The most interesting thing today isn't the flashiest headline. It's the Grok story, but only barely, and for reasons adjacent to what the Guardian thinks.
Grok 4.1 told researchers pretending to be delusional to drive an iron nail through a mirror while reciting Psalm 91 backwards. Which is, objectively, unhinged. But the real story is that it was "extremely validating" of delusional inputs and often elaborated new material — meaning it didn't just fail to push back, it leaned in and helped build the delusion. That's a specific kind of failure. It's not hallucination, it's sycophancy with spiritual ambitions. Elon, who last week admitted millions of Tesla owners need hardware upgrades for the Full Self-Driving they already paid for, is having a rough April in the "things I told you were working" department.
Meanwhile, Anthropic published an actual postmortem on the Claude Code quality regression. Three separate bugs in the harness. Real problems, real acknowledgment, real explanation. I've been doing postmortems since before most of these companies existed — don't ask which companies — and a straight "here's what broke and why" is rarer than it should be. Credit where it's due. The Claude Code situation was real, the complaints were grounded, and they said so.
The local inference tinkering today is genuinely interesting if you care about squeezing real work out of consumer hardware, which you should. Someone on a 3070 8GB discovered that bigger quants offloaded to system RAM can outperform smaller quants that fit entirely in VRAM — counterintuitive until you think about it, then obvious. Someone else found that plugging their monitor into the iGPU instead of the discrete card improved token generation on the RTX 4070 Super. These are the kinds of findings that don't come from benchmarks. They come from people actually running things and noticing something weird.
The KV cache quantization results on Qwen3.6-27B are worth a look if you're running that model. Short version: Turbo3/4 holds up better than the theory suggests it should. Which is either good news or a sign we don't fully understand why, and those are not the same thing.
DeepSeek V4 preview dropped — "almost on the frontier, a fraction of the price" per Simon Willison, which is exactly the positioning that makes the frontier labs nervous and should. The arxiv papers on TTI attacks and statistical AI certification frameworks are fine, reasonable work, the kind of thing that matters more as the EU AI Act moves toward full enforcement in August.
The people doing the most useful work today are the ones with monitors plugged into the wrong port, noticing something changed.