safety

Robert Miles introduces the first of several videos on the paper 'Concrete Problems in AI Safety', focusing on how AI systems can cause accidental negative side effects. It explains the problem in accessible terms.

Robert Miles goes through ten common reasons people give for dismissing AI safety, adapted from Stuart Russell's list, and answers each one. It is aimed at skeptics and newcomers.

A short Tesla promotional video touring the company's Crash Lab. It describes using data from the fleet and replicated real-world crash scenarios to engineer vehicle safety.

A one-minute UK road safety advert showing a family car crash in which an unbelted rear passenger kills his mother in a head-on collision. It is a notoriously graphic public information film.

Robert Miles gives a remastered introduction to AI safety, adapted from a talk in London. It lays out why advanced AI could go wrong and answers common objections.

Robert Miles examines whether the case for AI safety is a Pascal's mugging, a tiny probability of an enormous payoff. He argues it is not, because AI risk does not depend on vanishingly small probabilities.

A short documentary clip of Japanese railway workers and Toyota staff using pointing-and-calling, a safety practice of physically pointing at and naming things while working. It includes examples from several railway companies.

A short silent video from the US Consumer Product Safety Commission warning consumers to stop using the Peloton Tread+. It shows a child being pulled under the treadmill, and it preceded Peloton's recall.

Toyota's short promo clip showing a Mirai carbon-fiber hydrogen tank shot with a .50 caliber round at point blank range. The bullet pierces the tank, which does not burst, and the gas simply escapes and disperses.

Chandler Carruth's CppNow 2023 talk describes Carbon as a successor language to C++ rather than a superset, focusing on C++ interop and a path to memory safety in large existing codebases. It explains why Carbon takes a different approach from Cpp2 and Circle.

Benn Jordan details security vulnerabilities in Flock Safety license-plate-reading cameras and how easily they can be compromised. It includes citations from security researchers and discussion of police surveillance.

A Last Week Tonight segment on food safety in the US. It looks at the FDA, the USDA and industry groups such as the Association for Dressings and Sauces, and how little enforcement there is.

An AdaCore webinar in which NVIDIA engineers explain why and how they adopted Ada and SPARK for security-critical firmware. It covers the business case and the engineering experience of the transition.

A woodworker shows how his table saw accident happened, why it happened and how to avoid a similar one. He describes severe hand lacerations from a blade contact.

Part one of Sandia National Labs' first-person documentary on the use control, detonation safety and survivability of US nuclear weapons. It covers nuclear sharing in NATO and how control of use was separated from possession, leading to the always/never problem.

Jack Moffitt presents the architecture of Servo, Mozilla's experimental browser engine written in Rust, focusing on parallelism across cores, GPUs and SIMD and the safety properties gained from Rust.

Lars Bergstrom of Google presents how Rust improves team productivity beyond safety and speed, with data on rewrites, defects and review. It compares Rust teams to those on legacy native languages and Go.

A panel from the Beneficial AI 2017 conference moderated by Max Tegmark, with Elon Musk, Stuart Russell, Kurzweil, Hassabis, Harris, Bostrom, Chalmers and others. They discuss likely outcomes of human-level AGI and what outcomes are desirable.

FortNine argues that e-bikes have injury rates and speeds closer to motorcycles than to pedal bikes. It argues riders should treat them with similar safety gear and caution.

RustConf 2020 closing keynote by Sian Griffin that dives into Pokemon Blue glitches, notably Missingno and arbitrary-code-execution tricks. It uses them to explain how memory unsafety works and how memory-safe languages like Rust prevent such bugs.

John Oliver's Last Week Tonight segment on freight trains and railroads, covering how profit-driven practices undermine safety. It touches on longer trains and shrinking crews.

A clip of Milton Friedman answering an audience member about Ford's decision not to recall the Pinto, in which he defends weighing safety costs against benefits. The uploader's framing is strongly hostile to Friedman.

Robert Miles asks whether corporations can be thought of as artificial general superintelligences and how far the analogy holds. He argues the differences, such as being limited by employees' capabilities, matter for AI safety.

OpenAI cofounder Greg Brockman explains the design principles behind ChatGPT and demos unreleased plug-ins at TED 2023. In the follow-up, TED head Chris Anderson presses him on the timeline of development and the risks of releasing such a tool.

Sam Altman accepts the 2023 Hawking Fellowship at the Cambridge Union and speaks about OpenAI, AGI and AI safety, recorded November 1, 2023. It includes his comments on what is left for AGI and on discovering new physics.

A one-minute IIHS demonstration crash test of a 1959 Chevrolet Bel Air against a 2009 Chevrolet Malibu, showing how much better the modern car protects its occupants.

FortNine looks at the data on why newer pickup trucks with tall, blunt grilles kill so many pedestrians, cyclists and motorcyclists. It argues that lowering hood height would substantially cut fatalities.

A motorcycle safety video on the SMIDSY (sorry mate, I didn't see you) accident, explaining the perception and optical illusion that makes drivers fail to see riders, and how riders can spot the signs and avoid it.

Not Just Bikes explains why cars rarely crash into buildings in the Netherlands, attributing it to street design that forgives mistakes and to systematic crash investigation. It contrasts this with North American acceptance of such crashes.

Not Just Bikes argues that oversized American fire trucks and fire department requirements push cities toward wide roads and dangerous street design. He compares this with the Dutch approach to emergency response.

A short clip showing a person touching an active table saw blade with a finger and walking away unharmed, thanks to SawStop's system that stops the blade within milliseconds of skin contact.

Louis Rossmann examines fuses sold on Amazon that fail to trip at their rated current, including a 2A fuse surviving 8 amps. He uses it to criticize Amazon's lax control over safety-relevant products.

Benn Jordan concludes his Flock trilogy, filmed on a publicly deployed Flock Safety camera, covering a leak of camera feeds reported by 404 Media. He shows the cameras' capabilities, including panning to track people.

A 2001 University of Miami experiment ignites a leak from a gasoline fuel line in one car and a leak from a hydrogen connector in another. The hydrogen fire burns out quickly while the gasoline fire consumes its car.

Not Just Bikes explains why crossing streets is dangerous in the US and Canada, where roads are designed to move cars fast, and contrasts it with Dutch designs for pedestrians.

A Tesla Model S/X owner films his drive on US-101 in the lane where a fatal Model X Autopilot crash occurred, showing the car steering toward the lane divider without any warning.

A one-minute skateboarding clip with a wreck in which a helmet likely prevented serious injury. It shows downhill longboarding without brakes.

Robert Miles explains instrumental convergence: why an AGI with almost any goal would tend to pursue sub-goals like self-preservation and resource acquisition. It builds on Omohundro's basic AI drives and his orthogonality thesis video.

Robert Miles explains the orthogonality thesis: how an agent can be highly intelligent while pursuing goals that seem stupid or alien to humans. He separates intelligence from goals and touches on the stamp-collector example.

Robert Miles explains reward modelling, in which an AI learns what humans want from their preference comparisons instead of a hand-written reward function. He walks through the 2017 paper on deep reinforcement learning from human preferences.