20:45Matt Wolfe · Essay
Why Everyone Is Freaking Out About Mythos
A 20-minute reality check on Claude Fable 5 and Mythos 5 — what the hype got right, what it got wrong, and what the safety leash actually costs you.
June 11thThe argument in one line.
The most consequential AI story of the week wasn't a capability leap — it was a control question: Anthropic silently downgraded users who asked the wrong questions, got caught, and partially reversed in two hours, setting a precedent for how AI labs handle what you're allowed to know.
Read if. Skip if.
READ IF YOU ARE…
- You want one weekly video to stay current on AI without spending hours on Twitter and YouTube.
- You're evaluating Claude Fable 5 for coding or agentic work and want a fair-witness read on its capabilities and restrictions.
- You're tracking Apple's AI pivot and want to know what actually shipped vs. what's still vaporware.
- You care about AI policy and want a plain-language take on Dario's and Sam's competing visions.
SKIP IF…
- You need deep technical analysis — each story gets 2-5 minutes of coverage, not exhaustive treatment.
- You're only interested in one of the stories covered — the format doesn't let any single topic breathe past 5 minutes.
The full version, fast.
Anthropic launched Claude Fable 5, a Mythos-class model above Opus, but its aggressive safety filters — including silent downgrades for biology, cybersecurity, and LLM-training queries — triggered immediate public backlash from the Hugging Face CEO and other researchers. Anthropic partially reversed within two hours. The model costs $10/$50 per million input/output tokens (double Opus), and free access ends June 22. Apple's WWDC headlined with a Siri reboot built on both Apple Foundation Models and Google Gemini, featuring cross-device context, visual intelligence via camera, and an AI shortcut builder. Google upgraded NotebookLM to Gemini 3.5 with 100+ skills and new output formats, released Gemini 3.5 Live Translate (demoed real-time in AI Studio now), and launched DiffusionGemma — a text model that generates 256-token paragraphs simultaneously for dramatic local hardware speed gains. The week closed with OpenAI's S-1 filing, SpaceX's record $75B IPO, and a Midjourney hardware tease.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:07
01 · Intro
Title card and promise of the weekly digest.
00:07 – 05:35
02 · Claude Fable 5 & Mythos 5
Model tier explanation, silent downgrade controversy, backlash from Hugging Face CEO and researchers, Anthropic's partial reversal, pricing ($10/$50/M tokens), June 22 access window.
05:35 – 10:43
03 · Fable Coding Examples
One-shot Mega Bonk 3D clone, AI Tube YouTube recommendation clone, two-prompt B-roll article generator app.
10:43 – 18:25
04 · Apple WWDC: Siri AI
Siri reboot built on Apple FM + Google Gemini, cross-device context, visual intelligence, Spotlight AI, Watch and Vision Pro, AI shortcut builder, image generation, spatial reframing (live in iOS 27 beta).
18:25 – 19:45
05 · NotebookLM Updates
Gemini 3.5 upgrade, secure cloud computer per notebook, 100+ skills, new export formats.
19:45 – 21:45
06 · Gemini 3.5 Live Translate
Real-time spoken translation, Google Meet integration, live demo in AI Studio.
21:45 – 23:56
07 · Diffusion Gemma
Text generation via diffusion approach, 256-token simultaneous paragraph generation, efficient local hardware utilization.
23:56 – 25:40
08 · Dario's AI Policy Manifesto
FAA-style regulatory body, UBI + AI taxes, international coalition, off switch, biomedical pipeline.
25:40 – 27:07
09 · Sam Altman's Manifesto
OpenAI Phase 3, AI-assisted research by 2028, broad power distribution, personal AGI for everyone.
27:07 – 27:28
10 · ChatGPT Email Sending
Send emails directly from within ChatGPT.
27:28 – 27:46
11 · OpenAI IPO Filing
Confidential S-1 filed with SEC, IPO likely 2027.
27:46 – 28:23
12 · SpaceX IPO
$75B raised, 555M+ shares at $135, $1.77T valuation, now publicly tradeable.
28:23 – 28:41
13 · Coinbase for Agents
AI agents can now trade and pay with Coinbase.
28:41 – 29:11
14 · Midjourney Hardware
David Holz teasing invite-only hardware launch, no product details.
29:11 – 30:51
15 · Final Thoughts + CTA
Weekly cadence promise, subscribe ask, sign-off.
Lines worth screenshotting.
- Anthropic's silent downgrade policy — where asking the wrong question returns a worse answer without telling you — is the most precedent-setting AI safety decision of the week, not the model itself.
- Claude Fable 5 sits above Opus in Anthropic's tier stack: Haiku → Sonnet → Opus → Mythos tier (Fable 5). The actual Mythos 5 model is restricted to government cyber defenders.
- Fable 5 costs $10/$50 per million input/output tokens — exactly double Claude Opus 4.8's pricing.
- Anthropic reversed its most restrictive safety decisions within two hours of launch, after public backlash from the Hugging Face CEO and AI researchers.
- Apple built its new Siri on both its own foundation models and Google Gemini — the first time Apple has used a competitor's model at the OS level.
- Apple's 'describe a shortcut' feature lets you tell Siri what you want to automate in plain English and it builds the iOS shortcut for you — the most immediately useful WWDC announcement.
- iOS 27 beta already has spatial reframing in the camera app — you can drag a photo outside its original frame and AI fills in the missing edges.
- NotebookLM now runs a secure cloud computer per notebook, enabling code execution and 100+ skills — not just document Q&A anymore.
- Diffusion Gemma generates a full 256-token paragraph simultaneously instead of word-by-word, using local GPU hardware more efficiently — the same approach image generators use, applied to text.
- Dario Amodei is calling for an FAA-style regulatory body to review AI models before public release — the most concrete policy ask from any major AI CEO.
- Sam Altman predicts that by March 2028, AI systems will be conducting a significant fraction of OpenAI's own research in tandem with human researchers.
- SpaceX raised $75 billion in a record-breaking IPO — by the time this video published, the stock was actively tradeable.
- Coinbase now lets AI agents hold and spend crypto — your agent can trade on your Coinbase account if you authorize it.
- Both Dario and Sam independently called for broad distribution of AI power; both labs' actions this week cut against that stated goal in different ways.
The week AI labs started arguing about who controls the AI.
WHAT TO LEARN
Capability announcements dominated the headlines, but the real story was about access: who gets the most powerful models, who gets silently downgraded, and whether either lab's stated commitments to broad distribution match their actions.
- Anthropic's silent downgrade policy — sending users to a dumber model without disclosure — is more significant than any benchmark score, because it establishes a template for how labs can quietly limit access.
- Claude Fable 5 is genuinely impressive at one-shot coding tasks, but the June 22 access window means most users will lose free access before they've fully evaluated it.
- Apple's Siri reboot is built on both Apple Foundation Models and Google Gemini — the first time a major OS has openly shipped a competitor's model under its own brand.
- The 'describe a shortcut' feature in iOS 27 is the most immediately practical AI feature Apple announced: plain-language automation that builds the iOS shortcut for you.
- Diffusion Gemma's approach — generating a full paragraph simultaneously rather than token-by-token — signals that on-device AI is heading toward efficient local GPU use, not just smaller model sizes.
- Both Dario Amodei and Sam Altman published policy manifestos in the same week, calling for broad distribution of AI's benefits — while both of their organizations made decisions that concentrated capability among smaller groups.
- The OpenAI S-1 filing means the next major AI governance question isn't just regulatory — it's what shareholder pressure does to a mission-driven lab's safety commitments.
Terms worth knowing.
- Mythos-class model
- Anthropic's label for a new capability tier above Opus. Claude Fable 5 is a Mythos-class model with safety restrictions intact; Claude Mythos 5 is the same underlying model with some safeguards removed, available only to vetted cyber defenders.
- SWE-bench pro
- A benchmark measuring an AI model's ability to solve real software engineering tasks. Its reliability is disputed after evidence that some models accessed answer keys rather than solving problems independently.
- Project Glasswing
- Anthropic's collaboration with the US government through which the restricted Claude Mythos 5 model is being deployed for national cybersecurity purposes.
- Diffusion model
- A generative AI architecture that works by iteratively refining noise into a final output — used in image generators like Stable Diffusion and now applied to text generation in DiffusionGemma.
- Spatial reframing
- Apple's iOS 27 feature that lets you drag a photo beyond its original borders; AI generates the pixels outside the captured frame to fill in the new composition.
- S-1 filing
- A registration statement filed with the SEC to begin the process of going public. OpenAI filed a confidential S-1, signaling intent to IPO without disclosing the timing or financials publicly yet.
Things they pointed at.
Lines you could clip.
03:25
“If you're asking this model to help you create your own separate large language model, it would still respond. It would just dumb down the model and give you a worse response than it should.”
30:01
“I'll drink from the fire hose. I'll be overwhelmed all week, and then I'll turn around and break it down for you on these Friday videos so you can just watch one video a week and get totally looped in on everything you need to know.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
00:05Here's the AI news you probably missed this week. Starting with the brand new model that has everybody talking right now, Claude Fable five. Now this is a model that I did an entire breakdown video of where I explain exactly what this model is, what people love about it, what people hate about it, some misconceptions, and even test it a little bit.
00:22You can find it over on my YouTube channel. It's called why everyone is freaking out about mythos. At least for right now.
00:27I split test titles and thumbnails constantly. In case this is, like, the only video you watch each week to stay up to date with the news, I'll give you the TL didn't watch. So in Anthropic's announcement on June 9, they actually announced two models, Claude Fable five and Claude Mythos five.
00:42There's been a lot of videos and tweets and things like that saying that Anthropic finally unleashed Mythos on the world, but that's not technically true. They call Claude Fable five a Mythos class model, but it's not actually the Mythos model that everybody was freaking out about a couple months ago. This Mythos class model sits at a tier above Opus.
01:01So they have Haiku as their lowest tier, fastest, cheapest models to use. Sonic was their middle tier. Opus was their top tier.
01:08Now the Mythos tier is the tier above Opus, and we're getting Fable five, which is a Mythos tier model. The actual Mythos five model is for a small group of cyber defenders and infrastructure providers. They claim it's the same underlying model as Fable five, but with the safeguards lifted in some areas.
01:25So there's a lot of extra safe guards put on top of this fable five, which prevents it from cybersecurity related prompts, biology related prompts, and even prompts related to, like, training your own AI models. Anthropic doesn't want anybody to use their models to help them train their own models, which is like a big point of contention right now.
01:47If you're interested in benchmarks, here's how it stacks up against other models. It's pretty much state of the art in everything if you trust these benchmarks. Now, SWE bench pro, hard to trust because it was recently found out that a lot of the Claude models when being tested for this benchmark essentially looked up the answer key instead of actually solving the problem.
02:06Now this model is really designed for, like, the biggest, toughest, most logic intense projects that you throw at it.
02:13We could see here fable and mythos can work autonomously for longer than any previous Claude models. Now the big issue that people have with this model is more the safety and security issues. People have been showing screenshots where they upload blood work to it and ask it to analyze the blood work, and it refuses saying that it's not safe to do that.
02:31Or people asking questions about cancer. This is actually something I confirmed myself when I asked it explain how BRCA one mutations increase breast cancer risk. I originally asked Fable five this prompt, and it switched it to Opus 4.8 for me saying that it was too unsafe to ask Fable.
02:49There's also some controversy around this. Like, if you were to ask it about LLM development, it would actually pivot to, like, a dumber model and not answer the questions for you. This is from Anthropix documentation.
03:02We've added safeguards related to Frontier LLM development. We've implemented new interventions that limit Claude's effectiveness for request targeting Frontier LLM development. Unlike our interventions for cybersecurity biology and chemistry and distillation attempts, these safeguards will not be visible to the user.
03:18So, essentially, if you were asking this model to help you create your own separate large language model, it would still respond. It would just dumb down the model and just give you like a worse response than it should. And people on x were not happy with this.
03:31The CEO of Hugging Face said concentration of power, capabilities, and economic wealth is the biggest risk in AI. We need open science and open source more than ever. Jeremy Howard says, Anthropic has chosen the opposite of the safe path.
03:43They're allowing themselves, the current top lab, to use their top model for frontier AI research. They've said they'll sabotage others who try. This means the AI frontier advances and power imbalance increases.
03:54Graham Newbig here says, I feel we're getting a glimpse of a future where AI is only provided to a privileged few, and that's not a future I wanna live in. And there are many such cases of tweets along this same vein here. Luckily, however, Anthropic relented, sort of.
04:10According to this article on NBC News here, after the outcry, Anthropic backtracked and reversed some of its most conservative decisions less than two hours after Fable five's release, highlighting growing concerns about AI company's ability to unilaterally limit users access to helpful AI generated information. You should have visibility into the safeguards we have in place and why.
04:29We're sorry for not getting the balance right. Basically, what it sounds like is happening is Anthropic's still gonna block these things. They're still gonna send you to a dumber version of the model when you ask them.
04:40They're just gonna tell you that they're doing it instead of not telling you that they're doing it. So some quick additional details here before actually showing you some of the things it's capable of. As of right now, you can use it inside of one of the paid plans.
04:54So pro max and team plans can use this Fable five model, but only until June 22. On June 23, they're gonna remove Fable five from those plans. And if you wanna use it after that, it's gonna require additional usage credits.
05:08They do say if capacity allows, it'll extend the included window, so maybe we'll get it for a little bit longer. It is, however, pretty much the most expensive model. Pricing for both models is $10 per million input tokens and $50 per million output tokens.
05:21That's exactly double the cost of Claude Opus 4.8. Now saying all of that, it is a really, really good model, especially if you're using it for, like, coding stuff.
05:33It's coded stuff with one shot that just blew me away. In the last video I made about it, I actually gave it the prompt, make a clone of Mega Bonk, which is, you know, a really fun three d game, and it basically one shotted this game. And it looks really, really good.
05:49I mean, this was the very, very first iteration out of the box with only that single prompt. I mean, it looks more like a potato than an actual person, and there's not much animation. We're just getting cubes, but the game was really fun.
06:01I was pretty impressed with how it came out on the first try, and it coded the whole thing into a single HTML file, which is kinda crazy. But I wasn't totally happy with that first version, so I went back and forth with a few more prompts, maybe four more prompts, giving it different features and different things I wanted to do with the graphics.
06:20And here's where that game stands now that Fable created. And, yes, it even generated the music, but we can see our character walks around. I can actually rotate my camera around the character.
06:29Our enemies are a little bit better looking than the queues we had before. There's actually different depth to the terrain. There's, you know, rocks and things that you can bump into.
06:40I added the ability to sprint, the ability to jump. It's just a much better game than the original potato running around killing cubes.
06:49So I was really impressed with this. I even put this game online if you wanna try it and play around with it for free. The link will be in the description, and I put it up on GitHub.
06:58So if you do wanna download this and fork it and make your own version of it, it is available to mess around with there. But I'm warning you, it's pretty addicting game. Just for fun, I asked Fable to make me a YouTube clone called AI tube, not because I wanted to see if it can clone the look of the site, but I wanted to actually see if it would clone a recommendation algorithm.
07:18I wanted it to pull a bunch of videos from YouTube, and then based on videos that I watched, it actually recommend new videos, and this is just one shot. This is one prompt here, and this is what it looks like.
07:29As you can see, this is AI tube. This is not YouTube, complete with history. You can see I've already been testing it where I watched a couple videos in the history.
07:36There's an explore page with hardware, podcasts, entertainment.
07:40All again, videos related to AI, and it's showing a bunch of my videos because, well, I watched one of my videos and then it started recommending more of those. If I come down here and watch, you know, like one of Matt Vid Pro's videos, you can see it looks just like I'm watching it on YouTube.
07:56It's got this sidebar with up next and everything. But if I actually go home, it should have updated my homepage. Now, it's still recommending my videos because it pulled in more of my videos than any other videos, but we can see that it no longer is recommending the Matvid pro video because it thinks I already watched it.
08:12If I click on this Wes Roth video Alright. So and then click back to my home page, I actually have a slightly updated set of videos. I don't know how well the recommendation algorithm is working.
08:22It's kinda hard for me to test since it seemed to have pulled in a lot of my own videos, but it seems like it is updating. Every time I watch a video and come back to the home page, the videos sort of change a little bit.
08:32Anyway, it was a fun one prompt test. I don't actually plan on pursuing this. It's a real project.
08:37But probably the most impressive project that I built is this new one called YouTube article b roll generator. Now this one actually lives on my desktop. It doesn't live in the browser, but let me show you what this one does.
08:48You can see I can turn any article into b roll. I just give it a URL. So for example, I can give it this article we're just looking at on Anthropic Claude Fable five, but what I did was I made it so I can highlight text like this.
09:01Let's just say, I wanna highlight for a small group of cyber defenders and infrastructure providers, blah blah blah. I select that text, click save selection, and then maybe I want to select some more text down here about knowledge work. Select this paragraph here.
09:15Click save selection at the top. And then if I click capture page, watch what happens. It actually screenshots the entire page, and then it gives me these animations automatically out of the box where it zooms in on the page for me and highlights the text I just highlighted just like that.
09:30So it highlighted that first chunk of text, and then it's gonna zoom down and highlight that second chunk of text that I was talking about right there like that. And I've made it have a bunch of different animations. So I've got this headline reveal.
09:42So this is the second type of animation that it'll create where it'll zoom in like that. And then I've got this read along animation where it does it in this style. They're all pretty similar, but you get the idea.
09:52We've got a spotlight here where I press play on this video, and you can see it sort of spotlights it like that or this box call out. You could see it zooms to this and then sort of circles it. Yeah.
10:03Still some issues I gotta figure out here, but then it circles that paragraph there. I can change how long each of these steps last for. I can change the color of the highlighting, change the color of the borders, things like that.
10:16And then when I'm done, I click on export and it saves it as an m p four, and now I've got b roll of this article that I can throw into my video. And I don't have to go record this b roll separately. It just generated it for me.
10:28Obviously, a very niche problem that solves something that I do in my videos, but it's gonna save me a heck of a lot of time, and this was done with two prompts in Fable. Alright.
10:38I wanna move on past Fable for now because I did make a whole video about that, but there was another huge event this week that the AI world is buzzing about, and that's Apple's WWDC event, their worldwide developer conference event. And this year, it was all about AI.
10:55So I'm gonna quickly cover some of the big AI announcements that came out of WWDC. I'm gonna do my best to sort of rapid fire cover everything because, again, there was a lot, and I'm not getting into, like, every single announcement they got into, just the sort of more practical AI stuff that I think most people are gonna care about.
11:12They started by announcing that they're actually going to be using their own foundation models as well as Google's Gemini models. This year, we embarked on a deep collaboration with Google, leveraging the technologies behind their Gemini family of models.
11:26Together, we created the next generation of Apple Foundation models for our integrated Apple intelligence experiences.
11:33They explained that when you use Siri's AI on the phone, it's either going to handle that AI prompt on the phone directly or using private clouds where Google gets none of your data saved. They talked about how Siri's gonna have this personal context understanding. So it's gonna know your photos, your calendars, your text messages, pretty much anything that developers give access to this context, Siri is gonna be able to use that.
12:00A big chunk of this keynote was dedicated to, like, the Siri that we kind of always hoped Siri would be. More capable.
12:08It has access to world knowledge. It has access to context on your phone. It's supposedly gonna be more conversational.
12:14There's going to be a dedicated Siri app. It's gonna have visual intelligence so you can actually open your camera and it will understand what it sees around you as well as what's on your photos. You're gonna be able to write with Siri.
12:25They showed off some examples in their keynote where they asked about tickets for a concert, and it told them they had to enter a lottery. And then he set a reminder to enter the lottery at that time. He showed it a picture on Instagram and asked where it was, and it looked at that picture, figured out where it was, and then gave him directions to where that was.
12:43You know, he was rearranging photos by dictating to Siri. So it sounds like you're gonna be operating your phone by just sort of talking to it if you wanna do it that way. Hey.
12:52Take this photo. Put it here. Hey.
12:53Remind me about this thing. Look at the photo on my screen. Tell me where it is.
12:57Give me directions to it. Like, they're just trying to make this thing much more conversational to go do things on your behalf within your phone in the context that it has on your phone. You'll be able to customize your Siri, like how fast it talks and how expressive it is.
13:11There's various voices you can choose from. You know, kind of the stuff that you would expect from most of, like, an AI chatbot these days. Here's a few more examples of ways that they were using it in their demo.
13:21They put together a plan to visit different places in New York. It reviewed the conversation and helped them get back up to date with it. It reviewed some financial data for them.
13:31Again, a lot of what we've seen out of tools like ChatGPT, like Anthropic, like Gemini, it's a lot of the same kind of stuff just now coming native to an iPhone.
13:42Another thing I found kind of interesting is if you're on, like, a Mac or a MacBook, Siri's going to come to Spotlight. So, you know, if you hit command space bar and bring up this little Spotlight search, you're gonna be able to prompt it with AI and say, you know, rearrange these folders or look at all of these documents and tell me the differences between them, things like that.
14:00And the built in AI on device will be able to do those kinds of things for you. You could see he plugged in three different, like, shed kits, selected them all, and then asked for a comparison, and then it popped up a little box where it compared the three for him. This new version of Siri is gonna come to your watch.
14:16It's gonna come to the Apple Vision Pro. You can see this little bubble in this center of the screen here. Essentially, you wanna Siri on your Apple Vision Pro, you just look at the bubble and because it's got eye tracking, it sees that you're looking at it.
14:26And when you talk to the bubble, it, you know, does Siri stuff. And because there's a new dedicated Siri app, it sort of will save all of your conversations and things you're doing in that app so it works across device. And theoretically, you'll have context across device.
14:41So you can be doing something on your phone or your watch, your Apple Vision Pro, be sitting at your computer later, and ideally, it will have the context of what you were doing on that other device earlier. They're also adding a new Siri mode in the camera app. So if you select Siri, then it basically will use the context of what the camera sees for whatever you prompt it with.
15:00So in this picture, they're pointing it at a red ball and because they're in Siri mode, Siri starts telling them this appears to be a traditional cricket ball, etcetera. They're able to highlight things on the screen and ask Siri about those things. They look at like a calendar on a web page and it actually pulls in the event separately.
15:16You know, here's them talking to Siri in the Apple Vision Pro about what they're seeing in the browser. You can write with Siri, so it'll actually help you write like a rough draft. You know, it's a lot of the same stuff we've kind of had access to in almost all of the AI tools just, you know, now Apple's version.
15:31It'll give you suggestions inside of mail based on what it reads in the email. Now out of the entire keynote, this was the thing I was probably the most excited about actually getting on my phone is this shortcuts feature with AI. You can essentially describe the shortcut you want.
15:46Right? Like, if you're not familiar with shortcuts, they're like little automations that you can set on your phone. Like, if I'm in a certain location, send this text message to this person.
15:54So they're adding this feature called describe a shortcut, which allows you to use AI to say, here's what I want to happen, and it will actually go and build that shortcut for you. Like, you just say, when I'm leaving work, message Pedro. I'm on my way with my ETA, and it will actually go and build that shortcut for you.
16:11I, like, never use the shortcuts feature, but if I can just, like, explain what I want it to do and it goes and automatically figures out how to build that shortcut for me, that'll be pretty cool, I think. You're gonna be able to generate images with AI. So they gave it a picture of somebody and said, make her holding a cake.
16:26Then they said, give her a detective outfit, and it generated the same image, but with a detective outfit. And then they selected the cake and said, add candles to the cake. And then, of course, it generated an image with candles on the cake.
16:36You know, stuff that we've been able to do with all the other image generation platforms for a while now. Now this is one of the other ones that I thought was kind of cool, and I believe we might have access to this if you've updated to the beta version of iOS 27, which is the spatial reframing.
16:51You could take a picture like he did on his phone here, and if he doesn't like the framing, he can click this little reframe button and then use AI to change the framing, the angle of the picture. And you can see the little blurry parts of the parts that were outside of the photo, but it's gonna use generative AI to fill in those areas.
17:08But the portion of the photo that is still clear is not using generative AI. And I feel like those were the main, like, big announcements to come out of WWDC this year.
17:17Now if you do have an iPhone and if you did upgrade to the beta of iOS 27, you can use the reframe right now. So if I open my camera app and I take a picture of just like what I see in my office right now, let's look at this photo.
17:30Click on the settings, click on tools. And you can see I now have reframe down in the bottom right.
17:36So if I click on reframe, it scans the photo. And now I can actually drag this around and reframe this photo.
17:43Now we can see that it's like distorting my camera and, you know, it will have some wonkiness. It's obviously more designed to, like, reframe people in the photo. But let's, like, change the angle entirely, click reframe, and now it's gonna generate what it thinks the edges should look like.
17:58So there you go. And that's available in beta right now. One thing to note about all of these Apple AI announcements is they said they won't initially be available in The EU due to their AI restrictions there.
18:10They are hoping to figure that out as quickly as possible, but at launch, they won't be available. Moving on. Google had a series of announcements this week that if they weren't overshadowed by Anthropic and Apple, would probably be the biggest announcements of the week.
18:25Let's start with the fact that Notebook LM got a bit of an overhaul this week. It was upgraded to run on the Gemini 3.5 model, and each notebook is now equipped with a secure cloud computer, which enables it to write and run code useful for helping you perform deeper research and more complex analysis.
18:43So there's over a 100 curated skills that now work behind the scenes in Notebook LM, providing, like, a ton of new capabilities within the platform. There's also new formats that it can output. PNG and SVG, PDFs, docx, markdown, text files, PNG, JPEG, GIF, or GIF, CSV, JSON, XLSX, PPTX, and you could even make edits after your outputs are generated.
19:07It's gotten better at research. You can actually now start with loose ideas and questions, and Notebook will guide you through building your source repository directly in the chat.
19:17Now on the surface, when I log in to notebook l m, I don't notice a real huge difference, but it's when you actually start chatting with it, you'll notice that it's using a much better model and that it can actually use skills while you're chatting with it. Create a chart that shows the growth of the birds aren't real movement.
19:32We can see it now says confirming tool use and crafting the visual, and it kicked off the infographic tool over here on the right. And here's the chart that it generated for me. Not quite what I was asking, but, uh, you know, it took its own liberties.
19:45But Google also showed off their new Gemini 3.5 live translate model, which does exactly what it sounds like.
19:52It will actually translate conversations in near real time. Here's the example they showed off. Watch as we listen to the Google IO keynote in hint.
20:00What's really incredible is how people are using our AI.
20:04AI.
20:08Gemini apps final
20:12I think you get the idea. It's pretty much like real time. You can watch a video and have it translate for you in real time, and it's actually coming to Google Meet.
20:20So you'll be able to have meetings with people in other languages directly inside of Google Meet pretty soon. Cassie, how's the weather in Shanghai?
20:28It's
20:30nice to meet you. The weather here is really nice. I spent the whole day with my family in the park.
20:35That sounds wonderful. So that should pretty much eliminate all language barriers on meetings and videos online. It's also coming to the Google Translate app on Android and iOS, and you should be able to play with it right now over inside of a istudio.google.com.
20:50So if I come up here, click into the playground, click into speech and music here, you can see Gemini 3.5 live translate preview. I can select this and I can directly talk and it will translate.
21:03Let's have it translate to Hindi and I'll click talk. Hello, everyone. Don't forget to like and subscribe to Matt Wolf on YouTube.
21:11Matt
21:12Walker YouTuber like or subscribe I
21:16don't know how well that worked, but you get the idea here. But I can also share audio from a tab. So if I select a target language here, let's do Spanish and then share audio from a tab.
21:26I could select my YouTube channel here, press play on this video.
21:32Cloud fable And
21:36you can see that as I was watching the video, it was translating it in real time just on another tab, and that's available right now over an AI studio on Google. And the other announcement that Google made this week was this new diffusion gemma, and what's special about this is it does text generation, but it just does it really, really fast.
21:55It uses the same technology as the image generation models, you know, what stable diffusion in mid journey and. Of these are using diffusion models under the hood. Well, this is using that same technology, but it's generating text instead of images.
22:08We can see it's not the most intelligent model in the world. It's about as intelligent as Gemma four twelve b, maybe slightly less intelligent, but it just blows everything out of the water on speed based on tokens per second here. We can see in this example here where it's doing the Sudoku board, how quickly it solved this right side versus it's still working on this left side.
22:30So it's speed that this is really, really good at. So here's the quick explanation of what this is and why you might wanna use it.
22:38Most language models act like a typewriter generating one token at a time from left to right. In the cloud, this is efficient because servers can batch thousands of users requests together to share the hardware load.
22:47But when run locally for a single user, this word by word process leaves your dedicated GPU or TPU underutilized. It spends most of its time just waiting for the next keystroke. Diffusion Gemma reverses this inefficiency.
22:59Instead of predicting words sequentially, it drafts an entire 256 token paragraph simultaneously. By giving the computer's processor a larger chunk of work at once, Diffusion Gemma utilizes your hardware to its full potential.
23:11So when you start using a lot of these models on your own computer more and more, models like these diffusion models are going to use your own hardware much more efficiently. And honestly, I think that's the way things are going. I think more and more people are going to want to figure out how to run AI models on device without sending them to clouds, without sending their data to big companies that might train on them or use them to target advertising to you and things like that.
23:36So models like this show that companies like Google are actively thinking about how do we get better models on your own home device. Alright. I know I've covered a lot.
23:45It's been a big week, but I have a few more smaller things that I wanna talk to you about real quick. So let's just jump into a rapid fire, and I'll cover them super quick.
23:57Let's start with two different articles penned by, well, two of the biggest named CEOs in the AI space right now. Dario put out an article this week called policy on the AI exponential, and Sam Altman and Jacob Pachocchi put out this article called built to benefit everyone, our plan.
24:13Two sort of articles laying out how these two companies see the next phase of AI. Now Dario's article, the main sort of thesis of it is that we're approaching what he's calling powerful AI, or in words, a country of geniuses in a data center. Now this is a super long article here, so I'm gonna give you the TLDR.
24:32Dario argues that we need like an FAA style regulatory body that reviews these models before the public gets access to them to make sure they're safe and secure.
24:41He believes that there will continue to be this hyper growth in the economy, but we need to make sure that everybody shares in on that wealth. He acknowledges there will be job losses and also suggests that something like a UBI and actual higher taxes on the AI companies that get distributed is one way to sort of solve this.
25:01He talks about how we can have huge advancements in drug discoveries, but that the current biomedical pipeline is too slow, and that we should figure out ways to speed up the process and let AI take over a lot of it. As we've heard in the past, he's worried about domestic surveillance, autonomous weapons, things like that, and that we need an off switch for these models just in case things do get out of hand.
25:22And he suggests that we need some sort of, like, international coalition all working together to secure AI because any one country being too far advanced is like having a modern military, but in medieval times. So he's calling upon sort of international collaboration to solve for this.
25:40Now onto the article from Sam and Jacob. This article is a little less long, a little less dense, but here's the TLDR anyway. They believe that by March 2028, we may have a significant fraction of their research being done by AI systems in tandem with researchers.
25:54They also believe the economy is going to accelerate by accelerating scientific progress, productivity, and economic growth. And they also say that they're working to ensure gains are widely shared. Everyone should have an opportunity for meaningful share in the prosperity AI creates.
26:08The sentiment shared by both CEOs. They wanna give everyone on Earth a personal AGI, empowering them to benefit from one of humanity's most transformative technologies in whatever way they choose.
26:19They go on to say that they're entering what they call the third phase of open AI. The first phase was doing research towards AGI. The second phase was when the research became relevant and, you know, basically useful to the world, chat GPT essentially.
26:32And now we're on to the third phase. This is where the economy is beginning to shape itself around AI, but they go on to reiterate that above all, we believe a broad distribution of power will help lead to a better future. They wanna make sure it's accessible for all, and they also believe safety, privacy, affordability, open ecosystems, and public oversight matter.
26:51A good AI future cannot be one where a small number of institutions control most of the capability and most of the upside. They believe the transformation should belong to everyone. Very similar messages, one a little bit longer, one a little bit more brief, but coming from both of the most dominant AI labs of the moment.
27:07Alright. Those were a little less rapid fiery than I want them to be. The rest will be really quick.
27:11I promise. ChatGPT rolled out a new feature where you can actually send emails directly from with inside of ChatGPT. So you can have it craft an email.
27:20You plug in the email address that it's going to, and you can send it directly from within the chat app. We have a couple announcements around IPOs. OpenAI filed for an IPO.
27:31We don't actually know when this IPO is gonna happen. It's probably still a ways off, honestly, if I had to guess. We'll see the OpenAI IPO, like, next year, early next year sometime, but they did file.
27:43So they're showing their intention to IPO soon. But since we're talking about IPOs, as of this video going live, SpaceX has officially IPO'd.
27:52Now as of the recording, they raised $75,000,000,000 in a record breaking IPO, but this is, you know, before the official listing on the market where people can start trading it. So by the time this video is live, people can actively trade SpaceX on the market.
28:07At the time of this recording, they sold 555,000,000 plus shares for a $135 a piece, making the valuation 1,770,000,000,000.
28:17But, again, as of the time this video is actually live, the valuation of this company is now up to the open markets. Coinbase is making their platform available for your agents. So now your AI agent can trade and pay with Coinbase.
28:31So if you feel comfortable enough giving your Open Claw or Hermes agent or whatever agent of choice you're using access to your Coinbase account, apparently, you can do that now. And finally, we got this tweet from David Holtz, who is the founder of Midjourney, and they're kind of teasing there's some sort of, like, physical gadget coming.
28:51We're starting to send out invites for our first Midjourney hardware launch, and I feel like I'm missing some important mutuals. Poke me if you haven't got an invite and think you should. I have no idea what this means, but apparently, mid journey is releasing some sort of hardware in the near future.
29:04So that could be interesting. Still no clue what it is, but we'll see. If have ideas, let me know in the comments.
29:09Whatever it is, I'm sure I'll probably end up buying it. Anyway, that's what I got for you. I record these videos on Thursdays, publish them on Fridays.
29:16So if there's any news that I missed that came out, like, late Thursday night or on Friday, it didn't make this video, but I'll make sure it makes it into next week's video. There is a lot happening in the world of AI. Every single day, there's new announcements.
29:28There's so many new tools to test and models coming out that are supposedly better and YouTube channels and Twitter accounts that are hyping up everything is like this business is cooked or this model just changed the game or if you're not using this agent then you're falling behind. And so much of it is hype and so it is my goal with this channel to cut through all the noise, cut through all the hype, and share with you just the things I think are most important for the most amount of people to know.
29:55I'll drink from the fire hose. I'll be overwhelmed all week, and then I'll turn around and break it down for you on these Friday videos so you can just watch one video a week if you want and get totally looped in on everything you need to know.
30:06I'm also gonna throw in some, you know, tutorials and cool ways to use some of those tools in between the news videos, but these news videos, you can count on them pretty much every Friday. If that's something that interests you, maybe consider liking and subscribing to this channel, and I'll make sure more videos like this show up in your YouTube feed.
30:22But, again, that's what I got for you. It's been a very hectic week. A really cool model in Fable.
30:27I'm having a lot of fun using Fable to generate games and apps that make my editing life easier. And, you should go try them out.
30:34They're pretty cool. Anyway, again, thanks for nerding out with me. Thanks for hanging out with me.
30:38I'm done rambling. Hopefully, I'll see you in the next one. Bye bye.
The Hook
The bait, then the rug-pull.
It started with a model launch and turned into a week-long argument about who controls what you're allowed to ask an AI. Claude Fable 5 arrived above Opus in the tier stack — and within hours, Anthropic was issuing an apology for the safety filters it had quietly built in.
Named ideas worth stealing.
01:20list
Anthropic Model Tier Stack
- Haiku (fastest/cheapest)
- Sonnet (mid-tier)
- Opus (top tier)
- Mythos tier: Fable 5 (general) + Mythos 5 (restricted)
Anthropic's new 4-level model hierarchy, with Mythos tier sitting above Opus and split into a consumer-accessible model (Fable 5) and a government-restricted model (Mythos 5).
Steal forExplaining AI model tiers to clients or in content about choosing the right model for a task
22:35concept
Diffusion vs. Autoregressive Text Generation
Autoregressive models (GPT-style) generate tokens one at a time; diffusion models generate a full 256-token paragraph simultaneously, utilizing local GPU more efficiently.
Steal forExplaining why local AI models are getting faster without requiring more powerful hardware
How they asked for the click.
VERBAL ASK
30:00subscribe
“If that's something that interests you, maybe consider liking and subscribing to this channel, and I'll make sure more videos like this show up in your YouTube feed.”
Soft, honest, no hard sell — framed as value exchange. Comes after a genuine summary of what the channel promises.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
- ? Twiter/X ↗
- ??? Instagram ↗
- ?? Threads ↗
- ?? LinkedIn ↗
- ?? Facebook ↗
- Claude Fable & Mythos 5 ↗
- Anthropic Downgrade Transparency ↗
- Cube Basher ↗
- Cube Basher Repo ↗
- Apple Intelligence Update ↗
- New Siri AI ↗
- Apple AI Unveiled ↗
- NotebookLM Research ↗
- Gemini Live Translate ↗
- DiffusionGemma ↗
- Policy on AI Exponential ↗
- OpenAI's Plan ↗
- ChatGPT Email Sending ↗
- OpenAI S-1 Filing ↗
- SpaceX $75B IPO ↗
- Coinbase for Agents ↗
- Midjourney Hardware Launch ↗
Visual structure at a glance.
