California recently passed a law that will, in practice, cause AI chatbots to respond to any hint of emotional distress by spamming users with 988 crisis line numbers, or by cutting off the conversation entirely. The law requires chatbot providers to implement “a protocol for preventing the production of suicidal ideation” if they’re going to engage in mental health conversations at all, with liability waiting for any provider whose conversation is later linked to harm. New York is considering going further, with a bill that would simply ban chatbots from engaging in discussions “suited for licensed professionals.” Similar proposals are moving in other states.

If you’ve been reading Techdirt for any length of time, you know exactly what’s happening here. It’s the same moral panic playbook we’ve seen deployed against cyberbullying, then against social media, and now against generative AI. Something terrible happens. A handful of tragic stories emerge. Lawmakers, desperate to show they’re doing something, reach for the most visible technology in the room and start passing laws designed to stop it from doing whatever it was supposedly doing. The possibility that the technology might actually be helping more people than it’s hurting, or that the proposed fix might make things worse, rarely enters the conversation.

Professor Jess Miers and her student Ray Yeh had a terrific piece at Transformer last month that actually engages with the data and the incentive structures here, and their central argument may seem counterintuitive to many: the way to make AI chatbots safer for people in mental health distress might be to reduce liability for providers. For many people, I’m sure, that will sound backwards. That is, until you actually think through how the current liability regime shapes behavior — as well as reflect on what we know about Section 230’s liability regime in a different context.

Meta has returned to court in the US this week for the second phase of a lawsuit brought by Raúl Torrez, New Mexico’s attorney general, following a March verdict that found the company liable for child safety failures and imposed a $375m fine. On Monday, the state petitioned for a legal sanction against the company, a monetary penalty 10 times the original amount, and a sweeping, drastic overhaul of Meta’s child safety protocols.

In the second part of the landmark case, known as the remedies phase, the state is asking for Meta to be declared a public nuisance and for the judge to order the company to pay $3.7bn in an abatement plan. The money would fund programs for law enforcement, mental health services and educators. The state is also requesting that the judge force a series of design changes to Meta’s platforms aimed at improving child safety, including universal age verification, de-encryption of children’s messages, a guardian account linked to every child’s account, and a child safety monitor tasked with holding Meta to account for five years.

The New Mexico department of justice argues that these changes would make Meta’s social networks safer for underage users in the state. Meta, however, says the proposed reforms are unfeasible and could ultimately force it to shutdown its platforms in the state altogether.

New Mexico is not exactly a heavyweight, but it will be interesting to see how this plays out.

The good news, for me at least, is that the computer thinks I have a nice personality. According to an app called MorphCast, I was, in a recent meeting with my boss, generally “amused,” “determined,” and “interested,” though—sue me—occasionally “impatient.” MorphCast, you see, purports to glean insights into the depths and vagaries of human emotion using AI. It found that my affect was “positive” and “active,” as opposed to negative and/or passive. My attention was reasonably high. Also, the AI informed me that I wear glasses—revelatory!

The bad news is that software now purports to glean insights into the depths and vagaries of human emotion using AI, and it is coming to watch you. If it isn’t already: Morphcast, for example, has licensed its technology to a mental-health app, a program that monitors schoolchildren’s attention, and McDonald’s, which launched a promotional campaign in Portugal that scanned app users’ faces and offered them personalized coupons based on their (supposed) mood. It is one of many, many such companies doing similar work—the industry term is emotion AI or sometimes affective computing.

Some products analyze video of meetings or job interviews or focus groups; others listen to audio for pitch, tone, and word choice; still others can scan chat transcripts or emails and spit out a report about worker sentiment. Sometimes, the emotion AI is baked in as a feature in multiuse software, or sold as part of an expensive analytics package marketed to businesses. But it’s also available as a stand-alone product, and the barrier to entry is shin-high: I used MorphCast at no cost, taking advantage of a free trial, and with no special software. At no point was I compelled to ask my interlocutors if they consented to being analyzed in this way (though I did ask, because of my good personality).

So, there’s a piece in Jacobin arguing that data center moratoria are a “terrible idea” making the rounds on social media and beyond. It’s pretty easy to see why this makes for some good discourse; naturally, there’s going to be frisson among AI optimists when a perceived opponent—here, the nation’s most influential socialist magazine—makes a case for aligning with the tech industry’s goals.

While I’m pretty unconvinced on all but one or two of the points that the piece itself raises, and I think it seriously misconstrues the class politics of data center fights, I do think it’s worth litigating this idea. Because I do believe we should be thinking about what a broader and more engaged politics of resisting, regulating, and ultimately governing AI might look like. It’s a good occasion, in other words, to ask:

• Who is fighting data centers?
• Why are they fighting them?
• Are anti-data center movements a dead end—or a starting point?

The hardware efficiency gains are honestly the most interesting part of the paper. The main reason DeepSeek-V4 is so cheap to run comes down to how they completely bypassed the quadratic cost of standard attention for massive context windows.

They built a hybrid attention architecture that interleaves Compressed Sparse Attention and Heavily Compressed Attention. Standard models keep every single token in the KV cache which absolutely kills memory. CSA fixes this by compressing the KV cache of multiple tokens into a single entry and then uses a sparse routing mechanism to only compute attention over the top-k most relevant compressed blocks. HCA takes it a step further by compressing an even larger number of tokens into one entry but computes dense attention over them. So, a 1.6T parameter Pro model only uses a third of the compute FLOPs and 10% of the KV cache memory compared to DeepSeek-V3.2 at a one million token context.

They also aggressively pushed low-precision formats applying FP4 quantization-aware training to the Mixture-of-Experts weights and the attention Query-Key paths. MoE models are notoriously memory bound because you have to constantly shuttle massive expert weights into the GPU cores. Dropping these to FP4 slashes the memory bandwidth bottleneck and lets the model run way faster during inference without ruining accuracy since they handle the quantization dynamically during training.

On the infrastructure side they wrote a custom fused kernel using TileLang that overlaps communication and computation. When running expert parallelism across multiple GPUs you usually hit a wall waiting for the network. DeepSeek slices the experts into micro-waves so the GPU is crunching matrix math on the first wave while the network is simultaneously pulling the data for the second wave. They basically hid the network latency behind the compute time which means you do not need super expensive interconnects to get peak hardware utilization out of the cluster.