InApp : Tech Radar

The most valuable private companies in the world are running into a wall. They spend years prepping for an IPO, but once they actually hit the public market, they can't just slide into the S&P 500. There's a regulatory lag that keeps them on the sidelines, and it's creating a weird disconnect between a company's actual market cap and its eligibility for the index. I've watched this happen a few times now. A company goes public with a valuation that dwarfs almost everything else in its sector, yet it stays invisible to the passive funds that drive the bulk of market volume. It's a frustrating quirk of the system. We're basically pretending these companies aren't giant just because they haven't checked a specific box regarding profitability or listing duration. The real problem is that this isn't just a clerical annoyance. It changes how these companies price their shares and how they handle their first few quarters of public life. It ...

When the world's top computer science students start failing basic math, the problem isn't the AI. It's the erosion of the mental models required to actually use it. Look at the numbers from UC Berkeley. In spring 2026, 6% of students in CS 61A received Fs. For a course that serves as the gateway to a degree at one of the best engineering schools on earth, that's a loud signal. These aren't students who can't handle the material. They're students who have outsourced the "thinking" part of coding to a LLM and realized too late that they've forgotten how to debug their own logic. I've seen this cycle before with calculators and IDEs, but this feels different. We're not just automating syntax. We're automating the struggle that actually creates a programmer. If we stop valuing the friction of learning, we're just training a generation of operators who can't fix the machine when it breaks. I want to look at where exact...

We're confusing the ability to mimic human language with the presence of a human mind. It's a mistake we're making on a massive scale, and no one is leaning into it quite as hard as Anthropic. The company is a giant in the AI space, but their real specialty might be anthropomorphism. Take "Claude’s Constitution," an 84-page document that reads more like a moral manifesto for a sentient being than a set of alignment constraints for a probability engine. Then there's CEO Dario Amodei, who recently admitted in an interview that he's "open to the idea" that AI could be conscious. No. Absolutely not. The distance between a sophisticated token predictor and a conscious entity isn't a gap we can bridge with more parameters or a better prompt. It's a fundamental category error. I want to look at why we're so eager to believe the ghost in the machine, and why that impulse is actually dangerous for how we build these tools. The Illu...

Speed is a feature, but only if the model doesn't sacrifice the logic of your codebase to get there. We've all seen the "fast" models that hallucinate a library that doesn't exist just to finish a sentence. It's frustrating. The goal isn't just to get code on the screen faster, it's to get code that actually compiles without a ten minute debugging session. Microsoft is trying to solve this with MAI-Code-1-Flash. It's a model built from the ground up using clean, licensed data, which is a nice change from the legal grey areas we usually deal with in LLM training. More interestingly, it's designed specifically for the GitHub Copilot harness. The idea is that the model shouldn't just act as a fancy autocomplete, but as part of an agentic workflow that understands the environment it's actually operating in. The real question is whether this specialization actually translates to better code, or if we're just getting the wrong an...

Modern email clients have stopped being tools and started acting like assistants who won't stop interrupting you. We've traded actual control for "smart" features that usually just create more noise. It's a frustrating trade. I saw it happen this morning in Gmail. I opened the web UI to check for feedback on a project, but as soon as I clicked compose, a colorful animation hijacked my focus. It was there to highlight a new "help me write" button. I didn't ask for a writing coach, I just wanted to send a message. The problem is that these features are rarely about efficiency. They're about engagement metrics. When the interface starts fighting you for attention, the tool is no longer serving the user. I'm curious if there's any way to actually get a clean, quiet inbox again, or if we're just stuck with a UI that treats every single email like a chance to pitch a new AI feature. The Illusion of the Smart Inbox Priority inbo...

Most people looking at the Dav2d error logs see a generic bandwidth warning and move on. They assume it's just a standard rate-limiting hiccup, a temporary bottleneck caused by a sudden spike in traffic. But if you look closer at the specific error—the one citing the 160,000 daily file action limit—you'll see something else is happening. The system isn't just hitting a ceiling; it's hitting a wall built by a very specific structural logic. Dav2d handles data distribution in a way that makes certain types of requests fundamentally more expensive than others. When you hit that limit, it's usually because the way the data is being requested is triggering a cascade of file actions that the architecture wasn't designed to sustain. I spent the last week digging into these logs to figure out why this limit feels so much tighter than it should be. It turns out the way we're thinking about data distribution in this context might be fundamentally flawed. Core Pr...