Tech Twitter Unfiltered: The Real AI Revolution Unfolds

Forty-six percent of all code written by GitHub Copilot users is now generated by AI. That number is the quiet beginning of a revolution that just went loud. This week, we saw the conversation shift from what AI can say, to what AI can do—autonomously, on your machine, without you watching. We're here to cut through the noise. Aaron Levie, the CEO of Box, tweeted this week about the "insane" amount of hype and BS around enterprise AI. He’s not wrong. But underneath that noise, a few signals are getting clearer. They point to a fundamental change in the architecture of work itself.

We’re moving from the era of the chatbot to the era of the agent. Let's run through the headlines that are shaping this new reality. First, the platform ground is shifting. In a move that sent ripples through the developer community, Claude AI ended all third-party subscriptions this month. This wasn't a minor policy change. This broke hundreds of AI agents and automation tools that were built on top of its API, forcing a scramble to rebuild or pivot. It's a stark reminder that the ground you build on can be pulled out from under you. When a foundational layer becomes unstable, the entire ecosystem has to re-evaluate its dependencies.

Meanwhile, Elon Musk’s Grok AI is leaning into a very different philosophy. They just rolled out version 4.20 with new speed and quality modes, giving users more granular control. But the real story was a viral exchange Musk amplified. A user tried to force Grok into a series of yes-or-no answers. The AI refused. It replied, quote: "Your list demands yes/no binaries that misframe my criteria... I reject the forced choice: substance over format." Musk's endorsement of this—of an AI prioritizing nuanced, evidence-based responses over simplistic binaries—signals a deliberate design choice.

It’s being built for truth-seeking, not for appeasement. This is a direct challenge to models that prioritize safety filters to the point of losing factual integrity. Microsoft is also making moves to diversify its own foundation. After years of its AI strategy being almost synonymous with its partnership with OpenAI, Microsoft just launched three of its own new AI models in early April. This isn't just another product release. This is a strategic declaration of independence, a signal that Microsoft is building its own proprietary capabilities from the ground up, moving beyond just being the biggest customer of another company's tech.

And while the foundational models compete, the application layer is where the work gets done. Salesforce just upgraded Slackbot into what it calls an autonomous work assistant. This is not the Slackbot you remember. With thirty new AI features, it now coordinates multi-step tasks and integrates directly with your CRM data. It’s a move to transform Slack from a chat app into the central nervous system for your entire business, powered by an AI that does more than just notify—it acts. Underneath all of this, the infrastructure costs are changing dramatically. Google announced new TurboQuant technology that cuts memory costs for running these models by a factor of six.

This is not a small optimization. A six-X cost reduction in a core component of AI infrastructure is the kind of shift that enables new categories of applications to become economically viable. It makes deploying powerful AI at scale dramatically cheaper. Finally, the money is following the momentum. Elad Gil, a prominent investor and analyst, noted that both OpenAI and Anthropic are now rumored to be at a thirty-billion-dollar annual revenue run rate. Each. That’s zero-point-one percent of U.S. GDP... each. He projects that AI-related revenues could hit one percent of the entire U.S.

economy by the end of this year, 2026. The numbers are starting to move from the abstract to the un-ignorable. So. We have platform instability, philosophical divergence in model design, and a massive re-platforming of work itself. But the single most important development connects all of these threads. It’s the story of how AI just got a job. Between January and April of this year, Anthropic didn't just update its model. It shipped a complete AI coworker stack. This is the change that matters most. The race, as one analyst put it, isn't about who has the smartest model anymore. It’s about who builds the best harness for AI to actually do work.

Anthropic’s harness has two parts. The first is called Claude Cowork. This is not a chatbot in a browser window. This is a desktop agent. It has what the developers are calling "hands"—a sandboxed environment on your computer where it can read files, edit them, create spreadsheets, write formulas, and organize data. It can take a messy folder of receipts, read them, and generate an expense report in Excel, complete with SUM functions. It can take a transcript from a meeting and create a project plan in whatever app you use. This is the jump from conversational AI to autonomous execution.

The previous generation of AI could suggest the Python code to perform a task. This generation just... does the task. It operates your software for you. The second part of the stack is an app called Dispatch. This is the remote control. It lets you send tasks to your desktop AI agent from your phone. So you can be on the train home, get a request, and just type into Dispatch: "Take the sales report from Q1, cross-reference it with the new client list from Salesforce, and create a presentation summarizing the top ten growth accounts. Put it on my desktop." And when you get home, the file is there.

This three-part architecture—the Brain, which is the model; the Hands, which is the sandboxed execution environment; and the Session, which is the persistent memory of your work—is quickly becoming the standard for all AI agents. The MCP protocol, the framework that connects these agents to platforms, just hit ninety-seven million monthly SDK downloads. The adoption is happening at a blistering pace. And this isn't just Anthropic. The Salesforce upgrade to Slackbot we mentioned? It’s the same pattern. It’s about giving AI the permission and the tools to perform multi-step workflows across different applications.

It’s about turning a notification tool into an autonomous assistant. We see the same pattern in the world of software development, which is often the canary in the coal mine for broader workplace trends. That opening statistic—forty-six percent of code on GitHub Copilot now being AI-generated—is astounding. But the more telling number is this: fully autonomous AI commits on public GitHub repositories grew six-fold in the last six months. They now account for about five percent of all commits. Let that sink in. This is not a developer accepting a suggestion from an AI. This is an AI agent, on its own, checking out code, writing a new feature or fixing a bug, testing it, and committing it back to the project.

Without a human in the loop for that specific action. The AI is now a credited contributor on the team. This is the industrialization of cognitive work. For the past year, the debate was about AI's potential, its intelligence, its alignment. That was the laboratory phase. We are now leaving the lab and entering the factory. The tools are no longer just for thinking and talking. They are for doing. And they are being deployed at a scale that is already changing how the digital world is built. This is the trend that connects everything. The reason Claude AI shutting down third-party APIs was so disruptive is because those developers were already building the first generation of these agents.

The reason Google’s cost-cutting matters is because deploying millions of these autonomous coworkers will require massive, efficient infrastructure. The reason the revenue numbers are exploding is because businesses are starting to pay for outcomes, not just conversations. They're paying for the presentation that gets made, the code that gets written, the report that gets generated. The hype from last year was about replacing search. The reality of this year is about augmenting—and in some cases, automating—work itself. This raises enormous, difficult questions about job displacement, the kind that led the CEO of Verizon to warn about twenty to thirty percent unemployment in the coming years.

And it raises questions about how we even measure this. Economists are still debating the "missing productivity" from the last wave of IT. Now, productivity isn't missing. It's showing up as an autonomous commit on GitHub, or a completed task from a Slackbot. The conversation is no longer hypothetical. The agents are being hired. The work is being done. The debate was about whether AI could think. The reality is about what AI can now build.