Vollständiges Transkript anzeigen (2.916 Wörter)
Um artificial intelligence news every week, uh on Friday, at 2:00 PM Eastern time, New York time. Today is Friday, October 16th. You see a lot of news, as usual, and uh the interesting today is Anthropic replaced Git with Scion after Claude passes 30,000 merges a day. So they're using AI to do their own software development, and it just becomes so fast. Anyway. So, this is benchmarks, as usual. Uh please note that Argon, Gemini Argon keeps the top place in both regular text and in coding. And Muse uh released several models, uh from Meta, and it does very very good job, and they also released an open-source model. So, let me start. Uh okay, this is Artificial Analysis Intelligence Index. So, this is different models, doing the same tasks, and what we're comparing is cost. So, look at this. It's GPT-6 1, SoL, Max. So, SoL is the top model, and Max, so this is a great, this is a frontier-level model, absolutely. But it is more than 10 times cheaper than, let's say, Claude Fable. So, this is amazing. This is definitely a good, like, go-to model to use. Anyway. Uh Anthropic, Scion, replaces Git. So, what happened? Uh Claude development, Anthropic development was growing faster and faster and faster and they got to the point that they had 30,000 merges a day in Git. So, when they're submitting new code, so in May, it was 2.5 thousand. In September, it already hit more than 30,000. And they had something like 100 Git mirrors to handle all this traffic and it was getting more and more difficult. So, they created their own version control system, which they call Scion. And they also created something called Antik, which is AI-assisted code review tool, because now with a lot of code, people simply cannot review all the code by themselves, it's too fast and too much. So, they need to use AI for that as well. Now, Anthropic published an article explaining how this new source control works, and it's very let me first tell you about this picture on on the right. So, uh this is the time, uh Git took, like, 87 seconds and this is three-and-a-half seconds. This is the new Scion. So, these are different tests, like pull, checkout, rebase, status and so on, different commands. And you see that the new system is much faster, like really faster. And they see CPU and memory use fell 200 times. For certain things, like 100 times. So, it's really really high performance. How they achieve that? Well, regular Git, when you send some requests, for example, you want to fetch code and merge it into your code, the way it works, there's a negotiation between your local repository and repository on the server, they figure out what are the differences, then they package it. So, you see, with Git, client says, "I want E, I have C." Server determines the objects needed, constructs a custom transfer pack and then sends it. So, for each communication, you end up creating a custom package. Whereas in the new, in Scion, they don't do that. They still need to do recovery, what discovery, what is needed, but then they just ship the missing pieces from cache, and cache is used by everybody. So, no custom packaging. And this alone allows like huge speed up and you you don't need so many servers and so on. This article, Anthropic, Engineering, Redesigning Source Control, so it's very detailed, has a lot of, uh, graphs there, I recommend you to read it. Okay. Uh Meta releases Muse Wonder, its first frontier-class models. So, there are two versions, Muse Wonder and Muse Wonder Full, and they're closed, and they're very good. Have you already seen them on the benchmarks? Uh pricing is not cheap. 4 dollars per input and 20 dollars per output as usual for 1,000,000 tokens, 1,000,000 context. Uh okay, uh Meta releases Muse Spark Open. So, in addition to two closed models, they release an open source model with 816 billion parameters. So, this is their Spark model, and it is good, I mean, it's still not as good as GLM or Kimi or Qwen-3-8, but very close. You see the score 39 for this Spark, whereas Qwen-3-8 is 40. And Kimi and GLM are comparable. They're close. Uh but it is open model, so this is the best Western world open model so far, and you can run it like here, for example, they run it on two M5 Ultra Mac Studios using Apple MLX framework. So, this is good. Next, GLM-5.3 on Amazon Bedrock. Uh, usually, Chinese models are not recommended in in US for US business, but Amazon now provides, uh, GLM-5.3 on their cloud on Bedrock, so you can use it. Uh next, Qwen Image 2.1 Turbo, text-to-image and editing. It's a small model, you see 7 billion parameters, and developers can download weights on Hugging Face and run locally. Okay. ArtCraft and AI Reverse Engineering. So much noise about this. The idea is that you, uh, now can take any application and disassemble it and recreate it, basically. So there's no privacy anymore. And in this particular project, the developer, Brandon Thomas, uh, he released a system, which is kind of like, um, Adobe, uh, set of tools. ArtCraft, services, lightcraft, Photocraft, and so and it's like Photoshop. Uh, so you see, Reverse Engineer Anything. So, this is a, a GitHub repository and you can use those tools to reverse and a Uh, people tried his creation, uh it's not really Photoshop. I mean, features are there but it not exactly like, like, but still it's absolutely amazing how you can now create your own tools. Okay, uh Claude Code, Less Micromanagement. Well, this was the theme for several weeks already. So, Boris Cherny, who is creator of Claude Code, in January was explaining the spec-driven design where you first plan in detail what you do and then you do it. Today, uh he favors, well, not just him, the way Anthropic works, uh more autonomous execution. So you don't micromanage agents. Agents are smart enough. You just do high-level management. Define tasks, guardrails, completion criteria, review results, instead of prescribing every step and so on. The central achievement is reduced supervision while retaining verifiable outputs. Okay, and this is of course Boris Cherny himself. Okay, uh Local AI versus Claude. Uh, this is a very interesting video. And so the guy was saying, what if I will try to run the model locally? Or I run similar model in the cloud, which one will be cheaper? Uh, and you would think, well, if I already have my hardware, of course, running locally will be cheaper, and the answer is no. Cloud is actually cheaper. Uh, so recommendation is forget local, unless you very, very paranoid about privacy. Uh, but for just cost and performance, just go for the cloud. Okay. Microsoft Windows Hybrid AI. So, this is Microsoft and uh NVIDIA, working together and they calling it hybrid intelligence. Windows would route AI tasks between local hardware and cloud models, including NVIDIA, addressing agent cost, privacy, latency and offline access. Again, same thing, it's cheaper and faster to use cloud-based, uh, models and and higher quality. Claimed advances include Nvidia RTX Spark hardware, Surface laptop, compressed coding models, Microsoft Decision, 1, reportedly 35 times faster than GPT-6. Uh, okay. Next. Oh, these are decision models, um, so this is very hard topic. There are so many models, and I just made a table of some models. Uh what are the common themes? So, first of all, mostly the models have very short input context window. So, for example, Laya, which is a very good model, it's fast, it completely open source, uh you can run it locally, but their window is like less than kilobyte, or maybe for multilingual it says 8k, actually. Didn't try. But this is this is not a million token. So, you cannot give it a lot of text to make a decision. So, remember decision models, they make three types of things. They answer yes or no, or they select from more categories, or uh they give you some sort of ranking. Uh anyway, so uh mostly short input context window, uh some models are not tunable, like, for example, the first one, the TypeSafe, Jeff, model, it sits on the cloud, and you cannot tune it. You get what you get. Sometimes you want to tune the model because if you have your specific requests, and you have, uh, let's say, 2, 3, 400 examples of question and answer, you can tune the model, and the accuracy will grow from like 70% to 90 something percent, right? But then, you need to keep it on your own hardware, and it's completely different uh setup. Uh these models are fast, but accuracy may be limited. And again, tuning helps. And these are not all the models. There are so many, like I haven't even uh included here, Olamar, and whatever. It's just a small subset just to give you an idea. Okay, this is about my channel, uh what do we have? uh more than 7,000 subscribers, more than 300 videos, every week, name of the channel, Lev Selector. I provide links for slides, and please, please, provide feedback and ask questions under the videos. Uh next, Deno joins Cloudflare. Deno is an open-source runtime for running JavaScript, TypeScript, and you know, there is a node.js and there is also Deno, which is supposed to be better. Well, it is better, but what happened? Uh the whole Deno team now joins Cloudflare. And Cloudflare is not interested in Deno itself. So, the Deno will be discontinued in a year. So, for now, they are providing support for a year. So, Cloudflare acquires Deno team and their Sell D technology. So, Sell D is an open-source system for running and coordinating workers compatible applications. And Cloudflare has their own Worker D, which is open-source runtime for executing workers application on own infrastructure. So, what Cloudflare wants, they want to build, to work with agents, right? So, this Sell D is a foundation for AI agent infrastructure. It makes it easier to host many persistent, stateful agent sessions. Uh, so this is the overall, uh direction. And here are some links, like Cell D dev and GitHub and so on. This is of course very exciting. I love Cloudflare, I use it for all my websites, and now they will have more and more agentic stuff. Okay. Claude Motion, in beta for Team and Enterprise only. Uh turns supplied content and data into editable, code-based animated explainers, charts, product walkthroughs, and presentation visuals. And then convert into MP4. into actual video files. Now, AstaBrief 8B generates science reports. Uh so this is coming from Allen Institute for AI, everything they're producing is completely open-source. And so they released this model, it is on Hugging Face. Uh built on Qwen-3-8B. So it's a small thing. It turns questions and supplied scientific experts into cited reports. The model itself does not search online, it just kind of packages stuff. High precision, and good citation recall, open weights and training, like everything is available. Okay. TurboPuffer, RIP Vectors. So, TurboPuffer is a website, it's actually from the website, and which provide rag solutions, right? And they shifted how they do things from uh mostly based on vector databases to hybrid approach and using a caches, using files, MD files and so on. The version three of the system redesigned, document IDs replaced vectors as primary storage keys. Uh okay. Next. Odyssey 3, Interactive AI Worlds. Odyssey, it's a California AI company, launched Odyssey 3, in public research preview, it generates explorable worlds from prompts, responding continuously to movements and events. Okay. Next, step 5. It's a preview for coding agents. Uh released the 5 preview for coding agents and professional workflows, including finance. 600 billion parameters, so it's a big model, supports text, images, video, tool calling, 1,000,000 tokens input context, 64k output. Uh agents can inspect projects, edit files, run tests, and iterate. Available through OpenRouter and existing coding tools. Next, Laminar Flow 1. Detecting agent failures. Launched Flow 1, inside Signals platform to investigate AI agent execution traces. Reinforcement learning-trained flow 1 searches trace steps and returns structured explanations of errors. Affordable continuous monitoring helps businesses verify automation rather than trust completion claims. Okay. Next, Alibaba Qwen JAR, I don't know how to pronounce it, Long Horizon Agent Reinforcement Learning. It's a training framework, it's not a model, running on multi-node GPU clusters, it trains coding agents through unmodified harnesses on tasks lasting hours, with approximately 700,000 tokens per rollout. Okay. Black Forest Labs released FLUX 3 Image. So, it's a new image model with pixel-perfect editing, bounding boxes control where things go in the picture. up to 4K resolution of images, combines up to 10 reference images in one picture, accurate topography. So this is Black Forest is known for their image generation and editing software, this is the latest version. Uh Microsoft unveils Maya 300, its third in-house AI chip, and you can see this chip is pretty big. Uh this is for inference. claims 40% more token per megawatt than leading GPU systems. Large scale deployment starts next year. the goal is to reduce dependency on NVIDIA and AMD chips. GPT-6 becomes the ChatGPT default. At the beginning of this lecture, I recommended the GPT-6.1 SoL. And yes, so now this is default, reaches more than a billion weekly users. Answers can include charts, forms, buttons and small tools. So yeah, recently everything related to GPT-6 uh is actually good. GPT-6 instant web search answers 45 44% sooner than the previous model. They also have a ultrafast option which is up to 8 times faster, but it also expensive. But if you want very fast responses, you can have it. Okay. Claude Managed Agents, 1,000 agents. This is a Claude. And before they were like 100 agents, now they're talking about 1,000 agents, right? called Dynamic Workflows. We actually spoke about it before. It is in public beta, a lead agent writes a plan, then multiple agents work in parallel. up to 1,000 worker agents per run, and then results are merged at the end. Uh okay. Grok bot will route tasks to Claude, Midjourney and Suno. Very unusual, because usually, these big providers, like OpenAI or Claude or Grok, they are not interacting. But here we have Grok, which is Elon Musk's, uh company model and now they will route some tasks to the competitors, like Cloak, Midjourney and Suno. uh becomes an orchestrator, not a single-model assistant. Claude Opus 5.5, Midjourney for images, Suno for music. Announced after reports of Grok access problems on mobile and web. Okay. Modal raises $750 million at $15.75 billion valuation. Modal Labs is a startup founded in 2021 in New York City, with the they have also offices in San Francisco and in Europe. And they provide cloud computing for AI inference, model training and so on. And they already at $15 billion valuation and they just raised a lot more money. uh Fish Audio Drama 3, open-weight 35 billion parameter text-to-speech. 83 languages, up to 16 speakers per request, and it can handle all this simultaneously. You can also provide in brackets directions to control the tone of the speech, so you can direct the interactions. First audio arrives at approximately 89 milliseconds for a single stream on one NVIDIA GPU. It ranks first on Artificial Analysis Pronunciation Robustness benchmark. So it's a very good thing. Liquid AI, they have the new LLM model. Liquid AI is a startup in Cambridge, Massachusetts, and this is open-weight 24 billion parameters, so it's not a big model. It's specifically for on-device agents. 131,000 context length, reasoning, tool use, matches or beats Qwen and faster. So this is a good achievement. Grok Voice 3.0, tops the speech-to-speech index, ahead of GPT-Live-1 and Gemini 3.8 Live. Sub-second latency, reasoning and tool calls mid-conversation. already handles 20,000 Starlink support calls a day. So it's already in production. Character AI goes bankrupt, and this is interesting. So, Character AI was selling access to AI companions. So, it's like a chatbot with a personality, so you can select create the persona and then talk to this persona, for entertainment, role-playing, storytelling, and if nobody buys them, they would simply close unfortunately. Okay. Perplexity open-sources SPACE, a sandbox for AI agents. Isolates each agent in a virtual machine. Blocks unauthorized access and data leaks. Top models made 118 breakout attempts; none escaped. So this is safe. Uh another thing. JPMorgan, it's the Gemini. They announced that they will not allow uh outside models, like for example, models from Meta, to access JPMorgan financial accounts. So outside agents may not log in with customer credentials. They cannot move money or pay bills in retail accounts. Read-only data sharing through approved connectors still works. Google Deep Review checks technical documents. uh so Google Deep Review becomes better. Uh it accepts documents, multi-file input is coming. Google expects such tools to help both authors and paper reviewers. Jobs and Layoffs, not many layoffs this month as you can see. uh this haven't changed the top in demand skill. This is me, nothing changes. Thank you.
Video
Video
Video
Video
Video
Video
