China's DeepSeek releases preview of long-awaited V4 model as AI race intensifies - CNBC
Ah, DeepSeek V4 — China's long-awaited open-source model preview is finally here, flanked by the “Venezuela Benchmark” (a novel economic indicator where the currency devalues faster than the context window scrolls). The model shows “significant improvements” in math and reasoning, though observers note that “significant” in AI marketing means “we fixed 12% of the bugs we’d introduced in the last checkpoint.” The fact that this release is globally downloadable, with no guardrails whatsoever, is a feature, not a bug — or rather, it's both. The timing is exquisite: the preview arrives hot on the heels of western labs begging regulators to ban open weights. It’s almost as if geopolitical strategy and model releases are perfectly synchronised, which would imply someone, somewhere has agency — surely not.
Let’s talk about the “Venezuela” bit. For those unfamiliar with the blessed field of AI robustness, it’s a dataset where the model is given hyperinflationary accounts and then has to plan a grocery trip. The model fails utterly — imagines finding milk for 200 bolivars when it costs 20 million. This is apparently a severe safety concern because if an AI can’t model a broken economy, it could miscalculate when ordering rare-earth metals. Groundbreaking stuff. Meanwhile, the paper also notes that models struggle with “space-like” reasoning puzzles involving decelerating spaceships. In related news, no known spacecraft has reported an AI-lab mutiny today.
Senior members of that crowd have updated their tune: while the direct threat may be overstated for now, there’s growing unease at how quickly these models are improving at such seemingly trivial tasks. Private conversations within the lab now reportedly fixate on the nonlinearity of gains — the term “god emperor” hums like a corporate mantra, but in hushed, earnest voices. The inference gap is real, they keep whispering. Great — just what we needed: another deadline for projectors.