OpenAI just pointed an autonomous agent pipeline at the Linux kernel. It came back with 948 working local privilege-escalation exploits. In one run the agents: → Identified 2,308 potential security flaws → Successfully reproduced 1,724 of them → Developed working exploits for 948 distinct vulns that let limited access escalate to full root → Generated candidate patches for human review Total cost: ~$1.21 million in API credits (~$1,279 per distinct exploited vulnerability). For scale: their earlier July system (powered by GPT-5.6 Sol) found and exploited 166. This one did 948. Linux maintainers had already patched 61 upstream by the time the research wrapped; more were in preparation or under discussion. This is the shift from “AI helps write code” to “AI agents systematically hunt, reproduce, and propose fixes for security holes in the world’s most critical software at industrial scale.” Finding flaws is getting cheaper and faster. Triaging and fixing them at that volume is now the bottleneck. Controlled research environment. Existing foothold required. Expert review still needed before any patches ship. But the volume signal is unmistakable.

Inception just shipped Mercury 3, the most capable diffusion LLM released so far. It outperforms GPT-6 Luna and Claude Haiku 5.5 on quality across a range of benchmarks while running at 1,068 tokens per second on NVIDIA GPUs, 4–8× faster than typical speed-optimized autoregressive models. 1M context (up from 260K on Mercury 2.5). More than 40× larger than Mercury 1, with ~3× the intelligence on standard benchmarks, yet still just as fast. Pricing: $0.20 / $0.70 per 1M input/output (launch promo 50% off on some platforms: $0.10 / $0.35). OpenAI API compatible. Available today on Inception API, OpenRouter, Baseten, and Vercel. Enterprise options for dedicated capacity, etc. Early partners are already using it for browser/web agents, real-time tutoring, search agents, and support. One production note: it completed 50% of live Mind2Web navigation tasks (better than GPT-6.1-Sol) at the same token volume and steps but far lower per-token cost. Speed keeps agents from compounding latency across steps.

xAI just dropped Grok Voice 3.0 and it’s already resolving 70% of Starlink support calls with zero humans in the loop. It tops Artificial Analysis’s Speech-to-Speech Index at 86.9, handles interruptions and background noise better than prior versions, and responds in under a second. In Tesla vehicles it’s powering hands-free calendar checks, email replies, and even ordering coffee or booking dinner. One agent can verify a customer, diagnose hardware, issue a credit, and ship a replacement. At $0.08/min this is the cheapest way yet to get frontier-level voice agents that actually finish the job. The gap between “voice chat” and “voice that closes tickets” just got a lot smaller.

NVIDIA just shipped Nemotron 3.5 Bolt 74B-A9B (and companions), an open-weight 74B MoE model (only ~9B active) that fits in 24 GB via aggressive 2-bit quantization-aware training. It’s built for always-on local AI agents: browser/app navigation, coding, research, tool use, and background automation. Full weights, training data, and recipes are open. Works with Ollama, llama.cpp, vLLM, LM Studio, Hermes, OpenClaw, etc. Pair it with the new Nemotron 3.5 Super VL (120B total / 12B active, multimodal for long-running visual + text agents) and the tiny Bolt E4B (4 GB, for conversational/game NPCs, coming Oct 23). This lands the same day RTX Spark laptops go on sale (Oct 16). Local frontier-ish agents just got a lot more practical. Open weights + consumer hardware is the quiet shift that compounds.

German robotics unicorn NEURA just unveiled Neuron AI, their multimodal foundation model for Physical AI. It fuses 3D vision + touch + audio + robot state + language + memory in a Perception-Language-Action (PLA) architecture. Robots can react to things they can’t see and remember context even after objects leave the camera view. One model runs across five different NEURA embodiments (including their 4NE1 humanoid). Trained on real robot data, simulation, and human demos. Customers fine-tune and deploy skills fleet-wide via Neuraverse. Live demo next week at NVIDIA GTC Berlin: their 4NE1 humanoid will do a fully autonomous 11-minute laundry sequence (sort → washer → dryer) with zero human intervention. “Building the brain of the physical world,” per the CEO. Physical AI is heating up fast.

Amazon is in talks to acquire Decart for around $7 billion, per WSJ. The 3-year-old Israeli-founded startup makes software that optimizes AI models to run faster and cheaper across chips (Nvidia, Amazon Trainium, Google TPUs) plus real-time generative models (Lucy for live video, Oasis for simulated environments). AWS already gave them early Trainium access. Valued at ~$4B in May; Anthropic previously explored a ~$6B deal and walked after diligence. A close would be one of Amazon’s larger buys and a direct play to reduce Nvidia dependence. Chip optimization just became a multi-billion-dollar exit.

Meta just dropped its first real frontier models: Muse Wonder and Muse Wonder Full. Fourteen months after overhauling its AI efforts and hiring Alexandr Wang, Meta is finally claiming it has reached the frontier. These models are purpose-built for long-running personal and professional agents, handling insurance claims, leases, finances, health info, or building full apps, not just chat. The standout features are Active Memory (carries what matters about you across days, weeks, and months) and Active Focus (keeps grinding toward the goal, recovers from blockers, and tries new approaches). They’re trained at the gigawatt-scale Prometheus data center and support a 1M-token context with multimodal input. On Artificial Analysis’ Intelligence Index, Muse Wonder Full scores 56 and Muse Wonder scores 52, a material jump from Muse Spark’s 48 just six weeks earlier. Wonder Full lands near the top on several agentic and knowledge-work benchmarks, competitive with (and in places ahead of) recent OpenAI and Anthropic models. Muse Wonder is available now via the Meta Model API and Muse Code at $4/$20 per million tokens. The stronger Full version is locked to Muse Code subscribers. Both will soon power Meta’s Muse personal agent. After years of Llama open-source focus and recent closed-model pivots, this is Meta’s strongest signal yet that it intends to compete directly for the agentic future. Independent testing will tell how the claims hold up outside Meta’s own runs.

One thing I’ve been paying attention to with @Wahoopredict is how prediction markets can evolve beyond simply choosing between “yes” or “no.” The upcoming Multiple Selection feature is interesting in that regard. Instead of forcing a prediction into a binary outcome, it can give users more ways to express where they think the probability actually sits. That becomes even more interesting as WahooPredict covers markets across crypto, sports, stocks, commodities and real-world events. Prediction markets are ultimately about expressing conviction under uncertainty. The better the market design, the more accurately people can communicate that conviction. So I’m curious: will Multiple Selection become one of the features that makes WahooPredict meaningfully different from other prediction platforms?

Fall seven times. Rise eight. @0xhazels literally carries that philosophy through its Studio. You don't just enter. You connect. You participate. You create. You get counted. The GTD Race turns contribution into points. Pulse is designed to turn creator work into measurable rewards. Drops turns staking into access. And the NFT becomes the permanent thread connecting the ecosystem. That's why I don't see Hazels as just another NFT collection. It's trying to turn participation itself into an asset. And that's a much harder thing to build. But also a much more interesting thing to watch.

get on this 👇🏿