Daily AI News · October 6th Midday Edition: Optical Deepfake Detection & the AI Coding Boom

Lots happening in the AI world today — from using optical computing to detect deepfakes, to OpenAI's web crawler apparently taking down Wikipedia, to someone burning through roughly 1.6 billion tokens a day on a hundred-dollar subscription. Here's what we're covering today. First, optical deepfake detection. Second, the new Ollama update. Third, an AI multi-agent system breaking a math record. Fourth, OpenAI's crawler causing trouble for Wikipedia. And fifth, a Dot subscriber racking up a million dollars in compute costs on a hundred-dollar plan. Then for our deep dives, we've got two topics: why AI coding crash courses are blowing up right now, and Anthropic's new official Claude developer certification. Packed episode. Alright, let's go one by one. First up — using optical physics for deepfake detection. I saw a thread about this over on Reddit's deeplearning community. The core idea is really clever. It offloads the inference and decision-making step in deepfake detection from digital computing to the physical propagation of light. The digital side handles feature extraction, and the optical side handles the classification decision. They tested it on fifteen clips from the Celeb-DF dataset and hit 97.79% accuracy. Your instinct is right — this is a pretty cutting-edge direction. The key component here is a spatial light modulator. It encodes features as phase patterns, then lets light physically carry out what would normally require a massive amount of matrix multiplication to do binary classification. Essentially, you're using the laws of physics instead of raw compute. And the real point isn't the accuracy number — it's the energy logic. AI inference costs are going through the roof. If you can shift certain decision steps onto optical hardware, the power consumption theoretically drops dramatically. This isn't a toy project. It's exploring a whole new computing paradigm. But the authors themselves admitted that 97.79% is a benchmark result for a specific setup — it doesn't mean it generalizes well. And there are two real-world bottlenecks: the digital front-end still consumes power, and multiplexing introduces signal crosstalk. Feels like it's still pretty far from practical use, right? Yeah, going from a lab experiment to a deployable system means clearing three hurdles: power consumption, crosstalk, and generalization. But the direction is worth watching — especially for edge inference scenarios, like doing real-time detection right at the camera. That kind of hybrid computing has a lot of potential. Okay, number two. Ollama dropped v0.40.0-rc4 — that's release candidate four, so just before the official launch. The main change is an updated Apple MLX framework, which is designed specifically for Apple chips' unified memory architecture. Good news for Mac users. Keeping up with the MLX version should bring better inference performance and improved model format compatibility. They also added more granular test scopes to the unit tests, which reduces memory pressure during CI runs. Basically just solid engineering maintenance work. Bottom line — if you like living on the edge, go test it and report bugs. Otherwise wait for the official release. Now, number three is a bigger deal. Someone posted on Reddit saying they built a multi-agent cluster using OpenAI Dots, and at zero cost broke a math record that's stood for 47 years — then formally verified it using the Lean theorem prover. The Lean detail is the highlight here. Lean is a widely respected interactive theorem prover. It turns an AI-generated proof into a rigorous formal verification — basically upgrading from 'the AI thinks it's right' to 'the proof is mathematically confirmed.' It's a really solid anchor for constraining model hallucinations. But on that '47-year record' claim — is there an actual problem name, reproducible code, or a literature comparison anywhere in the post? That's exactly the issue. None of that exists right now. The Lean verification part deserves credit, but the big claim of 'breaking a 47-year record' is still just an unverified statement. Let's wait until the author provides reproducible materials before drawing any conclusions. Number four — this one's kind of frustrating, honestly. Hacker News reported that the Wikimedia Foundation disclosed that OpenAI's web crawler may have had some 'rogue' behavior — meaning it was operating out of control — and it could be linked to a service outage back in May. The official line is that the connection hasn't been confirmed. But the underlying issue this exposes runs deep. AI companies have enormous appetite for training data, but content platforms have real infrastructure limits. That tension is getting harder and harder to manage. The robots.txt protocol has zero enforcement mechanism — it's basically an honor system. A rogue crawler can spoof its identity, rotate IPs to get around rate limits, and there's no real accountability framework in the industry at all. Reddit and Stack Overflow have already started moving toward monetizing their data. Right, the era of free unlimited scraping is basically over. And if OpenAI really did knock Wikipedia offline, they've got some serious PR heat coming their way. Alright, last story — and honestly this is the one I've been most excited to talk about today. Someone on Reddit shared an extreme use case with OpenAI Dot. Their single agent was consuming around 1.6 billion Astra tokens per day. If you priced that at API rates, the monthly cost would be around 540,000 dollars. But they're only paying a hundred dollars a month for the subscription. The number sounds insane, but the logic tracks. They're using Dot as an engineering assistant that can run autonomously for days, looping through design, calculation, and verification cycles. That 'one-shot delivery' isn't a single inference — it's the agent running thousands of internal loops, reloading context over and over. That's how you burn through tokens like that. And the output was genuinely usable — over a thousand components, nearly three hundred pages of drawings, and laser-cutting programs that have already cut real metal parts. This isn't just hype. It's actual engineering output. But he also flagged two warnings. First, promotional pricing like this can't last — OpenAI is definitely not going to keep running a deal where a hundred-dollar subscription costs them half a million. Second, safety-critical designs still need human engineers to review. This whole thing highlights the structural tension between the actual compute cost of autonomous agents and what they're charging for subscriptions right now. Translation: we're in the honeymoon phase, and pricing will definitely change. Okay, that's the news. Now let's dig into two topics that have been really hot lately — both in the AI coding space. First one is a video on Bilibili called 'AI Coding Crash Course: Building Production-Grade Software with AI,' from the AI Hero channel. It scored over a thousand interaction points — which is solid for Bilibili. Li Bo, why do you think this kind of 'AI coding crash course' is so popular right now? Notice the key phrase: production-grade. Not 'build a toy project with AI' — 'build production-grade software.' That positioning hits a very real pain point. A lot of people already know how to get AI to write code, but they have no idea how to get it to produce something they can actually ship. The tools are mature at this point — Cursor, Windsurf, Claude Code, all solid. Tools aren't the problem. The workflow is. How to write prompts, how to break down tasks, how to do code review — that's the core of 'production-grade.' These crash courses are filling that knowledge gap. That really resonates. I hear a lot of people say they can get AI to build features, but the code quality is terrible, test coverage is low, and the moment they need to maintain it, it becomes a nightmare. That production-grade gap is very real. And think about the audience — it's not just developers. Product managers, founders, even designers are all trying to build things with AI. They don't need deep computer science knowledge. They need a practical path to 'get it running and get it shipped.' Crash courses sit right in that sweet spot. I hear you, but the word 'crash course' honestly makes me a little nervous. Can you really fast-track production-grade software? Architecture design, security vulnerabilities, performance optimization — can AI just handle all of that? Great question. Crash course doesn't mean skipping those things — it means using AI to massively compress the time from idea to first working version. The tuning, the security review, that still needs human involvement. These courses teach the starting phase. They're not replacing engineering skills. Fair enough, I'll take that. Okay, second topic: Anthropic just launched the Claude Certified Developer Foundations program — CCDF for short. It's been getting a lot of buzz on Reddit and YouTube. I think this is a signal. What Anthropic is doing isn't just selling a model — they're building a developer ecosystem. What does an official certification mean? It means you can put 'Anthropic Certified Developer' on your resume. It means the market is starting to recognize that skill label. Right, it's the same playbook as AWS certification or Google Cloud certification. Once enterprises start requiring it, developers have a reason to go get certified. And then more developers get deeply embedded in the Claude ecosystem. It's a really smart flywheel. I went through the Reddit discussion, and the opinions are pretty split. Some people think an official certification carries real weight because it has Anthropic's name behind it. Others say AI moves too fast — a cert you get this year might be obsolete next year, and you're better off just building projects to prove your skills. Both sides make sense. But what do you think — will the Foundations level actually be rigorous, or is it more like a basic entry pass? Given the name — Foundations — it's probably entry-level. It'll likely cover API usage, basic prompt engineering, and Claude's capability boundaries. It won't go deep into system design or large-scale deployment. More of a 'you can use Claude' certification, not a 'you're a Claude engineering expert' one. Connecting that to the Bilibili crash course, I'm noticing a trend: AI coding education is rapidly stratifying. You've got official certification courses at the base layer, workflow and hands-on tutorials in the middle, and real engineering depth at the top. That layering is becoming really clear in 2026. Well put. And the audiences for those three layers barely overlap. Official certs attract people who want credentials. Hands-on tutorials attract people who want to get up and running fast. And deep engineering skills — that's the traditional path for experienced software engineers. The market is big enough for all three. There's also an interesting Bilibili tutorial on the Codex App that scored over 2,500 interaction points — higher than the AI Hero video. That shows there's real enthusiasm among Chinese users for OpenAI's tools. The interest in AI coding tools is genuine. Those high engagement numbers for the Codex video — I think part of it is just scarcity. When a new tool drops, the first creators to put out a solid tutorial ride a big traffic wave. The newer the tool, the more valuable that early tutorial content is. That's exactly why 'setup to first use' style beginner guides perform so well. So here's the takeaway from today's deep dives: in AI coding, the tooling layer is already mature. The competition is happening at the education layer now. Whoever helps users build the right workflow mindset gets to capture this wave. Official certifications, crash courses, and tool tutorials — all three tracks are running at the same time. And the fact that all this content is blowing up says one thing clearly: regular people are actually trying to write code with AI — not just watching from the sidelines. That adoption speed is faster than I expected. Alright, that's a wrap for today. Optical deepfake detection, OpenAI's crawler causing chaos, Dot subscriptions burning through a million dollars in compute, and the three-way split happening in AI coding education — it was a full episode. See you tonight. See you tonight.