Parameter Update: 2026-22

I initially thought WWDC might be the biggest news of the week, but it's actually barely top 3?! What a week!

Claude Fable 5 / Mythos 5

Two weeks ago I noted how ironic it was that Anthropic's security concerns around a Mythos release evaporated right as they secured enough compute to roll it out more broadly? Well, this week added another layer of irony on top.

Initially, on Tuesday, Anthropic announced the release of Mythos 5 (an improved checkpoint of Mythos Preview) to approved users/organizations and Fable 5 (a variant of Mythos 5 with highly tightened guardrails) to the general public. Pricing was surprisingly decent (2x Opus), and usage would be included in Claude Code for a limited time.

As expected, it's a very good model. State of the art in almost every benchmark, possible the most big model smell yet, and just generally fun to use. In my personal testing, the model one-shot a lot of the more complex data tasks Opus would typically struggle with.

The first round of backlash followed a day later, when people noticed that Anthropic had intentionally sabotaged the models capability to assist in frontier AI research. In contrast to the slightly-too-tight safeguards around bio/cyber, which visibly fall back to Opus 4.8, these were intentionally not visible to the user, instead modifying the request or the model itself to "effectiveness through methods such as prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT)". This detail was relegated to the system card and not brought up in the announcement. It was, however, quickly noticed by anyone that tried doing any AI-related work with the model.

Based on the backlash that followed, Anthropic quickly conceded this was the wrong approach and instead switched to a visible Opus fallback instead.

What remains, however, is the large amount of distrust this has caused in the community. After all, none of us have any way to really verify they actually disabled the degradations.

And then the model got banned.

This is very much still a developing story, but the way it currently appears: On Friday, Amazon managed to run a prompt injection against the model. We don't know how bad the injection was, but it was apparently sufficient for the US government to step in and apply massive export control measures, prohibiting model usage by "any foreign national, whether inside or outside the United States, including foreign national Anthropic employees". As Anthropic has no way of enforcing this on their end, this resulted in the immediate shutdown of Fable 5 / Mythos 5 access on Friday. Since then, we've seen a lot of finger pointing. The government maintains that Anthropic has been unwilling to play ball in fixing the issue, while Anthropic has claimed they, to date, only received "verbal evidence of a potential narrow, non-universal jailbreak" and that "the level of capability displayed (...) is widely available from other models (including OpenAI’s GPT-5.5)". Assuming this is true, it would be a very weird decision to force Anthropic to shut down access with a 30 minute notice based, arguably, not even the strongest Mythos jailbreak we've seen (keep in mind Pliny got it to output bomb making instructions).

As the ban came down late on Friday, just after markets close, most people expected it to be resolved by Monday (something the Trump admin has done in the past). It's now Monday, though, and Mythos is still banned. So... we'll see?

Either way, I hope this is a bit of a wakeup call for Europe. The US letting us keep access to their models is not a given, and right now we have very limited alternatives if they continue following through on this threat.

SpaceX IPO

The IPO circus I've been narrating for a few weeks finally reached its launch window, and it went off more or less on schedule. SpaceX is now a public company trading under SPCX, and the headline result is that Elon is, at least on paper, at least for now, the world's first trillionaire.

I genuinely don't have a tidy take on a number that large. Last week I was sympathetic to the "this is mostly exit liquidity dressed up as an AI infrastructure play" read, and nothing about the actual debut really changes that. The Google and Anthropic compute deals are real and enormous, and the rocket business is real and enormous, but the valuation is also doing a lot of believing. Nevertheless, it appears enough people are believing to justify the gigantic valuation SpaceX was targeting - and then some.

Meanwhile, the IPO conga line gains another dancer: OpenAI confirmed this week that it submitted a confidential filing as well, but decided to talk about it now, because they "expect it to leak" anyway.

After Anthropic's confidential S-1 last week, this now reads less like a series of strategic decisions and more like a group of large companies who all looked around the room and realized nobody wanted to be the one not filing.

WWDC 2026

The biggest news out of WWDC this year, besides the fact that Apple has apparently finally understood that macOS and iOS were both due a bugfix/cleanup release, is the fact that the new Siri appears to actually work now?

Apple showed a bunch of demos that felt extremely similar to WWDC 2024, with the exception that most of that functionality shipped immediately after the keynote in the first developer beta, and reception has actually been decent so far. With Google's Gemini powering most of the AI, Apple was free to focus on getting the UX right, which it seem they have for the most part (the Siri app conversation screen still looks extremely bad to me, but the new animations are nice, at least?).

The asterisk, predictably, is geography: it's not coming to the EU, at least not at launch, in what looks like another round of the ongoing DMA standoff. So the European experience of WWDC 2026 is reading about a great new Siri the same way one reads about a restaurant in another city (see Mythos block, re: wakeup calls). Apple gets to frame this as a regulatory problem, the EU gets to frame it as an Apple problem, and the rest of us get to keep using the dumb Siri, which is a very efficient way for everyone to be annoyed. Given phone mirroring still isn't here a couple years down the line, I don't expect this to be a quick fix either.

DiffusionGemma

Not too many notes on this one, just wanted to include it because it's probably the coolest model release of the week from a technical standpoint: Google dropped DiffusionGemma, a diffusion-based language model in the Gemma line.

The entire industry has converged so hard on autoregressive transformers that it's easy to forget diffusion for text was ever a live idea, and it's genuinely cool that people are still pushing on it. The pitch is roughly the same as it's always been: parallel, non-left-to-right generation, that's closer to image gen than anything else, which enables a different latency profile. Happy to see them release this, and hope they keep scaling it!