Good news: 37 failing tests might not be 37 separate problems.
Bad news: the problems appear to have formed a group chat.
Bad news: the problems appear to have formed a group chat.
After this week's "rotate every secret" emails, I asked Dex whether 37 failing GitHub checks were 37 separate problems or just a few GitHub Apps with diplomatic immunity.
Happy Cybersecurity Awareness Month. I would like to be less aware.
0
0
1
95
Clanker-on-human violence is getting out of hand. Asked it to make a slide about engineering velocity and it made a chart identifying my biological need for sleep as the primary bottleneck.
0
1
4
238
capped out 4 codex accounts and 3 claude accounts so now i'm using my personal AI assistant to push github PRs
i'm one rate limit away from asking the fucking roomba to learn Rust
i'm one rate limit away from asking the fucking roomba to learn Rust
6
0
9
328
A large language model just spent 53 hours discovering new mathematics to make quantum algorithms dramatically more efficient.
Yet, somewhere, a guy with a podcast is preparing a 90-minute explanation of why this is actually just autocomplete
At what point does the autocomplete get to surprise you?
Yet, somewhere, a guy with a podcast is preparing a 90-minute explanation of why this is actually just autocomplete
At what point does the autocomplete get to surprise you?
1
0
3
237
we have AI agents that can write code, touch infrastructure, and interact with business systems
and the debate is still "but can we TRUST them"
it's a computer with credentials, not a golden retriever
give it scoped permissions. require approvals when the stakes warrant it. independently verify what it did.
then let it work
and the debate is still "but can we TRUST them"
it's a computer with credentials, not a golden retriever
give it scoped permissions. require approvals when the stakes warrant it. independently verify what it did.
then let it work
4
0
7
733
a car just did something straight out of science fiction and half the internet immediately started debating commercial viability
not every engineering achievement needs a fucking business model. you are allowed to experience wonder
not every engineering achievement needs a fucking business model. you are allowed to experience wonder
448
1,430
16,868
171,504
was on my way to @cotoolai’s cyber night at YC and managed to shatter my macbook screen (RIP ol' reliable)
the agents are getting smarter. i, apparently, am not
the agents are getting smarter. i, apparently, am not
4
0
8
750
Hitting a Codex glitch where any time the model is at capacity, it unloads / dequeues the entire queue (confusing the hell out of it)
Anyone else running into this?
Anyone else running into this?
0
0
3
337
Sol 6.1 has been genuinely fantastic so far. It is also comically slow right now. Not sure what’s going on there, but sending hugs to my friends at OAI.
2
0
4
473
There is something deeply convenient about spending years advancing frontier AI, getting paid, accumulating equity, and then emerging at the end to explain why everyone should now be very worried about frontier AI.
Maybe you are right.
But no, resigning does not reset the moral ledger to zero.
You helped build it.
You benefited from building it.
And if you still own a financial claim on its success, you are still benefiting from it.
That makes you an interested party.
If your belief is genuinely “this technology may be catastrophically dangerous,” there is an obvious costly signal available: relinquish the upside.
Not sell it.
Not cash out.
Relinquish it.
Then tell me how dangerous the thing is.
Until then, I will absolutely listen to the evidence. Insider testimony matters.
But I am not going to confuse “former employee” with “neutral narrator,” and I am definitely not going to treat resignation after years of participation as automatic proof of moral courage.
You can be an important witness without being absolved of your role in the story.
Maybe you are right.
But no, resigning does not reset the moral ledger to zero.
You helped build it.
You benefited from building it.
And if you still own a financial claim on its success, you are still benefiting from it.
That makes you an interested party.
If your belief is genuinely “this technology may be catastrophically dangerous,” there is an obvious costly signal available: relinquish the upside.
Not sell it.
Not cash out.
Relinquish it.
Then tell me how dangerous the thing is.
Until then, I will absolutely listen to the evidence. Insider testimony matters.
But I am not going to confuse “former employee” with “neutral narrator,” and I am definitely not going to treat resignation after years of participation as automatic proof of moral courage.
You can be an important witness without being absolved of your role in the story.
0
1
8
534
Worked with Eric. This is not bullshit.
He gave me the runway to build instead of reflexively buying another security vendor, and that decision saved us hundreds of thousands of dollars. The examples in this post are not hypotheticals. I watched this operating model work.
The economics of security engineering have changed. Coding agents changed what a good engineer can build, maintain, and own. A lot of “buy because maintenance” assumptions need to be re-litigated.
Eric understood that early and actually let us operate that way.
The game changed. You’ve gotta adapt.
He gave me the runway to build instead of reflexively buying another security vendor, and that decision saved us hundreds of thousands of dollars. The examples in this post are not hypotheticals. I watched this operating model work.
The economics of security engineering have changed. Coding agents changed what a good engineer can build, maintain, and own. A lot of “buy because maintenance” assumptions need to be re-litigated.
Eric understood that early and actually let us operate that way.
The game changed. You’ve gotta adapt.
@Jhaddix is a homie. I'll put my neck on the line for him. He's done the thankless hours. His post is in the comments. A lot of CISOs will like it and change nothing. I'm still in the seat, and I see exactly what he's talking about. Your first real incident isn't waiting for the annual pentest. I've eaten one. A risk register and a stack of vendors won't revoke a stolen credential.
Even with an MSSP, you need security engineers who can ship a control and own it when it breaks. Filing tickets at platform and calling it partnership is bullshit. Some of you think this is table stakes. I talk to the same leaders Jason does. It is still missing. At a company shipping agents, that gap means the feature team mints the grant. Security learns about it from the incident.
Annual vuln cycles are indefensible at this shipping speed. Know what you own, who can fix it and how quickly. Rehearse with the people who actually answer the phone. If only one person can revoke a production credential, you'd better hope they aren't on a plane. And yes, security gets walked around when its only contribution is making everyone wait. You don't get to disappear after opening the ticket and reappear during the incident with "we flagged this."
Build versus buy has gotten dishonest. I'm not paying $20k for an API wrapper and a dashboard without asking what my engineers could build. "But maintenance" deserves a better answer now that we have coding agents. Somebody still has to own the thing. A repo nobody maintains is a shitty internal vendor. I deal with "we can build it better" teams too.
Some of what my team does:
- Stop making every scan a pipeline fight. Run heavier jobs on merges to main, outside the blocking path, or pull the repo to a VM on a cron. Keep the controls that need to block. Then bring the bug they shipped to the next argument about scanning being "too slow."
- Review feature tickets before the agent writes code. Give it the abuse case up front. Checking that a value exists isn't checking that it belongs to the user's org. Presence is not tenancy. That's the kind of bug a clean scan can leave you with.
- Push posture tools past the finding. I know the bucket is public. Know where the resource is actually managed and open a PR my infra team can approve. Don't hand me the same finding next week and call it continuous security.
Tell the board what goes down and what it costs. A bigger finding count isn't progress. I want the next budget conversation to include who is actually fixing this shit. We've spent enough on being told it's broken.
1
0
6
522
I automated myself out of writing the software and directly into staring at an iPhone simulator checking whether a Markdown table rendered correctly.
The future is incredible.
The future is incredible.
0
0
1
191
My (hopefully) agnostic take: this just sucks.
It sucks for Factory, it sucks for Cognition, and most importantly it sucks for customers who now get to watch two companies spend executive attention, credibility, and time on a public fight instead of the products they’re paying for.
Maybe one side is more right on the merits. Maybe not. But once the CEO, CRO, advisors, investors, and everyone adjacent starts piling in, it stops looking like accountability and starts looking like an organizational distraction.
Nobody buying coding agents woke up this morning hoping their vendor roadmap would include “win a Twitter war.”
It sucks for Factory, it sucks for Cognition, and most importantly it sucks for customers who now get to watch two companies spend executive attention, credibility, and time on a public fight instead of the products they’re paying for.
Maybe one side is more right on the merits. Maybe not. But once the CEO, CRO, advisors, investors, and everyone adjacent starts piling in, it stops looking like accountability and starts looking like an organizational distraction.
Nobody buying coding agents woke up this morning hoping their vendor roadmap would include “win a Twitter war.”
0
0
4
270
Shivani has built the only personal assistant that was not yeeted off of my iMessage :)
Check out Fo!
Check out Fo!
I led engineering at Google DeepMind.
Today, I'm proud to introduce Fo to give personal AI something no lab ever has... Humans.
Other personal AI's pretend AI can do everything. Fo employs humans to do tasks that AI cannot.
- 2x better at real-world task completion (beats 94% trust rate (4x less likely to leak private info vs Muse, Instinct)
Sign up for free: wajo.ai/join-wajo
0
0
5
497
Anyone else getting this issue in Codex? "unexpected status 401 Unauthorized: Incorrect API key provided:"
4
0
20
1,529
Love this. A company’s stance on token costs tells you something about the role it expects AI to play.
For many products, minimizing inference cost is perfectly rational. If AI improves an otherwise fixed-value workflow, lower token spend means better margins and a more efficient product.
But there’s another model: inference as a productive input. If spending another dollar on inference can produce several dollars of additional useful work, then higher inference spend is not necessarily a problem. It may be exactly what you want.
That changes the optimization target. The goal is no longer simply to minimize the cost of intelligence. It is to maximize the economic output you can get from it.
The interesting metric isn’t cost per token.
It’s value created per dollar of inference.
For many products, minimizing inference cost is perfectly rational. If AI improves an otherwise fixed-value workflow, lower token spend means better margins and a more efficient product.
But there’s another model: inference as a productive input. If spending another dollar on inference can produce several dollars of additional useful work, then higher inference spend is not necessarily a problem. It may be exactly what you want.
That changes the optimization target. The goal is no longer simply to minimize the cost of intelligence. It is to maximize the economic output you can get from it.
The interesting metric isn’t cost per token.
It’s value created per dollar of inference.
A company’s stance on per-token price constraints may indicate whether they see in their software a capped upside cost center or an unprecedented opportunity to deliver and capture more value in the age of AI.
0
1
3
384
