Ben Kremer (@bkrem_) header
Ben Kremer (@bkrem_) profile picture

cooking… 👨‍🍳 // prev: staff eng @WalletConnect // you can do it like it’s a great weight on you, or you can do it like it’s part of the dance

Joined December 2013
Photo by Ben Kremer (@bkrem_)Photo by Ben Kremer (@bkrem_)Photo by Ben Kremer (@bkrem_)Photo by Ben Kremer (@bkrem_)Photo by Ben Kremer (@bkrem_)Photo by Ben Kremer (@bkrem_)Photo by Ben Kremer (@bkrem_)Photo by Ben Kremer (@bkrem_)Photo by Ben Kremer (@bkrem_)Photo by Ben Kremer (@bkrem_)
If this guy can ship a successful restaurant with comic sans, you can ship that imperfect v1
1
6
229
Shenzhen really said “what if we just built the futuristic utopia meme” (last slide)
3
8
827
Photos shot on Fujifilm XT-50
1
50
Agent-generated replies on here and elsewhere are already such a scourge. Whoever figures out how to have truly agent-free social spaces online will make a gazillion dollars
2
1
214
Finally took this for a spin, really good. • Rewrote tests that passed even when the code was broken 
• Removed 70 duplicate tests (−830 lines), coverage unchanged at 94% 
• Dropped test-only exports from prod code (Opus 5.5 high)
OpenClaw deleted around 400k LOC of its own tests without much change in code coverage. Modern models just love writing tests for every tiny change, even if they aren't useful. This skill helped. github.com/openclaw/openclaw…
2
101
Given all the security issues atm, I'd strongly recommend locking your GitHub Actions workflows with the official tool: gh extension install github/gh-actions-lock and then: gh actions-lock
18
11
1
260
17,265
This whole nvidia drama is starting to feel like someone going “oh that first one wasn’t enough to nuke Jensen? Let me drop another one”
1
4
158
If there's a path to software security becoming defense dominant again, this is it. The attacker only needs to find one gap, but the defender knows much, much more about the system. (E.g. the defender has the code.) The defender has to take advantage of that to find their own vulnerabilities faster and cheaper than an attacker would. So we're all going to do a lot of white box testing and code-level vuln scanning, spending tokens to find problems before adversaries do, but doing so more cheaply than an attacker because we have the code and other system knowledge they don't have.
after running our security agents 24/7 i feel better about where things are going they find quite things quickly and have access to internal details which help they should be more effective than an external agent attacker. the asymmetry still exists but not as bad as i thought
2
4
82
7,136
Ben Kremer retweeted
writing code in an interview is now larping in the literal sense. like you are straight up live action roleplaying a historical reenactment of what it was like to be a software engineer.
103
867
70
18,662
306,441
Observations from mainland China so far: - lots of AI generated ads in subway stations. Ads using (non famous) human models are actually the minority - people locked into AI generated mini dramas on Douyin (tiktok) Western aversion to AI media doesn’t feel present here at all
2
5
197
Adding the following to AGENTS .md does wonders to avoid this, also to catch it in (automated) code review: — Code MUST be DRY. Do not duplicate definitions or logic. When the same constant, type, validation, or block of logic appears in more than one place, abstract it into a shared helper/utility and reuse it. Flag copy-pasted logic and repeated literals/definitions that should be centralized.
I still regularly see frontier models make questionable programming choices like hardcode the same string value 10 times in a file instead of declaring a constant, which carries intent increasingly though I'm not sure anyone cares
6
13
305
12,651
Ben Kremer retweeted
Typical adversarial AI review: > Critical issue found [High] > If your server is hit with a nuclear bomb while processing this job, you'll lose data > Consider multi-region data redundancy for failover
25
54
4
1,833
34,675
TIL you can enforce pinned SHAs for actions on GitHub itself 🙌 See the thread’s original post why you definitely, unconditionally should do this - unless you like playing supply chain russian roulette.
Replying to @99florian
Oh fantastic, yes this is even better.
1
4
286
Not surprising that people end up hating AI when they’re on the receiving end of this
I told my Dot to "handle everything for our new office move" and after several days of it still not finalized I realized he had been spamming everyone with every minor update and 100s of questions trying to figure out movers insurance
1
81
Ben Kremer retweeted
She already knew the root password for SSH on Alpine
AAAAAAAAAAAAAHHHHHRHRGRGRGRRRGURBHJB EORWPSOJWPJORGWOIRGWSGODEWPGOHEPW09GJEDPOKSD!!!!!!!!!!!!!!!0924QU8T63095JRGHWPE09UJ0PWHRGW
48
855
27
13,255
307,697
First the privateKey == 4 madness, now this 😵‍💫 The last few weeks have taught me that there seems to be whole class of vuln labelled “too obvious for anyone to even think of trying it”. Great discovery and writeup by @kultekin
*sshole is out if you run sshd on alpine, openwrt or any musl box with password auth on: update now official writeup: sshhole.com
8
785
More things I don’t miss: Getting paged at 2am with a SEV1 incident and trawling logs *manually* to figure out wtf is going on, followed by an attempted fix that you’re not even confident about. All while shit is on fire and cortisol is through the roof. Reviewing huge noisy terraform plans and hoping you didn’t misinterpret something important. Having to write bash scripts by hand and dealing with the obtuse syntax.
the nostalgia for pre-AI swe is so funny to me, yeah man I really enjoyed fixing CI and merge conflicts by hand
2
1
9
551
Ben Kremer retweeted
Replying to @bkrem_
verifiability is all you need its interesting how this is reshaping software, feels like the last 20 years of tooling were ergonomics improvements for the programmer (languages with nicer syntax, less footguns), now they're attention improvements for the human in charge
1
1
2
46
I think we’ll see this move towards “agent-optimal” methods more and more now that we’re firmly moving beyond human code review. Before: agent reviews code in abstract like another human would (anthropomorphic) After: agent uses tools like formal verification (agent-optimal, but was mostly infeasible for humans before)
maybe a hot take but i now consider adversarial reviews mostly a waste of tokens. far better to use those tokens on empirical combined with formal verification. i have updated pstack to use the interrogate skill less, and eventually i think i may remove it altogether i have found very little value with code review in the abstract, agents tend to just get lost in the nitpicks and yagnis
1
5
256
Confirming that this is a goated prompt, so much mileage out of a handful of words. (Opus 5.5 high)
good prompt "find all of the implicit state machines in my app"
10
28
2,085
127,496