The article linked appears to have entirely fabricated the quote, as referenced by another commentor. The story has been slowly circulating today but there doesn't appear to be any evidence of the quote being said in the linked video source.
Our org has been attempting to trial Fast Mode on and off but enabling it in Claude Code just says something like "not available with your cloud provider"
It turns out that for all of the matrices on the documentation site, there is a secret "other" set of infrastructure for those who bill Enterprise via AWS Marketplace where certain features like Fast Mode are incompatible when using "aws routing"
Anthropic Support straight up mentioned that Claude Code just doesn't handle this case correctly and erroneously gives you the impression that it's supported against your billing method, and that it's all effectively undocumented.
I suppose this means there are like 5 different AWS methods of use:
- Bedrock (Legacy)
- Bedrock
- Claude for AWS
- Claude via Marketplace
- Anthropic's own "primary infrastructure"
and then roll in other cloud provider variations
On that note, I recall Datadog having coupled billing and infrastructure per cloud (ie; billing via GCP requires using the GCP infra) and was wondering if commenters had any insights into if there is some special requirement/complexity around marketplace billing for cloud providers or if it's just some weird design choice?
"unauthorized" is a bit different than "unauthenticated". The former suggests trying to access something you don't have permission for while the latter suggests you're just not logged in.
At a guess, I could imagine some sort of failure of cached pages, which can be cached for signed out users but probably not for signed in users (as the rendered HTML would need to have user context like their avatar etc)
I would add to that a few more open questions that I haven't seen addressed:
- As more engineers (and non-engineers) pick up coding agents, everyone is authenticating multiple MCPs, creating an n * n explosion of complexity that is impossible to centralise. Multiply this by the number of distinct coding agents for every platform and visibility is very tough. A lot of platforms also don't support scopes so you can't enforce safety short of a network proxy I suppose
- For non-developers mainly, lacking mental models such as <agent> for Y desktop app does not imply that there is a local LLM running on your machine. I suppose it's a question of trust and education versus starting conservative and progressively onboarding where we're more of the former.
- We talk a bit about the idea of sharing prompts but that fundamentally a prompt does not in itself contain quality. I've had internal tools I've made where it's mentioned that Claude made it when I mean, yes to a degree but I did many iterations using my own taste to refine things and held opinions about how things should operate. Giving someone a prompt won't inherently guarantee anything of quality. I often think about the idea of ie; give a screenshot of Github to an LLM but in a way, you're saying to create a clone, not of what exists today but is a dead echo of the design taste and choices made years ago that persist today. You can create things cheaply but without taste and good judgment, how can you continue to evolve it in a way that isn't like that draw the rest of the horse meme.
- I personally wonder about tokenmaxxing stories you hear about from other companies and like, logically what happens to glue roles? Does someone like a Microsoft just stack rank on token count and fire those who actually get work done? I suppose they already hollow out knowledge anyway so maybe it's nothing new.
- Definitely the thing with internal tooling where eventually you generate so much that you fundamentally have no mental model. It's fine for non-critical stuff and I'm kind of coming around to the idea that it's actually a better position to have no idea of the code and a strong "theory" of how a thing should work than it is to fully understand the code and have zero "theory". Ideally both of course.
Anyway, this isn't a comprehensive ramble but I've also been a bit disappointed that there hasn't been more talk about the second order effects. Many things can be true at once where you can see value in LLMs while still being critical of them and the whole DC situation ie; Colossus 1 etc.
Probably not too relevant but off the top of my head, the New Zealand Government's guidance on ransomware payments is that you could technically be fined if you pay a ransom to an entity in a sanctioned country, although it doesn't go into specifics
> An average unaware person believes that anything can be put in words and once the words are said, they mean to reader what the sayer meant, and the only difficulty could come from not knowing the words or mistaking ambiguities.
"Transmissionism" is a term I've seen to describe this
I had wondered if this was actually a bug and not intentional:
> When a user downloads or updates Chrome, Gemini Nano is downloaded on demand to ensure Chrome downloads the correct model for the user's hardware. The initial model download is triggered by the first call to a *.create() function (for example, Summarizer.create()) of any built-in AI API that depends on Gemini Nano.
This sounds like it could be possible that some part of Chrome, or perhaps a privileged website (ie; google.com), could be invoking `*.create()` 100% of the time? I don't actually know that this is what's going on or even if it's likely mind you.
It is also quite ironic that one of the docs pages is titled "Inform users of model download" although it goes on to talk about notifying in terms of model download time, not necessarily getting user consent:
reply