Plus: Reflection's open-weight Beam model, Manus raises $500M, and Nadella says assume every AI model is compromised
Sponsored by
Create scroll-stopping videos, product visuals, and social content with Pippit AI. Turn ideas, links, and images into ready-to-share marketing assets in minutes.
Try Pippit for freeToday, we've got OpenAI dumping 372 math proofs on the internet (and mathematicians dumping on OpenAI right back), an Nvidia-backed startup shipping America's answer to Chinese open models, Manus pulling half a billion dollars out of a deal Beijing killed, and Satya Nadella telling you to assume your AI is already compromised.
On this day in 2025…
The Guardian ran a story on "workslop" — the flood of confident, plausible, quietly-wrong output AI tools were handing US employees, with researchers pinning the blame not on the models but on employers who deployed them without training anyone. One year later, we're having almost exactly the same argument about AI-generated mathematics. Progress!
🧮 OpenAI Drops 372 Math Proofs — And Starts a Fight
On October 6, OpenAI published 372 significant proof claims across pure mathematics in a public GitHub repo, produced by an internal frontier model and shipped with Lean certificates, reasoning summaries, and compute estimates (the average result burned about three hours of ChatGPT Pro thinking). Thomas Bloom, who runs erdosproblems.com, confirmed 23 of them touch real Erdős problems and called all 23 genuinely interesting. Not everyone clapped: Terence Tao argued that dropping this many proofs at once makes mathematics less fertile, and a newly formed mathematicians' group accused OpenAI of trampling scientific norms. OpenAI says it consulted the Institute for Advanced Study first. Either way, this is the biggest AI-in-math moment yet.
🔓 Reflection AI Unveils Beam, America's Open-Weight Comeback
Beam is a 501B-parameter sparse MoE with just 23B active per token, built for coding and agentic work, with a 1M effective context and Apache 2.0 weights landing later this month. Reflection claims it matches far larger open models like GLM-5.2 while using three to four times less inference compute, and posts 80.9 on SWE-bench Verified. The subtext is geopolitical: a16z reckons 80% of developers using open models worldwide are building on Chinese ones. Beam is the Nvidia-backed pitch to change that.
💰 Manus Raises $500M+ After Beijing Killed Its Meta Exit
Butterfly Effect, the parent of AI agent startup Manus, closed over $500 million led by Boyu Capital and IDG Capital, with Tencent, HSG and ZhenFund following on — reportedly at a $4 billion valuation, roughly double the $2B Meta had agreed to pay before Chinese regulators forced the deal unwound. Manus has since shipped Manus 2.0 on a new in-house execution system called Cascade, plus Cue, an app where each personal agent gets its own email address, phone number and wallet. Investors clearly aren't spooked.
🚨 Nadella: Assume Every AI Model Is Compromised
Microsoft's CEO used a long weekend post to argue we can no longer treat AI as "a set of nested black boxes" whose advice we simply accept or reject. He wants containment, observability, tamper-proof human-readable evidence trails, and what amounts to an emergency brake — framing frontier models as something closer to an insider security threat. Convenient framing for Microsoft, which risks getting squeezed as labs turn their models into the primary computer interface. Still, the architecture argument is a serious one.
🛡️ Microsoft Execution Containers (MXC) went generally available this week and is MIT-licensed on GitHub. It's a policy-driven containment layer for agents: you declare the files and network destinations a workload may touch, and the OS enforces that boundary — crucially, from outside the agent's control, so generated code can't grant itself more access. GitHub Copilot, OpenAI Codex, OpenClaw, Replit, LM Studio and Unsloth already support it, with Claude Code, Perplexity, Manus and Raycast on the way.
🧠 llama.cpp added native support for decision models via a new /v1/systemone endpoint: send it a state (text, JSON, even a screenshot) plus typed questions, and it returns calibrated probabilities per option in a single forward pass — no output tokens generated. Median answer times run 3–43ms depending on model, which makes local request routing, content moderation and agent action selection genuinely practical.
Coming in Hot
Borrow distribution, not capital. Real people grow your app's reach.
AI image generator with top prompt accuracy
Multi-channel cold outreach with brand intelligence
AI search visibility platform for brand optimization
AI website builder for small businesses in 30 seconds
Build a website in seconds with AI
Managed cloud hosting for Claude Code and Codex agents
AI social media scheduler for creators and marketers
AI prompt generator for creative workflows
Conversational forms and surveys that convert
Submit your AI tool to the #1 AI tools marketplace.
What it does: ProductClank is a community-powered growth platform that helps startups promote products through creator campaigns, social engagement, and authentic word-of-mouth marketing.
Why It's Worth Your Time
Pricing: ProductClank offers a free starting option. Paid plans start at $29/month for Solo (800 credits), $79/month for Starter (3,000 credits), $199/month for Growth (12,000 credits), and $499/month for Pro (30,000 credits). One-off credit packs are also available.
Why it matters: ProductClank helps startups reach potential customers through real creators and community engagement rather than relying entirely on paid advertising. It combines campaign management, creator discovery, and AI content tools to support product growth.
What it does: Pippit AI is an AI-powered content creation platform that helps businesses turn product links, prompts, and images into marketing videos, product photos, and social media content.
Why It's Worth Your Time
Pricing: Pippit offers a free plan with daily credits and access to selected video and image generation features. Paid annual plans start at US$101/year for Starter (regularly US$168/year), US$246/year for Plus (regularly US$410/year), and US$2,196/year for Pro (regularly US$3,660/year). Paid plans include additional credits, faster generation, access to advanced AI models, and watermark removal. Features and credit limits vary by plan.
Why it matters: Pippit combines AI content creation, editing, and social media publishing in one platform, helping businesses produce marketing content with less manual work.
Cockroach Labs spent five months running their codebase like a hospital: bugs as patients, coding agents as the medical team, with triage and rounds. A genuinely useful read on structuring agent work instead of just throwing prompts at it.
After Trump declared AI should be called "super intelligence," a surprising number of tech's most powerful figures started complying — Musk promptly promising to rename SpaceXAI to SpaceXSI. Everyone else is still saying "AI."
