A Startup Claims It Broke the LLM Bottleneck, and BCI Trials Are Finally Taking Off

Let’s talk about two stories that actually matter this week. First up: a startup called Subquadratic just came out of stealth claiming they’ve cracked a bottleneck that’s been plaguing large language models for almost a decade.

Here’s the pitch: they’ve figured out how to slash the number of computations transformers need to generate answers. The result? A faster, cheaper LLM that uses way less energy than anything else on the market. That sounds like magic, and honestly, most experts are still skeptical. But Subquadratic has started sharing some receipts, which is more than most vaporware startups do.

How does it work? I’m not going to pretend I understand the math at a deep level, but the gist is they’re attacking the quadratic complexity problem that makes transformers so expensive. Standard attention mechanisms scale quadratically with sequence length, which is why long contexts cost a fortune. Subquadratic claims they’ve broken that scaling law. If true, it’s huge. If not, it’s another footnote in the long history of overhyped AI breakthroughs.

Will Douglas Heaven covered this for MIT Tech Review, and he captures the tension well. Some researchers aren’t convinced, and they have good reasons. We’ve seen this movie before. Remember when everyone thought sparse attention was going to save us? Or when ReLU was going to be replaced by something better? The field is littered with claims that didn’t pan out. But Subquadratic is at least putting their work out there for scrutiny, which is more than most.

The other big story: brain-computer interfaces are finally moving from lab curiosities to real-world trials. Jessica Hamzelou wrote about Casey Harrell, a man with ALS who’s been called “the first power user” of a brain implant. The guy is using it to maintain an income, reconnect with friends, read to his daughter. He called it “nothing short of revolutionary.” I don’t use that word lightly, but for someone with ALS, it probably is.

The numbers back up the hype. BCI trial volunteers have soared over the past couple of years. China just became the first country to approve a BCI for medical use. Engineers are cramming more features into these devices than ever before. We’re watching the technology edge from the lab toward the market, and it’s happening faster than most people realize.

Of course, there’s a long way to go. These are still invasive implants with all the risks that entails. But the trajectory is clear. We’re past the point where BCIs are just science fiction.

A few other things caught my eye this week:

Amazon workers who backed limits on data center expansion may be facing termination. The engineers say they’re under investigation, and the company isn’t messing around. They testified at meetings about pausing data centers, and now they’re looking at discipline, possibly firing. They’ve filed a complaint with Seattle’s Office for Civil Rights. This is the kind of story that makes you wonder how much pushback tech workers can actually afford to give.

A new fossil discovery has rewritten 150 years of evolutionary theory. Apparently early land vertebrates skipped the tadpole stage entirely. That’s a big deal if you care about how life crawled out of the water. Sponges may have been the first animals, which is a whole other rabbit hole.

And Bernie Sanders is proposing something genuinely interesting: direct public ownership of AI firms. His new legislation would create an AI sovereign wealth fund, funded by a one-time tax on AI companies’ stock. I don’t know if it’ll go anywhere, but it’s a more creative approach than the usual regulatory hand-wringing.

That’s the roundup. No grand conclusions here. Just a lot of interesting stuff happening.