# Thinking Machines Launches AI Model, AWS Invests $1B

An episode of [AI Chat: AI News on Claude, ChatGPT, Gemini, Grok, Perplexity](https://invisiblenetwork.ai/podcast/1669990482), 17 Jul 2026, 16 min.

Audio: https://content.rss.com/episodes/365073/2997967/ai-chat-podcast/2026_07_17_00_19_52_82dfcb5d-4f51-4c65-b426-a4da63bb7782.mp3

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! In this episode, we explore major developments in the AI landscape, including Thinking Machines' launch of the open weight model Inkling and OpenAI's creation of GPT-RED, an AI hacker designed to enhance security. Additionally, we discuss AWS's $1 billion investment in engineering teams for custom AI solutions, Apple’s partnership with Alibaba for AI in China, and Meta's strategy to monetize its AI computing capacity. Chapters 00:00 Introduction to Inkling 01:01 OpenAI's GPT-RED 01:59 AI Box's MCP Server 05:00 AWS's Custom AI Engineers 07:29 Apple's AI Launch in China 10:01 Meta's AI Compute Strategy Show Links • Get the top 80+ AI Models for $8.99 at AI Box: https://aibox.ai • How I Grow and Scale My Business with AI: https://www.skool.com/aihustle • Get the AI Chat Daily Newsletter: https://www.aichatdaily.com/newsletter

## Transcript

[0:02] Welcome to the AI Chat Podcast. I'm your host,
[0:04] Jaden Schaefer. Every day I cover cutting -edge
[0:07] AI news and talk with the leaders behind it,
[0:09] breaking down what it means for your life and
[0:11] business. Thinking Machines by Miriam Marotti,
[0:17] who famously left OpenAI after she was the CEO
[0:20] when Sam Altman got kicked out, have released
[0:22] their first open -weight AI model called Inkling.
[0:25] OpenAI has built something called GPT -RED. It's
[0:28] an LLM super hacker, and they built this to harden
[0:31] their own models. AI Box has shipped an MPC server
[0:34] that brings 80 different AI models into Cloud,
[0:37] ChatGPT, and Gemini. AWS is committing $1 billion
[0:41] to a forward -deployed engineering organization.
[0:45] OpenAI and Anthropic are both scaling theirs.
[0:47] Apple Intelligence has been cleared for a China
[0:50] launch with Alibaba's Quen and Meta plans to
[0:53] sell and resell their AI compute following what
[0:57] SpaceX is doing and kind of copying their playbook.
[1:00] Let's kick this off with Thinking Machines. This
[1:02] is a company that I'm really rooting for. They've
[1:04] raised, you know, over a billion dollars, but
[1:06] they haven't come out with anything super groundbreaking.
[1:08] But what I will say when it comes to a lot of
[1:10] these models, they're so expensive and they take
[1:13] so much time that when you think of something
[1:15] like even Anthropic, it felt like OpenAI had
[1:17] just run away from them and they were never going
[1:19] to catch up. And little by little, they pick
[1:21] their lane, they make their product better, and
[1:23] they're able to get market share until the point
[1:25] where Anthropic has now exploded and is making
[1:27] more money than OpenAI. It surpassed them in
[1:29] revenue. I think we might see a lot of that same
[1:32] strategy played out by a lot of different AI
[1:34] companies if they don't kind of shrivel up and
[1:36] die or get sold off for parts or, you know, an
[1:39] acqui -hire or something like that. And so Thinking
[1:40] Machines is one of these companies that I'm really
[1:42] excited about. I hope that they really go places.
[1:45] But this is run by Miriam Roddy, who is famously
[1:47] the CEO of OpenAI when Sam Altman left. And she
[1:51] has created this thing called Inkling. It's an
[1:53] open weight AI model. It has 975 billion parameters.
[1:57] And companies can download this and customize
[1:59] it themselves. So you don't have to go and pay
[2:02] for API access to OpenAI or... Anthropic, you
[2:05] can actually just go and fine tune this all on
[2:07] your own. You can fine tune it on your own data.
[2:09] And the bet that they want people to take is
[2:11] that that is going to be better than a one size
[2:14] fits all model that OpenAI or Anthropic or Gemini
[2:17] or any of these other big labs are selling. Inkling
[2:20] was trained on 45 trillion tokens of text, image,
[2:24] and audio and video. They did it in about nine
[2:26] months, which is way faster than OpenAI's five
[2:29] -year timeline or Anthropic's three years. In
[2:32] a test with Bridgewater Associates, The financial
[2:34] model trained on the hedge fund's own expertise
[2:37] scored 84 .7 % on financial reasoning benchmarks
[2:41] at roughly 1 14th the cost of the top models
[2:45] from Anthropic and OpenAI. So we're seeing some
[2:48] massive improvements when you're taking this
[2:50] kind of these kind of models and you're fine
[2:51] tuning it on your own data. The model activates
[2:54] only 41 billion of its 975 billion parameters
[2:57] per task, and users can dial up thinking efforts
[2:59] to trade speed for accuracy. So I'm rooting for
[3:02] them, but time will tell how well this model
[3:04] does. We just learned that OpenAI built something
[3:06] called GPT -RED. This is an AI model trained
[3:09] to attack their own systems, and they use it
[3:11] to catch security flaws before release. So the
[3:14] model cuts successful attacks on GPT -5 .6 from
[3:18] over 90 % down to 23%, which basically makes
[3:21] it one of OpenAI's most secure releases yet.
[3:24] One particularly interesting attack that it was
[3:27] able to discover is called a novel fake chain
[3:30] of thought, which basically is tricking a model
[3:32] into making up fake reasoning steps. So it's
[3:35] kind of similar to convincing someone that, you
[3:37] know, one plus one equals three. And then, you
[3:40] know, saying that I already checked the math,
[3:42] that's what equals and if it equals that, then
[3:44] therefore and you, you know, go trick them on
[3:45] the next thing. So they found that when they
[3:47] gave the same task as a human red teamer who
[3:50] tested GPT -5, in 2025, GPT -RED found more effective
[3:55] attacks than the humans. They also got it to
[3:57] go and hack like this third party vending machine
[4:00] agent, which is kind of funny. It does have a
[4:02] bunch of limitations, so it's not very good at
[4:04] back and forth conversation attacks, and it's
[4:06] not very good at exploiting images that are embedded
[4:09] into malicious text. The idea behind this is
[4:12] actually functionally working is that GPT -RED
[4:15] is automating how all of the security holes are
[4:17] found because it puts an attacker model against
[4:20] a defender. model, and they put them in these
[4:22] kind of simulated real world environments, and
[4:23] they're battling it out. OpenAI is not releasing
[4:26] this externally. So they're not giving this to
[4:27] other people to test, which is interesting, right?
[4:29] Because we had the whole anthropic mythos model,
[4:31] which was really good at security exploits, and
[4:33] they gave it out to all the top labs and said,
[4:35] hey, like, go harden all your security with this.
[4:37] So OpenAI is not giving this out externally beyond
[4:39] just using it for themselves. And I mean, it's
[4:42] got a huge massive drop in the success rate of
[4:45] a lot of these exploits. I think it shows AI
[4:47] powered red teaming can be just as effective
[4:49] or more effective than humans. I mean, you can
[4:52] just brute force way more tests than humans.
[4:55] And I mean, it's trained off of what humans are
[4:56] doing. But at this point, it's doing better than
[4:58] what a lot of the humans are doing. AI box, which
[5:01] full disclosure is my own startup has just released
[5:03] an MCP server and we are allowing people to plug
[5:07] 80 different AI models straight into cloud chat,
[5:09] GPT, Gemini or cursor. Basically, it lets you
[5:12] use any model inside of whatever assistant you
[5:15] already are used to using. Personally, I use
[5:17] clawed all day long, although I'm kind of switching
[5:19] over to ChatGPT with their new ChatGPT app. It's
[5:21] really amazing. But both either way, you get
[5:24] the AI Box MCP and it allows you to access all
[5:27] of the images that... OpenAI can generate. It
[5:30] allows you, if you're on Claude or ChatGPT, to
[5:32] generate all of the videos that Google V03 can
[5:34] make. And no matter what platform you're using,
[5:36] you can access what Eleven Labs can do with audio.
[5:39] I spent basically the entire day today getting
[5:42] ready for a big Facebook campaign that we're
[5:44] launching here at AI Box. And in the past, doing
[5:46] Facebook campaigns usually meant hiring a Facebook
[5:49] ads team. It meant creating tons of creative,
[5:52] which just takes a lot of time. And it meant,
[5:54] you know, spending a ton of time on landing pages.
[5:57] With Claude, and with the AI Box MCP that we
[6:00] built into it, I actually got almost all of the,
[6:02] so basically what I did is I went and recorded
[6:04] a bunch of just selfies of me talking about the
[6:06] product on my phone. I dropped those into a folder
[6:09] and I had Claude go and edit all of them. And
[6:11] then I was able to say, hey, I need to create
[6:13] a bunch of dynamic, so like image generated ads.
[6:15] I gave it the style that I wanted. I have these
[6:17] different ideas. Some of them, it looks like
[6:19] kind of like breaking news images. Some of them,
[6:21] they actually look like, it's like me doing a
[6:23] FaceTime call and there's like a text message
[6:25] popping up on the screen and it's some sort.
[6:27] of text message about AI box. So anyways, I had
[6:29] all these different creative ideas that I went
[6:31] and got from a bunch of people and to go and,
[6:33] you know, create a template on Canva and then
[6:35] switch out tons of different variations of the
[6:36] title and the backgrounds and stuff like that
[6:39] takes a lot of time. And before Claude couldn't
[6:41] do it because it couldn't do any of the image
[6:43] generation. But now that AI box is embedded inside
[6:45] of Claude and it can do the image generation.
[6:47] One in particular that I did is I was showing,
[6:49] uh, I was showing the ability to generate images
[6:51] inside of Claude and I needed that generated
[6:53] as an ad. So you had to have like basically,
[6:56] uh, the design of a Claude interface. And then
[6:59] you had to be able to show images generated and,
[7:01] you know, have like AI box logos and stuff. So
[7:04] what I did is I was like, hey, go create this
[7:06] mockup inside of Claude. But for the image part,
[7:10] just go use AI box to generate the image and
[7:12] pull that inside of, it's like the image inside
[7:14] of the image that it's generating. Is able to
[7:16] do that in a couple seconds. And I said, this
[7:18] is awesome. Go think of like 10 variations or
[7:20] like 10 different use cases of my product. And,
[7:23] you know, and in this case, it's the product
[7:25] is. showing that you can create images. This
[7:27] is kind of a meta example, I know. But I was
[7:30] like, go think of like 10 different variations.
[7:31] So it's like, okay, well, you could use it for
[7:32] like UGC content, you could use it for like,
[7:35] coming up with product images for logos for newsletter
[7:37] banners, whatever came up with all these ideas.
[7:39] And inside of the image ad that it's creating,
[7:42] it would go to AI box anytime it had to generate
[7:44] one of the actual graphics inside of its image.
[7:47] So like the you know, the graphic of the logo
[7:50] or the graphic of the UGC person, whatever the
[7:53] actual image was, it would hit AI box, it would
[7:55] generate the image, it would pull inside of its
[7:57] own thing. So I didn't have to do anything. I
[7:59] literally just said, go, you know, we have the
[8:01] concept, we have the template that I like, go
[8:03] create 20 variations, think of a bunch of good
[8:05] use cases and use AI box for all the images.
[8:08] And the cool thing is when when Claude was making
[8:10] this specific kind of like UI mockup. Um, because
[8:14] it's UI and it's not just like I went to chat
[8:16] GBT and got it to generate like one singular
[8:18] image, which isn't perfect. And the text might
[8:21] be funky and someone might be a little bit off
[8:23] or whatever, right? If you're just doing an image
[8:25] because Claude was kind of, uh, just creating
[8:27] it with code and then pulling an image into it.
[8:30] I said, okay, this is awesome. Now go put this
[8:32] actual graphic onto our website and animate it.
[8:34] So like where the chat bubble is, I'm like, have
[8:36] someone typing that out, have the bubble appear,
[8:38] have a loading screen that makes the image pop
[8:40] up. So it's so cool because Claude is able. to
[8:43] build these things. And because we're just pulling
[8:45] the image in with AI Box, it's not just a picture,
[8:49] but it's a full element that can be animated
[8:51] for the landing page. Then it can be turned into
[8:53] an ad for the Facebook ad and you can actually
[8:56] have the animated as a video ad. Anyways, so
[8:58] many possibilities. But if you're doing anything
[9:00] with Cloud, which doesn't have audio, video or
[9:03] image capabilities, just go get the AI Box MCP.
[9:06] It's like $7 .99 a month, I think, or $8 .99
[9:09] a month. And you get access to over 80 different...
[9:12] models and anything you need while you're using
[9:14] Claude, it can go and grab that and pull it in
[9:17] and save you so much time. No more back and forth.
[9:19] So go check it out. There's a link in the description
[9:21] to AI box dot AI slash MCP. If you want to get
[9:25] started with that. AWS is committing $1 billion
[9:28] to a new team of engineers who are going to be
[9:31] embedded inside of customers or in their customers'
[9:34] companies. And they're going to be building custom
[9:36] AI agents tailored to each business. And they're
[9:38] then going to hand off the working system when
[9:40] the project ends. There's been a ton of these
[9:42] companies being spun up by Anthropic and OpenAI.
[9:46] They're basically making these forward -deployed
[9:49] engineers. So you send the engineer into your
[9:51] customer's company. They build some sort of automation
[9:53] workflow. And I think it's kind of like OpenAI.
[9:55] And Anthropic's way of saying, hey, look, we
[9:57] know you guys want AI, but you also don't know
[9:59] how to use it. We are the experts. We'll just
[10:01] send some people in there to build this thing
[10:03] for you. They're directly copying the playbook
[10:05] that OpenAI spent $4 billion on and Anthropic
[10:08] spent $1 .5 billion on so far. And in those cases,
[10:14] OpenAI and Anthropic both partnered with some
[10:15] private equity firms to actually fund the staff
[10:18] for all of this FDE teams. Anthropic is doing
[10:22] it entirely with internal Amazon resources. instead.
[10:25] And I'm going to be honest, this might be an
[10:27] interesting strategy. If I was any of these big
[10:30] AI companies, and you know, like, let's say meta,
[10:33] for example, and you're thinking about doing
[10:34] layoffs, well, maybe instead of meta doing layoffs,
[10:37] they should build one of these forward deployed
[10:40] engineer teams, get a bunch of money and go and
[10:43] get, you know, meta embedded into their customers
[10:46] businesses and just deploy the engineers over
[10:49] there. Anyways, I'm not sure if this is Amazon
[10:50] getting around layoffs, but it is an interesting
[10:52] strategy, you can imagine. because OpenAI and
[10:54] Anthropic definitely didn't have any extra engineers
[10:57] they had to partner with people for this. AWS
[10:59] has a really big advantage, I would say structurally,
[11:02] because when the project ends, the customer runs
[11:05] AI agents on AWS's cloud, and that's basically
[11:08] just creating long -term revenue. OpenAI and
[11:09] Anthropic make money per token used. That's a
[11:12] lot smaller of a payoff per engagement, but the
[11:15] cloud is a really big win for AWS. Also, I say
[11:19] all of this like AWS is copying OpenAI and Anthropic
[11:22] like they pioneered. this. But this is actually
[11:24] something that Palantir has spent the last 20
[11:26] years doing. And so I don't think this is, you
[11:29] know, something that's just brand new in the
[11:30] situation. It's something that's been done in
[11:32] the past. And it's it's kind of something that's
[11:34] been successful. We'll see how big Amazon is
[11:37] able to scale this. Apple has just won regulatory
[11:40] approval to launch Apple Intelligence in China.
[11:42] They're partnering with Alibaba to power all
[11:44] of the AI features there. It's interesting, right?
[11:46] Because Apple said, hey, look, like they were
[11:48] going to have their own AI in Syria. And yeah,
[11:49] that probably would have been a big pain for
[11:51] them with China. But then they said, look, we're
[11:53] just going to say anyone with an iPhone can go
[11:57] and use whatever AI model they choose. So in
[12:00] China, that's going to be something from Alibaba.
[12:03] And in America, that's probably going to be Anthropic
[12:05] or OpenAI or whatever the model that you probably
[12:07] pay for that you have. premium of you could probably
[12:09] plug that straight into siri i think this matters
[12:12] a lot because china is apple's second largest
[12:14] market 20 .5 billion dollars in sales last quarter
[12:18] and i think being able to close that gap helps
[12:21] keep apple their number two smartphone position
[12:23] there's rivals like huawei alibaba's quen model
[12:27] is going to handle text and image understanding
[12:29] and generation across all of the different you
[12:31] know ipad and mac and vision os and apple's exploring
[12:35] a deal with baidu deep seek and ByteDance. I
[12:38] think they finally have settled on Alibaba after
[12:41] kind of talking to all of the other options.
[12:44] Alibaba's US listed shares went up 6 % when that
[12:47] was announced. This deal is basically going to
[12:49] give Quen some built -in distribution to the
[12:51] very new iPhone that will be sold in China. I'm
[12:54] going to be honest, I've tried some Quen models
[12:56] and been really impressed. Their text -to -speech
[12:58] model in particular is really good. But kind
[13:01] of the bigger story for me on all of this is
[13:03] that for not just Apple, but for any of the Western
[13:07] companies, China is not going to approve any
[13:10] sort of Western AI service if there isn't a local
[13:12] model partner that's locked in and that is powering
[13:15] everything. Meta is launching a cloud business
[13:18] called Meta Compute. They're going to be selling
[13:19] AI compute powered AI models to other companies,
[13:22] and they're doing this directly to compete with
[13:24] Amazon, Google, and Microsoft. And all this is
[13:27] right after Meta committed a whopping $182 .9
[13:30] billion to AI infrastructure, which is basically
[13:34] the SpaceX strategy. They're turning all of their
[13:36] excess data center capacity into revenue. It's
[13:38] interesting, right? Because Meta, I think, hopes
[13:40] that or would have hoped that their AI model
[13:42] was way more used and way more popular. But having
[13:45] all All of the extra AI infrastructure is, I
[13:47] mean, basically a goldmine. It's money that they're
[13:49] not spending and now money that they're actually
[13:51] making, similar to Grok, I think. Meta is going
[13:54] to sell both raw computing capacity, basically
[13:57] just like CoreWeave does, and they're also going
[13:59] to have access to their own AI models that they'll
[14:01] sell, including the recently launched Muse Spark
[14:03] model, which has gotten a lot of attention because
[14:05] it kind of finally pushes Meta to the forefront.
[14:08] The Ohio data center that is described by Mark
[14:12] as a Manhattan -sized project. is expected to
[14:15] open this year, and it's going to provide all
[14:17] of the excess capacity that Meta needs to resell.
[14:20] And SpaceX basically showed this model works.
[14:22] XAI signed deals in May with Anthropic, Google,
[14:25] and Reflection AI to buy Compute Time, and they
[14:28] basically turned all of their idle capacity into
[14:30] immediate revenue, and people paid a lot of money.
[14:33] I think Anthropic is paying over a billion dollars
[14:34] a month for access to that. For companies with
[14:37] deep pockets, this seems to be a good strategy.
[14:39] If at any point Meta's AI models get super, super
[14:41] popular, right, they can go and use their own.
[14:44] capacity. But if not, they're just going to make
[14:46] all of the money from all of the other AI companies.
[14:49] And I think at the end of the day, the demand
[14:50] for AI is not going to decrease. So it's a pretty
[14:53] smart bet. Guys, that was everything for the
[14:55] podcast today. Thank you so much for tuning in.
[14:58] If you enjoyed today's episode, make sure to
[14:59] go check out AI box .ai. Like I mentioned, where
[15:02] you can get our MCP that gives you 80 different
[15:05] AI models. And more importantly, if you're using
[15:06] Claude, you get images, audio, and video generated
[15:09] right inside of Claude. It's like actually magical
[15:12] to use. I've been using it all day. It takes
[15:13] two seconds to connect and it's only $8. Please,
[15:16] I beg you, if you have Claude, you have to try
[15:18] this out. It will just make your life so much
[15:20] easier. Also, if you want to get all of these
[15:23] different stories that I talk about on the podcast
[15:24] here straight into your inbox, go check out AIChatDaily
[15:28] .com. That's my website, my news site that is
[15:30] tied to this. I have deep dive articles on every
[15:32] story that I covered here. And I also have links
[15:35] over there where you can go and read more about
[15:37] it. There's a subscribe tab at the top of that
[15:39] website where you can get this all as an email
[15:41] or you can go get the deep dives on the site.
[15:43] Guys, thank you so much for tuning in and I will
[15:45] catch you in the next episode.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.