Agents Rise as Reality Sets In, Quantum Makes Waves

Good morning. It's been a fascinating week for artificial intelligence—one where hype meets hard reality, and some genuinely interesting technical work gets buried under the noise. The biggest news is also the most predictable: OpenAI has hired Peter Steinberger, the developer behind OpenClaw, the AI agent framework that's taken over developer culture these past few weeks. Sam Altman says it's about "the future being extremely multi-agent," which is corporate-speak for "we see this working and we want to control it." Fair enough. OpenClaw will remain open-source under a foundation, which matters more than the headline.

When AI Meets Reality

But here's what's more interesting: this week also delivered a much-needed reality check on what these tools actually do. One developer built an entire MCP server—a 21-tool collection that fixes LLM hallucinations around basic math, hashing, UUID generation, and date calculations. The point isn't that it's clever. It's that LLMs consistently fail at deterministic tasks. Meanwhile, enterprise adoption data from MIT shows 95% of organisations investing in generative AI are getting zero measurable return. That's not a rounding error. That's the core story nobody wants to talk about. The tools work beautifully for prototyping and exploration. For production reliability? We're still figuring it out.

One particularly grounded essay this week came from Alberto Pepe, who spent time genuinely using Claude daily and walked through both its genuine value and its real limitations. He made a point worth sitting with: the pricing model looks designed to hook everyone now and charge later. Free-tier access, cheap subscriptions, then the slow price creep begins once dependency sets in. That's not conspiracy thinking—it's just how software markets work.

Quantum Shifts

Away from the AI theatre, quantum computing quietly had several genuine breakthroughs. Swedish researchers found a way to cool quantum computers using noise instead of fighting it—turning a fundamental problem into a feature. Separately, a Sydney team reduced the qubit count needed to break RSA-2048 encryption from millions to around 100,000 by using better error-correcting codes. Neither solves quantum computing tomorrow, but both represent real progress on real problems. Worth watching more than the next GPT version number.

On the web side, developers are finally getting proper tools for agentic work. PlanetScale launched an MCP server that lets Claude and other AI tools query your database directly. Fly.io shipped writable Litestream VFS for edge computing. These aren't flashy announcements, but they're infrastructure that makes building actual things easier. That matters more than hype ever will.