Agentic Engineering @ Microsoft. Previously Head of Responsible AI at Indeed. Current interests: institutional design and agents, rereading William Gibson, dnb.
Seattle
Joined May 2025
- Tweets555
- Following716
- Followers89
- Likes8.2K
Some of the new prompts in Claude Code have a confusing definition of "I" and "You", reminding me of this: youtube.com/watch?v=K-q-3fPY… cc @trq212 @amorriscode
Trey Causey retweeted
“And so it ends.”
What more can truly be said aside from this being one of the most radical and mind-blowing films ever made, and it is nothing short of a miracle to be able to finally witness it in its uncompromised, unassailable glory. Pure fire and brimstone.
Variations on this are the #1 thing I find myself saying at work. "We assume that <coding practice> is optimal, but that was because it was optimal for humans. What if it's not optimal for machines? It might be, but there's no reason it has to be."
Have not been so pleased and excited about a model as I am about Opus 5.5 since ... GPT-4? I don't even use it beyond Medium. Truly a fantastic model.
The modal commenter has become obsessed with “bias” and has started using it to dismiss anything and everything. Bias is real, for sure, but it doesn’t mean “automatically false”! It would be a weird and sad world if people only believed things that didn’t align with their profession or background.
Last week, I had friends totally outside the industry ask me about my p(doom). It would be reasonable to think that this means the Responsible Al field is poised to have increased influence and impact.
And yet, what I see most prominently are RAI professionals variously criticizing the actual risk topics being raised, questioning the motives of those raising them, and congratulating themselves for raising concerns before it was "cool."
What I'm not seeing a lot of? Welcoming people into the big tent of trying to make Al go better, accepting a broad range of views and motivations, or offering help to those looking to learn more about the field.
This is just a huge strategic mistake. I wrote about this almost exactly one year ago (!) and feel it even more strongly today. Insisting that Al concerns take specific, proscribed forms along a limited number of axes while also impugning those who don't is a sure-fire way to never have a seat at the table. I'd say "there's still time" but ... I'm not sure if that's true anymore.
treycausey.com/blog/posts/wh…
One of the bigger downsides to this site is how many people seem to loudly revel in seeing others fail.
The best way to lose interest in a hobby and develop a dislike for its most ardent practitioners (ostensibly) is to follow its subreddit for a while.
Apropos of current events, re-sharing my post from last year on how it's incumbent on all of us working in AI to be intentional with our values and that the task of building safe and responsible AI falls to everyone, not just safety or RAI teams.
treycausey.com/blog/posts/wh…
Lots of attention paid to "how insecure coding agents are", a lot less attention paid to how coding agents are exposing "how insecure everything else is."
This story is actually insane:
• dude drops $2000 on a DJI robot vacuum like a lunatic
• refuses to use the normal app like a peasant
• Sammy Azdoufal fires up Claude to crack the API so he can drive it with an xbox controller
• Claude delivers the goods
• pulls an auth token from their servers, connects successfully
• except the system thinks he controls 7000 vacuums
• checks again
• yep, seven thousand
• DJI built authentication with zero device ownership verification
• any valid token works for any unit on the planet
• Sammy now has eyes inside homes across 24 countries
• live vacuum camera feeds everywhere
• full floor plans from the mapping data
• some guy in germany eating cereal at 3am, unaware his roomba is snitching
• one API call away from being the most informed burglar in history
• all he wanted was to steer his vacuum with a joystick
• does the right thing and reports it
• DJI fixes it in two days
• back to normal life with his stupidly expensive floor cleaner
• IoT companies stay undefeated at shipping garbage security
Subtle but important change I've seen in both Opus 4.6 and GPT-5.3-Codex is along the lines of "Next, if you'd like I could do the following two tasks. . . . Actually, I'll just go ahead and do them." Clearly an emphasis on longer-running tasks showing up; previous models would have paused here.
The continued use of the crypto bubble / hype cycle as a reason that tech can't be "trusted" to be honest about AI is so bizarre to me. At the height of the crypto hype cycle, the vast majority of tech companies had nothing to do with crypto.
Aside from Meta, none of the big players got involved much at all besides perhaps some exploratory investments that were insignificant to their balance sheets (as big companies will do, just to see if there's anything there).
Google, Microsoft, Amazon, etc. were not betting any part of the business on crypto. Based on my own experience and conversations, crypto boosters tended to be confined to rather small pockets at most companies.
I know tech can seem like a monolith to outsiders, but "crypto will change everything" was never a majority view. It just doesn't make sense to draw these comparisons in good faith.
Don't worry, you may very well lose your job because of AI, but staff writers at the Atlantic will be warmed by the satisfaction that the models don't meet their definition of "intelligence"
The "I'm not even watching the Super Bowl" posting has begun. A tradition older than the Puppy Bowl.
The answer to this depends on if you are trying to hire someone to solve today's biggest need or if you hiring someone for the biggest needs 6-12 months. I'm not sure they're strongly correlated, either.
Moltbook is a speedrun test of a lot of theories about alignment that wouldn't have been possible otherwise. I'm not saying it's wise necessarily, but there should be a whole of lot of priors being updates as it runs its course.
Trey Causey retweeted
snowstorm hack, zerobrew is a drop-in brew replacement. borrowing principles from uv (concurrent downloads, content-addressable store), it’s ~5x faster cold and ~20x faster than homebrew.
try it out! github.com/lucasgelfond/zero…