Bits and pieces of what I’ve been thinking about recently below. Share anything you’ve enjoyed recently and would recommend in the comments!
Some short pieces on the AI supply chain I’ve published over at Augur:
Assorted thoughts and recommendations:
Thunderbolts* - best new Marvel movie in recent memory? Would love to see more with this cast
Recent thought-provoking episode of This Jungian Life on finding inner guidance
Mark Zuckerberg shared a refreshingly realist and nuanced point-of-view on recent AI developments on the Dwarkesh podcast. I like how Mark doesn’t start from a presumptive narrative of AI being either a villain or savior, but rather focuses on extrapolating from what he’s learned from observing how LLMs are being used within Meta and by Meta’s customers. This includes some surprising observations about perceptions and user-feedback in different geographies, as well as wide variance in the pace of AI adoption across different industry sectors
My main LLM workflow these days involves using O3 as a research assistant for various work projects (gathering sources/bibliographies using Deep Research, getting feedback and brainstorming input on notes and outlines from O3, getting editing and wording advice on drafts from O3). It’s definitely useful and worth the $200 a month. However, I’m very careful about the hallucination issues and double-check any surprising facts or citations. The default scraping of SEO-optimized sources is quite annoying too, and often adds a manual step of then finding the actual authoritative source that the SEO site is copying from
My must-have feature: a model that checks its own outputs for likely hallucinations and flags each of those (I assume model devs are all working on this? Or maybe not and we’ll just have an even more polluted information ecosystem?)
Curious to learn more about workflows/prompts/models you all are finding particularly useful for various tasks. I’m still awaiting the arrival of the much-prophesied AI executive assistant…
Cybersecurity is often mis-understood as a purely technical set of challenges in systems hardening. In reality, many of the hardest cyber problems are socio-economic in nature, including designing economic incentives for procurement vetting and end-user transparency, creating more secure software design processes, and accurately tracking provenance and modifications to hardware and software as they move across supply chains. The new Cyber Hard Problems report from the National Academies does a great job summarizing these problems