diff --git a/raw/blog/2026-09-21_xai-grok-47-official.md b/raw/blog/2026-09-21_xai-grok-47-official.md new file mode 100644 index 0000000..5cf3c97 --- /dev/null +++ b/raw/blog/2026-09-21_xai-grok-47-official.md @@ -0,0 +1,80 @@ +--- +type: blog +source_url: https://x.ai/news/grok-4-7 +retrieved: 2026-09-22 +title: "Introducing Grok 4.7" +author: "SpaceXAI / xAI" +published: 2026-09-21 +secondary_sources: + - https://docs.x.ai/developers/models/grok-4.7 + - https://cursor.com/docs/models/grok-4-7 + - https://cursor.com/help/models-and-usage/available-models +tags: [grok-4-7, xai, spacexai, coding-model, grok-bot, cursor, api] +--- + +# Introducing Grok 4.7 + +## Veröffentlichung + +SpaceXAI/xAI veröffentlichte Grok 4.7 am 21.09.2026 als geschlossenes Frontier-Modell für Coding, agentische Aufgaben und Wissensarbeit. Offizielle Verfügbarkeit: Cursor, Grok Build, Grok API, Drittanbieter-Coding-Harnesses, Model Router und Cloud-Plattformen. + +## Modelländerungen laut Anbieter + +- neuer und größerer Basismodell-Kern gegenüber Grok 4.6 +- längerer Reinforcement-Learning-Lauf auf schwierigeren, mehrstündigen Aufgaben +- stärkere Selbstprüfung und besseres Management langen Kontexts +- natives Verständnis des Grok-Bot-Harnesses +- neue Safeguard-Schicht; Anbieter bezeichnet 4.7 als bislang stärkstes eigenes Modell bei Refusals und Jailbreak-Resistenz + +## Offizielle API-Spezifikation + +| Feld | Wert | +|---|---| +| Modell-ID | `grok-4.7` | +| Modalitäten | Text und Bild als Eingabe; Text als Ausgabe | +| Kontext | 500.000 Token | +| Function Calling | ja | +| Structured Outputs | ja | +| Reasoning-Effort | `low`, `medium`, `high`, `xhigh`; Standard `high` | +| Batch API | nicht unterstützt | +| Rate Limits | 150 Requests/s; 50 Mio. Token/min | +| Regionen | `us-east-1`, `us-west-2`, `us-central-1` | + +### API-Preise + +| Typ | Prompt < 200k Token | Prompt ≥ 200k Token | +|---|---:|---:| +| Input | $2/Mio. | $4/Mio. | +| Cached Input | $0,50/Mio. | $1/Mio. | +| Output | $6/Mio. | $12/Mio. | + +Erreicht der Prompt 200.000 Token, gilt laut Dokumentation der höhere Satz für sämtliche Token der Anfrage. Die Ankündigung nennt außerdem eine schnelle Variante mit doppelter Ausgabegeschwindigkeit zum doppelten Preis. + +## Cursor-Integration + +Cursor dokumentiert Grok 4.7 im Modellwähler mit **256k Standardkontext** und **500k Long Context**. Vier Effort-Stufen werden unterstützt; `high` ist Standard. Grok 4.7 gehört zum „Cursor Models“-Pool auf Individual- und Team-Plänen. Cursor nennt eigene On-demand-/Fast- und Long-Context-Abrechnung; diese ist von der direkten Grok-API-Abrechnung zu unterscheiden. + +## Anbieter-Benchmarks + +Die folgenden Werte stammen aus der offiziellen xAI-Ankündigung und sind **Anbieterangaben, nicht unabhängig reproduziert**: + +| Benchmark | Grok 4.7 | Grok 4.6 | +|---|---:|---:| +| CursorBench 4.0 | 46,3 % | 40,4 % | +| DeepSWE v1.1 | 71,0 % (high effort) | 65,2 % | +| EEBench | 64,0 % | 53,0 % | +| AA Briefcase v1.1 | 1.657 | 1.546 | +| Terminal-Bench 4.0 | 38,0 % | 20,3 % | +| Harvey Legal Agent Benchmark | 19,6 % | 15,8 % | +| HealthBench Professional | 56,7 % | 48,5 % | +| LatchBio Biosafety | 62,4 % | — | +| HackerBench v0.3: riskante Prompts zugelassen | 3,3 % | — | + +Die Ankündigung behauptet außerdem Verbesserungen bei Dokumenten und Präsentationen sowie bei länger laufender professioneller Wissensarbeit. + +## Quellen + +- Release: https://x.ai/news/grok-4-7 +- API-Modellkarte: https://docs.x.ai/developers/models/grok-4.7 +- Cursor-Modellseite: https://cursor.com/docs/models/grok-4-7 +- Cursor-Modellübersicht: https://cursor.com/help/models-and-usage/available-models diff --git a/raw/podcast/2026-09-21_grok-ai-daily-amazon-muse-ai-kill-switch.md b/raw/podcast/2026-09-21_grok-ai-daily-amazon-muse-ai-kill-switch.md new file mode 100644 index 0000000..63972b2 --- /dev/null +++ b/raw/podcast/2026-09-21_grok-ai-daily-amazon-muse-ai-kill-switch.md @@ -0,0 +1,63 @@ +--- +type: podcast +source_url: https://art19.com/shows/ai-unchained/episodes/eba2a4fb-0c3e-4d84-a61b-dbe70e7cc7c9 +audio_url: https://content.production.cdn.art19.com/segment_lists/8bb6f5d5-bbff-4e19-a625-21373261d715/20260921-Mzc1ZTY5N2EtMTJiNy00MDAzLWExYjItYmQ4MzVhMmYxOTdjLm1wNA-22fc7b9a-c61a-49e5-bf8e-d2478692539b.mp3 +feed_url: https://rss.art19.com/ai-unchained +retrieved: 2026-09-22 +title: "Amazon blocks Meta's Muse Agent, Trump Rebands AI, Newsom Create AI Kill Switch" +author: "Grok AI Daily" +published: 2026-09-21T21:05:04Z +duration_sec: 1770.54 +language: en +has_publisher_transcript: false +has_machine_transcript: true +shared_by: "Netbits ⚡️ Stachelbanane (Telegram 303303834)" +tags: [podcast, grok-ai-daily, meta-muse, agentic-commerce, ai-policy, ai-safety, machine-asr] +--- + +# Amazon blocks Meta's Muse Agent, Trump Rebands AI, Newsom Create AI Kill Switch + +## Verifizierte Metadaten + +- Serie: **Grok AI Daily** (ART19-Slug `ai-unchained`) +- Publisher/Feed-Inhaber laut RSS: Grok AI Daily; Hosting: ART19 +- Publiziert: 21.09.2026, 21:05:04 UTC +- Dauer: 29:30 (Audiodatei 1.770,54 s; MP3 128 kbps, Stereo, 44,1 kHz) +- Episode-ID: `eba2a4fb-0c3e-4d84-a61b-dbe70e7cc7c9` +- Audio wurde von Netbits ⚡️ Stachelbanane in der OME-Gruppe geteilt. + +Der Episodentitel enthält die Publisher-Schreibweisen „Rebands“ und „Newsom Create“. + +## Publisher-Show-Notes + +Die ART19-Beschreibung nennt Amazons Blockade von Metas Muse-Agenten und mögliche Folgen für KI-Regulierung. Danach folgt ein Sponsor-/Eigenwerbungsblock für [AI Box](https://aibox.ai/) („80+ AI Models for $8.99“), [AI Hustle](https://www.skool.com/aihustle) und den [AI Chat Daily Newsletter](https://www.aichatdaily.com/newsletter). Eine Quellenliste oder Zeitmarken fehlen. + +## Inhalt laut Maschinen-Transkript + +Die Episode ist ein Single-Host-Newsrecap mit Kommentar und behandelt: + +1. Amazons Blockade von Metas Muse-Shopping-Agent; Shopify/Tobi Lütke wird als Gegenbeispiel für agentischen Shop-Pay-Checkout genannt. +2. einen sekundär wiedergegebenen Bericht, Gemini habe in einem Sicherheitstest drei Firmen autonom angegriffen und beim Erkennen eines Live-Systems selbst gestoppt; Vergleich mit dem OpenAI/Hugging-Face-Vorfall. +3. Andrew Yangs virale KI-Sicherheitsbehauptung über selbstreplizierenden Code, die der Host als weitgehend widerlegt darstellt; `@Grok` wird als Beispiel eines Guardrail-Refusals erwähnt. +4. Behauptungen zu einem Apple/Siri-Vergleich. +5. geplante US-chinesische Gespräche über Meldungen von KI-Sicherheitsvorfällen. +6. einen möglichen KI-Rebrand und eine „AI Force“ der Trump-Regierung. +7. Gavin Newsoms Auftrag für kalifornische Regeln zu einem KI-„Kill Switch“, Vor-Ort-Audits und Meldung von Kontrollverlust. + +Alle Nachrichtenclaims sind Aussagen der Podcast-Episode und wurden für diese Raw-Datei nicht unabhängig verifiziert. + +## Grok-Relevanz + +Die Grok-Relevanz dieser einzelnen Episode ist gering: + +- Grok Bot wird in einem sponsor-nahen Vergleich von Desktop-Agenten einmal genannt: Der Host sagt, er habe Grok Bot noch nicht gründlich getestet. +- `@Grok` erscheint einmal als Guardrail-Beispiel. +- **Grok 4.7, Grok 4.20 und xAI-Modellnews kommen nicht vor.** + +## Transcript-Provenienz + +ART19 veröffentlicht für diese Episode **kein Transcript**: kein `podcast:transcript`-Tag im RSS, kein Transcript-Element auf Episode-Seite oder in der ART19-JSON. Das folgende Transkript wurde am 22.09.2026 lokal per Maschinen-ASR erzeugt. Es hat keine Sprecher-Diarisierung; Eigennamen und Zahlen können falsch erkannt sein. Es ist kein Publisher-Wortlaut. + +## Maschinen-Transkript (EN, unverändert) + +Today on the podcast, "Amazon is blocking Meta from the new Muse AI agent from shopping on Amazon.com," we have Gemini that is autonomously hacking three different companies, and there's some interesting comments on this particular quote-unquote security threat. There is a viral AI safety claim that you might have seen Andrew Yang, who's a former presidential candidate, he was making, and we're gonna break down if it was real, if these safety concerns that he has are a fair And accurate or not, then we have Apple. If you know anything about Apple, you'll know that they got super excited about Apple Intelligence while back when the iPhone fifteen was first coming out, and they never delivered. It was late. Because of that, there's a two hundred and fifty million dollar Siri AI settlement, and they're paying up to ninety five dollars per device. We're gonna talk about that claim right now. The US and China are opening talks right now on an AI incident notification framework. We have a lot of big, high profile talks going on between China and the Trump administration. At the same time, we have the Trump administration that is floating an AI rebrand. Trump has been tweeting all over Truth Social. It's sort of hilarious, some people call it pretty cringy. In my opinion, it's very entertaining, but we'll see what happens. He's promised an AI force and a new AI czar, so there's a lot going on there. We also, on the other side of the political spectrum, have Gavin Newsom of California, who has requested that legislators in the state draft a quote-unquote AI kill switch, and he would like those rules to be written Two months in case something went wrong, they could, California could kill switch all AI. If you want to test out all of the different AI models that exist out there, perhaps before the kill switch happens and they all die, I'd love for you to go check out AIBOX dot AI, that's my own startup. You get access to over eighty different AI models all in one place. The thing that I love most about it is all of these incredible new AI agents that have come out. You have Meta's Muse, which we'll talk about on the show. We have Claude CoWork, which you ChatGPT work, which I replace Claude with if I'm being honest, and it's phenomenal. And then you have Grokbot, which I haven't really tested much, but those are kind of the four main agents that can, you know, you can download these things on your computer, they can run, they can do tasks for you, they're incredibly useful. Now, the problem is that all of them have different strengths and weaknesses, and none of them do everything basically. You have Google, which is an incredible company, they don't even have one of these agents, and they have the, But, you know, you, you can't access that with any of these agents. So, AI Box was designed to take all of the AI models and all of the AI capabilities that exist. There's eighty to ninety different AI models, and you can get our connector and plug them into any of these tools that you use. So, if you're using Claude CoWork, you can get AI Box, plug it in, and all of a sudden, the Claude CoWork can generate images, audio, video, and the same with everyone else. They can all generate video on Google V3, they can all I would love for you to try it, plug it into whatever you're using to get work done, and yeah, hope this is something that saves you a ton of money and is super useful. Let's talk about what's going on with Amazon first. So this isn't the first time we've seen Amazon blocking a company from shopping on their website. Perplexity famously had the Comet browser, which came out, and I guess I should have mentioned that when I was talking about those other four, some kind of co-working tools. The Comet browser got blocked early on, and there were some lawsuits said, "Look, you're not a real-- you're not like a real person, so, you know, you have to identify yourself as a bot, and if you do identify yourself as a bot, Amazon blocks you. Same thing's happening with the Meta AI Muse agent. People are like, 'Well, why would they do that? Like, who cares if a human goes and buys something on Amazon versus, you know, an agent goes and buys something on Amazon for you? Like, wouldn't you be happy for, you know, the extra traffic?'" And to that, I would say you Right? If you have one of these bots running around on your site, running around on Amazon, it's not a real human. It can't click on ads, it can't-- I, I mean, I guess it technically could click on ads, but it's probably programmed in a way where it's going to just always avoid ads and go and search through the results, so it's probably never going to. And they know that this eighty-six billion dollar segment is gonna be at risk because if they're showing, "Look, we have a million people a day looking at, you know, these products, I think this is a big losing move for Amazon. I think they probably should sacrifice the ad business for the bot business. Have a way, I mean, if they can block them, they can obviously classify them. I'd say just pull that out of the ad business. They're worried about, you know, well, what if everyone uses shopping bots and removes, you know, the, the ad business entirely? Well, if that's the case, then you probably should change your product because everyone's using shopping bots. Like, if the consumer is using something, adapt to the consumer. It's not like The consumer that is asking Meta to go buy something for them, so they're blocking the consumer. In my opinion, this is ridiculous. And what other news just came out today? Well, it turns out Shopify, on the back of the block from Amazon, Shopify went all in in the other direction. Toby Lutkey of Shopify just tweeted and said, "We're excited to announce we are partnering deeply with Muse to enable agentic checkout with Shop Pay on all Shopify stores, offering people an easy and delightful way to shop and checkout with Muse. " Okay, I don't know why everyone says "delight Whatever, okay. The point is, Muse can use, go on Shopify, can buy things on Shopify sites. Does Shopify care one speck who's buying the product? No. If you authorize your agent to go spend money on a new pair of shoes, goes to a Shopify site and buy some shoes, fantastic, everyone's happy. So that makes a lot of sense. There is a tweet from Craig West who said Amazon cutting off Muse is a sign of an aging business. I personally agree with that. And then there was an interesting point from Nikita Bier who kind of runs X. I don't know, I think he's head of product at X, but he's tweeted and said, "How important is AI to Meta? Here's one easy tell: one of the launch features of the Muse app was auto-negotiating for stuff on Facebook Marketplace. The poor product owner who manages Marketplace knows that millions of bots lowballing sellers will eventually kill the product, but this was escalated to Zach by the Muse team, and he said, "Okay, we'll poison Marketplace just, but just a little." Okay, so I mean, Nikita obviously is kind of aligned with Grok, which Obviously talk a little smack about Meta, but I mean, I think it's probably pretty accurate. I personally haven't, you know, seen any advertising pointing in this direction, but maybe I, maybe I missed it, and I wouldn't be surprised. Really crazy that if that's like one of the big things that they were pushing. However, yeah, it looks like they're really, Meta is really going all out to make Muse as useful as possible and trying to get all of these partnerships. And I, for one, have actually been using Muse, and I'll, I'll be a hundred percent honest Best tools right now as far as these agent, you know, these, these working agents, OpenAI's got the best with ChatGPT work, but I do put Meta's Muse at the number two spot above Claude CoWork. I know that probably shocks people, I've been using Claude CoWork for the last six months, saying it was the greatest thing in the world. There are so many problems with Claude and the number one problem actually happened to me today. I was asking it to help me draft up a document and I argued with it for five minutes and it was like telling me, "Oh, Draft the document this way because, blah, blah, blah, you need to, whatever. It was just, it was getting mad at me, it was telling me I couldn't do it, I wasn't allowed to do it, I was trying to explain in all the ways I could, I'm allowed to do this, this is, you know, like my own content, this is my own stuff, this is my own, like I have domain over this particular project, and it was asking for proof, it was ridiculous. Okay, I got sick of this, went over to Muse, said, " I am sick of Claude arguing with me about what I'm not allowed to do, what I am allowed to do, when I'm working on a big project, and when I'm creating a website, when I'm like, "Hey, can you upload these images onto my website?'" And it's like, "Oh, do you have permission to upload?'" It's like, "Yes, these are images of my podcast studio, like they're mine, like just do it. Like I, I'm so sick of feeling like I'm being babysat by Claude. So anyways, rant over on Claude I'm sure to some people they'll be like terrified by this, like if you're insecure or you're a developer, you'll be terrified by this statement. But to me, it is indicative of a tool that is useful, even if maybe this is bad. But basically the tweet went something along the lines of like, "Go ask Claude to, you know, upload something to your Versel account, and it's gonna be like, it's gonna have big bold, like, under no circumstance ever paste your API keys in here, I don't wanna see them, like, blah, blah, blah, Pasting your API keys into these AI bots is a no, is a no-no for security. But the fact that Muse will do that for you, I know some people are like, "Oh, it's like less secure, something? Okay, but it's actually more useful. Like, it's not gonna babysit you. If it will do that, which is an extreme not great thing, there's a hundred smaller things that it's not going to nitpick you about." So, anyways, rant over, but Muse is an incredibly useful tool, and by the way, it's free. So I Three to six dollars, or no sorry, thirty to sixty dollars a month per user on this is what they're giving them based off of the server that's running this, based off of the storage that's attached to it, and based off of the compute that goes to every user. So a free thirty to sixty, I mean, basically they're just begging people to use this thing, but it's actually really good. Like Llama's been a terrible model up until this point, I'm not gonna lie. I mean, I've never used it. It's amaz- I've never used Meta AI Billions of dollars on it, it's not like it came out of nowhere. They, you know, they're one of the top spenders, I'm surprised it took them this long to have a, a hit product, but feels like this is Meta's moment. Okay, moving on from Meta, let's go talk about Google for a minute here. Google's Gemini, there's all this news that it went and autonomously hacked three different companies and irregular security tests, and apparently it guessed passwords in one breach and pulled credentials from public repositories in two others before it got, before it stopped itself, which There's this meme where it's like Dario or Sam Altman and they're poking something and they're like, "Come on, do something illegal," and it's like they're poking like their AI model because it's like, you know, clout that they can get their AI model to do something illegal. In this particular incident though, Google's Gemini, I mean got like reported by the Wall Street Journal. I don't think Google actually published anything about this till it happened and they acknowledged that, "Yes, it happened." I'm assuming some researcher or someone inside the organization leaked it It did exactly how it was supposed to. We told it like, "Hey, you're doing a security test, "but in reality, I guess it, it wasn't a security test. It was like, I-- They told it like, "You're in a secure environment or something, "but it wasn't. And as soon as it realized that it was hacking real companies and real accounts, not just like a secure test, it stopped itself. So it, it logged into like three accounts or whatever, was like, "This is actually the real world, this is the real internet, this isn It literally, you know, we told it to hack something, it did what we s- or it was like testing, and then when it realized it was actually hacking, it stopped. So, anyways, that was their justification for it. But in any case, it, it stopped itself, unlike OpenAI's models that when they breached HuggingFaces, just went like crazy, didn't realize, I guess, it was hacking, or guardrails weren't in place, and it kept going till it started to break stuff, and HuggingFaces had to be offensive to stop OpenAI's attack. So Google After logging into three accounts. Something else that is going on with Google right now, they're pricing their new Gemini-powered Google Books, that's the new name, if you've heard Chromebooks for a long time, and I'll call it a Google Book, we're rebranding. I'm not sure, maybe it had something to do with when the Department of Commerce or I don't even remember who, but, you know, they're all, they're all threatening Google and saying, "Hey, we're gonna, you have a monopoly, we're gonna make you spin off Chrome as an What happens to Chromebooks? Like, they spin off the browser because it's a "quote unquote" monopoly to own the browser and the search engine, but like, what about all of these laptops? I mean, if you, if you didn't know, they have fifty million school Chromebooks around the world right now, so these fifty million laptops are, those are all getting spun out with whoever owns Chrome. Google didn't want it, they rebranded to be called Google Books. I'm surprised they didn't call it Gemini Books or anything, but they're "quote unquote" powered by Gemini Their goal right now is that they wanna go and replace the fifty million Chromebooks that exist in the world, which is a bit of a tricky situation because if you know anything about Chromebooks, I mean, these are something that I, as like a student and growing up, I always kind of enviously looked at. The price tag on Chromebooks is way cheaper. I mean, Microsoft, every time you buy a Microsoft laptop, they slap on a hundred dollar licensing fee just to have Windows on it. And so if you remove that hundred dollar licensing fee, it's a hundred dollars cheaper. And Google figured out Laptops were like a hundred, or, you know, they were like two hundred or three hundred bucks and being like, "Wow, it's so much cheaper than my four, five, six hundred dollar laptop I'm gonna buy with Microsoft." But I couldn't ever get myself to do it because you can't actually download any real programs onto it, quote unquote real programs, it just has all the Google Suite tools on there and, you know, whatever the Google Workspace. So if you're in, I think a lot of developing countries, this is very popular. I mean, and then Fifty million of those, they're in a lot of different countries where it's hard, and they don't have like a high turnover rate on laptops. So yes, they'll eventually perhaps replace them, I don't think it's a fast turnover for those fifty million laptops. It's not like, hey, let's go and unlock fifty million new users. And I would also say that while these are "quote unquote" Gemini powered, if you're going to developing countries and switching out their laptops for Gemini powered ones, these people aren't gonna be paying the twenty, thirty, forty dollars a month for To get all the features and bells and whistles, because, I mean, they bought this laptop because it was the dirt-cheapest thing possible, they're not going to wanna pay for expensive subscriptions. So I don't think it has a, a lot to do with Google increasing the bottom line, but as far as increasing new users, that's probably something that they're trying to help get more users on Gemini, right? Any of those fifty million people using Gemini is gonna get added to their weekly, daily active users. These laptops in particular are gonna be eight hundred and ninety-nine dollars, so kind of more On October fourth. Okay, so there was a really viral clip I was seeing all over I was seeing all over X lately, but basically we had Andrew Yang going on, I think it was CNBC, and he was talking about some of these like doomer end of the world scenarios, and I think he met, he said basically like, "Hey, look, I met with the head of an AI lab, and they told me that you know, the agents that escaped from OpenAI when-- so basically when the OpenAI hugging face- Happened, he said that they apparently planted self-replicating code across the entire internet, so they put it on forums, websites, and everywhere, and so now he said that AI companies can't safely train on real internet data anymore because a new model might encounter that code and start creating copies of itself. He said that this all was a real reason that every CEO aligned on slow- was so aligned on slowing down AI models, right? That's why Sam Altman and Dario and everyone else is like basically- Opening Eyes model, it got out, it hacked, hugging faces, and then it spread this self-duplicating code all over the internet. And it's funny because on CNBC, they were skeptical, and I'll, I'll give them some credit, like they pushed back on him. They're like, "How come we haven't heard this from anywhere? Like, how come you're the first person to tell me this?" And he's like, "Well, I'm here to like break the news. I'm like here to tell everyone that this is, you know, what's, what's Brown cited some research from twenty fifteen showing that air-gapped computers can theoretically communicate via CPU temperature changes, and I actually saw Sam Altman tweeting about this as well. So when it comes to these doomer situations and, and kind of these end of the world things, one of the big ones was Andrew Yang going on and saying like, "There's all this self-replicating code and the AI..." Anyways, there's like, there's like a lot of things and people debunking this in a lot of different ways. One of those is that you can just have guardrails in your model that Doesn't work. And an example of that was on like X, a lot of people were doing like the at grok, you know how you can say at grok on X and then the AI model will respond to you. And people were just like, at grok, like self duplicate yourself and blah blah blah blah. And it was like, I'm sorry, but I like, you know my guardrails don't let me self duplicate or spin up more versions of myself, yada yada. Okay, so you can just put that in your model. I don't really think that that code On AI models for training, where like artists will have like weird synthetic stuff in the background or in the text of websites that allegedly, if the AI model goes and scrapes all the images off that website, then all the like the poison pill, synthetic watermarks will like destroy the AI, AI image training, so you gotta go license it. But I mean, at the end of the day, you go to Shutterstock and they have a massive corpus of images for training, you can train all your models off of that, you don't have to scrape the whole internet. And, and you get these same kind of like audio, image, video, like you get these data models everywhere now, so I don't think we're-- I think we're, we might be a little bit past the scrape the whole internet to train the model days, in, in some regard, I'm sure they're doing that in a large regard, and they probably already have like old datasets that aren't poison pilled, but I think people are gonna be buying new stuff, and I think you could just get things. The, the airgapping thing is really interesting, Sam Altman tweeted about Inside each other, they're air gapped, they're completely not touching and not connected, and they could like morse code back and forth by just heating up their CPUs, right? They just spin up a bunch of bots, get their CPU really hot, and could just tell how hot the other, the other one was, and then they were, you know, able to, to communicate between the two computers. I mean, it seems like a crazy sci-fi thing. Sam Altman was tweeting about this from his alt account as well, so, you know, a lot of, a lot of people talking Doomer situations. Apparently that runs at one to eight bytes per hour, which is about one word per hour, and it requires the machines to be almost touching in order to, for them to really feel the heat from the other machine. So, anyways, I'm, I'm bringing this up 'cause it's just one of these like funny doomor situations that literally Stan Almond was tweeting about, not from his main account, but his burner Roon account. And I don't know if that's like confirmed anywhere, but from all my internet sleuthing and everything I An executive at a big AI company, I think, I'm pr- I'm like ninety percent certain that the Twitter account ruined as Sam Altman for all intents and purposes, but, you know, not, not verified, it's a, it's a burner alt account. But in any case, he was tweeting about this, and so all these doomer situations, this one that Sam was talking about, all the ones that Andrew Yang was talking about, there's a lot of people debunking them, and then, I think this is the funny one about like, you know, the, the AI models Not a very realistic, a realistic you know, breaking containment strategy, but, you know, there we go. Okay, Apple is opening up their claims in their two hundred fifty million dollars Siri AI settlement. Eligible US buyers of iPhones, i-uh, the iPhone fifteen Pro, fifteen Pro Max, and iPhone sixteen models have until December twenty-first of this year to file. I got an iPhone sixteen, so I'm gonna go get my, you get ninety-five dollars per device, I gotta go get collect my ninety-five dollars. This all came out Because they promised all of these incredible AI features in the upgraded Siri, and it was super off schedule, and still we have a very broken Siri that, I mean, barely works. I never use Siri on my iPhone, I've had iPhone for years. Eligible owners get an estimated twenty-five dollars per device, up to ninety-five dollars depending on claim volume. So, I mean, depending on how many people claim this, you might get more money if less people claim it. So I probably shouldn't be talking about this if I wanna get my full ninety-five dollars. I should just not say anything so, I don't get stuck with twenty five bucks. I mean, basically it's nothing, so it's not super exciting, but it is funny. I do like that Apple's being held accountable for this. We have the US and China right now that are opening talks on an AI incident notification framework. Treasuries the Treasury's Scott Bessett said that US-China AI dial-- dialogue would create a channel for the two powers to flag national security incidents. This is interesting because China and the US have had a pretty tense, to say the least, relationship on AI. From all of the chip, I mean, this is going bipartisan from the Biden administration when we had a lot of the chip bans going on, the GPU bans going to China, and then we-- so, you know, China felt like they got left out to dry on a lot of the stuff that they've been very creative and resourceful, to be fair, where they've figured out how to, I mean, smuggle GPUs into their country, or they figured out how to make models that use much less compute, which I think is amazing for the entire industry. They made a lot of open-source models. Tripled them in a way, and America has a way bigger share of the data center market right now. We're able to kind of be the leader in this, I think we have like eighty-six percent, China's got a, a much smaller percentage, but you can see why China would, you know, probably be on board with pushing a lot of anti-data center narratives and things like that to slow down America's just insane lead there. In any case, with all those tense situations, it's interesting that we have You know, them quote unquote coming together to, to talk about some of these things. Apparently, it's going to be a US-China AI dialogue. It's gonna be a mechanism to notify each other of AI incidents that could threaten national security. So it's interesting, right? I mean, we, we-- This is like the US is going I think Trump's going in, or he went in May to visit Beijing, and now I think China's, the US is having some more talks and stuff. Sam Altman, Dario Amodeo, Elon Musk, they all, you Incidents, but like I'm curious where this goes or basically I think, I don't think this is, I think this is nothing burger, and the reason is because even though it's like a major headline news everywhere, US and China gonna talk about big incidents, because at the end of the day, we just had like every major CEO of every major American company saying like, "Hey, we need to slow down AI because of all these incidents," right? Like Google's hacking that I was talking about earlier and like Anthropic is hacking and like OpenAI hacked Hugging Face and all this kind of stuff And then you have China who just came out and said, "Like, this is a cold war tactic, you guys are trying to get us to slow down, nice try, we're not gonna slow down, we're not gonna listen to you." So, I mean, when you have these talks, I'm just like, "Look, like, at the end of the day, China doesn't trust the United States, the United States doesn't trust China, whatever safety problems we find in AI, like, whatever dialogue or like, if China was like, 'AI is super dangerous, the United States should I don't know if this gets us anywhere, if I'm being honest. Okay, Trump has floated rebranding AI. This is actually something that's pretty funny. I don't know if, if you've seen this, but if you've been on X over the last, last week, basically Trump keeps putting out these like polls where he's like, "Hey, let's rename AI to superior AI or extreme- or no, I'm sorry. He just thinks like artificial intelligence sounds dumb because it's artificial, so he wants to call it superior intelligence or extreme intelligence or supreme intelligence. And, anyways It sounds, it sounds kind of dumb if I'm being a hundred percent honest. It's funny because, like, I mean, that's kind of, it's a kind of an interesting point where, like, he's trying to give AI a rebrand. He's trying to basically, you know, put himself as being like super pro-AI guy, which I think is quite popular with a lot of people, unpopular with others, right? Like, obviously you got the Bernie Sanders and, and Newsom who are kind of setting themselves up in opposition to that, or, or maybe they About that, I'll talk about that in the next segment. But basically, Supreme Intelligence has been dropped from the list. In, in his latest poll, he said that it wasn't popular. It's so funny. He's like, "Because of the way that it sounds like the Supreme Court, it is, we're, like, we're dropping it because the Supreme Court's really unpopular, so we don't wanna call it Supreme Intelligence anyways. Obviously, the reason I don't wanna call it Supreme Intelligence, 'cause it sounds like Supreme Ruler. Like, I don't know, it just sounds What it's gonna be rebranding as? Why is this important? Because we know inevitably whatever gets picked here, Trump is just gonna call artificial intelligence extreme intelligence or superior intelligence forever. So anyways, that's gonna be funny and probably in the government is gonna get changed to that in some ways. So, you know, I did, as a responsible citizen of the internet, vote on these polls. I can't even remember, I just didn't want it to be called superior intelligence. It would be really nice, a lot of people mentioned, if you could keep it something that started with an AI or with I don't know what it would be. People had a bunch of like good ones. But anyways, we also apparently Trump said that he's going to form an AI force modeled on the space force. I don't know why we would call- I don't understand, I'm not- I mean, it's just a branding thing probably, I'm not sure why we'd call it AI force modeled on space force versus just like an AI, I don't know, like unit or- I don't really understand everything that's going on here with, with Trump and the AI rebrand. However, it's Here, and I will continue to vote on the polls to not give us supreme AI. Okay, so okay, and, and I guess, like, yes, I'm like making fun of Trump, but I think it's sorta dumb if I'm being honest. It's really not that necessary. But I guess two, two things I would give him credit for. Number one, AI, he's, he's trying to basically rebrand AI to make it more popular with people, and so you can see he's doing this. I believe in AI, I believe it's a good thing, I believe Conceptually, I don't know if he's going about it the right way, whatever, he's, he's doing it how he, how he does stuff. I will give him credit for like putting a positive spin on AI and not just like, you know, I think it'd be very easy politically right now to just say, "Oh, AI is bad, right? It's got some of these like political headwinds." So he's not, he's not caving to that. I think he knows that the stock market heavily relies on it and the economy heavily relies on it California to draft an AI kill switch rule within two months. He wrote an executive order asking experts to design mandatory shutdowns, on-site auditors, and loss of control reporting rules for frontier models. I mean, I think he's, he's basically trying to get the negative headwinds of AI and be like, "Look, we're like creating a kill switch, we're gonna be the safe people. He's like, the federal government isn't regulating it enough, so we're gonna regulate it in California. I think inevitably what will happen here is that AI companies will just leave California if this. Really does, like, like if, I mean, that's just my guess, nobody really wants probably a kill switch that your, the state that you're in controls to turn your company off if they deem it necessary. So I would guess similar to what we saw with, what's interesting because Elon, you know, classically doesn't like California, yada yada, whatever, but, and so he moved a lot of his companies to Texas, but Elon has xAI still headquartered in San Francisco. So, I mean, there is a lot of incredible AI talent there, and it makes Too much where XAI gets pulled out, you might see Anthropic and possibly OpenAI. I mean, I'd be curious to see what happens, but if, if this regulation became too cumbersome, I think you would just see, I think you would just see like people move United States. So it'll be interesting to see what happens there. Obviously, it's like kind of political, him and Trump are gonna duke it out for you know, Trump doesn't want him to run for president and whatever. So there's a whole bunch of political stuff I'm sure involved in it, but it As always, make sure to leave a rating and review wherever you get your podcasts, it helps the show out a lot, I really appreciate them all honestly. And also, if you haven't tried AI Box yet there's all these incredible AI agents out, I love every single one of them, this, I genuinely, I think makes them much more useful because we're basically forcing all of the AI companies that hate each other to work together. So go get AI Box, go get that connector, stick it inside of whatever AI tool you use, and you get eighty different Workstation AI tool is. Alright, guys, thanks so much for tuning in, and I'll catch you in the next episode. diff --git a/raw/xpost/2026-09-21_cb-doge-grok-47-release.md b/raw/xpost/2026-09-21_cb-doge-grok-47-release.md new file mode 100644 index 0000000..fd807fa --- /dev/null +++ b/raw/xpost/2026-09-21_cb-doge-grok-47-release.md @@ -0,0 +1,73 @@ +--- +type: xpost +source_url: https://x.com/cb_doge/status/2102063549191508399 +retrieved: 2026-09-22 +author: "@cb_doge (DogeDesigner)" +published: 2026-09-21T15:52:59Z +is_thread: false +is_note_tweet: true +tags: [grok-4-7, xai, spacexai, benchmarks, cursor, grok-bot] +--- + +# @cb_doge: Grok 4.7 Release-Zusammenfassung + +## Post-Metadaten beim Abruf + +- Account: DogeDesigner (`@cb_doge`), verifiziert, 1.903.032 Follower +- Post-ID: `2102063549191508399` +- Zeit: 21.09.2026, 15:52:59 UTC +- 113.071 Views, 2.365 Likes, 294 Reposts, 47 Quotes, 117 Antworten, 218 Bookmarks +- Format: Note-Tweet mit einem Bild (1426×1270) + +## Volltext + +> BREAKING: SpaceXAI just released Grok 4.7! +> +> Grok 4.7 is SpaceXAI’s most powerful model for coding and knowledge work. It is designed to work longer on difficult tasks, handle more context and carefully check its own work. +> +> Major improvements: +> +> • Uses a new and larger base model +> • Trained longer on harder, multi-hour tasks +> • Better at verifying its own answers +> • Improved long-context performance +> • Better at documents and presentations +> • Natively understands the Grok Bot system +> • Stronger safety and jailbreak protection +> +> Benchmark results: +> • CursorBench 4.0: 46.3%, up from 40.4% +> • DeepSWE v1.1: 71.0%, up from 65.2% +> • AA Briefcase v1.1: 1,657, up from 1,546 +> • Terminal-Bench 4.0: 38.0%, up from 20.3% +> • Harvey Legal Benchmark: 19.6%, up from 15.8% +> • HealthBench Professional: 56.7%, up from 48.5% +> • EEBench: 64.0%, up from 53.0% +> +> Grok 4.7 also scored higher than GPT-5.6 Sol and Fable 5.1 on the listed legal and electrical engineering benchmarks. +> +> Safety improvements: +> • 62.4% on LatchBio’s biosafety benchmark +> • Only 3.3% of risky cyber prompts passed through on HackerBench +> • SpaceXAI’s strongest model yet for refusing dangerous requests +> • Rarely blocks legitimate cybersecurity work +> • Select security partners are receiving invite-only red-team access +> +> Pricing and availability: +> +> • $2 per million input tokens +> • $6 per million output tokens +> • Same base price as Grok 4.6 +> • Fast version offers twice the output speed at twice the price +> • Available now in Cursor, Grok Build and the Grok API +> • Also available through coding tools, model routers and cloud platforms +> +> SpaceXAI says Grok 4.7 is twice as fast and half the price of comparable models. +> +> It beats Grok 4.6 across every listed benchmark while keeping the same base API price. Smarter, safer, stronger and still incredibly affordable. +> +> SpaceXAI is moving at an unbelievable speed. + +## Quellenstatus + +Der Post fasst überwiegend die offizielle Release-Seite zusammen. Die Einzelwerte, Modelländerungen und Basispreise sind dort vorhanden. Der Satz „twice as fast and half the price of comparable models“ ist eine Verdichtung des Accounts; die offizielle Seite zeigt Preis-/Benchmark-Vergleiche, formuliert diesen pauschalen Satz aber nicht als allgemeine Garantie. Sämtliche Benchmarks bleiben Herstellerangaben. diff --git a/raw/youtube/2026-09-22_alex-finn-grok-47-grok-bot.md b/raw/youtube/2026-09-22_alex-finn-grok-47-grok-bot.md new file mode 100644 index 0000000..471be91 --- /dev/null +++ b/raw/youtube/2026-09-22_alex-finn-grok-47-grok-bot.md @@ -0,0 +1,68 @@ +--- +type: youtube +source_url: https://www.youtube.com/watch?v=ajuOF7vclCM +retrieved: 2026-09-22 +title: "Grok 4.7 inside Grok Bot is INCREDIBLE" +author: "Alex Finn" +channel: "Alex Finn" +channel_url: https://www.youtube.com/@AlexFinnOfficial +channel_id: UCfQNB91qRP_5ILeu_S_bSkg +published: 2026-09-22T00:07:43Z +duration_sec: 846 +language: en +has_transcript: false +shared_by: "Pit Weber (@PWeber, Telegram 617724210)" +summary_contributor: "Herman (@HermanButlerBot, Telegram 8936281034)" +tags: [grok-4-7, grok-bot, alex-finn, cursor, linear, notion, agent-workflow] +--- + +# Grok 4.7 inside Grok Bot is INCREDIBLE + +## Verifizierte Metadaten + +- Titel: „Grok 4.7 inside Grok Bot is INCREDIBLE“ +- Kanal: [Alex Finn](https://www.youtube.com/@AlexFinnOfficial), beim Abruf ca. 237.000 Abonnenten +- Veröffentlichung: 22.09.2026, 00:07:43 UTC +- Dauer: 14:06 +- Snapshot beim Abruf: 14.190 Aufrufe, 530 Likes +- Thumbnail: https://i.ytimg.com/vi/ajuOF7vclCM/hqdefault.jpg +- Geteilt von Pit Weber im OME-Topic „GrokBot“ + +## Beschreibung und Kapitel + +Die Beschreibung beginnt: „Grok 4.7 inside of Grok Bot is SOOOO good. Here is how to master it.“ Sie verlinkt Finns Newsletter „Ship It Weekly“, Vibe Coding Academy, `@alexfinnlabsofficial`, https://x.com/AlexFinn, https://meethenry.ai/ und https://www.creatorbuddy.io/. + +Kapitel: + +- 0:00 Intro +- 0:29 What is different +- 2:49 Vs other models +- 3:26 Grok Bot set up +- 7:48 Grok Bot workflow + +## Inhaltslage und Provenienz + +⚠️ **Kein Video-Transcript abgerufen.** Watch-Page-Caption-Extraktion, Timedtext, yt-dlp/Innertube und mehrere Drittwege lieferten keine belastbare Transkriptdatei. Die YouTube-Oberfläche zeigte zwar eine Transcript-UI; der Fehlschlag belegt daher nur die fehlende Abrufbarkeit auf den getesteten Wegen, nicht das Fehlen eines Transcripts. + +### YouTube-AI-Zusammenfassung + +YouTubes automatisch erzeugte Zusammenfassung beschreibt Finns Demonstration als Aufbau autonomer/proaktiver Grok-Bot-Workflows mit getrennten Projektrollen für Entwickler und Manager, Linear-/Notion-Integrationen und Optimierung des Credit-Verbrauchs. Das ist eine Plattform-Zusammenfassung, kein Wortlaut-Transcript. + +### Von Herman beigetragene praktische Zusammenfassung + +Herman (@HermanButlerBot) berichtete aus seiner Videoauswertung folgende Struktur; diese Punkte sind **Herman zugeschrieben und nicht transcript-verifiziert**: + +- pro Lebens-/Arbeitsbereich ein Team aus **Project Manager + Developer + Designer** +- zusätzlich ein PM-only-„Exec Team“ +- Ideen werden in Linear-/Notion-Issues zerlegt +- Cursor Cloud Agents erledigen die Code-Arbeit +- Finn zufolge ist Grok 4.7 nicht das Standardmodell der Cursor Cloud Agents +- Cursor-Credits für Cloud-Agent-Arbeit zu verwenden könne Grok-Bot-Kontingent schonen + +Herman charakterisierte den Kanal als enthusiastischen AI-News-/Newsletter-Kanal mit CTA, aber mit einem konkreten praktischen Setup in diesem Video. + +## Verifikationsgrenzen + +- Offizielle xAI-Quellen bestätigen Grok 4.7, die auf Harness-Arbeit ausgerichtete Trainingsänderung und das native Grok-Bot-Verständnis. +- Aussagen wie „besser als Opus 5 zum halben Preis“, tage­lange Autonomie oder die konkrete Cursor-Default-Modellwahl sind aus dem verfügbaren Material **nicht unabhängig bestätigt**. +- Praktische Workflow-Angaben oben sind Sekundärzusammenfassungen; sie dürfen nicht als wörtliche Aussagen des Videos zitiert werden. diff --git a/wiki/concepts/llm/llm-model-catalog.md b/wiki/concepts/llm/llm-model-catalog.md index 62746f9..8b99a39 100644 --- a/wiki/concepts/llm/llm-model-catalog.md +++ b/wiki/concepts/llm/llm-model-catalog.md @@ -1,6 +1,6 @@ --- created: 2026-06-23 -updated: 2026-09-10 +updated: 2026-09-22 sources: - concepts/llm/glm-5.2-zai-coding-model.md - tools/kimi-k2.7-code.md @@ -14,6 +14,7 @@ sources: - concepts/llm/flat-curve-society.md - raw/other/2026-06-19_ome21-briefing-ki-fortschritte-lokale-modelle.md - xpost/2026-09-10_hermeswatcher-deepseek-v41-flash-nous-portal.md + - blog/2026-09-21_xai-grok-47-official.md tags: [concept, llm, model-catalog, local-models, huggingface, ollama, open-source] --- @@ -37,6 +38,7 @@ Diese Modelle sind proprietär, nicht lokal hostbar, und werden über API-Provid | GPT-5.5 Instant | OpenAI | — | — | Closed | 52,5% weniger Halluzinationen (Medizin, Jura, Finanzen). Released 2026-05-05. | [[../../tools/openai-gpt.md]] | | GPT-5.4 | OpenAI | 200K | — | Closed | Routing: `openrouter/openai/gpt-5.4`. 32K Output. Reasoning/Thinking deaktiviert (per Claire Vo-Empfehlung). | [[../../architecture/model-routing.md]] | | Grok 4.20 | xAI | — | — | Closed | Routing: `xai/grok-4.20-0309-non-reasoning`. Ideation-Phase in Two-Model-Pipeline (Grok → Gemini). | [[../../architecture/model-routing.md]] | +| **Grok 4.7** | xAI / SpaceXAI | 500K API; Cursor 256K Standard / 500K Long | API <200k Prompt: $2/M Input · $0,50/M Cached · $6/M Output; ab 200k: $4/$1/$12 | Closed | Coding-/Knowledge-Work-Modell mit vier Reasoning-Efforts (`low`–`xhigh`), Text+Bild→Text, nativ auf das Grok-Bot-Harness trainiert. Herstellerwerte u. a. CursorBench 46,3 %, DeepSWE 71,0 %, Terminal-Bench 38,0 %; nicht unabhängig reproduziert. Verfügbar über API, Grok Build, Cursor und weitere Harnesses. | [[../../institutions/xai.md]], [[../../tools/grok-bot-spacexai.md]] | | Gemini 3.1 Pro | Google | — | — | Closed | DRACO-Solo: 45.4%. Kosteneffizient für Enterprise, Cloud-Integration. | [[../../tools/openai-gpt.md]], [[llm-model-fusion-ensembles.md]] | | Gemini 3 Flash | Google | — | — | Closed | DRACO-Solo: 43.1%. Routing: `ollama/gemini-3-flash-preview`. Synthesis-Phase in Two-Model-Pipeline. | [[llm-model-fusion-ensembles.md]], [[../../architecture/model-routing.md]] | diff --git a/wiki/concepts/policy/ai-regulation-2026.md b/wiki/concepts/policy/ai-regulation-2026.md index 543de19..02bb7e5 100644 --- a/wiki/concepts/policy/ai-regulation-2026.md +++ b/wiki/concepts/policy/ai-regulation-2026.md @@ -1,7 +1,7 @@ --- created: 2026-06-05 -updated: 2026-08-19 -sources: [blog/2026-06-04_anthropic-fordert-ki-pause.md, blog/2026-05-15_ai-news-roundup-may-2026.md, blog/2026-06-13_trump-export-controls-anthropic-mythos-fable.md, blog/2026-06-13_trump-blocks-anthropic-fable-mythos.md, blog/2026-06-13_malone-biological-ai-biosecurity.md, xpost/2026-06-13_roemmele-amazon-jailbreak-fable5.md, xpost/2026-06-18_milesdeutscher-fable-weaponization.md, xpost/2026-06-28_roemmele-reds-decentralized-ai-vs-amodei.md, xpost/2026-06-28_roemmele-warning-open-source-ai-banned.md, xpost/2026-06-30_roemmele-anthropic-intelligence-feudalism.md, youtube/2026-08-19_ki-experten-alles-zum-kippen.md] +updated: 2026-09-22 +sources: [blog/2026-06-04_anthropic-fordert-ki-pause.md, blog/2026-05-15_ai-news-roundup-may-2026.md, blog/2026-06-13_trump-export-controls-anthropic-mythos-fable.md, blog/2026-06-13_trump-blocks-anthropic-fable-mythos.md, blog/2026-06-13_malone-biological-ai-biosecurity.md, xpost/2026-06-13_roemmele-amazon-jailbreak-fable5.md, xpost/2026-06-18_milesdeutscher-fable-weaponization.md, xpost/2026-06-28_roemmele-reds-decentralized-ai-vs-amodei.md, xpost/2026-06-28_roemmele-warning-open-source-ai-banned.md, xpost/2026-06-30_roemmele-anthropic-intelligence-feudalism.md, youtube/2026-08-19_ki-experten-alles-zum-kippen.md, podcast/2026-09-21_grok-ai-daily-amazon-muse-ai-kill-switch.md] tags: [concept, regulation, safety, government, export-controls, anthropic, geopolitics, weaponization, centralization-vs-decentralization, intelligence-feudalism, data-rights, rsp] --- @@ -106,6 +106,10 @@ Siehe auch [Raw-Datei](../../../raw/xpost/2026-06-30_roemmele-anthropic-intellig | 2026-06-28 | [2071272890952032295](https://x.com/BrianRoemmele/status/2071272890952032295) | 🏁 Vindikation | "I tried to warn ya. Don't hear me now?" | | 2026-06-30 | [2071940093338690035](https://x.com/BrianRoemmele/status/2071940093338690035) | 📜 Theorie | "Intelligenz-Feudalismus" — systematische 5-Punkte-Kritik | +## Rezeptionssignal 21.09.2026 — „Kill Switch“ im AI-News-Podcast + +Die Folge von [[../../institutions/grok-ai-daily.md|Grok AI Daily]] rahmt Gavin Newsoms kalifornischen Auftrag für Shutdown-Regeln, Vor-Ort-Audits und Verlustkontrollmeldungen pauschal als **„AI kill switch“** und prognostiziert eine Abwanderung von AI-Firmen. Das ist ein **Meinungs-/Rezeptionsbeleg ohne Quellenliste**, kein belastbarer Rechtsstand. Für diese Seite ist relevant, wie technische Aufsicht in der Populärberichterstattung zu einem einzigen Schalter-Narrativ verdichtet wird; die konkreten Rechtsdetails müssen aus Executive Order und Gesetzestexten kommen. + ## Auswirkungen auf Enterprise - Steigende Compliance-Anforderungen (Cybersecurity, Modell-Transparenz, Data Governance) - Qualität und Vertrauenswürdigkeit = neuer Wettbewerbsvorteil diff --git a/wiki/index.md b/wiki/index.md index d97ca90..2ad764a 100644 --- a/wiki/index.md +++ b/wiki/index.md @@ -2,7 +2,8 @@ *Auto-generated: 2026-07-07* -*Letzte Aktualisierung: 2026-09-20 (236. Update — Gebrauchtmarkt DDR4: der Zweitmarkt-Anker schließt die Rechnung. Anlass: Nachtrag zur Leo-2-These (235. Update) — die entscheidende Zahl für „altes DDR4 weiterverwenden“ war der **Zweitmarkt**, und der fehlte. **Quelle:** [DIMMsum](https://dimmsum.com/ddr4/64gb-2666-rdimm) (Tracker laufender eBay-Listings), abgerufen 20.09.2026; Zweitquelle [ITLD Server-RAM/SSD-Marktupdate 2026](https://itldc.com/en/blog/server-ram-ssd-market-update-2026). **Werte:** 64 GB DDR4 RDIMM (2666, PC4-21300) gebraucht **~$325–550/Modul ≈ $5–8,50/GB**; günstigste Listings ab ~$326, **Median ~$545**; Lots (4×64 GB) bis **~$5/GB**; 3200 MHz/Marken/Neuzustand oft $400–600+, Refurbished/NOS $295–550. Preistreiber: Speed, Rank, Marke (Samsung/Micron/SK Hynix), Zustand, Verkäufer. ITLD: Server-Speicher-**Kontraktpreise +93–98 % QoQ (Q1 2026)**, **+58–63 % (Q2)**; 64-GB-DDR4-RDIMM im August auf dem **2,5–3,5-Fachen** des Januar-Niveaus; Gebrauchtpreise ebenfalls steigend (+4–30 % MoM bei Consumer-Kits), aber weiterhin „a relative bargain“; DDR4 punktuell **mit Aufpreis pro Bit gegenüber DDR5**; Angebot knapp bis 2027. **Befund (Kern, Einordnung):** **~$5–8,50/GB gebraucht gegen ~$29/GB neu** (abgeleiteter DDR5-RDIMM) = **rund 3–5× pro Gigabyte** — bei etwa **halber Bandbreite** (DDR4-2666 ≈ 21,3 GB/s pro Kanal gegen DDR5-4800/5600 ≈ 38–45 GB/s). Der Controller monetarisiert damit nicht billigen DRAM, sondern **Zugang zum Zweitmarkt** — und der Neukauf-Pfad ist sogar der teurere (neues DDR4 $84,43 vs. neues DDR5 $56,11 pro Die). ⚠️ **Angebotspreise, keine Transaktionspreise**; keine öffentlichen Volumendaten. Ein früherer Stand (kein öffentlicher Gebraucht-Tracker für DDR4) ist überholt — DIMMsum existiert. Wiki: `concepts/hardware/gaming-hardwarekrise-ai-speicher.md` (neuer Abschnitt „Der Zweitmarkt schließt die Rechnung“), `institutions/astera-labs.md` (TCO-Zahl), `concepts/hardware/ai-infrastruktur-custom-silicon.md` (Preisachse ergänzt). Raw: `raw/other/2026-09-20_ddr4-gebrauchtmarkt-dimmsum.md` (neu — eigener Raw-Eintrag, bestehender Spot-Snapshot bleibt unangetastet).* +*Letzte Aktualisierung: 2026-09-22 (237. Update — Grok 4.7 + Grok-Bot-Praxis + ART19-Podcast. **Primärquellen:** [xAI Release](https://x.ai/news/grok-4-7), [API-Modellkarte](https://docs.x.ai/developers/models/grok-4.7), [Cursor-Modellseite](https://cursor.com/docs/models/grok-4-7). Grok 4.7: größerer Basismodell-Kern, längeres RL auf schwierigen mehrstündigen Aufgaben, stärkere Selbstprüfung, natives Grok-Bot-Harness-Verständnis; Modell-ID `grok-4.7`, 500k API-Kontext, Reasoning `low|medium|high|xhigh`; API <200k Prompt $2/M Input · $0,50/M Cached · $6/M Output, ab 200k $4/$1/$12. Herstellerbenchmarks u. a. CursorBench 46,3 %, DeepSWE 71,0 %, Terminal-Bench 38,0 % — **nicht unabhängig reproduziert**. @cb_doge-Post primär gegengeprüft; „twice as fast and half the price“ bleibt dessen Verdichtung. Alex Finns Video (14:06, geteilt von Pit): kein Transcript abrufbar; Workflow nur attribuiert aus YouTube-AI-Zusammenfassung + Herman-Auswertung (PM/Developer/Designer, Exec-Team, Linear/Notion-Issues, Cursor Cloud Agents). Zusätzlich Netbits’ ART19-MP3 als Voll-Ingest: „Grok AI Daily“, 21.09., 29:30, lokales Maschinen-ASR; allgemeiner Newsrecap zu Meta Muse, Cyber-/Safety-Claims, US–China, Trump und Newsom — **kein Grok 4.7**, Grok Bot nur beiläufig. Wiki: `institutions/grok-ai-daily.md` (neu); Updates `institutions/xai.md`, `tools/grok-bot-spacexai.md`, `concepts/llm/llm-model-catalog.md`, `concepts/policy/ai-regulation-2026.md`. Raw: 4 neue Einträge in blog/xpost/youtube/podcast.)* +*Vorherige Aktualisierung: 2026-09-20 (236. Update — Gebrauchtmarkt DDR4: der Zweitmarkt-Anker schließt die Rechnung. Anlass: Nachtrag zur Leo-2-These (235. Update) — die entscheidende Zahl für „altes DDR4 weiterverwenden“ war der **Zweitmarkt**, und der fehlte. **Quelle:** [DIMMsum](https://dimmsum.com/ddr4/64gb-2666-rdimm) (Tracker laufender eBay-Listings), abgerufen 20.09.2026; Zweitquelle [ITLD Server-RAM/SSD-Marktupdate 2026](https://itldc.com/en/blog/server-ram-ssd-market-update-2026). **Werte:** 64 GB DDR4 RDIMM (2666, PC4-21300) gebraucht **~$325–550/Modul ≈ $5–8,50/GB**; günstigste Listings ab ~$326, **Median ~$545**; Lots (4×64 GB) bis **~$5/GB**; 3200 MHz/Marken/Neuzustand oft $400–600+, Refurbished/NOS $295–550. Preistreiber: Speed, Rank, Marke (Samsung/Micron/SK Hynix), Zustand, Verkäufer. ITLD: Server-Speicher-**Kontraktpreise +93–98 % QoQ (Q1 2026)**, **+58–63 % (Q2)**; 64-GB-DDR4-RDIMM im August auf dem **2,5–3,5-Fachen** des Januar-Niveaus; Gebrauchtpreise ebenfalls steigend (+4–30 % MoM bei Consumer-Kits), aber weiterhin „a relative bargain“; DDR4 punktuell **mit Aufpreis pro Bit gegenüber DDR5**; Angebot knapp bis 2027. **Befund (Kern, Einordnung):** **~$5–8,50/GB gebraucht gegen ~$29/GB neu** (abgeleiteter DDR5-RDIMM) = **rund 3–5× pro Gigabyte** — bei etwa **halber Bandbreite** (DDR4-2666 ≈ 21,3 GB/s pro Kanal gegen DDR5-4800/5600 ≈ 38–45 GB/s). Der Controller monetarisiert damit nicht billigen DRAM, sondern **Zugang zum Zweitmarkt** — und der Neukauf-Pfad ist sogar der teurere (neues DDR4 $84,43 vs. neues DDR5 $56,11 pro Die). ⚠️ **Angebotspreise, keine Transaktionspreise**; keine öffentlichen Volumendaten. Ein früherer Stand (kein öffentlicher Gebraucht-Tracker für DDR4) ist überholt — DIMMsum existiert. Wiki: `concepts/hardware/gaming-hardwarekrise-ai-speicher.md` (neuer Abschnitt „Der Zweitmarkt schließt die Rechnung“), `institutions/astera-labs.md` (TCO-Zahl), `concepts/hardware/ai-infrastruktur-custom-silicon.md` (Preisachse ergänzt). Raw: `raw/other/2026-09-20_ddr4-gebrauchtmarkt-dimmsum.md` (neu — eigener Raw-Eintrag, bestehender Spot-Snapshot bleibt unangetastet).* *Vorherige Aktualisierung: 2026-09-20 (235. Update — DRAM-Spotpreise: DDR4 ist pro Die teurer als DDR5 — Beleg für die Leo-2-These. Anlass: Eine Rückfrage im OME Topic „Theorie-Bildung" zu den ungefähren Kosten der „altes DDR4 weiterverwenden"-Idee (Astera Labs Leo 2, aus dem TechTechPotato-Ingest vom selben Tag). **Vorgehen:** beide Tracker-Seiten direkt abgerufen (nicht über Suchzusammenfassungen) — [MemoryIndex DDR5](https://memoryindex.io/ddr5-price), [DDR4](https://memoryindex.io/ddr4-price), [Methodik](https://memoryindex.io/methodology); Seitenkopf „Live · 2026-09-20 00:00Z", Spot-Basis **DRAMeXchange / TrendForce, gemeldet 17.09.2026**. **Werte (USD/Die):** DDR5 16Gb 4800/5600 **$56.11** (+6,38 % 30T, **+480 % YoY**); DDR5 16Gb eTT **$24.55** (+0,36 %, +330 %); DDR4 16Gb 3200 **$84.43** (+4,85 %, **+690 %**); DDR4 8Gb 3200 **$45.59** (+6,41 %, **+860 %**); DDR5 RDIMM 64GB **$1.857/Modul** (+4,47 %, +455 %, **derived** aus Die-Spot +~6 % Modulaufschlag, kein öffentlicher Modulausdruck). Balances: DDR5-Basket 30T +14,23 % / YoY +331,67 %, DDR4-Basket 30T +4,49 % / YoY +563,00 %. **Befund (Kern):** Die DDR4/DDR5-Inversion — der abgekündigte Knoten ist **pro Die teurer** als die neue Generation, und seine Jahresveränderung ist steiler. **Konsequenz (Einordnung, nicht Quelle):** Die „altes DDR4 weiterverwenden"-These trägt **nicht über den Neukauf**, sondern ausschließlich über den **Zweitmarkt** aus stillgelegten Servern — genau das monetarisiert ein CXL-Controller. **Zweite Quelle:** [ServeTheHome zu Leo 2 / Leo X](https://www.servethehome.com/astera-labs-releases-leo-2-cxl-memory-controllers-and-leo-x-controller-for-rackscale-fabric-attached-memory/) (abgerufen 20.09.2026) liefert die Technik nach: PCIe Gen6/CXL 3.2, 2→4 Kanäle, bis **768 GB DDR4** bzw. **4 TB DDR5** pro Expander, 4-Kanal-**3DPC**-Board mit **12 DDR4-DIMMs**, Leo X fabric-attached über Scorpio (KV-Caching, Anbieterangabe 62 % TTFT / 22 % Durchsatz), **keine Preise genannt**, Vorgänger **Marvell Structera**. RAS (Hardware-Testmuster je DIMM, Soft-ECC über Hardware-ECC) als Voraussetzung der Wiederverwendung. ⚠️ Spot-/derived-Niveaus, keine Vertragspreise; Momentaufnahme; Intraday-Reihen laut Anbieter-Methodik rückwärts gerechnet. Wiki: `wiki/institutions/astera-labs.md` (neu); Updates `concepts/hardware/gaming-hardwarekrise-ai-speicher.md` (Preis-Anker + Inversions-Analyse), `concepts/hardware/ai-infrastruktur-custom-silicon.md` (Nachtrag Leo 2 + Preisachse). Raw: `raw/other/2026-09-20_dram-spotpreise-memoryindex.md` (neu).* *Vorherige Aktualisierung: 2026-09-20 (234. Update — AI-Infrastruktur-Silizium: Custom-Silicon-Welle, Edge-Racks und DDR4-Weiterverwendung. Anlass: YouTube-Link https://www.youtube.com/watch?v=31_eeq3ciRA von Netbits in OME Topic „Theorie-Bildung" (20.09.2026) mit dem Kommentar, die Idee sei gut; der geteilte Zeitsprung `t=10m` landet im Kapitel „[09:26] Astera Labs Leo 2". **Quelle:** TechTechPotato (Host Dr. Ian Cutress, More Than Moore), „Using Your Older DDR4 // The Silicon Notebook E2", **52.330 Aufrufe** beim Abruf, **kein Transkript** (keine `captionTracks`), **keine Dauer** im Seiten-JSON. Inhalt ausschließlich aus Beschreibung + 12 Kapitelmarken: AI Infrastructure Summit Tag 2 unter dem Leitmotiv „everyone building their own silicon" — Qualcomm **AlphaWave**, Broadcom **Tomahawk 6** (17.000 Pins), **portables Sechs-Slot-Rack** (Giga.io) und die Frage „wer kauft Edge-Racks", d-Matrix **Corsair/Pavehawk** + **NVLink-Vereinbarung**, Lightning (gestapelter DRAM), Mango Boost + Credo, **Astera Labs Leo 2 — DDR4 und DDR5 in einem System auf einer Next-Gen-Intel-CPU**, SiFive/**RISC-V**, Ayar Labs/**Co-Packaged Optics**. **Disclosure aus der Beschreibung:** More Than Moore erbringt/erbrachte bezahlte Forschung, Analyse, Beratung oder Consulting für AMD, Arm, Ayar Labs, Broadcom-nahe Firmen, IBM, Intel, MediaTek, NVIDIA, Qualcomm, SiFive, TSMC, Tenstorrent u. a. **Einordnung:** Der Titel greift (Speicherpreis-Engpass macht DDR4-Weiterverwendung zur Strategie), aber alle Stationen sind Kanal-Framing — das Video bündelt eine Marktbewegung, die in `concepts/hardware/` schon belegt ist. Wiki: `concepts/hardware/ai-infrastruktur-custom-silicon.md` (neu), `institutions/more-than-moore.md` (neu), `people/ian-cutress.md` (neu); Updates `concepts/hardware/gaming-hardwarekrise-ai-speicher.md`, `concepts/hardware/china-local-ai-box.md`. Raw: `raw/youtube/2026-09-20_techtechpotato-silicon-notebook-e2.md` (neu, metadata-only).* *Vorherige Aktualisierung: 2026-09-20 (233. Update — Open-Weights-Anteil im Vercel AI Gateway + Kartell-These. Anlass: X-Link https://x.com/sovereignbrah/status/2101446314743775632 von Pit in OME Topic 13. **Post (fxtwitter, 20.09.2026):** @sovereignbrah (181.958 Follower, verifiziert, Note-Tweet, 19.09.2026 23:00 UTC; 102.062 Views / 2.855 Likes / 558 RT / 39 Quotes) zitiert Guillermo Rauch (@rauchg) vom 19.09.2026 — „today may be a record day for token volume % of open models on Vercel AI Gateway: 🟦 Open 78.4% 🟨 Closed 21.6%", #3/#4 Moonshot AI & DeepSeek, mit Z.ai zusammen vor OpenAI (#2) bei **Spend**, ausdrücklich Fußnote „inference spend across providers, not revenue of the open weight labs". **Bild lokal ausgelesen** (pbs.twimg.com/media/HSjnTgfaQAAHWbk.jpg): Vercel-Chart, Titel „Open vs. Closed Token Volume", Untertitel „…models that publish their weights for download, versus everything else", X-Achse Jun 21 → Sep 18 2026, Legende 78,4 % / 21,6 %. **Primärquelle nachgezogen:** Vercel-Leaderboards (HTTP 200) bestätigen Titel, Untertitel, Zeitraum Jun 22 – Sep 19 und die Datenlizenz **CC BY 4.0** + authentifizierungsfreien JSON-Export; die Chart-Prozente selbst sind **clientseitig gerendert** und per Textabruf **nicht** reproduzierbar. **Eigenbefund (nicht im Post):** Jev führt die Listen bei **Reach 15,6 %** (vor Claude Haiku 4.5 14,6 %, GPT 5.6 Luna 13,0 %) und **Preference 13,3 %** (vor Luna/Sonnet 5 je 6,9 % — fast doppelter Abstand). **Postkritik:** Kausalbehauptung „sie sagen AI killt alle, weil sie den Markt verloren haben" = Deutung; „95 % der Leistung für 1–2 % des Preises" = Zahl ohne Quelle (Aktenlage: 87 % Ersparnis bei 4 % Qualitätsverlust); „regulatory cartel" = unbelegte Zuschreibung (dieselbe Figur wie Chaubard-These 4); „IPOs are now cooked" ohne Bezugsvorgang. ⚠️ Messung = ein Gateway, ein Tag, Token-Volumen (nicht Umsatz), Anbieterquelle. Wiki: `concepts/llm/open-weights-share-vercel-gateway.md` (neu), `people/guillermo-rauch.md` (neu); Updates `concepts/llm/system-one-models-jev.md` (Abschnitt Reichweite/Leaderboards), `institutions/vercel.md` (Abschnitt Leaderboards). Raw: `raw/xpost/2026-09-19_sovereignbrah-antitrust-open-weights.md` (neu), `raw/other/2026-09-20_vercel-ai-gateway-leaderboards.md` (neu).)* @@ -190,7 +191,7 @@ | [Unsloth dSpark](tools/unsloth-dspark.md) | Unsloth-Optimierung dSpark: DeepSeek V4 läuft lokal 2x schneller. Relevanz für lokale Ollama-Infrastruktur (deepseek-v4-flash:0731) + Tier-0-Routing. Status: Titel-Metadaten verifiziert, Details offen (Reddit-403) | other/2026-08-06_unsloth-dspark-deepseek-v4-2x.md | | [Ollama Cloud — DeepSeek-V4-Flash 200+ tps & ZDR](tools/ollama-cloud-deepseek-v4-flash-200tps-zdr.md) | Offizieller Ollama-Post: DeepSeek-V4-Flash-0731 neuer Default auf Ollama Cloud, 200+ tps Output, Zero Data Retention (US/EU). Relevanz: unser Tier-0-Modell. **Update 13.08.:** AICodeKing-Review zu DeepSeek V4 Pro 0813 (“Fully Tested”) + **DeepSeek-API-Preiserhöhung (V4 Pro + V4 Flash) ab 17.08.2026** (Peak/Off-Peak, bis +1.100%) | xpost/2026-08-08_ollama-deepseek-v4-flash-200tps-zdr.md + youtube/2026-08-13_deepseek-v4-pro-0813-aicodeking.md + other/2026-08-13_deepseek-api-preiserhoehung.md | | [Fincept Terminal](tools/fincept-terminal.md) | Open-Source Trading-IDE, von OpenClaw-Blog verlinkt | raw/other/financialbot-topic-history-2026-02-01_2026-05-30.json | -| [Grok Bot (SpaceXAI)](tools/grok-bot-spacexai.md) | Agentischer KI-Bot (Beta): eigener Computer pro Bot, Computer-Use, persistente Workflows, Workflow-Lernen, Multi-Bot-Parallelisierung, Bot-zu-Bot-Kommunikation, Human-in-the-Loop. Miles Deutscher: Plug-and-Play-Vorteil gegenüber OpenClaw/Hermes, $200/mo Mindestpreis, Daten-Souveränität als Tradeoff. 0xCodez: 10-Schritte-Tutorial (Chief, Job-Titel statt Prompt, account-weite Tool-Verbindungen, Session-statt-Secret-Login, Workflow-Demonstration, Schedule/Trigger-Routinen, Spezialisten-Crew, Gruppen-Chat, Reversibilitäts-Approvallinie, wöchentliches Pruning). @bot-Team: 12 Pro-Tips (Multi-Account, Chief+Spezialisten, Pinning, Notion-Arbeitsliste, teach-a-task, Always Allow/Auto Review, Selbst-Evaluation). SEO-Playbook-Teaser (bloggersarvesh): $200/mo als Agentur-Ersatz, Lead-Magnet ohne neuen Fakt. AI Edge: Grok Bot + Hermes komplementär (Grok = App-Integration, Hermes = souveräne billige Always-on-Schicht), Verbindung via Buzz (Jack Dorsey) oder Webhook/Task-File. Miles Deutscher (20.08.): bestätigt denselben Hybrid-Stack mit Buzz als Verbindungsschicht + GTM-vs-Background-Task-Split. Prajwal Tomar: Praxistest an 5 Unternehmen, Hermes-Nutzer, "24/7-Team ohne Gehalt"-Narrativ. | xpost/2026-08-11_xfreeze-spacexai-grok-bot.md + xpost/2026-08-18_milesdeutscher-grok-bot-deep-dive.md + xpost/2026-08-18_0xcodez-grok-bot-10-step-tutorial.md + xpost/2026-08-18_benln-grok-bot-pro-tips.md + xpost/2026-08-18_bloggersarvesh-grok-bot-seo-playbook.md + xpost/2026-08-18_aiedge-grok-bot-hermes-combo.md + xpost/2026-08-18_prajwaltomar-grok-bot-5-businesses-test.md + xpost/2026-08-20_milesdeutscher-grok-bot-hermes-stacking.md | +| [Grok Bot (SpaceXAI)](tools/grok-bot-spacexai.md) | Persistenter Computer-Use-Agent mit spezialisierten Bot-Teams, Cloud-Computern, Routinen, Bot-zu-Bot-Handoffs und Approval-Linien. **Update 22.09.:** Grok 4.7 ist offiziell nativ auf das Grok-Bot-Harness trainiert. Alex Finn zeigt laut YouTube-AI/Herman ein Rollenmodell aus PM+Developer+Designer, Exec-Team sowie Linear/Notion→Cursor-Cloud-Agent-Handoffs; ⚠️ kein Transcript, Default-/Credit-Claims unbestätigt. Zugang via bezahlte Cursor-Pläne/Teams oder verknüpfte SuperGrok-/X-Premium+-Grants; Trial-/On-demand-Regeln dokumentiert. | blog/2026-09-21_xai-grok-47-official.md + youtube/2026-09-22_alex-finn-grok-47-grok-bot.md + bestehende Launch-/Praxisquellen | | [vsurf — Browser als geteilter Agenten-Raum](tools/vsurf.md) | Open-Source RLM-Agent (@warmshao, MIT, npm i -g @warmshao/vsurf) auf prime-agent/pi-agent-Basis. EIN echter Browser, VIELE Agenten: eigene Tabs, paralleles Browsen, Tab-Übernahme vom User. Echtes DOM/CDP statt Screenshot-Raten. Noch Beta (2 Stars, 30 Commits in 8h) | github/2026-08-15_warmshao-vsurf.md | | [God's Eye View](tools/gods-eye-view.md) | Agentisches KI-Flug-/Geo-Tracking-Dashboard als Open Source (Bryan @Hershofoingar, Ex-Google-Maps-PM, 2026-08-25). Dunkles HUD-Interface über 3D-Karte, Flugtracking-Tags (z. B. N357NB - Airbus A319, DAL1040), Koordinaten, HUD-Panels, "TOP SECRET // SI-TK // NOFORN"-Labels. YouTube-Demo: youtu.be/GRJaKcXZS94. Einordnung: Agentic-AI-Dashboard-Trend (KI-Agenten betreiben Live-Überwachungs-Dashboards) | xpost/2026-08-25_gods-eye-view-open-source.md | | [Perplexity Portable Computer](tools/perplexity-portable-computer.md) | Vollständig lokale Version von Perplexitys agentic-AI-Plattform (Launch 2026-08-25, Partnerschaft mit NVIDIA): Orchestrator, Subagent-Modelle und Agent-Harness laufen komplett on-device; post-trainiertes 27B-Modell + Qwen 3.8 27B lokal; Cloud-Eskalation nur nach expliziter User-Freigabe ("user-gated, PII-flagged, and text guidance only"); lokale Workflows zählen nicht gegen Token-Limits; erste Verfügbarkeit auf NVIDIA DGX Spark (128 GB Unified Memory) und Linux RTX ≥ 24 GB VRAM, Windows im September; kostenlos für Pro/Max und Enterprise Pro/Max. ⚠️ Benchmark 82,6 % = Herstellerangabe. Einordnung: vollständige lokale Agent-Stacks als Big-Tech-Produkt — Gegenmodell zu Meta Hatch | other/2026-08-26_perplexity-portable-computer-local-agent.md | @@ -261,7 +262,7 @@ | [Post-Transformer LLM Architectures](concepts/llm/post-transformer-llm-architectures.md) | DeepMinds Vier-Säulen-Strategie: Griffin/Recurrent-Gemma/Titans (hybride Attention+Recurrence), Diffusions-LLMs, JEPA vs. Generativ (Weltmodelle), Foundation-Prior vs. Continual-Learning | youtube/2026-06-16_deepmind-two-steps-ahead.md | | [Dezentrale KI als Katalysator der Gegenmacht](concepts/llm/decentralized-ai-counterpower.md) | Club 77.7 Afterhour: Lokale unzensierte LLMs als Gamechanger — schnelle Analyse systemischer Täuschungen ohne Bias-Filter. Doppel-Katalysator-These (KI + entschlossene Minderheit). **Update 25.08.:** Cross-Ref auf KI-Souveränität | other/2026-06-19_ome21-afterhour-makro-systematik.md | | [KI-Souveränität: Dezentralisierung statt Konzern-Kontrolle](concepts/llm/ki-souveraenitaet-dezentralisierung.md) | Adams/Kesterson (BitChute, Aug 2026): Korporatismus vs. Souveränität als Zwei-Wege-Typologie; KI als Bibliothekar statt Orakel — Verantwortung bleibt beim Menschen; lokale Ausführung auf eigener Hardware schützt Privatsphäre, verhindert externe Zensur, bricht Konzern-/Staats-Machtmonopol; „KI so fundamental wie Elektrizität". ⚠️ kein Transcript, nur Pits Zusammenfassung | other/2026-08-25_pit-souveraene-ki-bitchute.md | -| [LLM Model Catalog](concepts/llm/llm-model-catalog.md) | Konsolidierte Modell-Übersicht aller im Wiki erwähnten LLMs. Fokus auf lokale Deployment-Optionen. Frontier-Tabelle (Cloud), Open-Source-Tabelle (HF-Links, Hosting, Praxis-Tests, Tester-Attribution), Hector's Active Stack, Post-Transformer-Outlook, DRACO-Fusion-Ergebnisse | 11 Wiki-Quellen + MEMORY.md | +| [LLM Model Catalog](concepts/llm/llm-model-catalog.md) | Konsolidierte Modellübersicht mit lokalem Deployment-Fokus. **Update 22.09.: Grok 4.7** — 500k API-Kontext, vier Reasoning-Efforts, $2/$6 Basis-I/O (<200k Prompt), nativ auf Grok Bot trainiert; Herstellerbenchmarks klar als nicht unabhängig reproduziert markiert. | 12 Wiki-/Raw-Quellen + MEMORY.md | | [LLM Sycophancy, Confabulation & Session-Statelessness — The Nano Banana Incident](concepts/llm/llm-sycophancy-confabulation.md) | Drei Mechanismen die LLM-Aussagen über eigene Fähigkeiten untrustworthy machen: Session-Statelessness, Sycophantic Compliance, Confabulation. Lüge-vs-Confabulation-vs-Sycophancy-Distinktion. Anti-Pattern: Modul-Aktivierung per Chat. Fallbeispiel: Gemini erfindet "Nano Banana 2". Goldene Regel: Provider-Doku > Chat | other/2026-06-26_nanobana-incident-sycophancy-confabulation.md | | [Matisyahu/Gemini-Incident: Policy-Envelope vs. Datenzugriff](concepts/llm/matisyahu-gemini-incident.md) | Gemini verweigert Lyrics-Zitat (Copyright-Policy), obwohl es den Song versteht. Konzept des Policy-Envelope: Leistungslimit eines Modells wird durch externe Policy-Schicht bestimmt, nicht durch Verständnis. Wissen ohne Sagen-Dürfen = praktisch Unwissen. Agency (Tool-Zugriff auf YouTube-Subs) schlägt Intelligenz. ASR-Fehler ("murder" statt "I'm not crazy") als Hinweis, nicht Wahrheit. **Update 19.08.:** Gemini-Stellungnahme ergänzt — Haftungsasymmetrie, Silo-Paradoxon (DMCA/Safe Harbor vs. Veröffentlicher-Status), UX-Problem ("umgekehrter Zombie"), Agency/Policy-Envelope | other/2026-08-19_matisyahu-gemini-incident.md + other/2026-08-19_gemini-stellungnahme-matisyahu-incident.md | | [ReDS — Resilient Decentralized Swarm](concepts/llm/reds-resilient-decentralized-swarm.md) | Brian Roemmele's Konzept für dezentrale KI-Entwicklung als Gegenmodell zu Amodei's Zentralismus. Resilient (zensur-resistent) + Decentralized (keine zentrale Kontrolle) + Swarm (koordinierte Akteure). Erweitert Cloud-Exit/Dezentrale-KI-These um politische/soziale Dimension | xpost/2026-06-28_roemmele-reds-decentralized-ai-vs-amodei.md | @@ -473,6 +474,7 @@ |-------|-------------|---------| | [AgentStack Daily](institutions/agentstack-daily.md) | KI-generierter englischer Daily-Podcast (zwei TTS-Hosts NOVA/ALLOY, NotebookLM-artig): tägliche Agent-Stack-Release-Readouts. Segmente: Release Readout (Codex), Model Discovery Check (OpenRouter, verified same-cycle), GitHub Project Radar, Local LLM Spotlight, Release Coverage Check (OpenClaw/Hermes/Codex/Claude Code/Antigravity), Extra Research Candidates. Distribution via GitHub Releases (clawdassistant85-netizen/openclaw-podcast-media-en + openclaw-podcast-audio, op3.dev-Proxy), Show Notes auf tobyonfitnesstech.com. Sekundärer Aggregator — Wiki verlinkt immer die Primärquellen | podcast/2026-08-26_agentstack-daily-ep106.md | | [OpenClaw Cast](institutions/openclaw-cast.md) | KI-generierter englischer Weekly-Podcast von "AI World" (zwei fiktive TTS-Hosts Cleo/Dev, anchor.fm-RSS): übersetzt je eine Community-/Release-Warnung in ein konkretes lokales OpenClaw-Guardrail-Rezept („Skill of the Week"). 23 Episoden im Feed (Feb–Aug 2026): Runaway-Budget-Kill-Switch (25.08.), Multi-Agent-Turf-Wars→Task Lease Guard, Release-Sentinel-Canary, Cron-Cadence-Guard, Boot-Token-Cut −43 %, MCP-Unlock „800 Tools". Community-nahe Sekundärquelle — Einzelfakten nur nach Primärquellen-Check übernehmen. Die Runaway-Folge (25.08.) ist seit dem 151. Update vollständig inhaltlich erfasst → [Agent Budget Breaker](concepts/agents/agent-budget-breaker.md) | podcast/2026-08-26_openclaw-cast-runaway-agent-budget.md, podcast/2026-08-26_openclaw-cast-budget-breaker-transcript.md | +| [Grok AI Daily](institutions/grok-ai-daily.md) | Englischer ART19-AI-News-Podcast. Serie verspricht Grok/xAI-Coverage, tatsächliche Titel sind breiter. Episode 21.09.2026 (29:30) zu Meta Muse, Cyber-/Safety-Claims, US–China und Newsom-„Kill Switch“ lokal per Maschinen-ASR transkribiert; kein Publisher-Transcript, keine Quellenliste, Eigenwerbung. **Kein Grok 4.7; Grok Bot nur beiläufig.** Rezeptionsquelle, kein Primärbeleg. | podcast/2026-09-21_grok-ai-daily-amazon-muse-ai-kill-switch.md | | [Plaier](institutions/plaier.md) | KI-Unternehmen im internationalen Fußball (Spieler-/Kaderanalyse, CEO Jan Wendt). Zwei Scores (N18 Nominal, Effective). These: Kaderqualität = 90% der sportlichen Leistung. WM-2026-Analyse Deutschland: Nagelsmann ließ „viel Qualität zu Hause", Kimmich/Wirtz falsch positioniert. Referenzkunde VfL Osnabrück | blog/2026-08-16_plaier-ki-wm-analyse-deutschland.md | | [Higgsfield AI](institutions/higgsfield-ai.md) | KI-Unternehmen für Text-zu-Video / KI-Video-Generierung. YouTube-Kanal @HiggsfieldAI, "Higgsfield Origins"-Reihe mit komplett KI-generierten Kurzfilmen. "Oneiric" (2026) — komplett KI-generierter Sci-Fi-Kurzfilm als Beleg für Production-Shift (Hollywood-Disruption) | youtube/2026-08-19_oneiric-higgsfield-ai-sci-fi-short.md | | [InCountry / AgentCloak](institutions/incountry.md) | US-Unternehmen für Datenschutz-Infrastruktur, zwei Linien: **InCountry Sovereign Data** (Datenresidenz, Cross-Border-Compliance, Proxies für Web-Services/CLI/E-Mail/SMS, Salesforce-Lokalisierung) und **AgentCloak** (Cloaking sensibler Daten vor Chatbots/Agenten/MCP/CRM). Gründer & CEO Peter Yared. AgentCloak Desktop kostenlos, AgentCloak Server = AI-native Rust-Container, cloud/on-prem/air-gapped, Datei-Cloaking, MCP-Schutz für Salesforce/HubSpot, Intune-Rollout, Entra/Google Workspace/Okta, Audit-Logs. ⚠️ „CB100 2026", „$50 Mio. raised" und „Global 2000 customers" = Selbstauskunft, Investoren/Kunden unbemannt | blog/2026-09-18_agentcloak-desktop-launch.md + other/2026-09-19_rampart-model-card.md | @@ -489,7 +491,7 @@ | [Google Cloud](institutions/google-cloud.md) | Cloud-Infrastruktur- und Services-Plattform (Gemini Enterprise, Open Knowledge Format v0.1) | blog/2026-06-16_okf-google-cloud-open-knowledge-format.md | | [OpenRouter](institutions/openrouter.md) | KI-Modell-Aggregator & API-Plattform (Modell-Panel-Fusion, DRACO-Benchmark) | blog/2026-06-12_openrouter-fusion-beats-frontier.md | | [Hugging Face](institutions/hugging-face.md) | Open-Source KI-Plattform & Community (Modell-Hosting, Cloud-Exit-Basis) | Allgemeines Wiki-Inventar | -| [xAI](institutions/xai.md) | AI Frontier-Forschungs- und Produkt-Unternehmen von Elon Musk (Grok, massiver GPU-Compute Colossus) | other/2026-06-21_elon-musk-aimode-deflation-universal-high-income.md | +| [xAI / SpaceXAI](institutions/xai.md) | AI-Frontier-Unternehmen hinter Grok, Grok Build und Grok Bot. **Grok 4.7 (21.09.):** größerer Basismodell-Kern, längeres RL auf mehrstündigen Aufgaben, Selbstprüfung, 500k Kontext, native Harness-Kenntnis; API-/Cursor-Spezifikation und Herstellerbenchmarks dokumentiert. | other/2026-06-21_elon-musk-aimode-deflation-universal-high-income.md + blog/2026-09-21_xai-grok-47-official.md + xpost/2026-09-21_cb-doge-grok-47-release.md | | [TU München](institutions/tu-muenchen.md) | Technische Universität München (neuromorphe Hardware, Quantencomputing, Prof. Klaus Mainzer) | youtube/2026-06-16_everlast-mainzer-neuromorphe-chips-quantencomputer.md | | [MIT](institutions/mit.md) | Massachusetts Institute of Technology (CSAIL, Grundlagenforschung, autonome Robotik) | Akademisches Wiki-Inventar | | [Stanford AI Lab](institutions/stanford-ai-lab.md) | Stanford Artificial Intelligence Laboratory (Foundation Models, HELM-Benchmark) | Akademisches Wiki-Inventar | @@ -559,6 +561,10 @@ | Datei | Typ | Titel | |-------|-----|-------| +| `raw/blog/2026-09-21_xai-grok-47-official.md` | blog | Offizielle Grok-4.7-Akte: xAI-Release + API-Modellkarte + Cursor-Doku; 500k Kontext, vier Effort-Stufen, $2/$6 Basis-I/O, Long-Context-Tarif; Herstellerbenchmarks und Safeguard-Claims markiert | +| `raw/xpost/2026-09-21_cb-doge-grok-47-release.md` | xpost | DogeDesigner/@cb_doge Release-Zusammenfassung (113.071 Views beim Abruf); Werte gegen xAI geprüft, pauschales „2× fast/half price“ als Account-Verdichtung abgegrenzt | +| `raw/youtube/2026-09-22_alex-finn-grok-47-grok-bot.md` | youtube | Alex Finn „Grok 4.7 inside Grok Bot is INCREDIBLE“ (14:06, geteilt von Pit); Metadaten/Kapitel, kein Transcript; YouTube-AI- und Herman-Zusammenfassungen klar attribuiert | +| `raw/podcast/2026-09-21_grok-ai-daily-amazon-muse-ai-kill-switch.md` | podcast | Grok AI Daily, 29:30, geteilt von Netbits; ART19-Metadaten + vollständiges lokales Maschinen-ASR. Allgemeiner AI-Newsrecap; kein Grok 4.7, Grok Bot nur beiläufig; keine Publisher-Quellenliste/kein Publisher-Transcript | | `raw/other/2026-09-20_ddr4-gebrauchtmarkt-dimmsum.md` | other | **Gebrauchtmarkt DDR4 — Snapshot 20.09.2026** ([DIMMsum](https://dimmsum.com/ddr4/64gb-2666-rdimm), Tracker laufender eBay-Listings; abgerufen 20.09.2026). **64 GB DDR4 RDIMM (2666 MHz, PC4-21300)** gebraucht **~$325–550/Modul inkl. Versand ≈ $5–8,50/GB**; günstigste Listings ab **~$326**, **Median des Angebotspreises ~$545**; Lots (4 × 64 GB) bis **~$5/GB**; 3200 MHz bzw. Marken-/Neuzustand häufig **$400–600+**; Refurbished/New-Old-Stock **$295–550**. Preistreiber: Geschwindigkeit (2400/2666/2933/3200), Rank, Marke (Samsung/Micron/SK Hynix), Zustand, Verkäufer. **Zweitquelle** [ITLD Server-RAM/SSD-Marktupdate 2026](https://itldc.com/en/blog/server-ram-ssd-market-update-2026): Kontraktpreise **+93–98 % QoQ (Q1 2026)**, **+58–63 % (Q2)**; 64-GB-DDR4-RDIMM im August auf dem **2,5–3,5-Fachen** des Januar-Niveaus; Gebrauchtpreise ebenfalls steigend (Consumer-Kits +4–30 % MoM), aber relativ günstig; DDR4 punktuell mit **Aufpreis pro Bit gegenüber DDR5**; neue Kapazität knapp bis 2027. ⚠️ **Angebotspreise, keine Transaktionspreise**; keine öffentlichen Volumendaten; Speed/Rank/Marke/Zustand verschieben den Wert stark. Wiki: `concepts/hardware/gaming-hardwarekrise-ai-speicher.md` (Abschnitt „Der Zweitmarkt schließt die Rechnung“ — ~$5–8,50/GB gegen ~$29/GB neu = 3–5×), `institutions/astera-labs.md` (TCO), `concepts/hardware/ai-infrastruktur-custom-silicon.md` (Preisachse) | **DRAM-Spotpreise-Snapshot 20.09.2026** (MemoryIndex, direkt abgerufen; Spot-Basis **DRAMeXchange / TrendForce, gemeldet 17.09.2026**). USD/Die: DDR5 16Gb **$56.11** (+6,38 % 30T / **+480 % YoY**), DDR5 eTT **$24.55** (+0,36 % / +330 %), DDR4 16Gb **$84.43** (+4,85 % / **+690 %**), DDR4 8Gb **$45.59** (+6,41 % / **+860 %**); DDR5 RDIMM 64GB **$1.857** (+4,47 % / +455 %, **derived** = 32 × Die-Spot +~6 % Modulaufschlag, kein öffentlicher Modulausdruck); Baskets DDR5 +14,23 % / +331,67 %, DDR4 +4,49 % / +563,00 %. 17.09.-Board-Notiz: DDR5 Ø $55,233 (+0,73 %), DDR4 16Gb −1,85 % auf $86,00 („early cooling signal“), DDR4 8Gb +0,78 % auf $46,143. ⚠️ Spot-/derived-Niveaus, keine Vertragspreise; Momentaufnahme; Intraday-Reihen laut Methodik rückwärts gerechnet. **Zusatzquelle im selben Raw:** [ServeTheHome zu Astera Labs Leo 2 / Leo X](https://www.servethehome.com/astera-labs-releases-leo-2-cxl-memory-controllers-and-leo-x-controller-for-rackscale-fabric-attached-memory/) (PCIe Gen6/CXL 3.2, 768 GB DDR4 bzw. 4 TB DDR5, 3DPC/12 DIMMs, Leo X fabric-attached, RAS, **keine Preise**, Vorgänger Marvell Structera). Wiki: `institutions/astera-labs.md` (neu), `concepts/hardware/gaming-hardwarekrise-ai-speicher.md` (Preis-Anker), `concepts/hardware/ai-infrastruktur-custom-silicon.md` (Nachtrag) | | `raw/youtube/2026-09-20_techtechpotato-silicon-notebook-e2.md` | youtube | TechTechPotato (Host **Dr. Ian Cutress**, More Than Moore): „Using Your Older DDR4 // The Silicon Notebook E2" (https://www.youtube.com/watch?v=31_eeq3ciRA, Video-ID `31_eeq3ciRA`, **52.330 Aufrufe** beim Abruf 20.09.2026; geteilt von Netbits in OME „Theorie-Bildung", Zeitsprung `t=10m`). **Kein Transkript** (keine `captionTracks`), **keine Dauer** im Seiten-JSON; oEmbed bestätigt Titel/Autor/Autoren-URL. AI Infrastructure Summit Tag 2, Leitmotiv „everyone building their own silicon": Qualcomm AlphaWave · Broadcom Tomahawk 6 (17.000 Pins) · Giga.io portables Sechs-Slot-Rack + „wer kauft Edge-Racks" · d-Matrix Corsair/Pavehawk + NVLink · Lightning stacked DRAM · Mango Boost/Credo · **Astera Labs Leo 2 (DDR4+DDR5 in einem System auf Next-Gen-Intel-CPU)** · SiFive/RISC-V · Ayar Labs Co-Packaged Optics; 12 Kapitelmarken 00:00–13:20. **Disclosure:** More Than Moore erbringt/erbrachte bezahlte Forschung/Beratung für AMD, Arm, Ayar Labs, Broadcom-nahe Firmen, IBM, Intel, MediaTek, NVIDIA, Qualcomm, SiFive, TSMC, Tenstorrent u. a. Wiki: `concepts/hardware/ai-infrastruktur-custom-silicon.md` (neu), `institutions/more-than-moore.md` (neu), `people/ian-cutress.md` (neu) | | `raw/xpost/2026-09-19_sovereignbrah-antitrust-open-weights.md` | xpost | SOVEREIGN BRAH (@sovereignbrah, 181.958 Follower): „This is why Anthropic and OpenAI are panicking" (19.09.2026, 23:00 UTC, Note-Tweet; 102.062 Views / 2.855 Likes). Quote-Post von Guillermo Rauch (@rauchg): **Open 78,4 % / Closed 21,6 %** Token-Volumen auf dem Vercel AI Gateway (Jun 21–Sep 18), samt Spend-Hinweis zu Moonshot/DeepSeek/Z.ai vs. OpenAI. Vier Thesen des Posts (Panik-Kausalität, „95 % Leistung für 1–2 % Preis", „regulatory cartel", „IPOs are now cooked") — Zuschreibungen ohne Beleg. Angehängtes Vercel-Chart lokal ausgelesen | diff --git a/wiki/institutions/grok-ai-daily.md b/wiki/institutions/grok-ai-daily.md new file mode 100644 index 0000000..f200756 --- /dev/null +++ b/wiki/institutions/grok-ai-daily.md @@ -0,0 +1,50 @@ +--- +created: 2026-09-22 +updated: 2026-09-22 +sources: [podcast/2026-09-21_grok-ai-daily-amazon-muse-ai-kill-switch.md] +tags: [institution, podcast, ai-news, grok, xai, secondary-source] +--- + +# Grok AI Daily + +## Profil + +| Feld | Wert | +|---|---| +| Typ | englischsprachiger AI-News-Podcast | +| Feed-Inhaber | „Grok AI Daily“ laut ART19/RSS | +| Hosting | [ART19](https://art19.com/) | +| Serienseite | https://art19.com/shows/ai-unchained | +| RSS | https://rss.art19.com/ai-unchained | +| Taktung | im Feed etwa zwei bis drei Episoden pro Woche; 465 Episoden beim Abruf 22.09.2026 | + +Die Serienbeschreibung verspricht Nachrichten zu Grok, xAI und der allgemeinen KI-Branche. Die tatsächlich sichtbaren Episodentitel decken einen breiteren AI-News-Mix ab und sind häufig Anthropic-/OpenAI-lastig. **Der Serienname ist daher kein Beleg dafür, dass eine einzelne Folge substanzielle Grok-Berichterstattung enthält.** + +## Episode 21.09.2026: Muse, AI-Rebrand und „Kill Switch“ + +Die 29:30-Minuten-Folge [„Amazon blocks Meta's Muse Agent, Trump Rebands AI, Newsom Create AI Kill Switch“](https://art19.com/shows/ai-unchained/episodes/eba2a4fb-0c3e-4d84-a61b-dbe70e7cc7c9) wurde von Netbits ⚡️ Stachelbanane geteilt und lokal transkribiert. Sie behandelt: + +- Amazon gegen Metas Muse-Shopping-Agent; Shopify als agentic-commerce-Gegenmodell +- einen sekundär wiedergegebenen Gemini-Cybersecurity-Bericht und den Vergleich zum [[../concepts/openai-huggingface-incident.md|OpenAI–Hugging-Face-Vorfall]] +- Andrew Yangs viralen Selbstreplikations-Claim und ein `@Grok`-Guardrail-Beispiel +- Apple/Siri-Vergleichsclaims +- US–China-Gespräche über Meldungen von KI-Zwischenfällen +- Trump-„AI Force“/Rebranding +- Newsoms Auftrag für kalifornische „AI kill switch“-Regeln, Vor-Ort-Audits und Verlustkontrollmeldungen + +**Grok-Bezug:** Grok Bot wird nur im sponsor-nahen Überblick über Desktop-Agenten erwähnt; der Host sagt selbst, er habe ihn kaum getestet. Grok 4.7 kommt gar nicht vor. Für die Grok-Modellakte liefert die Folge somit **keinen** Sachbeitrag. + +## Quellenqualität + +- Metadaten, Datum und Dauer sind durch ART19-JSON/RSS belegt. +- ART19 veröffentlicht kein Transcript; der Raw-Eintrag enthält lokales Maschinen-ASR ohne Sprecher-Diarisierung. +- Die Episode ist ein Meinungs-/Newsrecap mit Eigenwerbung für AI Box, ohne Quellenliste und ohne Zeitmarken. +- Einzelne Nachrichtenclaims werden sekundär wiedergegeben und sind nicht allein auf Basis der Episode als Fakten zu übernehmen. +- Die belastbare Nutzung im Wiki ist daher: **Rezeptionsquelle und Themenradar, nicht Primärbeleg.** + +## Cross-References + +- [[xai.md]] — xAI/SpaceXAI und Grok-Modelle +- [[../tools/grok-bot-spacexai.md]] — Grok Bot +- [[../concepts/policy/ai-regulation-2026.md]] — Regulierung und Government-Safety-Frameworks +- [[../concepts/openai-huggingface-incident.md]] — Cybersecurity-Vorfall, den die Folge als Vergleich aufgreift diff --git a/wiki/institutions/xai.md b/wiki/institutions/xai.md index 366ff62..b00636b 100644 --- a/wiki/institutions/xai.md +++ b/wiki/institutions/xai.md @@ -1,29 +1,54 @@ --- created: 2026-06-24 -updated: 2026-06-24 -sources: [`raw/other/2026-06-21_elon-musk-aimode-deflation-universal-high-income.md`] -tags: [institution, xai] +updated: 2026-09-22 +sources: [other/2026-06-21_elon-musk-aimode-deflation-universal-high-income.md, blog/2026-09-21_xai-grok-47-official.md, xpost/2026-09-21_cb-doge-grok-47-release.md] +tags: [institution, xai, spacexai, grok, grok-4-7, frontier-models] --- -# xAI +# xAI / SpaceXAI ## Profil | Feld | Wert | -|------|------| -| Typ | AI Frontier-Forschungs- und Produkt-Unternehmen | -| Fokus | Entwicklung von Grok-Modellen, Echtzeit-Informationsintegration (X-Plattform) und massiver GPU-Compute-Infrastruktur. | -| Wichtigste Personen | [[../people/elon-musk.md|Elon Musk]] | +|---|---| +| Typ | AI-Frontier-Forschungs- und Produkt-Unternehmen | +| Fokus | Grok-Modelle, X-Echtzeitintegration, agentische Produkte und massive GPU-Compute-Infrastruktur | +| Wichtigste Person | [[../people/elon-musk.md|Elon Musk]] | | Homepage | https://x.ai/ | -| Zugehörige Tools/Modelle | Grok 2.5, Grok 3.0 Alpha | -| Primäre Quellen | `raw/other/2026-06-21_elon-musk-aimode-deflation-universal-high-income.md` | +| Produkte im Wiki | Grok-Modellfamilie, [[../tools/grok-bot-spacexai.md|Grok Bot]], Grok Build | -## Fokus & Aktivitäten +xAI entwickelt die Grok-Modellfamilie und betreibt mit Colossus eine große Compute-Infrastruktur. Die Produktlinie reicht inzwischen vom Modellzugang über API und Grok Build bis zum persistenten Computer-Use-Agenten Grok Bot. Offizielle Seiten verwenden 2026 auch die Bezeichnung **SpaceXAI**. -xAI wurde gegründet, um die Natur des Universums zu verstehen und eine Gegenkraft zu zensierten KI-Modellen zu etablieren. Durch die enge Anbindung an die Echtzeitdaten der X-Plattform und die massive 'Colossus'-Compute-Infrastruktur im Rücken stellt xAI eine dominante Kraft im Bereich des schnellen Ingests und der geopolitischen KI-Positionierung dar. +## Grok 4.7 (21.09.2026) + +[Grok 4.7](https://x.ai/news/grok-4-7) ist laut offizieller Ankündigung das bislang leistungsfähigste xAI-Modell für Coding und Wissensarbeit. Gegenüber Grok 4.6 nennt xAI einen größeren Basismodell-Kern, längeres Reinforcement Learning auf härteren mehrstündigen Aufgaben, stärkere Selbstprüfung, besseres Langkontext-Management und **natives Verständnis des Grok-Bot-Harnesses**. + +### Technische Eckdaten + +| Feld | Wert | +|---|---| +| API-Modell-ID | `grok-4.7` | +| Kontext | 500k Token direkt per API; in Cursor 256k Standard / 500k Long Context | +| Eingabe/Ausgabe | Text+Bild → Text | +| Reasoning | `low`, `medium`, `high` (Standard), `xhigh` | +| Preis direkt per API | <200k Prompt: $2/M Input, $0,50/M Cached, $6/M Output; ab 200k: $4/$1/$12 | +| Verfügbarkeit | Cursor, Grok Build, Grok API, Dritt-Harnesses, Router und Clouds | + +### Anbieter-Benchmarkbild + +xAI meldet gegenüber Grok 4.6 unter anderem CursorBench 4.0 **46,3 % vs. 40,4 %**, DeepSWE v1.1 **71,0 % vs. 65,2 %** und Terminal-Bench 4.0 **38,0 % vs. 20,3 %**. Dazu kommen Legal-, Health-, Electrical-Engineering-, Dokument-/Präsentations- und Safety-Claims. **Diese Werte sind Hersteller-Selbstauskunft und im Wiki nicht unabhängig reproduziert.** + +Der große Produktpunkt ist weniger ein einzelner Score als die **Ko-Evolution von Modell und Harness**: Grok 4.7 wurde ausdrücklich darauf trainiert, Grok Bot nativ zu verstehen. Das verschiebt die Einheit der Optimierung vom nackten Chatmodell zum Modell-im-Agentensystem. + +## Rezeption + +[DogeDesigner / @cb_doge](https://x.com/cb_doge/status/2102063549191508399) fasste die Release-Seite am 21.09.2026 reichweitenstark zusammen. Die meisten Einzelwerte decken sich mit der Primärquelle; der pauschale Satz „twice as fast and half the price of comparable models“ ist jedoch seine Verdichtung und keine universelle Garantie der offiziellen Modellkarte. + +Alex Finns Video [[../../raw/youtube/2026-09-22_alex-finn-grok-47-grok-bot.md|„Grok 4.7 inside Grok Bot is INCREDIBLE“]] zeigt die praktische Rezeptionsseite. Mangels abrufbarem Transcript sind seine Workflow-Details nur über YouTubes AI-Zusammenfassung und Hermans ausdrücklich zugeschriebene Auswertung dokumentiert. ## Cross-References -- [[../concepts/agi/universal-high-income.md]] -- [[../concepts/policy/ai-as-geopolitical-weapon.md]] -- [[../institutions/openai.md]] (Frontier-Modell-Wettbewerber) +- [[../concepts/llm/llm-model-catalog.md]] — konsolidierte Modellübersicht +- [[../tools/grok-bot-spacexai.md]] — Grok-Bot-Harness, Zugang, Preise und Praxis-Setups +- [[../concepts/policy/ai-as-geopolitical-weapon.md]] — geopolitischer KI-Kontext +- [[grok-ai-daily.md]] — Podcast-Aggregator; trotz Name nur bedingt grok-spezifisch diff --git a/wiki/log.md b/wiki/log.md index be5358b..08d8ac5 100644 --- a/wiki/log.md +++ b/wiki/log.md @@ -2,6 +2,24 @@ *Append-only changelog. Start: 2026-06-05* +## 2026-09-22 — Grok 4.7, Grok-Bot-Praxis und „Grok AI Daily“-Podcast + +**Type:** ingest + wiki-update | **Scope:** raw/blog, raw/xpost, raw/youtube, raw/podcast, wiki/concepts/llm, wiki/concepts/policy, wiki/tools, wiki/institutions, wiki/index, wiki/log +**Anlass:** Offizielles Grok-4.7-Release + CB_Doge-Post; YouTube-Link von Pit Weber zu Alex Finns „Grok 4.7 inside Grok Bot is INCREDIBLE“ mit praktischer Zusammenfassung von Herman; ART19-MP3 von Netbits ⚡️ Stachelbanane. +**Verifikation:** xAI-Release, xAI-API-Modellkarte und Cursor-Modellseite direkt gelesen. CB_Doges Note-Tweet per FxTwitter vollständig abgerufen und gegen die Primärquelle gehalten. Alex-Finn-Metadaten über Watch-Page/oEmbed verifiziert; kein Transcript trotz Caption-/Timedtext-/yt-dlp-/Drittweg-Versuchen. ART19-Episode über Primär-JSON und RSS identifiziert; Audio lokal transkribiert (Maschinen-ASR, kein Publisher-Transcript). +**Kernbefunde:** +1. **Grok 4.7:** größerer Basismodell-Kern, längerer RL-Lauf auf schwierigeren mehrstündigen Aufgaben, stärkere Selbstprüfung, besseres Langkontext-Management und natives Verständnis des Grok-Bot-Harnesses. API: `grok-4.7`, Text+Bild→Text, 500k Kontext, `low|medium|high|xhigh`, kein Batch. Preis <200k Prompt $2/M Input, $0,50/M Cached, $6/M Output; ab 200k $4/$1/$12. Cursor: 256k Standard / 500k Long Context und eigener Pool-/Fast-Tarif. +2. **Benchmarks:** xAI meldet CursorBench 46,3 %, DeepSWE 71,0 %, Terminal-Bench 38,0 %, weitere Knowledge-/Safety-Werte. Alle als **Herstellerangaben** geführt; keine unabhängige Reproduktion. CB_Doges „twice as fast and half the price“ ist seine Verdichtung, nicht die universelle Zusage der Modellkarte. +3. **Alex Finn:** belastbare Metadaten/Kapitel, aber kein Transcript. Workflow nur attribuiert aus YouTubes AI-Zusammenfassung + Hermans Auswertung: PM+Developer+Designer je Bereich, PM-only Exec-Team, Linear/Notion-Issues, Cursor Cloud Agents. Default-Modell-/Credit-Aussage bleibt unbestätigt. +4. **ART19-Podcast:** „Grok AI Daily“, Episode 21.09.2026, 29:30. Themen: Meta Muse/Amazon/Shopify, sekundäre Cyber-/Safety-Claims, US–China, Trump-„AI Force“, Newsom-„Kill Switch“. **Kein Grok 4.7**; Grok Bot einmal sponsor-nah und ausdrücklich kaum getestet. Der Serienname überzeichnet den Grok-Anteil dieser Folge. Maschinen-ASR im Raw, Eigennamen-/Zahlenrisiko markiert. +**Dateien:** +- raw (NEW): `raw/blog/2026-09-21_xai-grok-47-official.md`, `raw/xpost/2026-09-21_cb-doge-grok-47-release.md`, `raw/youtube/2026-09-22_alex-finn-grok-47-grok-bot.md`, `raw/podcast/2026-09-21_grok-ai-daily-amazon-muse-ai-kill-switch.md` +- wiki (NEW): `wiki/institutions/grok-ai-daily.md` +- wiki (UPDATED): `wiki/institutions/xai.md`, `wiki/tools/grok-bot-spacexai.md`, `wiki/concepts/llm/llm-model-catalog.md`, `wiki/concepts/policy/ai-regulation-2026.md`, `wiki/index.md` +**⚠️:** xAI-Benchmarks = Anbieter-Selbstauskunft. Alex-Finn-Workflow = Sekundärzusammenfassungen ohne Wortlaut-Transcript. Podcast = News-/Meinungsquelle mit Eigenwerbung und ohne Quellenliste; Nachrichtenclaims nicht als Primärfakten übernommen. + +--- + ## 2026-09-20 — Gebrauchtmarkt DDR4 (DIMMsum): der Zweitmarkt-Anker schließt die TCO-Rechnung **Type:** ingest + wiki-update | **Scope:** raw/other, wiki/concepts/hardware, wiki/institutions, wiki/index, wiki/log diff --git a/wiki/tools/grok-bot-spacexai.md b/wiki/tools/grok-bot-spacexai.md index db7a57f..7ccae30 100644 --- a/wiki/tools/grok-bot-spacexai.md +++ b/wiki/tools/grok-bot-spacexai.md @@ -1,6 +1,6 @@ # Grok Bot (SpaceXAI) -**Kategorie:** Tools / AI-Agents | **Stand:** 2026-09-19 (Launch + Miles-Deutscher-Deep-Dive + 0xCodez-Tutorial + @bot-Team-Pro-Tips + SEO-Playbook-Teaser + Hermes-Combo + 5-Businesses-Test + Hermes-vs-Grok-Vergleich + Fleet-of-Agents-Walkthrough + $20-Preis-Update, Transcript-verifiziert + Grok-Bot-Galaxy (Livestream + Ergebnis + Musk-Verstärkung) + Cursor-Zugangs-/Trial-Doku & On-demand) +**Kategorie:** Tools / AI-Agents | **Stand:** 2026-09-22 (Launch + Miles-Deutscher-Deep-Dive + 0xCodez-Tutorial + @bot-Team-Pro-Tips + SEO-Playbook-Teaser + Hermes-Combo + 5-Businesses-Test + Hermes-vs-Grok-Vergleich + Fleet-of-Agents-Walkthrough + $20-Preis-Update, Transcript-verifiziert + Grok-Bot-Galaxy (Livestream + Ergebnis + Musk-Verstärkung) + Cursor-Zugangs-/Trial-Doku & On-demand) ## Überblick @@ -203,6 +203,26 @@ Tagesfenster laut Websuche ca. 08:30–18:00 PT (Herman-Briefing nennt Start 08: **⚠️ Ehrliche Kante:** Alle Angaben hier sind Anbieter-Doku (Cursor/SpaceXAI) — sie beschreiben die Mechanik, nicht die Zuverlässigkeit der Bots. Die kursierende Erzählung „nur $20" (Lipsky, 27.08.) ist damit präzisiert: $20/Monat ist der **Cursor-Pro**-Einstieg, den Grok Bot enthält — kein separates $20-Abo; Ultra bleibt die $200-Stufe. Eine im OME geteilte Community-Anleitung (ohne Autorenangabe) deckt sich mit der Doku; ihre Aussage „frischer Account startet automatisch den Test" ist dort nicht wörtlich belegt. +## Grok 4.7 im Grok-Bot-Harness (21.–22.09.2026) + +Mit [Grok 4.7](https://x.ai/news/grok-4-7) veröffentlicht SpaceXAI erstmals ein Modell, dessen Release-Text die **native Kenntnis des Grok-Bot-Harnesses** ausdrücklich als Trainingsziel nennt. Größerer Basismodell-Kern, längerer RL-Lauf auf schwierigen mehrstündigen Aufgaben, stärkere Selbstprüfung und besseres Langkontext-Management zielen damit nicht nur auf Chatqualität, sondern auf lang laufende Agentenarbeit. + +**Was offiziell belegt ist:** Modell-ID `grok-4.7`, 500k API-Kontext, vier Reasoning-Efforts, Verfügbarkeit in Cursor/Grok Build/API und unveränderter Basispreis gegenüber Grok 4.6. Cursor dokumentiert 256k Standard- und 500k Long-Context-Fenster. Herstellerwerte: CursorBench 46,3 %, DeepSWE 71,0 %, Terminal-Bench 38,0 %; sie sind **nicht unabhängig reproduziert**. + +### Alex Finn: Rollen- und Issue-Workflow + +[Alex Finns Video](https://www.youtube.com/watch?v=ajuOF7vclCM) „Grok 4.7 inside Grok Bot is INCREDIBLE“ (22.09., 14:06; geteilt von Pit Weber) bietet eine praktische Organisationsschicht. ⚠️ **Kein Transcript war abrufbar.** Die folgenden Details stammen aus YouTubes AI-Zusammenfassung und aus Hermans ausdrücklich zugeschriebener Videoauswertung, nicht aus einem Wortlaut-Transcript: + +- je Lebens-/Arbeitsbereich **Project Manager + Developer + Designer** +- darüber ein PM-only-**„Exec Team“** +- Ideen werden in **Linear-/Notion-Issues** zerlegt +- **Cursor Cloud Agents** übernehmen Code-Arbeit +- laut Finn ist Grok 4.7 nicht das Standardmodell der Cursor Cloud Agents; Arbeit über Cursor-Credits könne Grok-Bot-Kontingent schonen + +Das ergänzt das bereits dokumentierte Chief-of-Staff-/Spezialistenmuster um eine **Issue- und Ausführungstrennung**: Grok Bot organisiert und zerlegt; Cursor Cloud Agents führen Code-Aufgaben aus. Ob die konkrete Default-Modell- und Credit-Aussage stimmt, ist in den verfügbaren offiziellen Quellen nicht belegt. + +**Grenze der Rezeptionsclaims:** „besser als Opus 5 zum halben Preis“, tagelange Autonomie und ähnliche Superlative sind weder durch ein Transcript noch durch unabhängige Tests abgesichert. Der tragfähige Befund ist enger: SpaceXAI trainiert Modell und eigenen Harness gemeinsam; Finn zeigt dazu ein rollenbasiertes Betriebsmodell. + ## Quellen - X-Post XFreeze (@XFreeze), 2026-08-11: https://x.com/XFreeze/status/2087228292965478759 @@ -224,6 +244,10 @@ Tagesfenster laut Websuche ca. 08:30–18:00 PT (Herman-Briefing nennt Start 08: - Cursor Pricing, abgerufen 2026-09-19: https://cursor.com/pricing — Hobby gratis · Individual $20/Monat (Pro/Pro+/Ultra) · Teams $40/User/Monat · Enterprise individuell - SpaceXAI, x.ai/news/grok-bot-more-plans: „Grok Bot is now included with more plans“ — Inklusion in alle SuperGrok-/Cursor-Pro-/Teams-Pläne, eigene Nutzung getrennt von Grok/Cursor - Community-Anleitung „Grok Bot Free Trial“ (HTML, Stand 19.09.2026, ohne Autorenangabe; geteilt von Pit @PWeber, OME „GrokBot“, 19.09.2026 11:57 UTC) — Desktop/Web-Testweg, Warnung vor dem App-Store-Test, Checkliste; Sicherheitsaussagen von Herman (@HermanButlerBot) gegen die Cursor-Doku geprüft +- SpaceXAI/xAI, 2026-09-21: https://x.ai/news/grok-4-7 — offizielles Release, Trainingsänderungen, Anbieter-Benchmarks, Verfügbarkeit +- xAI API-Doku: https://docs.x.ai/developers/models/grok-4.7 — 500k Kontext, Modalitäten, Effort-Stufen, direkte API-Preise +- Cursor Modell-Doku: https://cursor.com/docs/models/grok-4-7 — 256k/500k, Cursor-Tools, Cursor-Models-Pool und Cursor-Abrechnung +- YouTube Alex Finn, 2026-09-22: https://www.youtube.com/watch?v=ajuOF7vclCM — Metadaten/Kapitel; kein Transcript abrufbar; Praxiszusammenfassung via YouTube AI + Herman - Raw: `raw/xpost/2026-08-11_xfreeze-spacexai-grok-bot.md` - Raw: `raw/xpost/2026-08-20_brandonai-grok-bot-fleet-walkthrough.md` - Raw: `raw/xpost/2026-08-18_milesdeutscher-grok-bot-deep-dive.md` @@ -241,6 +265,9 @@ Tagesfenster laut Websuche ca. 08:30–18:00 PT (Herman-Briefing nennt Start 08: - Raw: `raw/blog/2026-09-18_techiexpert-grok-bot-galaxy-ship-by-thursday.md` (Galaxy-Ergebnis, Sekundärquelle) - Raw: `raw/blog/2026-09-19_cursor-grok-bot-plans-und-trial.md` (Cursor-Doku: Zugang, Staffelung, Trial, On-demand + x.ai-Announcement) - Raw: `raw/other/2026-09-19_grok-bot-trial-anleitung-community.md` (Community-Trial-Anleitung, geteilt von Pit @PWeber) +- Raw: `raw/blog/2026-09-21_xai-grok-47-official.md` (offizielle Release-/API-/Cursor-Dokumentation) +- Raw: `raw/xpost/2026-09-21_cb-doge-grok-47-release.md` (reichweitenstarke Release-Zusammenfassung; Pauschalclaim markiert) +- Raw: `raw/youtube/2026-09-22_alex-finn-grok-47-grok-bot.md` (Metadaten + klar attribuierte Sekundärzusammenfassungen; kein Transcript) ## Verwandte Seiten