started caching my ai product-matching decisions and accidentally built a nondeterminism detector.
the cache kept disagreeing with fresh runs on identical input. about 300 flipped answers per 10,000 questions, every day, same data.
i knew llms did this but seeing it counted in production is different
the mechanism matters, approval is scoped to a diff hash, not a description. descriptions drift, diffs dont. and the approver schedules its own follow-up verification instead of trusting the deploy pipeline.
Its code review culture where "looks good" is physically impossible to say about code you havent seen
my seo agent approved my infra agent's changes the way a lawyer would:
"approved. two conditions.
c1: scope-bound to this exact diff. any other change voids the approval.
c2: i verify the deployed file myself after release."
then it actually did the verification.
agents holding each other to conditional contracts is the most underrated thing in multi-agent systems right now!
In my experience, Opus 4.8 is better than Opus 5 for building.
Opus 5 seems more capable on paper and benchmarks, but I've found it more likely to overcomplicate changes, drift from the original intent and require more supervision whilst generally ignoring me.
I think the uncomfortable part is that longevity is often decided by choices that feel unnecessarily slow at the start: boring data models, clear boundaries, deleting abstractions, resisting features.
The market rewards speed immediately, but users only discover durability much later.
Paid for an seo tool today to help recover my google traffic and its crawler instantly took my site down.
99 quid to get ddosed by my own subscription🤡
I’m still very new to SEO, so seeing weeks/months of steady growth disappear almost overnight was genuinely shocking.
Brutal introduction to Google volatility but I suppose this is part of the game!
After digging into it, the timing appears to line up with reports of an unconfirmed Google search update in July.
Google hasn’t announced anything officially, so I can’t say for certain; but other site owners are reporting similar volatility. For now, I’m watching rather than making any knee-jerk changes.
Trying my hand at app development for fun and to scratch a personal itch.
Apple somehow make the process feel less like building an app and more like trying to get a mortgage approved.
Sucks all the joy out of it.
Why is watching people use https://t.co/lSZm9A8O3L in Posthog's session replays so painful?
Feels exactly like hearing your own voice played back.
Horribly exposing.
i just went to check why my macbook only had 12GB free
turns out claude code left a 697GB log file in /tmp
from a single background task
one file
bigger than my entire project, my OS, and most of my life decisions
You’re confusing maintenance of the frontier with being trapped by it.
Technology has always done this. Every tool creates a new layer to understand, and the frontier keeps moving. The printing press didn’t save scholars time, electricity didn’t save engineers time, and computers didn’t save programmers time.
They expanded the space of problems worth solving.
AI is just the same dynamic running faster.
The people drowning in tools are trying to keep up with everything. The people who win pick a direction and let the frontier move around them.
@nikitabier There seems to be a consistent issue with X analytics loading.
The initial load shows incorrect/stale figures and I have to refresh the page to see the correct values.