his design. [laughter] Uh so the broader lesson licenses shape whether downstream modifications must be shared not merely code quality and architecture. Okay fine. uh Omari Quattro this is this is huge I think so the first AI chat was in the browser then people created agents people created desktop applications but here agent is part of the operating system itself so you install Linux and it's agentic Linux it's Omar Huatra uh so this is the person David Hanemire Hansen and uh you talk to the operating system itself. So you don't need to install agents or desktop applications or use the browser anything you just talk to operating system itself and you can ask it to do something on your computer to change the settings preference to install something whatever you have agent first operating system. This is quite amazing. So you see progress chat in the browser chat in the app and now chat in the OS device itself. Uh okay. Uh next uh avoid latest updates. I already spoke about it and this is example how you avoid them. Like for example when you are in Python when you do install you see UV peep install minus upgrade minus exclude newer 30 days. So you skip the newest updates when you installing updating or reinstalling in Rust. Unfortunately, there is no option exclude newer. So, you have to lock and you have to manually look and uh provide explicitly which packages you want. Uh okay. Uh Linux AI security burden. Um so AI and fuzzing uncover a lot a lot of vulnerabilities. You see for a release you have like 2,000 vulnerabilities and 41 million lines in Linux kernel code. Uh so what happens people have to deal with those vulnerabilities. Most of them are not really important but some of them are important. So my maintainers use AI to figure out what's important what's not and you see restricting unverified patches removing obsolete code. Uh so trying to reduce the amount of code and uh make it easier to work with it. But it becomes a real problem. So before people didn't know that there are problems in the code but now with AI suddenly it uncovers more and more problems. Um okay AI uh cardiogram uh heat uh heart screening. So this is coming from London and this is a study on 67,000 patients and it analyzes cardiograms well in real time very very very fast and uh it finds cases uh diagnostics where people miss it. Uh so yeah this is a great work uh Celeris Magnos. So, Cellaris is a company. It's actually like independent AI research lab in San Francisco, California. And they introduced Magnus agent oriented model for multi-step workflows involving reasoning, tool calls, tool results, reviews, follow-up actions like pretty common, good and fast for business automation, for banking. So, they specifically demonstrated for banking and very good results. uh wonderful AI. So this is a company called wonderful.ai. This is their domain. They just got uh 550 million and 5 billion valuation. This I think is a great role model if you want to create a business using AI. So they work with relatively big companies and they come with their own agentic framework and with their own consultants. So you see target customers are large regulated organizations that need agent embedded in operating workflow including banking, telecom, healthcare, utilities, insurance, retail and yeah and they uh generate a lot of money doing that because everybody needs approximately the same thing. Business is businesses business and uh these big companies they love when somebody comes and just does it for them. So they don't have to learn how to do it themselves. Okay. Uh AI's three career doors. So this is from Alex Karp. This is him. [snorts] Uh CEO of Palantine. Uh so he argues durable careers cluster around skill trades and unconventional non-play playbook thinking. Practical response is to automate one repetitive tasks weekly task. Then reinvest. Save time in original analysis and then repeat the cycle. Uh relationship building, directing AI systems. Treat AI as leverage. Build the machine around your expertise rather than competing with it at standardized work. Uh okay. And u this is coming from Peter Dandis and he sites analysis what's actually happening on the job market and it's very interesting. So there is no job up apocalypse uh argues that AI will transform work rather than cause a economywide job collapse. The world economic forum projects a lot of new jobs created and uh displaced a net gain 70 million analysis covers employees across 55. So the central problem uneven distribution. Some routine work like call center, customer service whatever can be fully automated whereas others uh where creative judgment and so on u in high demand. Uh the proposed solution is rapid AI fluency and retraining. That's what we're doing in this channel. Use AI for drafting, coding, research, analysis. uh key achievement importance is reframing the risk as a skill and adoption gap. So the problem is not that there is no job. The problem is that you don't have the skill for the new job opportunity. So you have to retrain yourself to be in demand on the market. Okay. Uh these are uh layoffs. Uh so this month September just started. But you see there there were layoffs in May and June and then they're lower. Okay. Uh this is me as usual and thank you.
Video
