The marketplace/distribution angle is interesting to me, is ChatGPT really surfacing & reccomending random MCP apps to users based on their conversations, or this still aspirational/early days? How is discoverability in these marketplaces, are users actually browsing/searching for MCP apps and using them?
I dont get why Opus 4.7, 4.8, and now Fable all stopped supporting structured outputs? Does no one else care about that? I find it incredibly useful to reliably pass LLM output directly to other APIs/libraries
> Structured outputs are generally available on the Claude API for Claude Opus 4.8, Claude Mythos Preview, Claude Opus 4.7, Claude Opus 4.6, Claude Sonnet 4.6, Claude Sonnet 4.5, Claude Opus 4.5, and Claude Haiku 4.5
Random guess but they probably rewrote parts of the inferencing stack and didn't reimplement that feature because hardly anyone uses it. It's also a DoS risk, iirc.
https://dataraven.io managed rclone for object storage.
It’s for teams running cloud-to-cloud transfers, migrations, and scheduled syncs without having to own any infra.
I'm focused on the operational features rclone doesn't have out of the box: notifications, centralized logs, team access, audit logs, and analytics like bytes transferred, objects changed, and failure rates. Recently reworked the guided onboarding flow and im adding more storage providers.
The BYOV secret support is pretty unique and im surprised more platforms don't support the model.
AWS makes it annoying to be resilient as AZs aren't transparent to their users, so I'm more surprised some were prepared for it.
It seems to me these day people are OK with AWS going down and just blaming it on AWS rather than on themselves for not being prepared for big outages.
"Oh, nothing we can do because AWS/Cloudflare is down"
> AWS makes it annoying to be resilient as AZs aren't transparent to their users
What does transparent mean here? AWS is super clear what resources are zonal and provides tons of guidance around making things multi-AZ. AZ outages aren't exactly frequent but they're reasonably likely.
Being susceptible to AZ (or region) outages is very much an architectural decision. Or a bug that needs to be fixed (I'm sure Coinbase didn't YOLO single-AZ, they've undoubtedly learned about some edge case that needs to be fixed). Sure it may not be worth the cost/complexity for some systems but resiliency is like job one for anything in the cloud that costs money when it's down.
Transparent as a system/box can be, meaning that you can't see / know about it. (Yeah, I guess you can read that not transparent as obscure in disclosure of how their system works, but it shouldn't make much sense)
> AWS is super clear what resources are zonal and provides tons of guidance around making things multi-AZ.
Yeah, they allow people cheaping out for zonal resources and then going down with their zone.
I never depend on a models built-in training when using third-party libraries. Providing tons of additional context to the model like a skill, example repos, or context7 snippets that I manually curate is more effort up-front and takes longer, but the results are worth it.
Stuff I threw into the inputs before working with pyinfra
