☀️ TRENDING AI NEWS
🤖 Alibaba releases Qwen3.6-27B, a dense open-weight model that outperforms 397B MoE models on agentic coding benchmarks
🏢 Pentagon requests $54B for AI-powered autonomous drone warfare program in 2027 budget
🛠️ OpenAI launches custom workspace agents for Business, Enterprise, and Edu teams in ChatGPT
⚠️ Meta installs keystroke-tracking tool on employee computers to train its AI models
A 27-billion-parameter model just beat a 397-billion-parameter one. That sentence would have been absurd 18 months ago. Today it's a Tuesday for Alibaba's Qwen team - and it tells you everything about where open-source AI is headed right now.
Let's get into it.
🤓 AI Trivia
Which company originally developed the table tennis robot named Ace that made headlines this week for beating elite human players?
🏓 Boston Dynamics
🏓 Sony AI
🏓 DeepMind
🏓 Sanctuary AI
The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇
🤖 Alibaba's 27B Model Just Humbled a 397B Giant
Alibaba's Qwen team has released Qwen3.6-27B, and the benchmark numbers are turning heads. The model outperforms a 397-billion-parameter MoE architecture on agentic coding tasks - which means this relatively compact model is beating something 14 times its size where it counts most for developers.
Hybrid Architecture Does the Heavy Lifting
The secret sauce is a hybrid architecture that blends Gated DeltaNet linear attention with traditional self-attention. It also introduces a Thinking Preservation mechanism - a technique designed to keep the model's reasoning coherent across long agentic tasks, which is where smaller models typically fall apart.
For developers who have been nervous about running large models locally or paying for API access, Qwen3.6-27B is the kind of open-weight release that genuinely shifts the decision. If you're tracking coding agents and developer tools, this is worth a close look. And if you're building something fast, 60sec.site can help you spin up a landing page for your project in under a minute using AI.
⚠️ The Pentagon Wants $54B for Autonomous Drone Warfare
The US Department of Defense has released its 2027 budget request, and the headline figure is striking: over $54 billion earmarked for the Defense Autonomy program - a more than hundredfold increase from previous funding levels. The budget represents the clearest signal yet that military AI is no longer a side project.
Autonomous Drones at the Center of the Strategy
The budget specifically outlines funding for AI-powered autonomous drone warfare. Military experts quoted by The Guardian are raising concerns that the Pentagon is moving faster than its own safety protocols can keep up with - essentially deploying systems before fully understanding their failure modes.
This sits in uncomfortable proximity to the ongoing debates around AI safety and AI regulation. When a senator warns of an AI-fueled financial bubble and the military is requesting $54B for autonomous weapons in the same week, the scale of what's being built becomes hard to ignore.
🏢 Meta Is Recording Everything You Do at Your Desk
Meta is installing a tool called Model Capability Initiative (MCI) on US-based employees' computers. It runs in the background of work apps and websites, recording mouse movements, clicks, keystrokes, and periodic screenshots. All of it feeds back into training Meta's AI models.
Your Workflow Becomes Training Data
The idea is to capture real-world, expert human behavior in work contexts - how engineers navigate codebases, how analysts move through spreadsheets - and use it to make AI agents better at those same tasks. It's a logical play, but it raises immediate questions about employee trust and privacy rights.
Employees reportedly weren't consulted before rollout. This is the tension at the core of AI agents right now: to make them truly useful, companies need training data that mirrors real expert behavior. But gathering that data requires watching people work - constantly.
🤖 Sony's Ping-Pong Robot Just Beat the Best Humans
Sony AI's robot, named Ace, has become the first machine to beat elite table tennis players under official tournament rules. Ace won three out of five matches against top-ranked human competitors - a result that researchers are calling a genuine milestone in robotics.
What Makes This Different From Past Robot Athletes
Previous table tennis robots, like Omron's FORPHEUS, were impressive demos but only competed with amateur players or under modified rules. Ace is different: it operates under standard regulations, meaning it has to handle the full speed, spin, and unpredictability of elite human play. The visual tracking and real-time reaction systems required to do that represent a significant engineering leap.
It's a useful reminder that AI progress isn't only happening in language models. Physical intelligence - the ability to perceive and act in the real world at speed - is advancing fast too.
⚠️ A Wall Street Law Firm Filed AI Hallucinations in Court
Elite Wall Street firm Sullivan & Cromwell has apologized to a New York federal judge after admitting a major filing contained errors generated by AI hallucinations. The filing was part of a high-profile case involving the Prince Group. The firm's co-head of global litigation, Andrew Dietderich, addressed the court directly.
The Cautionary Tale Keeps Writing Itself
This isn't the first time a law firm has been caught submitting AI-generated citations that don't exist - but Sullivan & Cromwell is about as blue-chip as legal practices get. If it's happening there, it's happening everywhere.
The incident is likely to reignite conversations about legal technology oversight and what counts as acceptable AI use in high-stakes professional contexts. For anyone in enterprise AI deployments: this is exactly the kind of story your legal team is going to bring up in the next meeting.
🏢 Thinking Machines Lab Signs Multibillion-Dollar Google Cloud Deal
Mira Murati's Thinking Machines Lab - the AI research company she founded after leaving OpenAI - has signed a multibillion-dollar infrastructure deal with Google Cloud, according to TechCrunch. The deal is powered by Nvidia's latest GB300 chips, which represents a significant hardware commitment for a lab that's still in relatively early stages.
What's notable here is the tech partnerships angle: Google Cloud is increasingly positioning itself as the infrastructure partner of choice for frontier AI labs, locking in relationships early. This follows a pattern we're seeing across the industry - compute access is becoming the strategic moat, and cloud providers know it. For more on the AI infrastructure arms race, check out our coverage at Daily Inference.
🌎 Trivia Reveal
The answer is Sony AI! Ace was developed by Sony's AI division and made history as the first robot to beat elite table tennis players under official tournament rules - winning 3 out of 5 matches. Previous robots like Omron's FORPHEUS only competed with amateurs or under modified conditions.
💬 Quick Question
Meta is literally recording employee keystrokes to train its AI. If your employer told you they were doing this - would you be fine with it, or would it be a dealbreaker? Hit reply and let me know your honest take. I read every response.
That's it for today. See you tomorrow with more from the frontier - stay sharp out there. 👋