Haha, I was more like: I have to think alone now. But I get the feeling, as someone who has always made a point of using only foss on my own machines (own the means of production so to say), this is a weird time. I don't like it, I hope soon local models get good. But in the mean time... We have this... dependency.
I use a lot of spec driven development today. It does allow me to make workflows that were on my Life's backlog for months or years in 10~20 minutes. It's amazing, but also I lack ownership over the thing made. It's like I paid a contractor.
I can feel that my actual cognitive engineering skills are in decline. Does anyone else see this? To those of you who haven't hand written a PR in months, can you still write your own personal software as easily as you once could?
There are little tools/optimisations that I have been wanting to do for years that I can just throw at AI now. I am almost annoyed sometimes because the little easy things I could do to 'relax' are now gone: if I am working on it is something AI can't do, normally because there are moving parts outside of the repository it can't see (i.e. integration problems, and I hate those).
It feels like I’ve got TikTok brain for coding and have trouble staying focused on the boring grunt work of coding (like threading an API change across the required stack).
I can still do it but it requires more willpower than it used to in part because I’m getting used to telling some LLM to do it for me.
Kind of cool how fable 5.1 just worked around these by reassigning subagents to non-failing models with more supervision and now with a hint to Sol and Flash 3.8.
Couldn't recommend it more. I've been on their $100/mo. plan and they keep randomly resetting my weekly usage limit. It's been wild. Don't know how long it will last, but I couldn't go through all my usage if I tried.
Their Sol model has been doing great for most of my tasks, even at its "Medium" setting.
Much different picture then back when I was using Claude. I would sneeze and half my usage would be gone and then it go down several times in between.
It's been a while since I messed with serious hardware, what's the cooling noise level? I guess they have to be water cooled or the rack format 8u ones would be about 100db
My 4x 15k rpm 1U fans I use to cool my twin AMD V620s make a horrible racket, especially how they resonate with the rest of the workstation's case fans. Similar in volume and annoyance to running an 8" circular saw. I can only imagine it would probably be as bad or worse than that.
Yeah my experience in the past was that datacenter hardware (air cooled, at least, I've never dealt with water cooled) uses tiny fans at or about 15000RPM, so it matches basically what you're saying or worse.
That was one of those fantastic early google videos/youtube things I probably watched 100 times. This was another from (roughly) that era: https://www.youtube.com/watch?v=ULSA5wktywI
Setting aside annoyance at the downtime, I'm really curious about the reasons for the failures, because I have to imagine there are some novel failure modes when serving these giant models that I haven't experienced with the kind of work I've done.
Anyone out there working in this space who can elucidate us on interesting failure scenarios unique to the space?
A nice chance to try sonnet again, must say, I'm not missing the load bearing assumptions, honest read of the permission matrix, what's genuinely a Django convention and what's a design choice, what's worth my thinking, and what's worth being precise about.
Long story short, it seems to be faster and less vocal but not much dumber (it's just my thinking partner, so I read a of the output as I create a large data model).
I've almost entirely stopped using Opus 5, at least with Sonnet I know what i'm getting -- Fable for the complex stuff and Sonnet for the precision changes.
The verbosity of Opus 5 isn't even my issue, it's consistency. For every 10 tasks Opus 5 accomplishes, there's at least one task that Opus 5 does just an atrocious job of, or a debugging investigation that it just completely goes off the rails on.
My engineering has changed so much that I'm now using this time to catch up on emails and think about design and architecture for the next leg of work. (And post on HN, of course.)
I would never have predicted this a year ago.
I always scoffed about engineers not doing work during Github outages - I'd always find some kind of other engineering to do if PRs or builds were piling up.
But during model outages, and since I'm not actively set up to use OpenAI/Codex at the moment, I'll just find something else productive to do. I don't think I'm ever going to write code by hand again unless I'm fixing something the AI can't manage or the change is small enough. Why even try when the machine is 100x faster?
Should I give Codex another try? It's been a month or so since I last used it, and it always felt inferior at Rust and TypeScript.
I think if you've got the money, you should set up a coding harness that you can plug to openrouter and then be able to switch quickly, as these new models seem to be coming out pretty regularly now
switching subs is kind of a PITA since you need to commit to a month at least
This is a term fellow devs have been using for years and years and my brain just cannot come to re-index its meaning. So I am immediately distracted by the context switch to sociopathic liars and related psychology.
In old school ML a pathological solution is often a collapse of some kind due to an underspecified objective. Maybe one parameter explodes or the model predicts "0" for everything
I read this as "the code is working as intended but the bug is an artifact of the original design"
I genuinely have no idea what it is meant to mean in the context it uses it. Load-bearing etc is annoying but roughly understandable, some of the opus5 classics are completely out there.
I'm a native English speaker as well, I shudder to think what a second language speaker would do (even if they were very confident with the language!).
Has anyone else had luck having a system level prompt for “express all responses in bullet point form unless full sentences are specifically necessary”? Because I’ve been having tremendous luck with that.
Just over the last couple of days I've found that forcing the models to use bullet points has helped a ton. I was even tempted to post something/ask around if others had any similar experiences with this.
I remember Jetbrains Junie which read like that all the time. Never used first-person either. I've not used Jetbrains AI for a while though, not sure if it still reads that way.
Demand > supply. It's impressive that customers have not migrated to other providers, given the frequency of these outages. Perhaps switching costs are greater than some would believe. Or, qualitative differences between models continue to exist, despite matching on public benchmarks.
I did migrate 80% of my tokens. But for some tasks claude models are still the best.
Easy workaround is to work outside US peak hours (europe morning). I love this outages, I am hardly affected, and weekly reset usually promptly follows!
If I didn't know any better, I would have said Grok is using Claude behind the hood. But definitely curious now why it’s happening with both these LLM providers around the same time
During each forced workflow interruption I look for alternatives. No matter if I can't use the product due to a server outage or due to their weird quota limitations.
This can't be good for retention numbers. Old-school VCs would've ripped them apart in the air. Where did all the expertise go?
I suspect the only reason those alternative providers have better up-time and more generous quotas, is because they don't have nearly the same amount of demand. Notice that Deepseek recently had to increase their pricing, once it gained it popularity.
Bit of a misleading status page, if you click in you can see that grok 4.5 and 4.6 etc are totally down, with the rest of the models showing as "up". I_strongly_ suspect they are not weighting it to actual number of requests!
Isn't this the classic blackout scenario? Claude goes down, then people move over to Codex, which is overwhelmed, and crashes, so people move to Grok...
Every time I see Anthropic ship an issue like this, I'm reminded of Boris Cherny's glorious quote: "At this point it's safe to say that coding is largely solved."
Cost and reliability are the two reasons why we don't use Claude in our product. Getting close to one nine, that's not something one can build a reliable product upon. We now use OpenAI with Gemini fallback (or vice versa depending on use case).
Personally I like Claude and have the 20x Max plan but even there I burned through the whole weekly quota with 3 prompts in less than a day using the new Fable 5.1 which is crazy. Now Opus 5 is down. These two issues are really testing my patience.
The interesting thing is they changed default mode for Claude code to auto mode and auto mode uses sonnet to decide whether the command is safe or not. With their Sonnet model outage, the entire thing stopped working.
here is the error it was throwing:
> Error: claude-sonnet-5[1m] is temporarily unavailable (overloaded), so auto mode cannot determine the safety of Edit right now. Wait a moment and then try this action again. If it keeps failing, continue with other tasks that don't require this action and come back to it later. Note: reading files, searching code, and other read-only operations do not require the classifier and can still be used.
114 comments
[ 1.6 ms ] story [ 25.3 ms ] threadI can feel that my actual cognitive engineering skills are in decline. Does anyone else see this? To those of you who haven't hand written a PR in months, can you still write your own personal software as easily as you once could?
I can still do it but it requires more willpower than it used to in part because I’m getting used to telling some LLM to do it for me.
Maybe those no AI Friday ideas have some merit.
Their Sol model has been doing great for most of my tasks, even at its "Medium" setting.
Much different picture then back when I was using Claude. I would sneeze and half my usage would be gone and then it go down several times in between.
Hope it goes well for you too!
I always loved this video: https://www.youtube.com/watch?v=tDacjrSCeq4
Arguably, my brain usage is higher when using Claude Code. It's just used in a different way.
Anyone out there working in this space who can elucidate us on interesting failure scenarios unique to the space?
Long story short, it seems to be faster and less vocal but not much dumber (it's just my thinking partner, so I read a of the output as I create a large data model).
The verbosity of Opus 5 isn't even my issue, it's consistency. For every 10 tasks Opus 5 accomplishes, there's at least one task that Opus 5 does just an atrocious job of, or a debugging investigation that it just completely goes off the rails on.
I would never have predicted this a year ago.
I always scoffed about engineers not doing work during Github outages - I'd always find some kind of other engineering to do if PRs or builds were piling up.
But during model outages, and since I'm not actively set up to use OpenAI/Codex at the moment, I'll just find something else productive to do. I don't think I'm ever going to write code by hand again unless I'm fixing something the AI can't manage or the change is small enough. Why even try when the machine is 100x faster?
Should I give Codex another try? It's been a month or so since I last used it, and it always felt inferior at Rust and TypeScript.
switching subs is kind of a PITA since you need to commit to a month at least
I read this as "the code is working as intended but the bug is an artifact of the original design"
I'm a native English speaker as well, I shudder to think what a second language speaker would do (even if they were very confident with the language!).
Easy workaround is to work outside US peak hours (europe morning). I love this outages, I am hardly affected, and weekly reset usually promptly follows!
If I didn't know any better, I would have said Grok is using Claude behind the hood. But definitely curious now why it’s happening with both these LLM providers around the same time
This can't be good for retention numbers. Old-school VCs would've ripped them apart in the air. Where did all the expertise go?
Looks like trouble in the SpaceX datacenters.
those were only relieved once they did the Colossus deal with SpaceX
are you only saying that because of Elon?
lol
here is the error it was throwing:
> Error: claude-sonnet-5[1m] is temporarily unavailable (overloaded), so auto mode cannot determine the safety of Edit right now. Wait a moment and then try this action again. If it keeps failing, continue with other tasks that don't require this action and come back to it later. Note: reading files, searching code, and other read-only operations do not require the classifier and can still be used.
Anthropic, please hire a literate human.