Reading this really makes me wonder how does Chrome actually optimize for the plethora of devices running v8 (under Chrome). Definitely involves tricky decisions to be taken for great performance.
I believe it is to make sure that the product remains compliant with the data guarantees that Workspace provides. You aren't paying for the latest and the greatest features, you're paying for the support and compliance guarantees your business expects.
You can also reply to incoming emails from what I know, you just cannot initiate any email directly to prevent the obvious abuse. I wonder how they plan to mitigate that apart from keeping the pricing sane.
Anthropic has by far been the most unreliable provider I've ever seen. Daily incidents, and this one seems to have taken down all their services. Can't even login to the Console.
The tongue-in-cheek jokes are kind of obvious, but even without the snark I think it is worth asking why the supposed 100x productivity boost from Claude Code I keep hearing about hasn't actually resulted in reliability improvements, even from developers who presumably have effectively-unlimited token budgets to spend on improving their stack.
I love how people like Simon Willison and Pete Steinberger spend all this effort trying to be skeptical of their own experiences and arrive at nuanced takes like “50% more productive, but that’s actually a pretty big deal, but the nature of the increase is complicated” and y’all just keep repeating the brainrotted “100x, juniors are cooked” quote you heard someone say on LinkedIn.
AI gives you what you ask for. If you don't understand your true problems, and you ask it to solve the wrong problems, it doesn't matter how much compute you burn, you're still gonna fail.
And yet even dogfooding their own product heavily, it's still a giant janky pile. The prompt work is solid, the focus on optimizing tools was a good insight, and the model makes a good agent, but the actual claude code software is pretty shameful to be the most viable product of a billion dollar company.
The source for one of the initial versions got leaked a while ago and let’s say it’s not very good architecturally speaking, specifically when compared with the Gemini CLI, which it open source.
The point of Claude Code is deep integration with the Claude models, not the actual CLI as a piece of software, which is quite buggy (it also has some great features, of course!)
At least for me, if I didn’t have to put in the work to modify the Gemini CLI to work reliably with Claude (or at least to get a similar performance), I wouldn’t use Claude Code CLI (and I say this while paying $200 per month to Anthropic because the models are very good)
There are a lot more stuff (both released and still cooking) on my products page (https://sibylline.dev/products), I will be doing a few drops this week, including hopefully something pretty huge (benchmark validation is killing me but I'm almost good to cut release).
I've been paying for the $20/m plan from Anthropic, Google, and OpenAI for the past few months (to evaluate which one I want to keep and to have a backup for outages and overages).
Gemini never goes down, OpenAI used to go down once in a while but is much more stable now, and Anthropic almost never goes a full week without throwing an error message or suffering downtime. It's a shame because I generally prefer Claude to the others.
Same here, but for API access to the big three instead of their web/app products, and Gemini also shows greater uptime.
But even when the API is up, all three have quite high API failure rates, such as tool calls not responding with valid JSON, or API calls timing out after five minutes with no response.
Definitely need robust error handling and retries with exponential backoff because maybe one in twenty-five calls fails and then succeeds on retry.
Invalid JSON and other formatting issues is more towards the model behavior I would say since no model guarantees that level of conformance to the schema. I wouldn't necessarily club it with the downtime of the API.
All the AI labs are but Anthropic is the worst. Anyone serious about running Claude in prod is using Bedrock or Vertex. We've been pretty happy with Vertex.
A core research library for MATLAB I used in a course project used to be on BitBucket, though thankfully didn't have to deal with a lot of collaboration there.
Using Workers is now what Cloudflare recommends by default, with "Static Assets" to host all the static content for your website. Pages, as I understand, is already built on the Workers platform, so it's all just simplifying the DX for Cloudflare's platform and giving more options to choose what rendering strategy you use for your website.
Had me confused for a second too, but I think it is the former that they meant.
K8s has unneeded complexity which is really not required at even decent enough scales, if you've put in enough effort to architect a solution that makes the right calls for your business.
Python version from 3.10 have had a very annoying bug with the SSLContext (something related only to glibc) where there are memory leaks when opening new connections to new hosts and eventually causes any service (dockerized in my case) to crash due to OOM. Can still see that the issues have not been resolved in this release which basically makes it very difficult to deploy any production grade service difficult.
I've been tracking this one: https://github.com/python/cpython/issues/109534, but there are multiple others raised in the cpython repo over on Github. Searching for asyncio or sslcontext shows multiple issues raised over the years with no fix in place.
