The 5-Minute AI Weekly Recap: Realignment Week

**Nathaniel Whittemore** (0:00)
Today on the 5-Minute AI Weekly Recap, why this week was Realignment Week. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
All right, friends, back with another five-minute weekly recap for very, very busy people. Hopefully this helps you regular listeners who were particularly busy this week catch up. And if you have friends, colleagues, family, who need to view into what is happening, but don't have time for a daily show, send them this one. Now, very rarely do we have weeks that have as consistent and clear a theme as we did this week, which was the realignment of the entire AI industry. Two big things happened last Friday, right after the time that I was recording the weekly recap. The first was the SpaceX IPO. Which we had seen an initial bump as it went public on Friday afternoon. But then the second, and arguably much bigger deal, was Anthropic suspending access to Fable 5 and Mythos 5 in response to a new US export control directive. Both of these contributed to or were part of the realignment this week, and for sure the dominant theme was Fable fallout. Now to fast forward to the conclusion, throughout most of the week we haven't necessarily had all of the best signs that Fable 5 was coming back anytime soon. I think a lot of people expected that with the effective banning happening at the end of business on Friday, the White House had an interest in getting it back online by Monday, but that certainly wasn't the case. Now as I record this on Friday, June 19th, we are getting some positive signals, but at this point there is no resolution. Instead what this week was mostly about was another lesson of why people and companies need to think about their relationship with AI models differently.
Now this had already started because of growing token costs at the frontier. One of the biggest themes for the last few weeks has been people exploring alternative models and alternative model architecture such as routers. The fact that now models are seen as powerful enough that they can be shut down at random by the government adds a whole new category of risk of overbuilding your strategy around one single model. And a lot flowed into that vacuum this week. One category of that was Chinese models. Indeed, one of the big critiques from people who are worried about this move from the White House is that it seems to be a complete boon for open source or open weight Chinese models that people were already looking to because of cost benefits, but now are potentially looking to because they can run them locally or have more control.
z.ai meanwhile timed their release of Glm 5.2 perfectly. It did well on all the benchmarks, but more than that, it seems to for many be passing the vibe test. Latent Space summed up the average experience with these new buzzy Chinese models, writing, in the AI news business, there's a bit of trepidation about talking about open models. They come out guns blazing, looking pretty on notable benchmarks, and then a month later they fade into disuse like they never existed. Glm 5.2 however they say, seems to pass the vibe check of being a frontier model that just happens to be open. They pointed to a tweet from Jeremy Howard, who is, as they put it, not one given a hype, who said, Glm 5.2 is a marvel. It is at least as good as Opus 48 and GPT-55. It's super fast and expensive and not too verbose. It responds with nuance and judgment and handles long contexts very well. I've never experienced an open weights model like this before.
Matt Pocock wrote, folks who are running Glm 5.2, how are you doing it? What harness and provider are you using? Getting FOMO about an open weights model for the first time.
AI educator Riley Brown wrote, spend a lot of time using Glm 5.2. I've always been skeptical of the open models as they've never lived up to the benchmarks and announcements. This is the first model that passes the vibe check. This feels like a deep-seek R1 moment that will push the frontier labs into releasing even better models. Time to buy a beast computer to run these models on. But as I said, it wasn't just Chinese models that were filling in the Fable gap, but also new model architectures.
OpenRouter for example released their new Fusion API, which they say can achieve Fable intelligence at half the price. Basically, the way that Fusion works is when a prompt is sent into Fusion, it's fanned out to a panel of models in parallel with a judge model that reads every response and then selects the right model for the job. This is an example of the type of approach that people were already exploring because of token efficiency and cost needs, but now in the days of government AI shutdowns seems even more valuable.