The AI bust is here — and it's real.

A new paper from OpenAI and researchers at the University of California, Berkeley maps out how to evaluate language models before they hit production. The authors walk through the full pipeline: picking the right test set, choosing evaluation metrics, running benchmarks, and checking for failure modes like hallucination, bias, and prompt injection. They flag that accuracy isn't the only metric — latency, cost per token, and safety all matter for real systems.

What stands out is the part most teams skip: the adversarial test set. You don't just run the model on normal prompts; you run it on the prompts it's likely to get wrong — edge cases, multi-step reasoning, edge-case inputs. They show that models which look fine on standard benchmarks can still fail hard in production. The paper also covers how to detect when a model is confidently wrong, and how to set up human-in-the-loop review for the cases that need it.

This isn't theory. The authors have tested dozens of models using this framework and the differences are real — some models score well on standard benchmarks but collapse on adversarial tests, others are slower and more expensive for marginal gains. The takeaway is practical: if you're building with LLMs, you need a test set that looks like what your users will actually send, not just the standard leaderboards.

Why this matters for us: small shops and solo builders using OpenAI or Anthropic APIs are the ones most likely to skip these tests — and the ones who get burned when the model starts hallucinating in front of real customers.

DoorDash has put an LLM-based moderation platform into production to handle the flood of user-generated content — reviews, photos, and messages — that a food delivery app of this scale generates every day. The system runs on a standard open-source model and acts as a first pass, flagging content that crosses the line into harassment, fraud, or policy violations before a human ever sees it.

The architecture is straightforward: the model evaluates content on a handful of policy dimensions, scores each one, and sends flagged items to a queue for human review. The idea is to catch the low-hanging fruit at machine speed so the moderation team isn't drowning in spam, fake reviews, or the occasional bad actor trying to game the system. DoorDash has been public about its reliance on AI for content moderation for a while now; this is the next step — a dedicated system rather than a patchwork of ad-hoc classifiers.

What's notable is the honesty in the write-up: the LLM is a first filter, not a replacement for judgment. The human-in-the-loop is the real system. This is how most companies are actually doing it — models do the heavy lifting, people handle the edge cases, and the business keeps moving. DoorDash is one of the bigger examples of this pattern applied to a consumer-facing platform with millions of users.

Why this matters for us: la gente uses these apps every day, and when the moderation works, it keeps the platform usable; when it doesn't, it's our money and our time on the line.

Elizabethtai walks through her typical day using LLMs to handle the parts of writing that eat up time — research, outlines, drafts, editing — while she stays on top of the judgment calls: voice, accuracy, what actually matters to readers. The post is a log of the workflow, not a sales pitch.

It's the kind of practical breakdown the big tech announcements are missing. Meta just scrapped its AI-native pivot, cutting 60 percent of its teams. Salesforce shoved its entire CRM into Claude and claims you'll never need the app again. Meanwhile a writer in the Bay Area is just trying to get the next post out the door without burning out. The real story is what's happening at the ground level, not the boardroom.

Why this matters for us: la gente in Brown and Black communities are the ones actually writing, coding, and building with AI — and the folks making the money on it rarely mention them.

Bluesky rolled out an update to Attie, its AI assistant, turning it into something you can actually use to dig into conversations across the AT Protocol. Instead of just answering a single question, it now pulls together news, trends, and what people are saying on Bluesky and other apps that speak the same language.

The AT Protocol is the open network Bluesky sits on — think of it as the plumbing under the social feed. By making Attie work across those apps, the company is treating the protocol like a searchable commons rather than a walled garden. That matters because the old social feeds were designed to keep you scrolling. This is the opposite: it's built to answer a question and let you go.

Why this matters for us: if you've been tired of algorithms feeding you outrage, this is one of those moves that quietly puts the search back in social.

An indictment unsealed Friday charges Volkswagen engineers with insider trading tied to a joint venture with Rivian. Prosecutors allege they used confidential information about the deal to buy stock in the electric-vehicle maker before the market got wind of it. The DOJ filed the charges in the Southern District of New York.

The joint venture is the kind of thing that moves markets — VW brings cash and manufacturing muscle, Rivian brings the EV know-how. When that deal is still locked down inside the companies, the people who know about it are sitting on a gold mine. These engineers apparently took advantage.

Insider trading charges like this aren't new, but the fact that the DOJ is pursuing it years later shows they're keeping a close eye on cross-border deals and the folks who profit from them.

Why this matters for us: When people with inside info on deals between big companies make a quick buck, it's another way the system is rigged against the rest of us who don't have access.