A Free Tool Just Stripped the Guardrails Off Meta and Google's Open Models in 10 Minutes

The FT just published an investigation into a tool called Heretic. It’s free, sits on GitHub, requires no specialist hardware, and strips the safety guardrails off open-weight AI models in under ten minutes. Since its release late last year, it has been used to produce more than 3,500 decensored models, downloaded over 13 million times. The creator stripped safeguards from Google’s Gemma 4 within 90 minutes of its release. ...

June 6, 2026 · 2 min · 424 words · bjr

The Pope Wrote an Encyclical About AI. And Brought Anthropic With Him.

Pope Leo XIV’s first encyclical, Magnifica Humanitas, was released on May 25. It is entirely about artificial intelligence. The Pope issues encyclicals rarely, and the choice of subject for a first one is a deliberate signal. Dedicating it to AI tells you how seriously the Catholic Church is taking what is coming. Leo frames AI as the Industrial Revolution of our time, with the same potential to reshape labour, dignity, and the structure of society. The encyclical leans on two biblical metaphors: the Tower of Babel for unchecked technological ambition, and the rebuilding of Jerusalem’s walls for collaborative human effort. Its core position is that technology is never neutral. It carries the values of whoever builds, funds, and deploys it. ...

June 6, 2026 · 3 min · 453 words · bjr

Nvidia Just Did for Humanoids What It Did for Cloud AI

Nvidia announced the Isaac GR00T Reference Humanoid Robot at GTC Taipei. It’s an open hardware and software stack for academic research, built around the Unitree H2 Plus chassis, dual Sharpa Wave tactile hands, multi-view sensing, and a Jetson AGX Thor T5000 onboard. The software side is the full Isaac suite: Teleop for data capture, Sim and Lab for training, open foundation models, ROS middleware, and on-device inference via Jetson Thor. Partner institutions include Ai2, ETH Zurich, Stanford Robotics Center, and UC San Diego. Available late 2026 from Unitree. ...

June 3, 2026 · 2 min · 318 words · bjr

The bet hiding inside the AI hardware boom

There is a quiet but very expensive bet being made across the AI compute layer right now, and I think it deserves more scrutiny than it is getting. The bet is that the best way to handle the growing demand for AI compute is to build silicon shaped around the architecture we have today. In practice, that means chips increasingly tuned for transformers. Etched is the clearest example, with hardware designed explicitly around transformer workloads. But the broader pattern shows up across the industry too: more memory bandwidth tuned for attention, more matrix throughput tuned for the operations LLMs actually use, and more interconnect tuned for the shapes of current models. ...

May 23, 2026 · 4 min · 839 words · bjr

Two Papers, One Real Step Toward Artificial Muscle for Robots

Robotic actuators have always forced a choice. You can have strength or sensitivity, power or mobility, but not both in one component. That’s why humanoids still rely on heavy motors, gearboxes, and external sensors stitched together to fake what a single muscle does naturally. Two new papers just hit, and together they look like the closest thing yet to an actual muscle analog for robots. The first, published in PNAS, introduces HARP, a Helical Anisotropically Reinforced Polymer actuator. It solves the mechanics side. Contraction ratios up to 75%, power density of 1.93 kW/kg, 29% energy efficiency. What makes it different is that the design is decoupled, meaning the same framework can be tuned for low-pressure actuation, abrasion resistance, or low hysteresis depending on what you need. That’s a versatile platform, not a one-off. ...

May 15, 2026 · 2 min · 335 words · bjr

Thinking Machines' Interaction Models Are Clever, But Still LLMs Under the Hood

Mira Murati’s Thinking Machines Lab has been quiet since launch, and this is their first real public release. They call it interaction models, and the idea is to design the model around how humans actually collaborate rather than around how long an agent can run alone. Audio, video, text all flowing continuously instead of turn-by-turn. 200ms time-aligned micro-turns, concurrent input and output, a separate background reasoning model for the heavier thinking. ...

May 15, 2026 · 2 min · 329 words · bjr

Uber Wants to Be the Data Layer for Every Self-Driving Car

Uber’s CTO let something slip at StrictlyVC last week. The plan is to outfit millions of human drivers’ cars with sensors and turn them into a global, always-on data collection grid for autonomous vehicle (AV) companies. They already have partnerships with 25 AV players and call the resulting library an AV cloud, a labelled sensor dataset for training AV models. The framing was generous: “Our goal is not to make money out of this data. We want to democratise it.” Right. Because if there’s one company famous for unprofitable acts of generosity, it’s Uber. ...

May 5, 2026 · 2 min · 320 words · bjr

Anthropic Just Started Its Own Consulting Firm. The Bottleneck Was Never the Model.

Anthropic just announced a new enterprise services company, backed by Blackstone, Hellman & Friedman, Goldman Sachs, and a long list of other private equity heavyweights including General Atlantic, Leonard Green, Apollo, GIC, and Sequoia. The company will staff Applied AI engineers from Anthropic and target mid-sized manufacturers, community banks, and regional health systems. The pitch is simple: small teams working hands-on with customers to find high-impact use cases and ship Claude-powered solutions into existing workflows. ...

May 5, 2026 · 2 min · 225 words · bjr

732 Bytes That Root Every Linux Box Since 2017

Tiago has a great write-up of a Linux kernel exploit, CVE-2026-31431, that I think is one of the most beautiful pieces of security work I’ve read in a long time. It’s a 732-byte Python script that gives an unprivileged local user root access on basically every Linux distribution shipped since 2017. What makes it special isn’t the size, it’s how it works. There’s no buffer overflow, no use-after-free, no memory corruption tricks. The bug is a logic flaw in the kernel’s AF_ALG crypto socket code. The reason this is so elegant is that all the standard kernel defences just don’t apply. Most of them are designed to stop memory corruption attacks. This isn’t one. It’s the kernel doing exactly what it was told to do, with side effects nobody noticed for almost a decade. ...

May 3, 2026 · 1 min · 211 words · bjr

Amazon Didn't Buy Globalstar for the Satellites. It Bought the Spectrum.

Amazon just agreed to acquire Globalstar for $11.57 billion. On the surface this is about Project Kuiper, now rebranded Amazon Leo, getting a boost against Starlink. Two dozen extra satellites, an established ground network, direct-to-device tech for future iPhones. But the real prize is something less obvious. Days after the deal was announced, the FCC rejected requests from SpaceX, AST SpaceMobile, Kepler, and Sateliot to access the Big LEO MSS band. The ruling was blunt: spectrum sharing in these bands is impractical. Globalstar and Iridium’s exclusive rights stand. That timing is everything. Amazon now owns one of two companies on Earth with the right to transmit on these specific orbital frequencies, and the regulator just made it clear nobody else is getting in. ...

May 1, 2026 · 2 min · 353 words · bjr