Daily AI Fresh News · October 20 Evening Edition: COBOL Comeback & Traffic Wars

Lots of new stuff in the AI world today, and it spans a pretty wide range. From ancient COBOL to website traffic defense, there's a bit of everything. Today we're mainly covering these: first, OpenAI's big COBOL leap; second, ChatGPT booking flights; third, n8n workflows for small businesses; fourth, the GPT-6.5 leak rumor; fifth, Blackwall's valuation quadrupling; sixth, the Bilibili Agent tutorial; seventh, Cloudflare's Clairic. And an eighth one: voice AI for beginners. Yeah, eight items—that's enough to keep us talking for a while. So let's start with the news roundup, one by one. Li Bo, you ready? Let's go. Today's batch of material is pretty interesting. First one: I just saw a YouTube video with a pretty explosive title. It says OpenAI pushed COBOL accuracy from 5.4% to 86.6%, and also threw in Anthropic's free bug scanner. Those numbers—I have to say, the point isn't the two version numbers, it's COBOL. Think about it—what era is COBOL from? Banks, insurance, government systems are all full of it. 5.4% basically means it can't do it at all; 86.6% means it can actually get to work. So you're saying the real value here is all those old systems still running COBOL? Exactly. The people maintaining those old systems are nearing retirement age on average. If a model can read COBOL, it's basically giving those systems a new lease on life—that commercial value is way more real than bumping up a few benchmark scores. Got it. And what about that free bug scanner from Anthropic? By the way, that combo is kind of interesting. Right—on one hand, letting the model understand old code; on the other, scanning bugs for free. That's clearly a play to win over enterprise developers' minds. The strategy is very clear. Okay, second item, also a YouTube short. The title is 'ChatGPT can't book your flight, but an AI agent can.' See, the contrast is pretty blunt: chatbots can only chat, agents can actually get things done for you. A lot of people still haven't wrapped their heads around this. From conversation to execution—that's the real watershed this year. Being able to chat isn't worth much; being able to get things done is. Alright, the third one is even more down-to-earth. There's a 60-second YouTube short about three free n8n workflows that every small business needs. That kind of content is actually super practical. Most small business owners have no idea what an agent is, but if you say it'll auto-reply to emails and auto-bookkeep, they get it immediately. So when it comes to real-world implementation, you still have to start from specific processes, right? Exactly. No matter how powerful a general large model is, it ultimately has to be packaged into this kind of ready-to-use workflow to make it into a small business's books. Then the fourth one—this one deserves a question mark. There's a YouTube video title screaming 'GPT-6.5 massive leak,' and it also throws in Claude Fable 5.5, Gemini 4 Carbon, MiniMax M3.1 Flash, and Kimi K3.1. That kind of title is obviously clickbait. A bunch of version numbers piled together sounds lively, but if you look closely, it's basically speculation and patchwork. So we'll just touch on it and not spread it further. Right, videos with version numbers flying everywhere—just take a look. When the real release happens, official channels are way more reliable. Fifth item, this one's from TechCrunch. Blackwall raised $100 million at a $2.2 billion valuation—that's quadrupled. What they do is help websites tell traffic apart: which is normal search engine and AI crawler traffic, and which is malicious scraping and DDoS. I think this is the most thought-provoking item today. You see, bots and human behavior are getting more and more alike, rule-based filtering is basically failing, so you have to rely on behavior analysis plus machine learning. So the real reason this space is hot is that AI crawlers have messed up the internet's traffic structure? That's exactly the logic. Content owners want legitimate crawlers in but also need to block malicious traffic—that balance is getting harder and harder, so money naturally flows there. Alright, that's it for the news. Next we'll dive into two hot topics: one is that 500-episode Agent tutorial on Bilibili, and the other is Cloudflare's Clairic. Sure, these two go well together—one teaches you how to use it, the other talks about how to look at it. Let's start with the Bilibili one. The title is 'Bilibili's most complete Agent zero-to-hero tutorial in 26 years, 500 episodes, from LLM to MCP to RAG, get a job after finishing.' It has over 870 interaction points—that's pretty hot for Bilibili's scale. Let me pour a little cold water on that 500-episode number. The logic of these tutorials is basically to break the entry barrier super low, using quantity to create a sense of security. But isn't that exactly what a lot of people need? Beginners are most afraid of not knowing where to start. The need is real, sure, but 500 episodes doesn't equal 500 knowledge points. A lot of content is repetitive, just padding the time. As for 'get a job after finishing,' I think you have to take that with a grain of salt. I don't know—I think don't be so harsh. Look, it connects LLM, MCP, and RAG together, at least giving newbies a complete path. There is a path, but my worry is that after watching, you think you've learned it. With Agent stuff, if you don't actually write code, watching 500 episodes or 50 episodes makes little difference. I agree with that. No matter how much input, without output it's all for nothing. So my take is, this kind of tutorial is great as an index, but dangerous as a bible. The real value is in typing along once, then building a small project yourself. Alright, let's talk about the second one: Cloudflare's Clairic. There's an article on dev.to with a really catchy title: 'Clairic doesn't write output tokens, it reads value—don't guess.' That title actually points to a key shift. In the past, when evaluating models, people looked at how many tokens it wrote, how many words it spat out. So what does 'doesn't write output tokens' mean? Does it not generate text? My understanding is it shifts the focus from how much content is generated to the value of the output. 'Read value, don't guess'—that's a phrase worth pondering. This is actually in the same vein as that earlier thing about running eight use cases through one gateway as a benchmark—both are rethinking how to evaluate models. Right, you see, someone previously used a unified gateway plus fixed use cases as a benchmark, and was forced to issue two corrections. That shows evaluation itself isn't reliable. So Clairic's approach is trying to pull evaluation from the crude token-counting method toward looking at actual value? I think that's the direction. But the hard part is right there: how do you quantify value? Who sets the standard? That's much harder than counting tokens. Then let's keep a disagreement. You think the direction is right but hard to implement; I think the direction is right and someone has to take the first step anyway. Alright, I accept that disagreement. But to sum up: model evaluation is shifting from competing on token count to competing on actual value, and Clairic has sort of put this issue on the table. Right, and that voice AI for beginners item actually confirms this trend—everyone's moving toward being able to listen, speak, and do real work. Exactly. Whether it's agents, traffic management, or evaluation methods, this year's main thread is clear: AI is moving from being able to talk to being able to get things done. Alright, that's it for today. Lots of information—see you tomorrow.