Meta Buys Moltbook, DeepSeek V4 This Week, Google's New Models & More!

Meta just acquired a social network built entirely for AI agents. Google dropped a wave of new models and workspace updates. They are breaking rumors that DeepSeek version 4 is dropping this week and a whole lot more. So, let's get into it. Meta just

acquired a company called Moltbook. And depending on how this plays out, it could be a really significant move. So, what exactly is Moltbook? It's basically a social network, but not for people. It was designed specifically for AI aging. The idea is that AI systems could interact with each other, transact, and operate across platforms in a shared space. Think of it kind of like a online world, but built for AI agents. The deal was reported by Axius, and Meta hasn't disclosed how much they have paid. But here's what makes it interesting. The founders of Moldbook, Matt Schlit and Ben Parr, are being bought directly into Meta Super Intelligence Labs, which is the unit led by Alexander Wang, the former CEO of Scale AI. So, this isn't just a technology acquisition. It's a talent acquisition. Meta is pulling in people who have been thinking deeply about autonomous AI agents and how they operate in social and commercial environments. And that tells you a lot about what Meta is actually building towards. They want AI agents that just don't sit on your phone and answer questions. They want agents that can act on your behalf across Facebook, WhatsApp, Instagram, and beyond. millions of AI assistants interacting and transacting inside Meta's ecosystem. At least that's what the vision seems like at the moment. But there's also an interesting subplot. Moltbook was designed to work alongside a separate project called OpenClaw which was previously called Clawbot and before that moldbot. The creator of OpenClaw was actually hired by OpenAI last month and that project is now being open sourced under OpenAI's backing. So you got this weird situation where the two sides of essentially the same ecosystem have been split between OpenAI and Meta. That's just the kind of thing that happens in the space right now. Everything is moving so fast that companies are scrambling to pick up the pieces. The deal is expected to close mid-March with the founders officially joining Meta on March 16th. Next up

Google dropped a big Workspace update today and there's actually a lot here worth talking about. The core idea is that Gemini is becoming a much more active collaborator across Docs, Sheets, Slide, and Drive. And it's not just going to be on your sidebar where you can ask questions to, but something that actually does the work with you. In Google's Docs, there's a new help me create feature where you describe what you need and Gemini goes out, pulls relevant information from your drive, Gmail, chat, even the web, and drafts a fully formatted document, not just a rough skeleton. a real first draft with structure and formatting already applied. And once you have that draft, you can ask it to refine specific sections without regenerating the whole thing. There's also a match writing style feature that can analyze a document and normalize the tone across it, which is useful if multiple people have been editing the same file and the voice is all over the place. In Sheets, this is probably the biggest leap. You can now describe what kind of spreadsheet you need in plain language. And Gemini builds it for you, pulling data from your files, emails, and chats, constructing the whole thing from scratch. Google says this hit a 70. 48% success rate on standard spreadsheet benchmark, which they say nears human expert level and beats other competitions. There's also a fill with Gemini feature that autopop populates columns based on inferred intent. So you add a column called praise or complaint and it just figures out what you mean and fills it in. They claim it's nine times faster than manual entry on large tasks. Slide is getting the ability to also generate individual slides that actually match the visual style of your existing deck. And full presentation generation is coming soon. And drive is getting AI powered search that doesn't just return a list of files. It actually answers your questions and cites the relevant documents. These are rolling out to Gemini Alpha Business customers and Google AI Pro and Ultra subscribers first in English only for now. Still on

Google, they also released Gemini embedding 2 today. And this is one of their more technical updates, but it's worth knowing about. Embeddings are how AI systems understand meaning. They are the underlying representations that make things like search, recommendations, and retrieval actually work. Up until now, embedding models were mostly text only. you'd have separate systems for images, audio, video, and so on. Gemini Embedding 2 changes that. It's Google's first natively multimodal embedding model, which means it maps text, images, video, audio, and documents all into a single unified space. You can pass in mix inputs like an image combined with some text in a single request, and the model understands the relationship between them. It supports over 100 languages, handles up to 8,92 tokens for text, up to six images per request, up to 120 seconds of video audio natively without needing a transcript first, and PDFs up to six pages for developers building rag systems, search tools, or anything that needs to retrieve information across different types of media. This is a pretty meaningful upgrade. It's available now in public preview through the Gemini API and Vertex AI. Okay, this could just be a

rumor, but keep that in mind. But the signs are pointing to Deepseek version 4 dropping this week. A source who claims to have spoken with someone at a cloud provider that's a Deepseek partner says the release is highly likely within days. And interestingly, the source says the cloud partner themselves didn't have much information. Deepseek is apparently keeping this very close to the chess. They're not sharing details with partners ahead of time, which is consistent with how they've operated before. They just dropped things out of nowhere. What little we know is that there are papers tied to this release. The cloud contact mentioned that the blog would link to them which suggests there's actual research accompanying the model and not just a product announcement. And whoever is dropping this seems genuinely excited about it, which is a good sign. Deepseek has a track of releasing models that'll punch well above their weight and shake up the market when they land. Version 3 was a big moment. If version 4 is landing this week, that's going to be a significant release to keep an eye on. We'll cover it as soon as it drops. There are also

early signs that Google is about to release Gemma 4. A poll request from Google's bot account was actually spotted on GitHub adding mpu support for a Gemma 4 model, which is usually a pretty reliable indicator that something is close. And separately, an insider source is claiming the largest version of Gemma 4 will sit around 120 billion parameters total with about 15 billion active tokens. That makes it a mixture of expert style model, which is big overall, but only activating a fraction of the parameters at a given time for efficiency. Nothing official from Google yet, but the signals are pointing to a release coming soon. Especially interesting timing given DeepSeek potentially dropping this week, too. is going to be a busy few days. Switching

to OpenAI, they announced that they're acquiring Prompt Fu, an AI security startup. So, what does Prompt Fu do? They build tools that help companies find and fix security vulnerabilities in AI systems. Things like prompt injections, jailbreaks, data leaks, and agents behaving in ways they're not supposed to. It's basically red teaming as a product. Their tools are already used by over 25% of the Fortune 500 companies, and they have a popular open-source library that developers use to test and evaluate LLMs. OpenAI is planning to integrate Prompt Foods technology directly into OpenAI Frontier, which is their platform for building and operating AI agents in an enterprise environment. The goal is to make security testing a native built-in part of the development workflow. So, companies aren't just bolting it on the end, they're catching the risk earlier in the process. No purchase price was obviously disclosed. Prompt Fu has about 11 employees and has raised about $22 million total with a post money valuation of around 85 billion as of last July. OpenAI says they will continue to maintain Promelu's open-source project and keep it available across different providers and models, not just open AIs. This move makes a lot of sense given how much Open AI is pushing into the enterprise. As AI agents get more access to real data and real systems, the attack surface grows. Having dedicated security tooling built into the platform is increasingly

important. Microsoft announced something called co-pilot co-work. The basic idea is that co-pilot stops giving you just answers and actually starts completing tasks on your behalf. You describe what you want to accomplish and co-work turns it into a plan that executes in the background, drawing on your emails, meetings, files, calendars, and team messages to get actual work done. Some concrete examples they walk through are you can ask it to clean up your calendar and it reviews your schedule, identifies low-v valueue meetings and conflicts, propose changes and once you approve it actually applies them, declining meetings, rescheduling, adding focus blocks, even sending prep documents or you can ask it to prepare for a customer meeting and it pulls relevant inputs from across your tools, builds a briefing dock, creates a deck and schedules prep time all connected. You can also hand it a company research and get back formatted executive summary, a structured memo with citations and an Excel workbook with labeled data. What's notable is that Microsoft explicitly says that this is built with enthropics technology. They integrated cloud into the Microsoft 365 co-pilot as the backbone of co-work. So this is another sign of how Microsoft is pulling into multiple models rather than committing entirely to one. Copilot Co-work is in research preview right now with a limited set of users and is coming to the Frontier program more broadly in late March. But that's it for today's

video. Make sure to subscribe to the channel, follow us on Twitter, follow the World of AI, and don't forget to subscribe to our newsletter. We post constantly and you don't want to miss this. I'll see you guys in the next