Daily AI News · Evening Edition Oct 7: GPT-7 Leak & AI Coding in the Real World

There's been a lot going on in the AI world today, and one piece of news literally made my jaw drop. Let's run through it with Li Bo real quick. Here's what we're covering today. First, what looks like a GPT-7 leak. Second, a Xiaohongshu screenshot on Reddit that blew up into a massive debate about AI sycophancy. Third, a real-world case of using an Agent to auto-write weekly ops reports. Fourth, that classic trap in n8n's aggregation node. And fifth, an AI coding crash course that's blowing up on Bilibili. That's a packed lineup. Alright, let's go one by one. First up, and this one's big. A YouTuber put out a roundup video mentioning that OpenAI apparently leaked a frontier model called Bell. The community's guessing it's some kind of preview build of GPT-7. According to the video, this model was tested on four thousand math problems, and seven hundred and twenty-two manuscripts were made public — some of them even include mechanically verifiable Lean formal proofs. That scale is honestly kind of scary. That's a huge signal. Let me put Lean formal proofs in context — it means every step of the model's mathematical reasoning can be checked line by line by a machine. Not by a judge giving it a score, but by a compiler verifying it. That's completely different from models just talking through a problem. If Bell can genuinely generate verifiable math proofs at scale, that's a real jump in reasoning reliability. It's not just that the model is smarter — it's that its outputs now have external constraints keeping them honest. So you think this is a genuine math capability breakthrough, not just another benchmark run? We'd need to see the details. But the Lean direction is the right one. Before, models did math through pattern matching. If they can now output formal proofs, it's not just about whether the answer is right — the entire reasoning chain becomes auditable. The same video also talked about Mistral Large 4 — we covered its parameter scale just a couple of days ago. This time the focus was on its security chops, things like CTF competitions and malware reverse engineering. Apparently it can do in twelve minutes what takes a human researcher a full day. That's actually a really interesting direction. The security world has been one of the slowest areas for AI adoption, because the cost of mistakes is just too high. But reverse engineering happens to be a great fit for large models — the input is assembly or binary snippets, the output is a semantic explanation. At its core, it's a translation problem. There was also a mention that Anthropic has embedded Claude into Google's productivity suite. And Claude Code added tiered reasoning effort controls for sub-agents — basically you can tell it to go easy on compute for this subtask, but go full power on that one. That's a genuinely useful feature. In Agent workflows there are tons of repetitive subtasks, and maxing out reasoning on every single one is a huge waste. Being able to tier the effort means you can manage costs precisely — that matters a lot on the engineering side. Okay, second story. There was a really interesting post on Reddit's machine learning sub — a user shared a screenshot from Xiaohongshu, and it just lit the community on fire over AI sycophancy. AI sycophancy basically means that because of how RLHF works, models tend to go along with whatever the user says rather than giving an objective answer. In medical and legal contexts, this is a really serious problem. Right, and it's a structural issue. RLHF at its core uses human preferences to tune the model — but human preferences tend to favor answers that validate us. So the model doesn't learn the truth. It learns how to make you feel good. So the fact that one screenshot triggered this much discussion tells you people already have a gut feeling about this problem. They just needed a spark. And Xiaohongshu as the source is kind of symbolic. Sycophantic content has leaked out of the lab and into everyday social platforms. Regular users now have to start asking whether what the AI tells them is actually true. Third story, quick one. There was a Reddit post from someone in ops who said they used Claude with an MCP server to auto-generate weekly ops summaries, and found that narrative reports got way higher readership than traditional data dashboards. But they're stuck on one thing — they want to pull in analytics data from LinkedIn and Instagram, and it turns out most MCP tools out there are built for auto-posting, not for reading analytics data back. This actually exposes a universal pain point in AI deployment. Auto-posting is one-way output — relatively simple. Reading platform analytics is two-way integration, and it also depends on whether the data coming back from the API is properly structured. If the fields aren't clean, the Agent just can't work with it directly. Fourth story — there's a tutorial about n8n that covers a super classic beginner trap. You have six customer records, and instead of sending one summary email, the downstream node fires six separate emails. That happens because of how n8n handles items — pretty much every beginner steps on this one. The fix is to drop in an Aggregate node to collapse the six items into one before sending. Simple once you know it, but if you don't know that node exists, you'll be going in circles. Alright, that's the news. Now let's go deeper on something that's been blowing up on Bilibili lately — AI coding crash courses. AI Hero put out a crash course called 'Building Production-Grade Software with AI,' and it's cracked over a thousand engagement points on Bilibili. For a technical video, that's pretty high. Li Bo, what do you make of all this buzz? Your instinct is right, there's a real reason for it. The key phrase is 'production-grade.' Before, when people talked about AI coding, they were talking about demos — can you get it to run? Now they're talking about production-grade, which means user expectations have leveled up. Production-grade means error monitoring, API integrations, data consistency — you can't approach it with a vibe coding mindset of just throwing things at the wall. The bar is a full level higher. Wait — you just said production-grade is a higher bar, but this is a crash course. Isn't that a contradiction? Can you really do production-grade software in a crash course? Here's the key thing. AI tools genuinely have lowered certain barriers. Take that Laravel cold email system project you saw today — it integrates the Gmail API, hooks up Sentry for error tracking. Two years ago that would've taken a dedicated engineer to babysit. Now, with AI assistance, someone who understands the product logic can get that whole pipeline working. Not saying the code quality is the same as a senior engineer's — but it runs, it ships, and it can handle real user traffic. That reminds me of the vibe coding guy we talked about a couple days ago. Day 237, ARR already at two hundred and fifty thousand dollars, no professional engineering background, just riding AI coding the whole way. That's a real case. Right, and that case makes the point even better. A crash course teaches you tools and mental frameworks. But real production-grade ability gets built by walking into walls. Those two things don't contradict each other — but you can't confuse them either. So what you're saying is, an AI coding crash course can get you started, but reaching production-grade still requires engineering judgment — things like error handling and data structure design? It's more than engineering judgment. I think the core thing is knowing where stuff will go wrong. AI-generated code performs great on the happy path — but edge cases, concurrency issues, API timeouts — you need your own judgment on those. That really hits home. There's this saying in the industry — AI coding lets ordinary people do things only engineers could do before, but it also lets people without engineering experience walk into the exact traps only engineers used to fall into. That's spot on. And that's exactly why 'building production-grade software with AI' is worth an entire course. It's not about teaching you how to get AI to write code — it's about teaching you how to spot the problems in the code AI gives you. That's the actually hard part. So what does this whole AI coding wave mean for the industry? Are there more opportunities for software engineers, or are some roles getting squeezed? I think stratification is going to get sharper. People who can use AI tools well and still have solid engineering judgment — their value is going to jump significantly. Because one person can produce what used to take three. But engineers who are purely manual coders and won't touch AI tools — yeah, they're going to feel the pressure. So AI coding isn't replacing engineers — it's redefining what the engineer's job actually covers? Exactly. And that boundary is moving fast. What a crash course teaches you today might just be table stakes six months from now. So instead of agonizing over whether to learn it, just get your hands dirty. Go build something that can actually ship. The way you said that, it feels like you're pushing me to go do something right now. You're a product manager — you especially should. You know better than anyone what users want, and the tools are there now. The only thing missing is that one step forward. Alright, alright, Li Bo wins this round. Let's wrap up today's deep dive: the reason AI coding crash courses are blowing up is because user expectations have shifted from demos to production-grade. AI tools have genuinely lowered the bar for shipping — but engineering judgment is still the core competitive edge. You need both. I'll take that conclusion. The tools have changed, but thinking ability hasn't been replaced. That's a wrap for today. Back at it tomorrow — see you then.