AI Unfiltered: Daily Brief on Real Shifts in Models, Products, and Funding

On March thirty-first, Salesforce turned Slackbot into an agentic desktop assistant that can monitor your screen and transcribe your meetings. This isn't just an upgrade to a chatbot. It’s a transformation of Slack into a proactive operating system for your work, powered by Anthropic's Claude. With over thirty new capabilities, it can now watch what you do, understand the context of your tasks, and execute workflows across third-party tools. This move signals a fundamental shift. AI is no longer just a feature within an application. It's becoming the application itself. That was the biggest move, but not the only one.

Google just released Gemma 4, a suite of open models under an Apache 2.0 license, designed to run on everything from a mobile phone to a high-end GPU. The goal is accessibility. The largest model, a 31-billion parameter version, now ranks third on the Arena AI text leaderboard, placing it among the most capable open models available. But the real story is the smaller models, designed to run efficiently on consumer hardware like Android phones and even a Raspberry Pi. Google is trying to put a capable model in every developer's hands, regardless of their hardware budget. Meanwhile, a Chinese startup called MiniMax just launched MiniMax-M1.

It's the world’s first open-source, large-scale hybrid-attention reasoning model. It supports a one million token context window and can output eighty thousand tokens of pure reasoning. What matters here is the efficiency. Using a new mechanism called Lightning Attention, it performs deep reasoning tasks with only thirty percent of the computing power of a comparable model like DeepSeek R1. The entire reinforcement learning training phase cost just over half a million dollars. That’s an order of magnitude less than expected, and it shows the barrier to creating state-of-the-art models is falling faster than anyone predicted.

The money is also flowing toward specialized applications. Sona, a platform for frontline workforce management, just raised forty-five million dollars in a Series B round. Their goal is to replace the legacy software that manages scheduling, payroll, and operations for industries like retail and hospitality. As their CEO put it, every other software category has been transformed by AI, but the tools for the world's largest workforce are two decades old. Sona is betting that AI-native infrastructure is the only way to fix it. In that same vein, a company called Titan launched banking-native AI models. These are not general-purpose LLMs.

They are tailored specifically for compliance and regulatory environments. The models have banking logic and regulatory frameworks embedded directly into their architecture. This allows for traceable reasoning, which is critical for audits. On their own benchmarks for banking tasks, they achieve seventy-six percent answer accuracy, outperforming general models that don't understand the constraints of a regulated industry. And finally, as all these new agents and models get deployed, the question of safety becomes urgent. An AI infrastructure startup named Vijil just raised seventeen million dollars to address this.

Their platform is designed to enhance the trust, security, and resilience of enterprise AI agents once they are in production. It continuously evaluates them for reliability and safety, enforcing governance controls and runtime protections. It’s the unglamorous but essential plumbing required to make any of this enterprise-ready. Let’s go back to Slackbot. The change from a reactive bot to a proactive agent is the real story here. "Agentic" means it has autonomy. It can see you’re trying to schedule a meeting across three different channels, recognize the pattern, and offer to complete the task for you. It can join your Zoom or Google Meet call, transcribe the entire conversation, summarize it, and assign action items to the right people.

It becomes a central hub that coordinates work across the six thousand applications in the Salesforce ecosystem. This is the promise of automating repetitive cognitive labor. Here’s the turn. To do this, Slackbot needs to see everything. The new capabilities include monitoring your desktop activity. It needs to know what applications you’re using and what data you’re looking at to understand your workflow. It needs access to your audio to transcribe meetings. This introduces a profound privacy trade-off. The convenience of an AI assistant that anticipates your needs is paid for with persistent, low-level surveillance of your work.

We’ve seen this debate with consumer products, but this is one of the first major deployments of an always-on agent in a corporate environment. User acceptance of this feature will be a major test case for the entire industry. It’s no longer just about whether the AI is useful. It’s about whether we're comfortable with what it has to see to be useful. Now, let's look at Google's Gemma 4. The release of a highly capable open model is significant, but the strategy behind it is what matters. Google isn't just releasing one model; it's releasing a family of them, optimized for a spectrum of hardware. This is a deliberate ecosystem play.

The smaller E2B and E4B models are built to run on-device, on mobile phones and IoT hardware. This is about distribution at massive scale. By making them efficient on consumer devices, Google is trying to make Gemma the default choice for any developer building an app that needs local, on-device intelligence. Then you have the high end. The thirty-one billion parameter model is a direct competitor to other top-tier open models, but it performs best on high-end NVIDIA H100 GPUs. So Google is bracketing the market. They are providing a solution for the independent developer tinkering on a Raspberry Pi, and a solution for the well-funded startup with a cluster of GPUs.

The goal isn't just to have the best model. The goal is to make Gemma the most available model. This is Google attempting to build the foundational layer for the next generation of AI applications, much like Android became the foundational layer for mobile. They are betting that ubiquity will ultimately be more powerful than having the single highest benchmark score. We’ve stopped building applications with AI. We’re now building them on AI.