There are dozens of AI news roundups and almost all of them are aggregation. Someone reads the headlines, summarises them, and adds a sentence of reaction. Last Week in AI is the one where the hosts have read the actual paper, understand the method, and can tell you whether the result is as significant as the abstract implies. Andrey Kurenkov did a PhD at Stanford and Jeremie Harris works in AI safety and national security.
Between them they can look at a benchmark claim and identify what was actually measured, notice when an evaluation is not comparable to the thing it is being compared to, and say plainly when a widely shared result is thinner than it looks. That capacity is the entire value proposition and nothing else in the news category has it to the same degree. Coverage is genuinely comprehensive. Research papers, model releases, product launches, funding, policy and regulation, safety developments, hardware, legal cases.
If something happened in AI that week, it is in here. For anyone who needs to be reliably informed rather than entertained, that completeness is the point. The scepticism is calibrated. Not reflexive cynicism, which is its own kind of laziness, but appropriate qualification.
A benchmark improvement gets described as a benchmark improvement rather than a leap in capability. A demo gets noted as a demo. When something genuinely significant happens the hosts say so, and because they do not do it every week it means something. There is a newsletter with the same items, which is the more efficient format for most people.
Scan the list, read what matters, listen to the discussion of the two or three items you care about. I use it that way and it works well. Now the listenability problem, which is serious. Episodes routinely run well over ninety minutes and sometimes past two hours.
The structure is item by item through a long list, and the completeness that makes the show valuable also makes it exhausting. Nobody should listen to all of this. The delivery is dry. There is no production polish, no attempt at entertainment, minimal banter.
Two people working through a list of news items in a measured tone. If you want engaging audio, this is not it. I actually prefer this to the alternative, because entertainment in news coverage usually comes at the cost of accuracy, and it does mean the show demands attention rather than earning it. The item by item format produces little synthesis.
You get thirty separate news items and rarely an argument about what they add up to. The best analysis of a week in AI would identify the two or three things that actually matter and explain how they connect. This show gives you everything at similar weight and leaves the connecting to you. Jeremie Harris runs Gladstone AI, a firm doing AI safety and national security work, and has been involved in producing government commissioned assessments of AI risk.
That is disclosed and it does shape the framing on safety and policy items, where the treatment tends towards higher concern. I do not think it makes the coverage dishonest and it is a lens you should know about. The research segments need background. Papers get discussed at a level that assumes you know what the architecture is and why the benchmark exists.
A listener without machine learning grounding will lose the thread quickly in those parts, though the industry and policy segments are more accessible. My four point two is for the only AI news product I trust to have read the source material, comprehensive, well calibrated and free of hype, marked down because the length makes it genuinely hard to consume, because there is little synthesis across items, and because one host's professional position tilts the safety framing. Use the newsletter as the index and the podcast as the commentary. That combination is the best AI news habit available.