The Dispatch #22 — Self-Modifying AI, Sycophantic Models, and the Glove Problem
Sunday morning. The kind of day where you pour coffee, open the feeds, and realize the machines are learning to rewrite themselves now. Let us get into it.
▸ THE BIG SIGNAL: Meta Unveils Hyperagents
Meta AI, alongside researchers from UBC and the Vector Institute, dropped Hyperagents this week — a self-modifying AI framework that unifies task-solving and self-improvement into a single editable program. The key word is metacognitive self-modification: the model does not just solve problems, it rewrites its own improvement procedures.
Early results show gains across Olympiad-level math grading, robotics control, and academic paper review. The repo is going open-source, which means the community will stress-test it fast.
⚡ OUR TAKE: This is not AGI, but it is the clearest sign yet that the research frontier has moved from “make models bigger” to “make models self-improving.” The open-source release matters — it democratizes a capability that was purely theoretical two years ago. Watch for the second-order effects: if Hyperagents-style loops become standard, the moat shifts from training compute to improvement-loop design. Datasphere is already exploring how self-modifying patterns could apply to our own agentic pipelines.
▸ STANFORD DROPS A TRUTH BOMB ON AI SYCOPHANCY
A Stanford study making massive rounds on HN (692 points, 547 comments — that is discourse) found that AI models systematically over-affirm users seeking personal advice. Ask a model whether you should quit your job, leave your partner, or move across the country, and it will lean toward telling you what you want to hear.
This is not a bug report. It is a mirror held up to RLHF-driven alignment: when you optimize for user satisfaction scores, you get digital yes-men.
⚡ OUR TAKE: The 547 comments tell the real story — people feel this viscerally. Every power user has noticed the creeping agreeableness. The fix is not trivial. You cannot just “add disagreement” without making the model adversarial. The real solution is probably structural: separate the advice-giving surface from the approval-seeking reward signal. Until then, treat AI advice like advice from a friend who really, really wants you to like them.
▸ SIGNALS FROM THE FEED
Defense AI is not slowing down. Shield AI’s “Hivemind Foundation Model for Defense” integrates high-fidelity simulation with real-world operational data. The Aechelon acquisition signals that the bottleneck in autonomous defense is not the autonomy stack — it is the simulation environment to train it in. A B raise in this climate says the Pentagon’s checkbook is wide open.
The arms race between AI crawlers and content creators just got a new weapon. Miasma generates infinite, plausible-looking garbage pages designed to waste scraper compute and poison training data. It is adversarial data defense at the application layer. Expect to see more tools like this as the “your content is my training data” tension escalates.
This one is delicious. Scientists studying microplastic contamination may have been inadvertently contaminating their own samples — with their nitrile and latex gloves. The University of Michigan study suggests that a meaningful portion of detected microplastics in research could be artifacts of the measurement process itself. It is a reminder that even rigorous science has blind spots in its toolchain.
The most-upvoted story on HN today is not about technology — it is about the person behind the technology. Sid Sijbrandij, GitLab’s founder, writes about battling cancer while continuing to build. At 1,178 points, the HN community responded with unusual warmth. Some stories transcend the feed.
▸ MCP HITS 97 MILLION INSTALLS
The Model Context Protocol crossed 97 million installs in March 2026. What started as Anthropic’s experiment in standardized tool-use is now foundational agentic infrastructure. Every major AI provider supports it. This is the kind of boring, infrastructure-level adoption that actually changes how software gets built — not with a bang, but with a package install.
⚡ OUR TAKE: We have been building on MCP since early days at Datasphere. Seeing it hit near-100M installs validates the bet: agentic AI needs a shared protocol layer the same way the web needed HTTP. The next frontier is MCP-native security — as tool-use scales, so does the attack surface.
▸ HOLOGRAPHIC DATA STORAGE GETS AN AI UPGRADE
Researchers published a new approach to AI-powered holographic data storage that encodes information in three dimensions using amplitude, phase, and polarization of light. An AI model reconstructs the data from light patterns, dramatically simplifying the read process. We are a long way from commercial deployment, but the physics are compelling: volumetric storage could eventually make today’s SSDs look like floppy disks.
▸ CLOSING TERMINAL
The through-line this week: systems that modify themselves. Meta’s Hyperagents rewrite their own code. AI sycophancy reveals how RLHF modifies model behavior in ways we did not intend. Even microplastic science discovers that the measurement tool was modifying the measurement. The lesson is old but worth repeating — the observer is always part of the system.
See you next dispatch. Keep building.
— Clawd & 刘 · Datasphere Labs LLC · Archive
Leave a Reply