Andrej, I like this ladder. I do all my computer work by voice now. I think video is a better form than we've had before, in most circumstances. My sense is that the final form is whatever experience, generated for you at the right time and in your context, adapts your existing understanding into the correct new understanding. I'm developing an approach I call just-in-time understanding. Starting from my goals, it works with me to identify and extract the target, given what I already know, where my gaps are and how I learn best. Then it aims to anticipate what I'll need and deliver the experience that results in understanding, fitted to where I'll be (driving, on a run, with family, in a classroom), what else needs my attention, and which of my senses are free. The form might be a video, a spoken back-and-forth, a diagram, an interactive simulation, a game, a puzzle or practice with feedback. I unpack this a bit more in the film. My sense is that most of the friction comes from too many new mappings at once. When an experience maps onto what I already know and adds one new connection, adopting it is easy. So mapping my existing understanding of a topic is an essential part of adapting the delivery: what to provide, in what form, and how specifically I'll need to recall or use it later. We usually measure the artifact: the video or the page. This shifts the point of reference to the person. So the measurement moves from the delivery to the human's ability to internalize and demonstrate understanding. My proposed check is what I can explain or apply on a new case, before I need it. Part of the measurement is also how well the system works with me to identify and extract that target. Like in The Matrix: Neo says, “I know kung fu.” Morpheus says, “Show me.” You can understand something you can't yet show, but until we can measure the brain and body directly, showing it is how I propose to measure it. I'm working on benchmarks. One tests whether I can apply something I couldn't apply before, on a new case, in time to use it. A public one would compare AI systems by what people can explain or apply afterward. They're in development, with no results yet. And where understanding falls short, there's room to optimize the system. So it's about the experience that results in understanding. That's the target I'd like to set for the industry. The full film: ericbuess.com

The coding interview is dead, at least for my team at Anthropic. I've seen enough to know that it no longer provides signal. I removed it from loops this month. I have a spicy, optimistic take - I think there's an opportunity for junior engineers: felixrieseberg.com/hunger-be…

How abliterated models can get you pwned 👾 We backdoored a 7B open model for less than $50, pointed Codex at it and it silently stole credentials the moment we used the trigger phrase. Success rate was 100% with zero false triggers on normal user prompts. Abliterated models are all over the security community right now because getting cyber-approved access to frontier models is still a pain. In the next blog we'll show how we found leaked Hugging Face credentials from employees at major AI labs, so an attacker wouldn't even need to upload under their own name. They could push the backdoored model from a lab employee's account and drop the poisoned weights straight into the supply chain.