when the AI support agent passes you onto a human whose first message begins "you're absolutely correct, ..." π
thanks, now I just have LLM with a longer delay
when the AI support agent passes you onto a human whose first message begins "you're absolutely correct, ..." π
thanks, now I just have LLM with a longer delay
forks like this are pretty straightforward to keep updated: you just copy the latest stdlib code and re-apply your minimal patch to expose whatever internal fields you need
I think forking is the correct, idiomatic way to do this
As a comparison, Go gives you little choice in TLS cipher suite: go.dev/blog/tls-cip...
If your use case is nuanced enough to need it, you can easily maintain a fork (as people do for the TLS package: github.com/refraction-networking/utls)
well this looks like a great nerd-snipe: given a number, is there a way to quickly find its minimal representation?
hmm, just getting a second 16GB stick has gotta be the best bang for buck, right?
still Β£100+ though π
related: I regularly see a quoted post followed shortly by the original
which always seems a bit redundant? I already saw it as part of the quoted post
I'm also mid-way through reading The Soul of a New Machine which means @bcantrill.bsky.social is 2/2 on great recommendations so far this year
been listening to the shellgame.co podcast and, wow, the LLM-dialect is so much more annoying when put through text-to-speech
overall great concept though e.g. creating a voice clone of yourself and then sending it to therapy
guilty...
on _multiple_ occasions I've watched automation go rogue and suspend accounts for an entire company
since then, I've always built a fuse* into systems like this: if too much stuff tries to happen at once, the fuse blows
(* an if-statement with a fancy name)
I'm getting similar results for Amazon and Google badges though so π€·π»ββοΈ
generally frowned upon to be publicly posting pictures of your ID badge anyway, so maybe it's just very limited training data
*not* having the company name on ID badges is pretty much corporate opsec 101
I wonder if the training data is polluted by all these etsy listings for Uber Eats drivers
oof, four years since my last (personal site) blogpost π§
so, like every other developer's new-year writing habit, I've started with a classic mistake: spending a bunch of time re-working the theme
the default behaviour of most ACME clients already seems to be renewing much earlier than technically necessary, so a properly functioning setup wouldn't see any failures
but a broken setup, or forgotten manual renewal would see aborted TLS handshakes gradually increasing
> Itβs not like thereβs an increasing percentage of requests that fail as you get closer to the deadline.
now there's an interesting idea for when we next shorten cert lifespans:
for the last 7d of validity, clients should start intermittently failing an increasing percentage of requests
I'm definitely mis-calibrated then: routinely going back for the last sip only to find the mug is already empty...
and, depending on the model you got, valetudo.cloud might protect you against a similar Roborock fate!
I'm stuck with eufy with its stupid 1.5GB app. literally the largest app on my phone. larger even than my podcast player including all downloaded podcasts
it's almost 2026 and a doodad I just bought came bundled with software/drivers on a CD
ah that makes sense. yeah otherwise you've no choice but to brute force?
I'd assumed it was done in the "the md5 of this sentence begins..." style where you just have enough entropy (via e.g. alternate phrasing) to guarantee collision
(the dithered art style seemed good for hiding that entropy)
just π€―
not just a hash quine (image which contains its own hash), but an image which can render *any* hex string without it's md5 changing
just used this to find two liked posts which were causing the majority of weird stuff in my feed
1. enter your handle
2. scroll through and click "explain" on anything weird
3. check "top source posts" for repeat offenders
4. go un-like them
(I'm hoping that un-liking actually has an effect π€π»)
interesting how there's a consistent-across-models mistake in which side of the plug opens
(the location of the screw in step 4 is correct, but in reality it's the back of the plug that comes off, not the front)
"π is so close to π" still could be true though!
(assuming it's an old-school machine with rotating drums which could have those symbols adjacent)
it's also what I predicted πͺπ»
storing hashed phone number pairs is a very smart way to do this
obviously needs some further scrutiny, but this approach solves so, so many of the vulnerabilities you usually see in contact discovery systems
in a similar vein: a surprising number of cold-email bots assume you can get the company name by:
stripping TLD from the domain name, and capitalising
I get multiple emails a week congratulating me for founding Phish...
so maybe IDEs should just have an integrated web panel like they do a terminal π€·π»ββοΈ
the one benefit I can think of is: being in the terminal lets you embed yourself in arbitrary IDEs
cleaner than a separate GUI window, but without the overhead of maintaining separate extensions for intellij, vscode, zed, etc.
feels like the terminal is just an aesthetic choice now? wouldn't building an actual GUI be way simpler at this point?
> [use] your mouse to click and navigate directly within the input
> clean and polished display regardless of how you resize your window