Resources:
Join the discussion on LinkedIn: Got something to say? Let us know on LinkedIn and network with other AI leaders
Upcoming Episodes: Check out the upcoming Everyday AI Livestream lineup
Connect with Jordan Wilson: LinkedIn Profile
Start Here Series in our Inner Circle Community: Join for free access
AI Developments with Direct Impact: Updates in Models, Security, and Productivity
This week, a cluster of pinpoint AI events delivered strategic developments for organizations monitoring AI adoption and risk. From advances in generative models, to enhanced enterprise features, major model leaks, and evolving global security threats—each offers actionable intelligence for decision makers responsible for digital transformation and organizational resilience.
OpenAI GPT-5.5 and Images 2: Concrete Improvements in Productivity and Visual Output
OpenAI’s new GPT-5.5 model is now available both inside ChatGPT and through its API, positioning itself as the company's most capable general-purpose model. GPT-5.5 matches the per-token latency of GPT-4 but delivers superior performance on coding, research, tool use, and multi-step tasks with high ambiguity resolution. Notably, it ties or beats human experts on the current GDP benchmarks 85% of the time, showing pronounced improvements for coding, computer operation, and knowledge work.
On the visual AI front, OpenAI’s Images 2 model brings sharper media generation to both free and paid users, producing high-resolution, compositionally consistent images with reliable typography and multilingual support. With the ability to generate up to eight coherent images per prompt, Images 2 supports presentations, infographics, and product visuals at quality levels that reduce or eliminate manual corrections previously required for text or layout errors. This efficiency accelerates creative workflows and internal or client-facing deliverables where speed and accuracy are critical.
Cloud-Based Agents: Bringing AI Automation to Teams
Workspace agents now operate as shared, cloud-running entities within ChatGPT’s team and enterprise plans. These agents automate long-running workflows, orchestrate processes across platforms like Slack, and facilitate permissions and approval flows within organizations. Unlike traditional approaches where agents needed to run on individual machines, these cloud-based agents operate independently, supporting asynchronous work and operational continuity.
Availability is currently limited to business and enterprise plans, and they are included at no extra cost until an upcoming May 6 deadline, after which usage-based pricing or limits are expected. Organizations with eligible plans should prioritize pilot deployments to assess time and resource savings, especially as these tools enable less technical staff to harness AI-driven automation.
Chronicle and Codex Upgrades: Meeting the Demand for Context-Aware Assistance
The Codex ecosystem received notable updates, including the Chronicle feature—an opt-in tool that records on-screen history, designed for research preview. This is akin to controversial "recall" systems but emphasizes enterprise adoption with clear user consent. Enhanced browser control, new Google Sheets and Slides integrations, and system-wide dictation features now make Codex a robust utility for digital operations, increasing operational coverage for knowledge workers and analysts across platforms.
Enterprise Security: Model Distillation Threats and Federal Warnings
A White House memo warns of “industrial scale” AI model theft by entities sourcing from US firms. Distillation—where outputs from advanced models are used to train less capable ones—enables international actors to create imitator models without embedded safety and security features. Federal intelligence sharing with US tech companies and the active pursuit of technical defenses heighten the regulatory pressures on organizations evaluating AI model sourcing. For procurement, the provenance of foundation models directly ties to compliance and risk management, particularly as federal policy around export, use, and cloud compute is evolving.
Google Gemini Across Workspace: Contextual AI Comes to Productivity Suites
Google’s latest Gemini rollout brings intelligence directly into Workspace apps, including Gmail, Docs, Sheets, and Slides. By leveraging stored work context—like previous edits, corporate template preferences, and historical communications—Gemini produces context-aware drafts, summary overviews, and polished outputs that reflect brand standards and reduce the manual overhead of document formatting and context switching.
The new AI Inbox and AI Overviews in Gmail enable users to quickly prioritize and act on important messages, while agentic orchestration within the Google ecosystem shrinks workflow friction and speeds up information retrieval. Early access phases are underway; business leaders should monitor rollout schedules and develop adoption roadmaps for maximum value extraction.
Vertex AI Becomes Gemini Enterprise Agent Platform
At Google Cloud Next, Vertex AI was rebranded as the Gemini Enterprise Agent Platform, now focused on broadening accessibility for non-technical staff via the Agent Studio and a robust Agent Development Kit (ADK). Support for persistent memory, multi-agent orchestration, and security protocols—including cryptographic agent IDs and centralized registries—align the platform with enterprise governance requirements. Over 200 models are now accessible, including Gemini 3.1 Pro, Lyria for music, and Gemma for advanced applications. This shift positions Gemini as a versatile and compliant layer for organizations integrating agentic AI into core operations while maintaining security and auditability.
Security Incident: Mythos Leak and Implications for Model Integrity
Anthropic’s top-secret Mythos model, described as capable of identifying zero-day vulnerabilities across all major operating systems and browsers, was recently accessed by a private Discord group after they identified its location by extrapolating from naming conventions. The access was possible due to a contractor’s privileged credentials, not a technical hack.
Claims from the group include utilization for benign coding tasks and reference to additional unreleased models. While Anthropic is investigating the scope, initial evidence points to the importance of vendor management and strict credential hygiene in third-party relationships—especially when pre-release models are involved. Benchmark comparisons after GPT-5.5’s release further question the uniquely high-risk claims around Mythos, suggesting a blend of marketing positioning and actual capability.
Additional Developments: Copilot Expansion, AI-Powered Productivity, and Market Movements
Microsoft Copilot’s agentic capabilities are now live in Word, Excel, and PowerPoint, supporting greater automation in legacy productivity suites and promising further user-driven time savings.
Anthropic and DeepSeek both reported updates to their LLMs, though DeepSeek’s V4 model underperformed against expectations and frontier benchmarks.
Mergers and investments in the AI landscape are accelerating (e.g., Google’s $40B+ commitment to Anthropic and Cohere’s possible merger with Aleph Alpha), highlighting intense competition and ensuring infrastructure improvements.
Reports indicate worsening uptime for some large AI providers, underscoring the need for redundancy strategies and careful vendor selection for mission-critical workflows.
Enterprise providers such as Meta are trimming staff in response to AI compute expenses, while also reportedly gathering employee interaction data to train internal AI systems—raising privacy and ethical considerations for those partnering or benchmarking with these technologies.
Strategic Takeaways for Digital Transformation and Risk
The past week’s developments exhibit tangible upgrades in generative model capability and integration, cloud-based automation, cross-suite AI intelligence, and intensifying concerns around security, provenance, and ethical use. Pragmatic organizations can accelerate internal adoption, safeguard proprietary data, and remain compliant by deploying new capabilities within trial windows, monitoring evolving vendor agreements, and including these fast-moving variables in regular strategy reviews.
Prioritizing pilots for new enterprise AI features, performing diligent risk assessments of model sources, and creating flexible infrastructure plans to accommodate provider-led changes will create direct operational and strategic value as each new release lands.
Topics Covered in This Episode:
- OpenAI GPT-5.5 Model Launch Recap
- Images 2: OpenAI’s Flagship Image Model
- ChatGPT Workspace Agents for Teams
- Codex Major Feature Upgrades Overview
- White House Accuses China of AI Model Theft
- Google Gemini Workspace Intelligence Release
- Google’s Gemini Enterprise Agent Platform Update
- Anthropic’s $40B Google and Amazon Funding
- Meta Layoffs Over AI Investment Surge
- Anthropic Mythos Model Leak on Discord
- Microsoft Copilot Agent Mode in Office Apps
- SpaceX AI Startup Acquisitions and Rumors
- Adobe Unveils AI Agents for Business
- DeepSeek V4 Open Source Model Release
- Cohere and Aleph Alpha $20B Merger News
- OpenAI and Microsoft AI Cybersecurity Partnership
- Meta Multi-Year AWS Graviton AI Deal
Episode Transcript
Jordan Wilson [00:00:17]:
OpenAI shipped about three months of updates over the course of three days. Google held its Cloud Next conference and finally rolled out the version of Gemini many have wanted for years, and the White House has accused China of industrial scale AI theft from US companies. All of these are pretty big AI developments, but the biggest story in AI news this week might have been the top secret anthropic mythos model that they said was too much of a cyber threat to release to the public, yet some randoms on a Discord server got access just by guessing the URL and went undetected for weeks. Yeah. There's a lot that happened this week in AI and when making decisions on how your company approaches and uses AI, you absolutely have to keep up to date with the latest happenings, news, and releases. But to do that, it's gonna take you hours every single day, unless you just tune in with us on Mondays as we bring you the AI news that matters. Let's get into it. If you're new here, welcome.
Jordan Wilson [00:01:24]:
My name is Jordan Wilson, and this is Everyday AI. It's a daily live stream podcast and free daily newsletter helping everyday business leaders like you and me keep up with the nonstop, updates that are happening from big tech. I tell you what's important, what's not, and how to use it to grow your company and career. So if that's what you're doing, make sure to go to youreverydayai.com. We're gonna be recapping both all of the news from today's story as well as everything else happening in the AI world. So if you are FYI, if you're listening to this today and you happen to be in San Francisco or, you know, Silicon Valley, make sure to hit me up. I'm gonna be out, in San Francisco for a couple of days at the Sage Future Future Conference. So I do in the show notes.
Jordan Wilson [00:02:07]:
I always have my, link to my LinkedIn. So go ahead and send me a message. Let me know where you're gonna be. Alright. Let's get into it. Yeah. If you are new here on Mondays, we do the AI news that matters. On Wednesdays, we usually go hands on and do more of a demo with a big release from that week.
Jordan Wilson [00:02:23]:
On Fridays, we do AI feature Fridays highlighting some of the, you know, AI features that roll out to large language models. And then Tuesdays and Thursdays, we kind of rotate the show. So, that's kind of our lineup for the week. But let's get started because OpenAI, I mean, they had their biggest week undoubtedly since, December 2024. That's when they had their kind of ship miss event where they, for the first time, brought out reasoning models, Sora, all of these other things. But OpenAI absolutely dominated this week, and it was not even close. Alright. So here's just a preview of what they did.
Jordan Wilson [00:03:01]:
So, this week included, well, launching the world's new top ranked model in GPT 5.5, a much stronger image model, also best in the world with images too. And what I think a lot of people have been asking for for a long time, agents that work inside of chat, GPT, and can access your company's data. And they had a bunch of codex updates as well. Alright. Let's start at the top. I'm gonna go through these pretty quickly, because if I'm being honest, I could probably talk for at least thirty minutes about just these updates. But we did go over these in a little more depth when we covered our Friday features, and that was episode seven sixty three from Friday. So this will be a faster version, although we could talk the entire episode.
Jordan Wilson [00:03:49]:
Alright. So let's talk about GPT 5.5. So it is now live both inside chat GPT, and it is available on the API side as well for builders. It is live in position as OpenAI's most capable general purpose model. And if you haven't used it, I suggest you do. It's really good. Alright. So it's, matches g b t five fours per token latency while scoring higher on a benchmark.
Jordan Wilson [00:04:16]:
So it's as fast but smarter. OpenAI says it excels at coding, tool use, research, and handling messy multi step tasks with better ambiguity resolution. GPD 5.5 also reportedly ties or beats human experts on, the g b GDP valve, benchmarks. So about 85% of the time, which is kinda nuts to think about with large gains in a jet set coding, computer operation, knowledge work, and early scientific research. Alright. So that's GBT 5.5. Then images two. My gosh.
Jordan Wilson [00:04:50]:
This one, I mean, I don't know. This might sound crazy. Obviously, people are gonna be using the model itself, GBT 5.5 more that powers, you know, probably millions of companies and different products. But I think the one that visually got the most attention was images too, because this is a step change in what was available before when it comes to AI media generation. And I'm not saying AI image, because it can accomplish a lot more than that. So images two is OpenAI's new flagship image model, and it was launched broadly, which does include access to free tiers as well. And it delivers stronger editing, better composition and layout, and near production typography, multilingual tech supports, two k resolution, a wide range of aspect ratios, and up to eight coherent images that you can, create just from a single prompt. So images two dramatically improved text rendering and it come, compositional consistency, making it practical for, well, things you might want in a photo, such as signs, if you're mocking up user interfaces, graphics.
Jordan Wilson [00:05:57]:
One thing I've been, testing it out and using for, it's really good at creating presentations, infographics, etcetera. A lot of times before with, ChatGPT's earlier version, especially images one, the original, you know, it didn't get the text correct. A lot of times, you know, it might just mess up a couple of letters. So to use it, you'd probably have to go and clean it up, and there's a lot of character inconsistencies, and this has definitely changed with images too. Alright. And I think probably, right. And I think the reason why I might tackle this one on Wednesday, I think there's just so many use cases now because of that quality jump. Right? Like before, you could use it for product images.
Jordan Wilson [00:06:39]:
You could use it for infographics. You could use it for, you know, updating your stock photography online. But there was still, you know, for the average person, if you didn't really iterate on it or if you didn't know what you were doing, I still think you were left with something that looked like AI. Right? Which is not necessarily a bad thing, and that that does obviously have scary implications for now what is possible. But now I think it's at the point where people even who don't have, you know, reps using, you know, AI image generators can go in and get something that looks literally top tier professional. So I think we're gonna, on Wednesday, do our deep dive putting AI to work going hands on with images too. Alright. Next, workspace agents.
Jordan Wilson [00:07:23]:
These are really, really cool. So if you've been using codex to build at all, well, these are powered by codex. But workspace agents are a new shared cloud running agent feature available inside of Chad GPT. Yeah. So these are available on the web. You don't have to be, you know, running these via codecs. So they are cloud running agent features for team plans only right now. So that's CHED, g p t business, enterprise, EDU, and teachers, and they can automate long running workflows.
Jordan Wilson [00:07:51]:
They can be scheduled. They can be deployed into Slack, and you can operate across tools with admin control for permissions and approvals. Alright. One of the downsides with this and still not a lot of clarity. Well, number one, are they replacing GBTs? Maybe, maybe not. I don't know. I'll I'll find out. Also, they are free for now if you have one of those paid plans.
Jordan Wilson [00:08:13]:
Right? So sorry. I should say they are included for now in those paid business plans. But OpenAI did say that is only until May 6. So we're not sure if at that point, there's just gonna be limits, or if they are gonna be just charged at an extra rate. So we'll see, but it's only gonna be available here for a couple of weeks. So I would encourage if you have one of those team plans like I do, start taking advantage of them. They're really, really cool. Right? I've been, running the similar version of these types, inside codex.
Jordan Wilson [00:08:44]:
Right? But bringing them into chat g b t, into the team's environment is just a way for, you know, nontechnical teams, well, to be able to use these, but then to obviously run them in the cloud. Because if you are, you know, building these agents in codex, they're easy. It's natural language. You don't have to have really any programming or technical skills. But the downside of using them in codecs is, well, your machine has to be on. Right? So the big, advantage here of the workflow agents is, number one, you can share them across teams. And number two, they don't have to be tied to a local device they run-in the cloud. Alright.
Jordan Wilson [00:09:15]:
And next, last year, so many updates here. Codecs saw a lot of under the radar upgrades. So there's the new Chronicle feature, right, which is very similar to the controversial, Windows Copilot recall feature. It essentially remembers everything that's been on your screen, that is opt in only, and it's in research preview. There's now, browser control skill, which works really, really well, a little bit different and better in some instances than computer use. There's a new Google Sheets and slide skills, and now there's also an operating system wide dictation. So, yeah, they literally, inside of codex, essentially just shipped an app that you can use that you can just dictate anywhere. Right? So that part is pretty cool as well.
Jordan Wilson [00:10:00]:
Alright. So, we won't keep talking. Maybe through in about six minutes. Not bad considering literally, I think just those updates right there, are going to fundamentally change how people work, period. Just those three. Huge week for OpenAI. So, congrats to the team there. Next.
Jordan Wilson [00:10:20]:
The White House, has accused China of industrial scale AI theft. So the office of science and technology director issued a memo last week saying Chinese entities are using thousands of proxy accounts and jailbreaking techniques to run industrial scale campaigns that distill advanced US AI models into less capable versions according to both the White House memo and reporting from the financial types. So, if this sounds familiar, yeah, you know, Anthropic came out, kind of exposing this, a couple of months ago. The big companies as well have hinted about this in February. I think was the month when OpenAI, Google, and Anthropic all came out and said, yeah. China is distilling our models, and now the US government is confirming this and getting involved as well. So if you don't know what distillation is, it is using outputs from stronger models to train weaker ones. So in short, what the US government is now saying, is that Chinese entities signed up thousands of proxy accounts, and essentially would put prompts into these strong models from in Tropic, from OpenAI, from Google's and others, and they would essentially just use it.
Jordan Wilson [00:11:37]:
Right? Like, cheating on their homework, and then they would use that to, train their own models. So that is what distillation means. So you can derive, information from the original more powerful model, to usually make a weaker model. So the White House warns those distilled models frequently lack the security and safety protocols embedded in The US versions, creating potential national security and public safety risks when released externally. The administration says it is sharing intelligence about these extraction campaigns with US based AI companies and will collaborate with the private sector to develop technical defenses and detection strategies. So The US has moved to curb technology transfer to China as the global AI race intensifies, and the policy context includes the controversy over The US export decisions, such as allowing NVIDIA to sell h 200 GPU chips to China after earlier controls on their h 20 chips. So this is not a surprise that this is happening, but this is also one of those reasons. I think there is this big wave of momentum about a year ago, when the, you know, deep seek model came out last January of, you know, everyone saying, oh my gosh.
Jordan Wilson [00:12:55]:
These, you know, Chinese AI models, you know, you you should be using them at all times. Are they powerful? Absolutely. Are they distilled from US models exclusively? Probably not. Right? But a good majority of them are. So here you have a pretty significant step here with the US government coming on, and they're not, you know, outright saying don't use these, but they are warning against them because they say they lack the security and safety protocols that US versions have. So, you you know, this is one of those, kind of I wouldn't even call it a gray area yet. It's not a gray area because a lot of the Chinese models, I think, were developed, and created and deployed in a responsible and ethical manner, but many of them were not. Right? So this is always one of those interesting points of contentions, in bigger enterprise companies in The US.
Jordan Wilson [00:13:47]:
When they look at some of these Chinese models and they're like, wait. We could, you know, cut our AI costs down by, you know, $500 a month by using, you know, model x from China. Why wouldn't we? So we'll see here. I don't know if this will get to the point where we see potential restrictions on AI model use, but we have seen that at state levels. We have seen that in certain departments in governments, them not being allowed to use certain models, out of China anyways. Alright. Our next piece of AI news, the version of Google Gemini, I think most people have wanted for the past two years, right, since Bard became Gemini. You know, everyone's like, wait, Why can't Gemini just work everywhere and bring your context no matter where you're accessing it? Because, obviously, you can access Google Gemini from a myriad of places.
Jordan Wilson [00:14:41]:
Well, now you can. Well, it's rolling out, not available to everyone now. But Google at their Google Cloud Next conference announced workspace intelligence into its workspace apps, and that lets Gemini access emails, chats, docs, sheets, slides, and drive to create context aware drafts, edits, and summaries that mirrors a user's past preferences and company templates. And this is a pretty big shift that aims to save companies time and reduce switching between files and programs. So Google says the system builds on its personal intelligence, right, which was, released for personal Gmail accounts a couple of months ago, which kind of ties it to Gmail in photos, but it now applies that same personalized access across the company's productivity suite inside of Google Workspace. So, when asked inside of a doc or a slide deck, Google Gemini can now use stored workspace context. So that's past edits, templates that your company has created or styled choices to produce drafts, apply image edits, handle comments, or even generate slide decks that match corporate visuals automatically. Google describes workspace intelligence as able to retrieve relevant emails, chats, files, and information from the web to produce polished outputs in your exact voice, brand, and templates, reducing manual formatting and context switching.
Jordan Wilson [00:16:08]:
Gmail. Yay. Alright. I'm I'm excited for when this one does finally roll out to workspace accounts, but Gmail now gains new tools, called the AI inbox and AI overviews. Right? They have been testing the AI overviews in certain workspace accounts, but not the full AI inbox, which had been relegated to, just Gmail accounts previously on their ultra pricey ultra plan. But that AI inbox turns the inbox, in your Gmail into a task centered view to to help users catch up faster while AI overviews produces short summaries of email threads to surface key points quickly. Alright. So this is being rolled out now.
Jordan Wilson [00:16:50]:
Alright. But no word on exactly when companies are getting it. It's just certain companies are getting it. Alright. So it is also going to be dependent on the feature, and they're rolling out to workspace customers over time. Alright. Our next piece of AI news, Anthropic just cashed a lot of checks. Alright.
Jordan Wilson [00:17:09]:
Well, I don't think they actually cash checks. Right? Like, if you get $40,000,000,000 from Google, I don't think they give you the big checks like you won the lottery. Although, maybe we should start to normalize something like that. It would be pretty funny and, you know, interesting to watch. But Google will invest at least $10,000,000,000 and then could put up another $30,000,000,000 for $40,000,000,000 new investment in anthropic if the company meets certain agreed upon milestones, which could value the company at more than $350,000,000,000 for the initial trench. So Google Cloud will provide anthropic as well with five gigawatts of compute, which is a ton over five years, mostly via their new and now updated, I believe their eighth version of their TPUs, their tensor processing units, which materially expand Anthropic's access to large scale infrastructure for training and serving generative models. So this new deal, you know, up to $40,000,000,000, new capital investments just from Google, follows Anthropic's February funding that implied a $380,000,000,000 post money valuation and comes alongside Amazon's also new recent additional $5,000,000,000 commitment. So, Alphabet, you know, shares.
Jordan Wilson [00:18:33]:
So Google parent company jumped on the news of their big investment in Anthropic. So why is this important? Right? Why do we talk so much about fundraising? It's not just because Anthropic and, OpenAI and, you know, SpaceX now, which includes x and crack and all these other things. Right? It's not just because these companies are racing toward an IPO, and that's obviously going to impact anyone's portfolio. Right? Because AI is driving the economy and, the stocks here in The US. But more importantly, if you are an anthropic user, if your team, uses if your company uses anthropic, the past month has been absolutely horrendous in terms of uptime. I actually put a chart out on this, but the uptime is not even what you would need for a bare minimum to be required for enterprise software. Right? So the up times are very bad right now for anthropic. So this is good news because if you are an anthropic user, well, I don't know if they'll change their rates or their rate limits, but, you know, pref hopefully, this means that their uptime, will increase as they're able to bring more compute power, on in the coming months and years, both with this deal, with Google Cloud to be able to bring on additional five gigawatts of compute, but also with this big capital infusion.
Jordan Wilson [00:19:56]:
Alright. One company going in the opposite direction, not growing but shrinking, at least when it comes to their people, is Meta. So according to Axios, Meta will lay off about 8,000 employees, which is about 10% of its workforce as part of an efficiency push tied to soaring AI related costs. So the reductions underscore investor pressure after Meta said their capital expenditure could rise at least 60%, this year versus 2025 to support its new Meta Superintelligence Labs and its core business. While its free cash flow is forecast to fall also about 83% year over year. So this follows earlier large skill cuts at Meta. So if you remember back kind of post pandemic, they cut more than 20,000 jobs there in 2022 to 2023. And this mirrors just moves across big tech, including Amazon's planned 16,000 cuts, Microsoft, offering buyouts to, big percentage of its staff, and some recent layoffs at Block, Salesforce, Snap, and others showing industry wide cost pruning amid AI investment.
Jordan Wilson [00:21:10]:
So the news also follows report that Meta may reportedly be collecting employee keystrokes to train their models, highlighting privacy and ethical trade offs as firms seek richer internal data to accelerate AI development. Yeah. It's one of those stories. Right? Where, we've seen a lot of headlines that, you know, essentially, you know, for from this meta key stroking, you know, accusation that came out, in the in the media, you know, that essentially people that are still working at Meta are, for all intended purposes, just retraining their eventual replacement. Right? Because they're collecting now every single thing that's happening. So even if, you know, employees aren't explicitly training models, well, they maybe are at least according to reports. Alright. A couple more pieces of AI news.
Jordan Wilson [00:22:02]:
Well, this one also from Google Next, and Google announced at their Google Cloud Next conference that Vertex AI has been rebranded and expanded into the Gemini enterprise agent platform, a unified service to build, scale, govern, and monitor agentic AI across enterprises. So the platform is designed for agent building and orchestration, offering visual agent studio and an agent development kit or ADK, plus a runtime for long running agents with persistent memory to retain context across sessions. So Google says their model garden now exposes over 200 models, including its own Gemini three one pro, Gemini three one flash image, Lyria three, its music model, and Gemma four, which I personally have been loving and using a ton. And the platform supports multi agent coordination, so tasks can be split across specialized agents and tools. In the updates, Google says that governance and security are central. Agents receive cryptographic IDs, a centralized registry catalogs, approved agents and tools, and a gateway plus monitoring and anomaly detection models help enforce access policies and surface suspicious behavior. Alright. So, you know, we are now seeing, you know, Vertex going from what was kinda before a more, technical platform.
Jordan Wilson [00:23:32]:
Right? You would think more of, oh, this is an IT esque platform to now it's being kind of, rebranded the Gemini Enterprise agent platform. So I'm guessing that, you know, this is gonna bring more users potentially onto the platform versus where I think maybe previously Vertex was just, you you know, your your ML people, your head of AI, you know, your IT type. But with this, you know, you might start getting even some of your nontechnical people now, using the platform, you know, bringing the, visual agent builder and the, agent coordination into the fold, for more, which I personally think is smart. I mean, we'll actually see from a from a user interface user experience. Right? Because Vertex, I I mean, at least for me personally, there was a little bit of a learning curve, to understand how to use Vertex. So we'll see what it actually means for end users. And if Google's, you know, if one of their ultimate goal is just to get more nontechnical people using the platform. Alright.
Jordan Wilson [00:24:33]:
In our last big AI news story of the week, you can't make this one up. So, yes, you know that mythos model that was too powerful to release to the world? Well, yeah. Some some people on, Discord got access to it and used it for weeks just by guessing the URL. So apparently, it wasn't too good at cyber. So according to Bloomberg, a private discord group claims it located and had access Anthropic's Claude Mythos preview model for more than two weeks simply by guessing the model's online location from past anthropic naming patterns, yeah, in a recent data leak. So Anthropic has described Mythos as powerful enough to identify and then exploit zero day vulnerabilities in every major operating system in every major web browser. And they offered, limited preview access in the invite only project glass wing program that was, they said, intended to help secure critical software. So Bloomberg reports the discord group's, access relied on not on any complex cyber techniques.
Jordan Wilson [00:25:49]:
It wasn't even a breach, but one member just had privileged credentials and an as an anthropic, contractor. And the group then provided evidence convincing Bloomberg that the access to the model was real. So the group said it used mythos for benign tasks such as building simple websites, and they claimed the access to additional unreleased anthropic models as well. Though, Anthropic is investigating and said there's no sign of wider compromise so far. So Anthropic did acknowledge the report to Bloomberg saying it is investigating claims of unauthorized access, to mythos through a third party vendor and that it currently has no evidence the incident affected Anthropic's own system or extended beyond that vendor's environment. So, we covered, this mythos, story. Let me get the, the episode number here. Pulling it up here, live.
Jordan Wilson [00:26:46]:
It was a couple weeks ago. Where was it? It was there we go. Episode seven fifty four, if you wanna go, listen to it. So at the time, right, I waited a couple days. I wanted to, you know, read some of the reports, kind of, you you know, see what other smart people were thinking. But I acknowledge at the time that it was probably a mix of a scary capable model and a little bit of marketing. Now I think it's probably mostly marketing and maybe a model that's highly capable. And also following this in an interview this week, OpenAI CEO Sam Altman called Anthropic's, promotion of Mythos fear based marketing.
Jordan Wilson [00:27:29]:
And after this leak, I'm I'm I'm I'm kind of starting to believe him on that. Right? I don't know. Can you actually believe that a a a model is so powerful, in cybersecurity, right, in these zero day, you know, vulnerabilities? Essentially, they're saying Mythos was able to find all of these bugs that have existed in the web, for decades. Right? And the smartest humans alone, the smartest, cyber, hackers couldn't find them, but only Claude Mythos could. Right? Because it was that good. Right? And when you look at that and you're like, okay. Well, that's convincing. And, you know, if that is true, which Anthropic did say it was true, right, then it does have scary implications.
Jordan Wilson [00:28:15]:
But even at the time, I'm like, okay. There's still a little bit of marketing because this has always been, you know, Anthropic's approach to things. Right? Its CEO has talked relentlessly about how AI is gonna take 50% of, you know, white collar jobs because it's so good, and he's been, you know, going on any news program to say that. And, you know, with the latest mythos, you know, they're saying it's it's too powerful to release. No. I'm saying probably not. Right? So not only, this Discord leak, but, we also saw GPT 5.5, just released a couple of days ago as well. And when it came to benchmarks, right, there were some benchmarks that GPT 5.5 was better than Mythos on.
Jordan Wilson [00:28:57]:
So it's like, okay. Is this true? Right? If if if if a model is way too powerful in some coding benchmarks and security benchmarks, g b d five five is either better or just barely neck and neck. It's like, is it too powerful at this point, or is it just marketing? Right? Is is anthropic kind of sitting on this thing knowing that they have an IPO coming up, knowing that they can't compete, with OpenAI on users. They can't. Right? And they I don't think they ever will be able to. So I don't know. Is it one of those things since it's so powerful they have to sit on it, or, are they trying to drum up excitement right before they go public? Alright. That's it for our big stories, but now we're gonna end with our what's new and what's next kind of our bullet point recap of everything else that didn't make, you know, top eight news story of the week.
Jordan Wilson [00:29:51]:
And also keep in mind, we do go over just kind of AI or large language model releases, feature releases on Fridays, on our Friday features section. So some of these, we might be talking about on Friday, and then anything that comes out, you know, between today and Thursday, we'll probably be diving into as well. Alright. So, here we go. Copilot's new agentic capabilities in Word, Excel, and PowerPoint are generally available. Yeah. Agent mode now generally available in Word, Excel, and PowerPoint. So we had SpaceX, you know, kind of bought the rights to acquire, you know, cursor more or less.
Jordan Wilson [00:30:29]:
So it could be up to a $60,000,000,000 deal. And there is a report going on that SpaceX, maybe, acquiring Mistral as well. So looks like SpaceX could be, you know, forming a super team before it goes public itself. Claude released live artifacts in Claude CoWork. Pretty impressive. I've been enjoying those. An AI robot broke a human record for a half marathon. Not scary at all.
Jordan Wilson [00:30:56]:
Adobe unveiled agents for business at its conference. Codex announced Chronicle. Oh, I already talked about that. Good. We can skip over that one. OpenAI released a privacy filter. An open weight model they released for detecting and redacting, PII in text. So that's pretty huge.
Jordan Wilson [00:31:13]:
I think developers and builders are really gonna like that because they'll be able to build more things that deal with PII. Anthropic says engineering errors caused month long claud code quality drop. Yeah. That one, I don't know. Maybe we'll see if I do an episode just on that. But, yeah, Anthropic has been saying that they don't degrade their models, and then, you know, they've been saying this for a long time. And then they're like, oh, psych. Yeah.
Jordan Wilson [00:31:36]:
We've been degrading it, and we did. Anyways, HeyGen released Hyperframes, which looks really good. An open source Apache two tool that converts HTML and GSAP into m p fours. DeepSeq v four was released and somewhat underwhelming on benchmarks. Right? I thought it was gonna be a top news story, but I looked at it. I'm like, it's not even, you know, not close to Frontier models, and it's not even the top open source model. So, yeah, I don't think, you know, there goes all the deep seek, buzz. Sorry, everyone that thought v four was gonna be a frontier model.
Jordan Wilson [00:32:07]:
It's not. Yeah. Distilling can only get you so far. Cohere, is reportedly merging with Germany's Aleph Alpha, creating a $20,000,000,000 transatlantic AI powerhouse. OpenAI and Microsoft expanded their partnership with, cybersecurity where Microsoft gains trusted access to cyber capable models. ServiceNow in IBM stocks fell 179% respectively as AI disruption fears rocked software and ETFs. Microsoft is reportedly offering buyouts to about 9,000 US employees amid massive AI infrastructure investment. OpenAI released chat g p t four clinicians offering free access to verified US clinicians.
Jordan Wilson [00:32:54]:
NotebookLM, I'm loving this update. They can now auto label your sources. Man, you know, notebook l m is one of my absolutely favorite tools. So, this update will be well received by power users such as myself. I'm always looking back at my sources because I bring them in so fast, and I'm like, what are all these? Alright. So that's great. Microsoft is, investing $18,000,000,000 by the end of the decade to expand AI data centers in Australia. Claude now plugs into popular consumer apps like Spotify, Uber, Instacart, TripAdvisor, Automotive, TurboTax, and more.
Jordan Wilson [00:33:32]:
The trial between Elon Musk and OpenAI over a $134,000,000,000 in damages is beginning on Monday. Google DeepMind released Deep Research Max, alright, which uses Gemini 3.1 pro agents for exhaustive research. Unfortunately, only available on the API side right now, except there is actually, an agent that you can use inside of Google's AI studio. I do believe you have to, use your own usage on that, but still, it's been impressive. Right? I've I've, checked it out. It's really good. So if you haven't used that, go check it out. Alright.
Jordan Wilson [00:34:10]:
Last but not least, Meta signed a multi year, deal for AI workloads with AWS Graviton. Alright. That is a wrap y'all. Don't spend hours every single day trying to keep up with what's happening in AI and how it impacts your business. Just join us on Monday as we bring you the AI news that matters. Alright. So I hope this was helpful. If you're listening on the podcast, appreciate your support.
Jordan Wilson [00:34:37]:
Please make sure to follow and subscribe to the show. Reach out to me on LinkedIn as well. Tell me how we can make everyday AI better. So thank you for tuning in. Hope to see you back tomorrow and everyday for more everyday AI. Thanks, y'all.