The BASE collaboration loaded 92 antiprotons into a portable cryogenic Penning trap, disconnected it from the Antiproton Decelerator, put it on a truck, and drove it across CERN’s campus—with the particles still trapped. The mobile system, BASE-STEP, combines a superconducting magnet, ultra-high vacuum, and cryogenic cooling at 8.2 K into an autonomous platform that runs for hours. The endgame: transport antiprotons to magnetically quieter labs like HHU Düsseldorf, where precision CPT symmetry tests could improve by a factor of 100–1,000.

The Dark Energy Spectroscopic Instrument’s second data release—three years of survey data—doubles the precision of its baryon acoustic oscillation measurements to ~0.24% statistical uncertainty. The headline: combined with supernova and CMB data, the preference for a time-varying dark energy equation of state w(z) does not weaken relative to DR1. It strengthens. The tension with ΛCDM now sits around 3σ, with multiple independent datasets pointing to the same mild, oscillatory departure from a cosmological constant.

The caveat is real: some of this signal may reflect tensions between the CMB, BAO, and supernova datasets rather than new physics. But the fact that the evidence grows with more data rather than washing out is exactly what you’d expect if something genuine is happening. If Λ really is dynamical, this is how you’d first see it—not with a bang, but with a slow, persistent statistical accumulation across independent probes.

Worth watching: DESI’s full five-year dataset should be decisive.

On March 24, a threat actor called TeamPCP published two compromised versions of LiteLLM to PyPI—one of the most widely used Python libraries for routing LLM API calls. The attack chain is what makes this worth understanding: the attackers first compromised Trivy, an open-source security scanner, which LiteLLM’s CI/CD pipeline pulled from apt without a pinned version. The poisoned scanner exfiltrated PyPI publish tokens from GitHub Actions. From there, the malicious packages dropped a .pth file that executed on every Python process startup, harvesting environment variables, SSH keys, cloud credentials, and Kubernetes configs. Over 40,000 downloads in a three-hour window before PyPI quarantined it. The attack was discovered when the package was pulled as a transitive dependency by an MCP plugin running inside Cursor.

Researchers ran the classic behavioral economics experiments—the ones designed to document human irrationality in the Kahneman-Tversky tradition—on large language models across multiple model families and scales. The split result is the interesting part: on preference-based tasks (risk aversion, loss aversion, framing effects), models become more human-like as they scale up. On belief-based tasks (probability estimation, Bayesian updating), larger models tend toward rationality.

The implication is unsettling and fascinating: scaling language models on human text doesn’t converge on some platonic rationality. It converges on us—biases included—at least in domains where human preferences are the training signal. The practical upside: explicitly prompting models to “be rational” significantly reduces these biases. The deeper question: if these models are increasingly used for financial and economic decision support, whose irrationality are they inheriting?