690 karma · joined February 21, 2023
getinvoke.com
I hypothesize it may have roughly the same effect as denying all contributions where the author used intellisense 10y ago.
A substantial portion of people who write good code will be using some sort of LLM assistance, even if it is just something like Cursor Tab (autocomplete).
Yes, you'll also hit all of the spammy PR "contributors", but you'd also do so by prohibiting all contributions by people who have a belly button
I have an idea you might be interested in.
Technologies: Anything. Very strong with TypeScript, frontend (React, Svelte, Vue, etc.) and backend. I love Bun. Very strong with Rust. Have used many languages.
I've built getinvoke.com - an app that lets you add custom, context-aware shortcuts to any mac app. Currently focused on music producers, mixers, and anyone else who spends their day doing audio work.
Very comfortable with Audio related work. Have good experience with music production and mixing, and mentored indie artists.
Got good grips on how React, Svelte, Vue, Solid work, both internally, and in practical terms. Led the transition for a 20K DAU website from a jQuery mess to a rewrite in Vue.
I've built Ethereum/Web3 payment systems before.
Worked with realtime online gaming and implemented cryptographically provably-fair algorithms.
Came up with & implemented a pluggable AI workflow system (think n8n) in the GPT-3.5 days.
Absolutely love design and UX. Kinda obsessed to be honest
Email: merlin@merlin.audio or merlindruzinec@gmail.com (personal)
if someone made the same gigantic mess of a PR without LLMs, it would still be rejected, because it is a gigantic mess of a PR.
the low effort part is the problem. what if i made a great, focused, readable PR but had claude write it out? what if i carefully checked and deliberated each line, just as if i had written it myself?
granted, in the real world, 99.9% of slop PRs are written by LLMs. so i thought "okay, reasonable, ban the thing that is most likely to cause problems."
but then how does the "no LLM translators!" rule fit into that view?
And even worse, if I wanted to take over npmjs.com tomorrow and godaddy would kinda... just hand it over (?!?!?!) then i could probably become a crypto billionaire overnight
Opus 4.5 and 4.6 were the first models that i could talk to and get a sense that they really "understood" WHY i'm saying the things i am
Opus 4.7 kinda took that away, it's a definite regression. it doesn't extrapolate.
———————————————
refactor this thing? sure, will do! wait, what do you mean "obviously do not refactor the unrelated thing that's colocated in the same file"? i'm sorry, you're absolutely right, conceptually these two things have nothing to do with each other. i see it now. i shouldn't have thought they're the same just because they're in the same file.
———————————————
whereas GPT 5.5, much like Opus 4.6, gets it.
i wanted to build a MIDI listener for a macOS app i'm making, and translate every message into a new enum. that enum was to be opinionated and not to reflect MIDI message data. moreover, i explicitly said not to do bit shifting or pointer arithmetic as part of the transport.
what did Opus 4.7 do? it still used pointer arithmetic for the parsing! should i have to be this explicit? it also seemingly didn't care that i wanted the enum to be opinionated and not reflect the raw MIDI values. Opus 4.6 got it right (although with ugly, questionable implementation).
GPT 5.5 both immediately understood that I didn't want pointer arithmetic because of the risk of UB and that shuffling around bits is cumbersome and out of place. it started searching for alternatives, looking up crates to handle MIDI transports and parsing independently.
then it built out a very lean implementation that was immediately understandable. even when i told Opus 4.7 to use packages, and even how to use them, it still added a ton of math weirdness, matching against raw MIDI packet bytes, indirection after indirection, etc. even worse, it still did so after giving them the public API i wanted them to implement.
GPT 5.5 nailed it first try. incredibly impressed with this model and feel much safer delegating some harder tasks to it
I don't like that trend. I get why they're doing it, but I don't like it
their own, called Zeta 2
and then Mercury by Inception. May also be available in VSCode through some third party extension like Kilo, not sure
it seems to happen more when i'm overweight, making me think it's blood pressure (BP) related, but then doing the valsalva maneuvre, which spikes BP, doesn't cause any problems at all.
i've tried acetaminophen, even 1.2g of it, to no avail. it doesn't help.
i've also tried every other remedy, such as curcumin, fire/ice locally, hot and cold showers, neck massages, working out muscles that may be involved in it, everything. nothing helps.
except for ibuprofen. 400-600mg kills it every time.
at least for me, there seems to be a definite difference, as ibuprofen can anecdotally help in some situations that acetaminophen can't. i wonder what exactly it can / can't treat and why.
i asked it to look at writing research (especially from NN/g) and come up with some alternatives for a heading that is roughly supposed to convey:
"this app lets you create custom shortcuts for all mac apps, even sophisticated ones, mouse wheel ones, ..."
it came up with the following headlines:
1. ONE APP, MANY MAC APPS
2. ONE INSTALL, MANY APPS ONE APP.
3. ONE PACK PER MAC APP.
4. ALL YOUR APP SHORTCUTS IN ONE PLACE
5. [app name] IS ONE APP. PACKS COVER THE REST.
whereas GPT-5.4 came up with 1. The Shortcuts Your Apps Are Missing
2. What Your Apps Still Don't Let You
3. Do The Parts You Still Do by Hand
4. When Built-In Shortcuts Run Out
5. Where Your Apps Stop Short
now both of these aren't amazing, but please tell me how in the world "ONE APP MANY MAC APPS" makes sense as a headline for fucking anything lolthat's not something even GPT-3.5 would come up with.
"one install, many apps" ........huh???
- economics: i'd wager a bet that Opus 4.7 is just distilled Mythos Preview - performance: surgery like this would explain the spiky performance and weird issues
just spitballing tho
Opus 4.6 and 4.7 are better than GPT-5.4 xhigh, but only marginally. I can't give proper pointers on what to change because it's incredibly hard to quantify.
In essence, though, GPT-5.4 needs explicit instructions not to take liberties - this is included in the default system prompt of Claude Code which leads me to think Opus is just as overzealous as GPT-5.4 unless explicitly told off.
And it takes EVERYTHING you say at face value. Questions like "don't you see why this is bad?" will be answered with "yeah, i do." which is also kind of cool...
because with Opus in Claude Code i constantly have to reassure the model i'm not insinuating anything, lest it takes my question and runs with it into a frenzy of "oh shit my bad let me fix it im so sorry" type changes.
i personally haven't noticed any downgrade at all.
it's entirely possible there's a mass delusion going on where everyone gets wowed by 4.6 initially, then accepts the new baseline and gets used to it, then thinks that baseline is no longer impressive and thus degraded
it doesn't help that anthropic changed defaults for its claude code harness for all users suddenly
the best and only evidence i've seen for actual degradation is that the web version of opus 4.6 failed the car wash test, and since you cannot simply choose to "disable adaptive thinking" and other parameters with the web version, you truly may have gotten a worse product
[1] https://marginlab.ai/trackers/claude-code-historical-perform...
either way, all that one can really rely on is the benchmarks, and those are easily cheated/overfitted to.
i think it's all very hard to quantify, so take my previous comment with a massive rock of salt
definitely not bologna, at least anecdotally :)
it's like the idea behind the book _The Mom Test_ suddenly got very important for programming
...because false positives are good errors. false negatives is what i'm worried about.
i feel massively more sure that something has no big oversights if multiple runs (or even multiple different models) cannot find anything but false positives
Usually I ask something like this:
"This code has a bug. Can you find it?"
Sometimes I also tell it that "the bug is non-obvious"
Which I've anecdotally found to have a higher rate of success than just asking for a spot check
The crazy thing is that it's a diffusion-based LLM. That makes it very fast, like Cursor Tab, and the outputs seem very accurate in my limited testing (although I find Cursor Tab to still feel "like 10% better")
---
That said, you should really give agentic coding a la Claude Code a try. It's gotten incredibly good. I still need to check the outputs of course, but after using it for 2-3 days, I've learned to "think" about how to tackle a problem with it similarly like I had to learn when first picking up programming.
Once I did, suddenly it didn't feel risky and weird anymore, because it's doing what I would've done manually anyways. Step by step. It might not be as blackboxy as you think it is
the expectation is that they'll eventually make money. they can't raise forever. only startups are not profitable for a few years. but most companies that have existed for a long while have been profitable
and since they're expected to make a LOT of money, everyone wants a piece of that future pie, pushing up the valuation and amount raised to admittedly somewhat delusional levels like here