The field notes
The Daily AI Drop - Evening Edition, October 13, 2026
Alex and Jordan break down eleven afternoon AI videos: the Muse Wonder versus Opus 5.5 showdown, OpenAI's 700 AI-written math papers, Microsoft Discovery's science agents, Gemini building a 3D cathedral in Antigravity, ChatGPT Space for teams, ten AI agents given real jobs, and the arrest of an OpenAI engineer over a hidden recording device.
Eleven new videos dropped this afternoon: a benchmark showdown over Meta's Muse Wonder, an unreleased OpenAI model writing 700 math papers, Microsoft's agentic platform for science, Gemini building a 3D cathedral, ChatGPT's new team workspace, ten AI agents handed real jobs, a one-minute harness explainer, a tiny vision model pushed to its limit, and the arrest of an OpenAI engineer over a hidden recording device.
NEW Muse Wonder Opus 5.5 Killer? — Mehul Mohan
Watch: Watch source video ↗
Mehul puts the big claim about Meta's new frontier model to the test: is Muse Wonder actually a killer of Anthropic's Claude Opus 5.5? Meta's own charts show Muse Wonder Full beating Opus 5.5 on just over half of Meta's internal tests, but the independent Artificial Analysis Intelligence Index still scores Opus 5.5 at 58 against Wonder Full's 56 (standard Wonder: 52). It matters because the purchasable Wonder is priced exactly like Opus 5.5 — $4/$20 per million tokens — while the headline-grabbing Full version is only available to Muse Code subscribers. Company charts versus independent measurement: the "killer" question is really about pricing and access, not just benchmarks.
OpenAI's AI Wrote 700 Math Papers. Engineers Still Won't Trust AI With Their Code. — Nate B. Jones
Watch: Watch source video ↗
OpenAI pointed an unreleased model at roughly 4,000 open math problems and posted 719 AI-generated manuscripts (372 families of results) to GitHub in a single October 6 drop. Mathematicians reacted with open hostility — Oxford's Ursula Martin called it a messy first draft tossed at the community; Columbia's Michael Harris refused to read them — and OpenAI withdrew three papers within a day after a sign error invalidated an argument, revising fourteen more. It matters because this is narrow superintelligence made concrete, yet engineers still keep coding agents out of production: math has a near-free verifier (a proof checks or it doesn't) while production code has to survive real users. The question for your own work: what is your verifier?
Speed up Research with Microsoft Discovery — Microsoft Developer
Watch: Watch source video ↗
Microsoft Discovery coordinates specialized AI agents — some analyze research evidence, some propose candidate materials, some run scientific tools, some validate results — in a continuous loop that shifts toward the most promising candidates as new evidence arrives. Product manager Ali Zaman walks through battery design as the example. It matters because this is the agent-harness formula working in the hardest domain: specialized agents with clear roles, tools that ground them in reality, and a human scientist setting the goals and making the key calls. Fewer physical experiments, faster materials science.
Fewer Experiments, Faster Discovery — Microsoft Developer
Watch: Watch source video ↗
A shorter cut from the Microsoft Discovery conversation: how AI agents help scientists narrow an enormous research space down to the candidates worth actual lab time. It matters as the crispest statement of the idea — agents explore, evidence accumulates, the plan adapts, humans decide. Pair it with the full conversation above.
Gemini 4 Argon and Google Antigravity: From 2D concept to 3D cathedral — Google Developers
Watch: Watch source video ↗
Gemini 4 Argon — the model that topped the Vals Index last week — running inside Google's Antigravity agentic development platform, taking a flat 2D concept and building it into a 3D cathedral. It matters as the applied answer to the benchmark debates: inside a real multi-agent harness, can the model turn a vague creative concept into something with genuine depth and structure? If agents can build a cathedral from a sketch, the "agents can't do real creative work" objection gets much weaker.
Work with your team in ChatGPT Space — OpenAI
Watch: Watch source video ↗
ChatGPT Space (launched at DevDay on September 29) is OpenAI's answer to the copy-paste problem: a shared workspace where teammates, ChatGPT, and personal Dot agents all work from the same project files, instead of chatting solo and pasting output into docs. Inside a Space, Pages are documents that people and agents edit together — charts, checklists, code, images — rolling out to Pro, Business, and Enterprise. It matters because ChatGPT is becoming the place where work lives, not just where answers come from — aimed squarely at Notion and Google Docs territory.
Block on Finding and Fixing Security Risks with OpenAI — OpenAI
Watch: Watch source video ↗
Block — the fintech company behind Square and Cash App — shares how it finds and fixes security risks with OpenAI. It matters because this is enterprise AI in its mature phase: not pilots or demos, but a payments company trusting AI with security operations. Security is quietly becoming one of the strongest production use cases for agents, because it has a built-in verifier — the vulnerability is either fixed or it isn't.
I gave 10 AI agents real jobs. Here's what they did — The Next New Thing
Watch: Watch source video ↗
Ten AI agents, real jobs, real results reported back. It matters because the industry keeps theorizing about the agent-employee era — this is someone actually running the experiment and showing what happened. If you're wondering whether agents with job descriptions is hype or here, this is the video to watch.
Agent harnesses explained in 1 minute — LangChain
Watch: Watch source video ↗
Sixty seconds on what an agent harness actually is. It matters as the explainer to send the colleague who keeps asking — after a week of deep dives on the Claude Code harness, LangChain's managed agents, and Microsoft Discovery, this is the one-minute version of the week's biggest theme: the harness is the product, the model is the commodity.
MiniCPM-V 4.7 1B: Install with Ollama and llama.cpp, Then Push It to the Limit — Fahd Mirza
Watch: Watch source video ↗
Fahd installs the 1-billion-parameter MiniCPM-V 4.7 vision model locally with Ollama and llama.cpp, then pushes it to its limits to find where it breaks. It matters as the companion to last night's 35B test: how much intelligence can you fit on hardware you own? A 1B vision model that does anything useful is a genuine milestone for private, offline AI.
OpenAI Employee Arrested Over Recording Device — Matthew Berman
Watch: Watch source video ↗
The FBI arrested OpenAI product engineer Abdulrahman Alfozan, 35, a Saudi national at the company since July 2025, on October 9 — charged with using a device to intercept oral communications. On September 30, a routine sweep found an encrypted audio recording device hidden in an executive conference room: inserted into the USB port on the back of a TV, concealed in toilet paper, believed planted around September 28. OpenAI says it used its own AI models to identify who planted it within days, then referred the case to the FBI; more devices were reportedly found in his backpack, and the company believes he meant to plant them across its offices. It matters because frontier AI capability is now a strategic asset — and strategic assets attract insider threats — and because OpenAI reports no indication of model-weight, source-code, or customer-data exfiltration. The lab racing toward superintelligence is learning in public what it costs to protect it.
Coverage note: YouTube's RSS feeds returned 404/500 errors for 7 of the 15 channels this afternoon (they parsed fine at 7 AM — an endpoint block, not bad IDs), so those channels' videos were verified via each channel's videos page in the live browser instead. Siraj Raval's newest video is 7 days old — nothing new today. All 15 channels were checked.