Google's Push for AI Dominance & More AI News You Can Use

0:00 Welcome to another week in AI.

0:01 Between Google releasing a firehose of new releases and Anthropic lifting

0:05 the ceiling on what's possible and best-in-class open-source models from China,

0:10 we have a lot of ground to cover in this week's episode of AI news you can use,

0:13 the show that pulls together all AI releases,

0:15 we filter for the ones that matter,

0:16 and then I personally get to present it back to you.

0:18 So, let's start with the story that made a lot of headlines

0:21 that I created dedicated video on cuz I wanted to talk about it.

0:24 Claude 3 Opus, their flagship model that is too dangerous to be released.

0:29 Obviously a PR move, but also obviously super impressive benchmarks.

0:34 In short, my take is I think this is going to be as good

0:38 as we expect because Claude has a track record of delivering the best models,

0:42 but what warranted that discussion in the separate

0:44 video was the second and third-order consequences

0:47 of us getting a gigantic model that is

0:49 just head and shoulders above everything we have now.

0:51 So, that's interesting and we'll talk about that more once it comes out,

0:54 but there's a few more Anthropic-related things that we need to talk about here.

0:57 First of all, last Sunday they cut off the ability

1:00 to use their Claude subscription with Open Claude or other agents.

1:03 In other words, people who were previously paying $200 a month

1:05 to play around there now will need to pay a thousand,

1:08 maybe even up to $3,000 a month depending on the usage.

1:11 For most people and what they're doing with those systems, that's not worth it.

1:14 For me as a content creator who's been tinkering

1:16 and developing more and more things that work there,

1:18 there's actually a bunch of great use

1:19 cases for content creation concretely that yeah,

1:22 even the increased price tag is well worth it.

1:24 Also, I want to course-correct on one thing I said last week.

1:27 It was a story that I called very last second and I mis- reported it to you.

1:30 So, my bad for that.

1:32 It's about the source code leak of Claude Code.

1:35 I said that it was an April Fools' joke.

1:37 It was also a Claude Code leak that was an April Fools' joke.

1:40 So, actually wasn't an April Fools' joke.

1:41 It was real and there were some interesting insights from it.

1:44 Nothing groundbreaking.

1:46 I want to highlight one interesting thing that had happened though.

1:48 So, if you're not familiar, Claude Code, which is Anthropic's agentic product,

1:52 basically leaked and users took the code and looked into it.

1:56 What were they doing?

1:56 How do they handle memories and more?

1:58 And one thing they found is that Claude Code

2:00 had a dreaming function where basically every day it stored

2:04 its experiences in certain files that it then later

2:08 would recall much like a human does when we sleep.

2:10 Great idea, right?

2:11 Cross-session memory.

2:12 Fantastic and keep working with it and it

2:14 remembers what you were working on yesterday.

2:16 And within a few days, Open Claude,

2:18 which obviously is the open-source competitor to Claude Code,

2:21 added a new feature, dreaming.

2:23 So, now if you use Open Claude, it can handle memory in a similar way.

2:26 And I think that's just so funny because

2:28 this is going to keep happening, you know?

2:29 You're going to see in today's episode.

2:30 We get new apps from Google,

2:31 we get new completely open models from China that are

2:34 almost as good as some of the best models available today.

2:37 And I just thought it was so funny that one

2:39 of these big tech companies is developing a proprietary system

2:42 that is closed and then a leak happens and within

2:44 a few days the open-source alternative implements one of its best features.

2:49 I think we can expect that moving forward.

2:51 It's going to be an interesting time.

2:52 Google, oh boy, did they feel like releasing a bunch

2:54 of stuff to become a part of this week's AI hype.

2:56 Let's have a look at that.

2:57 So, first of all, another interesting thing

2:59 that I couldn't get access to yet from Google, but if you're on the Ultra plan,

3:02 the $250 plan, and are in the US,

3:05 select users are getting this already and it's called AI Inbox by Gmail.

3:09 It's Google's attempt to build a email

3:12 assistant that gets you to inbox zero automatically.

3:14 Look, they shared this screenshot and I think this is one

3:16 of the things that nine out of 10 people would love to have.

3:19 An AI agent that helps them manage their inbox

3:22 in a contextual way where it knows them,

3:24 knows the previous interactions.

3:25 So, from what I can see here, this doesn't focus on actually replying to emails.

3:29 It's more like prioritizing the emails

3:30 that you get and creating daily personalized briefings.

3:34 Mind you, this will be just step one.

3:35 I mean, they're slowly rolling it out just to the highest-tier customers.

3:38 Within 6 months, everybody's going to have this on an affordable plan

3:41 and there's going to be more proactive versions of this too, of course.

3:43 And my Gmail agent is going to talk to your Gmail agent to figure out when

3:47 our Google Calendars have time slots so that we

3:49 can meet in person and talk about AI.

3:51 Utopia.

3:52 If you're on an Ultra account in the US, check your account.

3:54 You might have this already.

3:55 Okay, then next we have an application

3:57 that went pretty unnoticed across social media,

3:59 but I thought it was really worth mentioning.

4:01 I don't know about you, but I wasn't familiar with Google Vids.

4:04 It's a very simple video editing software,

4:06 but even if you're not into that, they're using that as a vehicle

4:09 for all these different applications that we've

4:12 been watching over the past few months.

4:14 Video generation model, music generation model,

4:15 they're all putting it into here.

4:18 For example, if I want to start out by creating

4:20 a AI avatar for my video, I can do that.

4:22 Now, I bring that up not because AI video

4:25 is like the biggest and hottest topic right now,

4:27 but because I love to keep an eye on how

4:29 these apps are evolving every week as you might know.

4:31 And as I pointed out before,

4:32 we're releasing these bundling and this is a great example of that.

4:35 It's the same experience as before,

4:36 but now I don't have to bring it over into video editing software.

4:39 All right, guys, welcome back to the channel.

4:41 I'm here in Soho and we need to talk about the latest updates in generative AI.

4:45 Damn, [clears throat] this guy's background looks better than mine.

4:50 And if you're on a pro account, you can even create songs right in here.

4:53 This is so easy to use and part of the Google

4:55 Suite that if you ever needed to create videos with AI avatars,

5:00 this does feel like the most frictionless way to do that today.

5:03 And if you want it, pro tip, you could start with a slideshow that you have,

5:06 just add a voiceover and a custom song right in here.

5:09 You don't even have to touch the AI avatar stuff.

5:11 Additionally, it also has avatars feature where you can pick one

5:14 of these avatars here or you can create a custom one.

5:17 For now, it does not have the ability to upload your own picture.

5:19 Just a question of time though.

5:21 The voice you select makes a difference in how viewers respond to your video.

5:24 Hey, what's up, YouTube?

5:25 Hey, look at that.

5:26 That's pretty good.

5:27 You can now easily add other videos in here, AI-generated or not.

5:30 And as we're talking about avatars,

5:31 there's two more releases this week that are actually interesting.

5:34 One of them is Pika Labs coming out with a product that is meant to give

5:38 AI agents a face so that they can join your Zoom calls and talk to you.

5:43 It's a bit weird.

5:43 This thing is very much in beta though, but if you want to try it,

5:46 they did open source it, but it only works for the PK API, of course.

5:50 I think I might actually try this one over the next week,

5:51 but the one that we'll try right now is Runway Characters.

5:55 All right, I got the updated app.

5:56 I'm on a free plan, by the way, and yep, I'm going to talk to the cat.

5:59 Ah, you're here.

6:02 [laughter] I was in the middle of something.

6:04 What's up, Mochi?

6:05 What do you want to know about me?

6:07 Okay, tell me your deepest and darkest secret.

6:10 A secret, you say?

6:11 The deepest and darkest?

6:14 Fine.

6:15 I once batted that glass off the counter just to watch it shatter.

6:18 The sound was quite satisfying.

6:21 Don't tell anyone though.

6:22 Now, All right, Mochi, what else can you do for me?

6:24 I can do all sorts of things.

6:27 I could tell you a story, critique your I don't know, man.

6:30 What am I supposed to say?

6:31 It's a talking cat with a female voice.

6:34 You try for free though, it's kind of fun.

6:36 Let's see what's next.

6:37 Okay, then next we have a quick look at SeaArt 2.0.

6:39 This one went viral when it came out.

6:41 Video generation model, it was really almost impossible to access.

6:44 Now it's available on major platforms and we

6:46 went ahead and looked at some examples.

6:49 And on our typical benchmarking problems, I always enjoy looking at these.

6:52 Let's start with the sloth on a floaty.

6:56 Okay, that's really good with the camera movement and everything.

6:59 This one is always tricky for it cuz there's a lot [music] of touching water.

7:02 Oh my god, so much camera movement, but extremely good.

7:04 What about the ghost?

7:06 Yeah, also accurate.

7:06 I mean, it's just clearly visible

7:08 from these benchmarks why this got hyped so much.

7:10 Okay, that hand was weird.

7:11 Overall, I get it.

7:13 If you want the highest quality video model,

7:15 it's probably the one and it also lives

7:16 at the top of various AI video leaderboards right now.

7:19 Next up is an interesting one and not

7:22 something that you might necessarily use immediately,

7:24 but I do want to talk about it because it got

7:26 so much attention on Twitter and it's a very interesting idea.

7:29 It's Andrej Karpathy's LLM Wiki idea and now

7:32 there's also tools around it that came out.

7:34 And here's the base idea.

7:35 He just tweeted that, "Hey,

7:36 I'm basically using LLMs to build personal knowledge bases

7:40 for various topics of research interest." In other words,

7:43 he just creates text files surrounding certain topics and then

7:46 he stores them and his agent searches over them.

7:49 It's like a personal wiki.

7:50 There's a bunch of details we could talk about here,

7:52 but where this got interesting was people testing

7:54 this versus traditional knowledge retrieval workflows like RAG.

7:59 And what some tests, and now this is not super verified,

8:02 but some tests found that it was 70 times more efficient than the RAG workflows.

8:06 Now, look, quick primer if all of these means nothing to you.

8:08 RAG was basically the workflow of the past

8:10 few years to deal with many, many files.

8:13 And this is a different approach than that workflow.

8:15 With RAG, you would store it in a vector database and there was

8:17 a whole process of putting stuff in there and then you had to retrieve it.

8:20 That's what the R stands for in RAG.

8:22 But now he's just saying, "Hey, just put it in text files,

8:24 build your own wiki, and let the agent see all of it.

8:27 No need to store it away." Now,

8:28 this has actually been the approach that tools like

8:31 Claude Code have been using over the past year.

8:33 Claude Code doesn't use RAG.

8:34 It just stores a bunch of text files and retrieves over them,

8:37 but he used this to build his own

8:39 research library and people really resonated with this idea.

8:42 Now, they resonated with it so much that it has 19 million views on Twitter

8:46 and that people built out various versions where you can start using it too.

8:49 All of it is in the super early and super

8:51 experimental phase and this repository is one of them,

8:54 but I'm featuring it here not because I

8:56 want you to start using this, just to show

8:58 you that these new ideas and approaches are

9:00 being generated in real time by real users.

9:03 And actually, for some of the more advanced users here,

9:04 I just want to briefly show you that I

9:06 really got into Obsidian recently together with my Open Claude.

9:09 I built out this entire homepage for all

9:10 the education that I do and it's incredible.

9:13 It's basically my agent's playground where it uses all

9:15 of this as context and then creates new files in here.

9:18 But basically, this is my own wiki on my own

9:21 life and educational style and all the different patterns

9:23 that I follow so that my agent can be aware

9:25 of it too and we just collaborate in one place.

9:28 Now, this is different from what Andrej Karpathy is sharing.

9:30 He's talking about a knowledge base like this to just for research

9:34 on a certain topic and this is really operating system for my work,

9:37 but I thought I'd share both of those with you

9:39 and we can dive deeper in the future if there's interest.

9:42 One more big story this week and this one will be pretty short because

9:45 it's state-of-the-art large language model from China

9:48 called GLM-5.1 and this is a big boy.

9:50 This one you won't be able to run on any MacBook.

9:52 Heck, even the big Mac minis have trouble with this one.

9:55 It has 754 billion parameters,

9:57 but on many benchmarks, especially relating to coding,

10:00 it almost matches Opus 4.6, the current go-to model for many people.

10:04 And the main thing that they market around this and that has been

10:07 communicated is that it can run for up to 8 hours on agentic tasks.

10:11 And mind you, this model is fully open source, fully freely available now.

10:15 You just need a machine or rented GPUs to run it.

10:18 Insane, especially following last week's Gamma

10:20 4 announcement from Google where they released

10:22 their own version that is actually small enough to run on some phones.

10:26 It's kind of the opposite of this.

10:27 This is the big boy, Gamma 4 is kind of the small version.

10:30 Both of them open source,

10:31 way more capable than anything we've seen before at that size.

10:34 So, yeah, just AI moving forward.

10:36 Most consumers just keep using the web apps,

10:38 that's where the best models are at.

10:39 One quick follow-up to Suno 5.5 last week,

10:42 team member Daniel actually got so into it this week that I just have to report

10:45 back that according to him it's actually way

10:47 better than what we talked about last week.

10:50 We made a bunch of AI music and used jingles that you'll

10:52 from here now to hear at the end of these videos.

10:54 You guys really really liked last week's one, so keep an ear out for that one.

10:58 OH, BROTHER, THIS GUY STINKS.

11:02 OKAY, that line makes me cringe on the inside, but we're keeping it in.

11:05 And with Microsoft releasing free in-house models,

11:07 this is their first move that directly competes with OpenAI,

11:10 not kind of, but like directly.

11:11 They have a transcription model, a voice model,

11:13 and an image model called MAI, with MAI representing Microsoft AI.

11:18 And then finally, with new Claude connectors, uh notoriously unreliable,

11:21 but now you can also connect to your Microsoft 365 account.

11:26 If you're using that over Google Docs, Google Sheets,

11:28 you can connect it to your Word documents if they live in the cloud, too, now.

11:32 Okay, that's pretty much everything we have for this week.

11:34 A lot of variety, and as per usual,

11:36 I would love to hear which one of those stories

11:37 would you want me to go deeper into.

11:39 I read the comments every week,

11:40 so leave one or upvote another one that represents your opinion,

11:43 and we can go in that direction.

11:44 All right, I hope this was helpful or inspirative.

11:46 My name is Igor, and I'll see you very soon.

11:48 AI news you can use,

11:51 breaking [music and singing] down the stories you actually care about.

11:55 AI news you can use, [music] Igor's here to help [singing] you figure it out.