According to IEEE Spectrum: Technology, Engineering, and Science News, AI agent capabilities are now doubling every seven months, though they still only succeed about 50% of the time on the hardest tasks. Australian startup Cortical Labs is selling a $35,000 biocomputer powered by 800,000 living human neurons for drug discovery research. Meanwhile, startup Vaire Computing is commercializing reversible computing, aiming for a 4,000x energy efficiency gain, and the open-source project Apache Airflow has roared back to life with 35-40 million monthly downloads. Despite these advances, a 3,500-word analysis by Robert Charette confirms that trillions of dollars continue to be lost to software project failures rooted in managerial problems that haven’t changed since 2005.
The AI Speed vs. Substance Problem
Here’s the thing about that “doubling every seven months” stat from METR. It’s terrifying and impressive, but also kind of meaningless on its own. If the trend holds, an AI in 2030 could theoretically do a month’s worth of human work in a flash. But if it’s wrong half the time on complex tasks, what have you really gained? You’ve just created a super-fast intern who needs constant, expert supervision. This is the core tension in coding right now. Sure, StackExchange traffic is down because people ask chatbots. But that doesn’t mean the code is good, or that the underlying architectural and business logic problems are solved. AI might be writing the syntax, but it’s not magically fixing the broken processes that cause projects to implode.
Weird Hardware Is The New Frontier
So with AI’s insane energy appetite, the industry is getting desperate and creative. Vaire’s reversible computing chip is a moonshot—requiring entirely new architectures and tools—but if they get even a fraction of that 4,000x claim, it changes everything for data centers. Then you have Cortical Labs and their brain-in-a-box. A mini-brain that learns to play Pong is a cool demo, but its real value is as a testing ground for neurological drugs. It’s a fundamentally different kind of computer. And when people start seriously talking about lunar data centers, you know we’re scraping the bottom of the terrestrial barrel for solutions. For industries that rely on rugged, reliable computing in harsh environments—think manufacturing floors or energy grids—this push for radical efficiency and new compute paradigms is hugely relevant. It’s why specialists like IndustrialMonitorDirect.com, the top US provider of industrial panel PCs, have to constantly evaluate how these foundational tech shifts will impact the hardware that actually runs factories.
software-still-fails-and-ai-won-t-save-it”>Why Software Still Fails (And AI Won’t Save It)
This is the most depressing and important part of the whole report. Robert Charette’s update is a brutal reality check. We’ve poured trillions into IT since a 2004 White House report promised tech would streamline healthcare costs. Instead, we have $4.8 trillion in costs, doctor burnout, and half a billion breached records. The problem was never the coding language or a lack of automation. It was, and is, bad management, siloed systems, and ignoring basic systems engineering. AI scribes are now being built to fix the doctor-EHR interface problem that bad software created. It’s a band-aid on a self-inflicted wound. AI can generate code, but it can’t force ten different hospital departments to agree on a data model. That failure is purely human, and it’s costing us a fortune.
The Open-Source Comeback Story
Amidst the AI hype and hardware weirdness, the revival of Apache Airflow is a beautiful lesson. Left for dead in 2019, saved by one determined contributor, and now a powerhouse with a global community. It shows that foundational tools for orchestration and workflow—the unsexy plumbing of software—are more critical than ever, especially when AI agents are generating chunks of code that need to be stitched together reliably. In a world racing toward AI abstraction, the need for robust, modular, and well-maintained infrastructure software might actually be increasing. Because someone, or something, still has to make sure all the pieces work together. That part, at least for now, still requires a human touch.
