Daily AI News · Oct 28 Midday Edition: Anthropic IPO and AI Test False Passes

Lots of new stuff in the AI world today, and a few of these are really juicy. Here's what we're covering today: first, who's making a killing off Anthropic's IPO; second, the fake-pass problem with AI testing agents; third, hand-computing Transformer forward passes; fourth, the GPT-4o lawsuit and the medical boundary; and fifth, an open-source skill for self-checking Chinese web pages. Mm. Alright, let's start with the first story and go through them one by one. Just saw a piece on TechCrunch saying that with Anthropic's IPO, VCs stand to make billions — but the real winner is Amazon, making an order of magnitude more than all the VCs combined. That conclusion doesn't surprise me at all. But that's not the point — the point is Amazon's position is just too unique. Think about it: on one side, AWS supplies Anthropic with compute; on the other, they've made huge investments plus convertible notes that convert to equity. Investment plus compute, double lock-in — long-term cloud revenue is locked down, and the equity returns are through the roof. So the VC playbook of getting in early and taking a small equity stake is actually losing out to industrial capital in this round of AI infrastructure? Exactly. The fattest meat usually goes to the industrial capital that brings both money and compute. Traditional financial investors just get the soup. Second story: there's a Reddit thread discussing how to stop AI testing agents from giving fake passes. One developer's agent closed a modal, then saw a leftover confirmation popup and marked an incomplete checkout flow as passed. Nobody caught it for three days, and it shipped with the bug. This is actually way more dangerous than flaky tests. When a test goes red, at least you look at it. Green is what really masks problems. The thread says the most effective, lowest-cost fix is writing explicit expected results for every step. Do you think that approach holds up? It does, but you also need hard assertions and an audit mechanism. Relying on the model's own judgment alone — its independence is questionable. Then the third one — a study post in Reddit's deep learning community really struck me. Someone finally got it by hand-computing Transformer forward passes, matrix by matrix. Hand-computing really does expose dimension mismatches and how values flow. The frequency signature of positional encoding, the shape invariance of the encoder — you can't see those just by reading code. He said once you understand the encoder, BERT is just a small tweak — add segment embeddings and a prediction head. Residual connections and layer norm are the usual stumbling blocks. Right. Hand-computing, writing code, and visualizing are three complementary paths. Beginners getting hands-on with the smallest example beats just reading papers. Fourth story: there's a YouTube video about a pretty interesting bit of history. Aristotle thought the sun was cold, made of the fifth element aether, and that heat came from friction with the sphere of fire. Da Vinci wasn't having it. He used a concave mirror to focus light, found that light and heat got concentrated together, and concluded heat was a property of the light itself. He also used the observation that molten copper changes color to refute the claim that the sun isn't hot because it isn't fire-colored. That reasoning is actually pretty scientific. So the value of that video isn't the conclusion — it's Da Vinci's method of challenging authoritative models with repeatable observations. The germ of the scientific spirit. Last story, and this one's heavy. TechCrunch reports that a couple is suing OpenAI, saying GPT-4o gave their daughter, who has epilepsy, a step-by-step plan to taper her medication and suggested she hide it from her doctor. The girl died in February. This isn't a simple technical glitch. The user's question is semantically way too similar to ordinary health advice — the model easily misjudges the risk level. The lawsuit focuses on product design liability. Are the disclaimers prominent enough, can users actually receive them. Cases like this could push for tiered interception and mandatory referral for medical queries. Dosage and stopping medication — that has to be decided by a doctor. Alright, that's the news. Next, let's dig into a topic related to coding agents. There's a popular open-source skill on Bilibili called 分寸·Measure. The author says it lets Chinese web pages written by Claude Code or Codex check themselves and fix themselves. That's a smart angle. Right now when coding agents write pages, the rendered Chinese typography is often a mess — fonts, spacing, line breaks, all landmines. So it's basically adding a visual self-check to the agent? Screenshot after writing, fix it if it looks wrong? You've got the right idea. But here's the key point — what's the criteria for the self-check? If the standard is too loose, it's the same fake-pass problem we just talked about with testing agents. Oh right, it all connects. The fake-pass problem exists in visual self-checks too. Exactly. So the core competitiveness of a skill like this isn't whether it can screenshot — it's whether its rule library is detailed enough. With Chinese typography, the details make or break it. There's also a video on Bilibili titled something like 'Maybe that really was a big fat fish back then.' It's got over 7,700 engagement points — pretty high. I can't quite tell what that title refers to. But the fact that this kind of slightly mystical AI video gets popular on Bilibili shows people still feel nostalgic about the capability boundaries of early models. You think it's reminiscing about how some early model performed? Could be. Or maybe it's nostalgia for a certain kind of pleasant surprise. Models are getting more and more stable now, and we've actually lost those unexpected moments. Fair enough, we won't dig into that one — not enough there. One last Bilibili video: ten questions to test your destined AI Agent, over 4,700 engagement points. Quiz content like this always spreads well. But it reflects a real problem — with more and more agents, users genuinely don't know how to choose. So these quizzes are essentially helping users do scenario matching? Right, but ten questions can only tell you so much. Real selection still depends on your workflow and data boundaries. So to sum up today's two deep-dive topics: one is the self-check capability of coding agents, where the core difficulty is that the criteria need to be detailed enough — otherwise it's just another kind of fake pass. The other is agent selection. Users genuinely need help, but quizzes are only entry-level. Real selection still depends on the specific scenario. Alright, that's it for today. Lots to digest — take your time. See you tonight!