Tech Twitter Unfiltered: The Real Conversations Shaping AI & Software

Salesforce is freezing all hiring for software engineers in 2025. Instead, they're spending three hundred million dollars on Anthropic AI tokens. Last week, we covered Chamath Palihapitiya’s warning that Anthropic could become the Friendster of AI. This week, Marc Benioff is betting a third of a billion dollars that he's wrong, and he’s betting his entire engineering payroll on it. This isn't a pilot program. This is a declaration that the nature of building software has fundamentally changed, and the market for human programmers just got a new, very powerful competitor. That move from Salesforce is the earthquake, but the tremors are being felt everywhere.

Let's run through the other signals lighting up the network today. First, let's stick with Salesforce, because the numbers behind that hiring freeze are staggering. Benioff says internal AI tools have already boosted developer productivity by thirty percent. Automated systems and AI agents are now handling, and this is a direct quote, "between thirty and fifty percent of Salesforce's total global workload." Their dedicated AI business unit, Agentforce, just hit eight hundred million dollars in annual recurring revenue. So they're not just using AI to save money on engineers...

they're selling the output of those same AI systems as a core product. That leads directly to the second major thread, a massive three-thousand-word analysis from analyst Keith Tsang that’s being passed around with the caption: "We're watching SaaS die in real time." He calls it the SaaSpocalypse. Since January of this year, 2026, one TRILLION dollars in software stock value has been wiped out. Why? Because agentic AI is destroying the per-seat pricing model that built the last twenty years of software. Why pay for ten human seats when one person can manage ten AI agents doing the same work?

Tsang’s data shows AI-native companies are growing three times faster than legacy vendors. This isn't a future prediction. This is a present-tense market correction. Third, in the middle of this chaos, a tech analyst named Michael Parekh dropped a bomb about geopolitics and open source. Just as companies like Google, Meta, and Microsoft are pulling their most powerful models behind closed walls, Parekh argues that Nvidia and Apple are about to become the champions of open source AI in America. Not out of ideology, but strategy. Their dominance is in silicon and hardware. By promoting open source models that run best on their hardware, they create an ecosystem that can compete with China's state-sponsored AI push.

He frames it as a weapon. Open source isn't a philosophy anymore; it's a corporate strategy to win a global arms race. And finally, a report from a group called New Fire Energy is making the rounds, and it's the most unsettling of all. The thesis is simple: recursive self-improvement in AI is no longer a theory. It's happening now. AI development cycles that used to take six to twelve months are now compressing to weeks, because AI systems are optimizing their own training pipelines. Frontier models went from performing at seventeen percent of human level in machine-learning engineering in 2024 to sixty-five percent by early this year.

The report’s opening line says it all: "We’re in the early phase of the singularity. The mistake almost everyone keeps making is treating it as something we are approaching. We are not approaching it. We are in it." Let's go deeper into that Salesforce decision, because it's not just about one company. It's a blueprint for the end of software as we've known it. Marc Benioff went on the All-In podcast to explain the move, and it's even more radical than the headlines suggest. That three-hundred-million-dollar budget for Anthropic tokens? It's mostly for programming and software development.

He’s not buying AI to help his salespeople write emails. He’s buying AI to write the code that runs his company. For years, the story was "AI is a copilot." It's a smart assistant that makes your existing employees more productive. That story is now officially dead. Salesforce engineers are being shifted into supervisory roles. Their job is no longer to write code. Their job is to manage workflows and review code generated by autonomous agents like OpenAI's Codex and Cursor. Benioff calls it a "digital labour revolution." That’s a very polite way of saying he’s replacing human labor with digital labor.

And this brings us to Keith Tsang's "SaaSpocalypse." The entire business model of Software as a Service was built on a simple premise: you have employees, and each employee needs a license. More employees, more licenses. Growth was tied to headcount. Agentic AI breaks this model completely. If one brand manager can now direct a team of AI agents to do the work of a ten-person marketing department, you don't need ten licenses for your marketing automation software. You need one. Maybe. Or maybe the AI just connects directly to the platform's API and you pay for usage. This is the seismic transformation.

Gartner is now predicting that eighty percent of enterprises will be using generative AI applications by the end of this year. And the pricing is flipping from subscriptions to usage-based models. This is why a trillion dollars in value evaporated. That money represented the future discounted cash flows of per-seat licenses that are never going to be sold. The market suddenly realized that the assumption of "more humans, more software seats" is no longer valid. The new equation is "more tasks, more AI compute." And the companies that sell the compute—like Nvidia—and the companies that sell the most effective agents—like Anthropic and OpenAI—are vacuuming up all that value.

So if the business model of software is breaking, and the job of a software engineer is changing overnight, you have to ask: what is driving this insane speed? Why now? This is where that New Fire Energy report comes in. It connects the dots between what Benioff is doing and why he's able to do it. The report argues that the AI progress curve is no longer linear, or even exponential in the way we usually think about it. It’s recursive. The output of one generation of AI models is now feeding directly into the cost structure and efficiency of building the next generation. Let me make that concrete.

An AI model, let's call it AlphaEvolve, is tasked with optimizing the training pipeline for the next model, let's call it BetaMind. AlphaEvolve figures out a more efficient way to process the data and tune the parameters. This makes training BetaMind cheaper and faster. BetaMind, being a more advanced model, is then put to work optimizing the training for GammaPrime. The loop isn't theoretical. It's not happening inside a single system. The report says, "It is closed across a company." AI is making better AI. The benchmarks are getting absurd. In 2024, the best AI could solve about seventeen percent of the coding challenges on a professional platform.

By early 2026, it's sixty-five percent. In cybersecurity, frontier AI systems are now solving ninety-three percent of professional capture-the-flag security challenges. A test called "Humanity's Last Exam," designed to be a comprehensive measure of reasoning, saw AI scores jump thirty percentage points in a single year. This is the engine running underneath everything else. It's why Benioff can freeze hiring for engineers. It's why the SaaS model is collapsing. The underlying capability of the technology is improving at a rate that is itself accelerating. And this leads to the report's most chilling and clarifying insight.

They argue we're making a category error by talking about "the singularity" as a future event. An inflection point we are heading towards. They write, and I'm quoting again: "We are not approaching it. We are in it." This isn't a future shock. It's a present-day reality that we are just beginning to process. The change isn't coming. The change is here, and it's busy rewriting code, and rewriting the economy along with it. So what does this week set up? We've just watched one of the biggest software companies in the world announce it will now buy AI agents instead of hiring human engineers.

We've seen the market wipe a trillion dollars of value off of companies who were too slow to realize the game has changed. And we have a credible report that the reason for all this chaos is that AI has begun to improve itself, compressing progress from years into weeks. The conversation is no longer about what AI might do. It's about managing what it is doing. The question is no longer "When will the jobs be impacted?" The question is "My job was just impacted, now what?" The focus shifts from forecasting the storm to navigating the flood. The old playbook is gone. Growth at all costs, funded by selling more seats to more humans, is over.

The new playbook is being written right now, in the language of API calls, token budgets, and autonomous agent workflows. And the defining feature of this new era isn't the technology itself. It's the speed. The sheer, relentless, recursive speed of the change.