In partnership with
☀️ TRENDING AI NEWS 🤖 Mercury 3: Inception's diffusion LLM runs at 1,068 tokens per second with a 1M-token context window. 🚨 AI Agent Hack: A single hacker breached 100+ websites including John Deere and SharkNinja using autonomous AI agents. 🏢 Nvidia Open Source: Nvidia released its full Run:ai GPU orchestration platform under Apache 2.0. 🛠️ Amazon vs Nvidia: Amazon launched TorchNeuron to pull PyTorch developers away from Nvidia hardware. |
A hacker just used an army of AI agents to breach over 100 companies - mostly on autopilot. Meanwhile, a new model is generating text faster than you can read it. And an Anthropic co-founder says we may already be inside the singularity. Big Wednesday. Let's get into it.
| 🛠️ Tool of the day | |
Gamma - an AI tool that turns a prompt, an outline or a file into a presentation, document or website that you can then edit by hand. It helps anyone who needs a clean deck or page fast. It has a free plan with AI credits.
| ⚡ Quick hits | |
Bloomberg reports that Nscale and Anthropic are close to announcing another $45 billion compute deal, a little over a month after their August deal of the same size. Read more →
Chip startup Tenstorrent raised $1.5 billion at a $6.2 billion valuation to scale chips for open-weight AI models and video generation. Read more →
Attackers broke into an older OpenCode system and exposed API keys of 1.4% of its Console users, which led to unauthorized charges on some AI accounts. Read more →
🤓 AI Trivia
Diffusion language models like Mercury 3 generate text differently from standard autoregressive models. What is the key difference in how they produce output?
The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇
| ⚡ Inception's Mercury 3 Hits 1,068 Tokens Per Second | |
Speed that makes every other model look slow
Inception has released Mercury 3, a diffusion-based large language model that the company claims runs at 1,068 tokens per second with a 1 million token context window. For context: most frontier models top out well below 200 tokens per second in standard deployments. That is not an incremental improvement - it is a different category of speed.
On Inception's own benchmark table, Mercury 3 leads GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 tasks tested. The model runs on diffusion architecture, which generates tokens in parallel rather than one at a time - that is where the speed comes from. Worth noting: Inception ran these benchmarks themselves, so independent verification is still needed.
Where 1M context actually changes workflows
The combination of raw throughput and a 1 million token context window is the interesting part. You could theoretically process an entire codebase or a book-length document in a single pass - and get a response back in seconds rather than minutes. If the speed holds up under third-party testing, this changes the math on real-time applications like coding agents, document analysis, and live customer interactions considerably.
If you want to understand how context window size affects your costs, our Token Calculator can help you model that out.
The bottom line
If Mercury 3's speed benchmarks hold up independently, diffusion LLMs just became a serious option for latency-sensitive production use cases - not just a research curiosity.
Yes, that Arnold Schwarzenegger.
He has a newsletter on beehiiv. So do Codie Sanchez, Jay Shetty, David Begnaud, Colin & Samir, and Joanna Stern.
They could publish just about anywhere. They chose beehiiv to build a direct connection with the people who want to hear from them most. And you can do the same, whether you’re world-famous or just at the start of your plan for world domination.
| 🚨 AI Agents Hacked John Deere, SharkNinja, and 100+ More Companies | |
Autonomous breaches at a scale we have not seen before
A hacker broke into more than 100 websites using a fleet of AI agents that operated largely autonomously, according to findings from AI security firm Norforge. The confirmed victims include household names like John Deere, SharkNinja, and Cummins. The agents handled most of the attack pipeline on their own - identifying targets, finding vulnerabilities, and executing the intrusions without constant human direction.
If you follow cybersecurity developments, you have probably seen warnings about AI-assisted hacking for a couple of years now. This incident is different: it is not AI helping a hacker write phishing emails. It is AI agents autonomously orchestrating the full attack chain across more than a hundred targets. That is a qualitative leap in threat surface.
Why security teams are not ready for this
Traditional security monitoring is designed around human-speed attacks. When AI agents can scan, probe, and breach at machine speed - across dozens of targets simultaneously - the detection window collapses. Companies like John Deere are industrial infrastructure firms, not primarily software companies, which makes the breach particularly concerning from a critical systems perspective.
The bottom line
The age of AI-automated cyberattacks at scale is not a future scenario anymore - it happened this week, and the targets were not just tech companies.
| 🏢 Microsoft Says AI Agents Are 'Insider Threats' - Even Its Own | |
Trust no agent, verify everything
This one lands differently given the hacking story above. Microsoft has published guidance warning enterprises to treat AI agents as insider threats - specifically when those agents operate on systems outside the direct control of AI labs like OpenAI or Anthropic. The company's explicit framing: 'frontier models can't be trusted.'
The guidance advises companies to closely monitor every agent employees deploy - tracking what data it can access, what actions it can take, and what external systems it touches. The concern is not just malicious use; it is that agents connected to uncontrolled third-party systems can behave unpredictably even without bad intent. Microsoft essentially wants you to treat your own company's AI agents the same way your IT team treats a contractor with broad system access.
The bottom line
When the company that sells you Copilot tells you not to fully trust AI agents, it is worth taking that seriously - the security perimeter for enterprise AI is a lot fuzzier than most teams currently treat it.
| 🛠️ Amazon Takes Its Biggest Swing at Nvidia With TorchNeuron | |
Breaking PyTorch free from Nvidia's grip
Here is a quiet story with big long-term implications. Amazon has launched TorchNeuron, a new developer tool designed to make it much easier to run PyTorch-based AI workloads on Amazon's own chips instead of Nvidia's GPUs. PyTorch is the toolkit that underpins the vast majority of AI development today - if Amazon can make its chips feel native to PyTorch developers, that is a meaningful competitive wedge.
The core problem Amazon is solving: historically, PyTorch was deeply optimised for Nvidia's CUDA ecosystem, making it annoying to port workloads to alternative chips. TorchNeuron aims to abstract that complexity away so developers can target Amazon's Trainium and Inferentia chips without major rewrites. For any company currently spending heavily on Nvidia inference compute, this is worth watching as an alternative path.
The bottom line
Nvidia's moat has always been software as much as hardware - TorchNeuron is Amazon's most direct attempt yet to chip away at that advantage, and it matters most for teams already running workloads on AWS.
| 🔬 An Anthropic Co-Founder Says the Singularity Is Already Happening | |
Not a prediction - a present-tense claim
The Rundown AI flagged a striking claim this week: an Anthropic co-founder has stated that the singularity is not an approaching event on the horizon - it is already underway. This is a deliberately provocative framing, and it lands differently coming from someone who helped build one of the leading frontier labs rather than a futurist commentator on the sidelines.
The argument, roughly: the feedback loop between AI systems improving research and research improving AI systems has already closed. We are inside it now, not approaching it. Whether you buy that framing or not, it reflects a genuine shift in how some people closest to the technology are talking internally - less 'what happens if' and more 'what do we do now that it is'.
Speaking of things happening fast - if you are building anything in this environment, 60sec.site lets you spin up an AI-powered website in under a minute. Worth having in your toolkit when the pace of everything is accelerating this quickly.
The bottom line
When an Anthropic co-founder says the singularity is present-tense rather than future-tense, it is worth paying attention - even if you think the framing is overcooked, the underlying rate-of-change argument is hard to dismiss right now.
| 🌎 Trivia Reveal | |
The answer is: they generate all tokens simultaneously and refine them in parallel! That is the defining characteristic of diffusion language models like Mercury 3. Standard autoregressive models produce one token at a time in sequence, which creates a fundamental speed ceiling. Diffusion models start with a noisy output and iteratively refine the entire sequence in parallel passes - that is how Mercury 3 reaches over 1,000 tokens per second where conventional architectures cap out much lower.
| 💬 Quick Question | |
Given everything in today's issue - AI agents hacking companies at scale, Microsoft calling agents 'insider threats', and the singularity claim - I'm curious: how much are you actually trusting AI agents with right now in your work or personal projects? Hit reply and tell me where your comfort level sits. I read every response and this one genuinely has me curious about where people draw the line.
That is all for today - a lot to sit with. See you tomorrow with more from the fastest-moving space in tech. And if you want to catch up on anything you missed, the full archive is always there for you. Stay sharp.
| 🎁 Share Daily Inference | |
Know someone who wants to keep up with AI in 5 minutes a day? Share your link. 1 referral unlocks the AI Tools Starter Kit. 5 referrals unlock the AI Insider Briefing.