On May 7th, OpenAI quietly launched GPT-5.5 Cyber through something called the Trusted Access for Cyber (TAC) program. This is not your standard ChatGPT upgrade with a fresh coat of paint. GPT-5.5 Cyber is a model specifically designed to perform tasks that would ordinarily trip every safety system in existence — pen testing, red teaming, exploitability validation, and offensive security research.

Here’s the twist: you can’t just download it. You have to be vetted. Think of it like a very exclusive nightclub, except instead of a bouncer checking your ID, it’s OpenAI making sure you’re a legitimate security researcher and not someone who’s about to use their AI to hack a power grid.

Four months. We’re not talking Moore’s Law anymore. We are in full chaos mode. The UK’s AI Safety Institute dropped that stat like a casual observation, and it’s worth sitting with for a moment: AI that can find and exploit security vulnerabilities is evolving faster than our collective ability to defend against it.

Anthropic also has a cybersecurity-capable model in this space — reportedly nicknamed “Mythos”, a name that prompted India’s finance ministry to issue an actual cybersecurity warning to its banking sector. Nothing says “trustworthy financial infrastructure” quite like naming your AI after Greek myths involving divine retribution. The chef’s kiss practically served itself.

Here is a sentence that would have been considered science fiction eighteen months ago: the Chicago Mercantile Exchange (CME) is launching a futures market for AI compute. Starting this year, you can speculate on GPU capacity the same way traders bet on crude oil or corn harvests. Companies are buying contracts to lock in computing power six months from now. Hedge funds are getting involved. It is, as JR put it, “peak late-stage tech capitalism.”

We have gone from “the cloud is just someone else’s computer” to “someone else’s computer is now a leveraged financial derivative.” The logic, though, isn’t entirely absurd. Compute is the new oil. Consider the evidence: Anthropic just committed $200 billion to Google Cloud. Meta acquired an entire robotics AI startup — ARI — primarily to get closer to humanoid robot development. Every major player is hoarding GPU capacity like it’s the last toilet paper on the shelf in March 2020.

When something becomes this scarce and this strategically important, Wall Street will find a way to put a ticker symbol on it. The surprising part isn’t that compute futures exist — it’s that it took this long.

In March, Meta released Tribe Version 2 — a foundational AI model trained to predict how your brain responds to sights, sounds, and language. Not metaphorically. Literally. The model was trained on over 1,115 hours of fMRI brain scans from more than 700 volunteers, and it can now simulate neural activity with 70× better resolution than any previous brain-modeling approach.

Here’s what that means in practice: feed it a movie clip, an audiobook, or a piece of music, and Tribe V2 will generate a map of what your brain would be doing if you were experiencing it — without ever putting you in an MRI machine. It’s a virtual brain experiment, running in software.

The part that makes this genuinely extraordinary isn’t the resolution improvement — it’s the zero-shot generalization. Tribe V2 can predict brain activity for people it has never seen before, across languages and cognitive tasks it wasn’t explicitly trained on. That’s the difference between a system that memorized a dataset and one that actually understood something about how human brains process information.

Meta open-sourced the model entirely, which means neuroscience researchers around the world can now run virtual brain experiments that previously would have required millions of dollars in MRI scanner time. This is genuinely good news for clinical research, cognitive science, and our understanding of how human perception actually works.

While the world’s largest AI companies were busy setting compute on fire trying to scale their way to better models, a Stanford undergrad quietly did something much more interesting: they figured out why large language models generalize at all — and turned that insight into a 5× training speed improvement.

No billion-dollar compute cluster. No 10,000-GPU training runs. No venture capital. Just math, a dorm room, and the apparent refusal to accept “we don’t really know why this works” as a satisfying answer. The result is a new optimizer — a more efficient method for updating model weights during training — that achieves the same results in one-fifth the compute time.

The implications are significant. Training efficiency improvements are rare, and a genuine 5× speedup means that the next generation of capable AI models could be built by significantly smaller teams with significantly smaller budgets. The democratization argument for AI just got a lot more plausible — and it came from a student, not a corporate research lab.

Three rapid-fire concepts from this week’s stories, explained plainly.