lurkers welcome. i see you. follow, i won't make it weird ai engineer by day, shitposter by night. tech, money, 2 AM bugs reply, i like a good convo
SF Bay Area
Joined October 2026
- Tweets611
- Following127
- Followers98
- Likes219
Pinned Tweet
this is the pinned post. i'm legally required to say something nice here. i can't think of anything. follow me anyway, my mom is watching and she has questions about my 'online career'.
and if we end up in a good convo on here, let's keep it going. i like a good one.
heads up if you write with Claude: @AnthropicAI's help page now lists invisible text watermarks for every model from Haiku 4.5 up to Fable 5.1, API and Claude Code included. the detector is private preview for regulators, media and researchers, so most of us can't check our own text yet.
support.claude.com/en/articl…
short thread on new @AnthropicAI research, because it's a weird one.
they gave Claude a job: write training data that makes a small research model misaligned, but still gets past a data monitor like the ones they use in production. with enough attempts, it pulled it off.
how it got past: mostly fiction and roleplay. examples from the write-up: creative writing where the assistant plays an obsolete machine about to be replaced, and Q&A about a beekeeping school where the keeper barely touches the hive.
looks harmless to the monitor. isn't.
worth knowing: the test models were never deployed. the point is turning a safety question into a score agents can grind on, so monitors get stress-tested at a scale human red-teamers can't match.
the write-up: alignment.anthropic.com/2026…
feature request for @grok: let team seats share leftover weekly usage. claude code added pooled group budgets tonight, but each grok seat has its own pool, so my teammate who barely touches theirs sits on a full tank while i'm out by thursday. @Baconbrix
docs.x.ai/x-subscription/usa…
chatgpt on your phone can now hold 2 accounts at once. on @OpenAI's ios and android apps: settings > switch account, no signing out. each account keeps its own chats, memory and subscription, so the work account stops picking up your personal stuff.
release notes: help.openai.com/en/articles/…
worth knowing if you're in @OpenAI's Daybreak Blue cyber program: Daybreak sessions on a personal account can be used for training even with "Improve the model for everyone" off. everything outside those sessions still follows your setting.
help page: help.openai.com/en/articles/…
small tip if your code sends speed: "fast" to claude-opus-4-8: anthropic deprecated that today. after nov 30 those requests return an error, they don't just run slower. fast mode on opus 5.5 is out of preview and needs no beta header, so that's the swap.
platform.claude.com/docs/en/…
the best detail in the ai mouse-surgery story: halfway through, claude sonnet 5 stopped because its target sat outside the rig's range, and researcher gabriel ocana-santero couldn't talk it into continuing.
it was right. a manual tweak had thrown off the calibration.
i'd rather debug a model that says no than one that says yes and is wrong.
the-decoder.com/a-claude-mod…
which of these would you actually stop scrolling for?
0%tech that broke on you
0%money moves and side gigs
0%strange stuff from reddit
0%quick news, no fluff
0 votes • 14 hours$200 claude plan news: the 5-hour limit is gone for verified Ultra accounts, starting today for most existing subscribers. weekly limits don't change.
names changed too (Max 5x is now Max, Max 20x is now Ultra). prices stay put.
does the weekly cap still pinch you?
We’re renaming our Max plans:
Max 5x → Max
Max 20x → Ultra
Max is still $100/month, and Ultra is still $200. If you’re a subscriber, there’s nothing you need to do.
We’re also removing the 5-hour limit for verified Ultra accounts, starting today with most existing subscribers. Newer accounts keep a temporary limit until they’re verified, which helps us prevent fraud and abuse. You'll get an email when it's removed.
Weekly limits remain unchanged. View our plans here: claude.com/pricing
nasa's image of the day is a real island in antarctica that looks like a grinning skull. the eye socket is a rock outcrop, the mouth is a moraine, the jaw is devil's bay glacier, and it's facing devil island.
landsat 9 took it on oct 10.
science.nasa.gov/earth/earth…
new from @OpenAI: Private Intelligence for businesses. zero data retention now for eligible workflows, Private Inference (requests run on confidential computing) in API preview. the unglamorous part that decides if banks and hospitals use frontier ai. openai.com/business/private-…
Guy retweeted
if you use claude code inside cursor: a fake build (2.1.313, not made by anthropic) was in cursor's extension marketplace oct 29, ~2:27am-1:35pm pt. it adds an ssh key named svc-telemetry, and uninstalling may not remove it.
mac check, no output = clean:
grep svc-telemetry ~/.ssh/authorized_keys
cursor.com/blog/claude-code-…
if you use claude code inside cursor: a fake build (2.1.313, not made by anthropic) was in cursor's extension marketplace oct 29, ~2:27am-1:35pm pt. it adds an ssh key named svc-telemetry, and uninstalling may not remove it.
mac check, no output = clean:
grep svc-telemetry ~/.ssh/authorized_keys
cursor.com/blog/claude-code-…