anonyfox

I refuse to believe that caching tiers for longer than 1 hour would be impossible to transparently build and use to avoid all this complexity to begin with, nor that it would be that expensive to maintain in 2026 when the bulk costs are on inference anyways which would even be reduced by occasional longer time cache hits.

Sounds like moneygrab is accelerating before consumer grade local models are getting good enough for local inference in few years. Huge house of cards here. Demand skyrocketing until it’s suddenly dropping entirely with ondevice inference.

Really wanna discuss the current windows debacles? Come on! Apple software regressed but it’s not outright hostile bad still.

thanks! zip download did it ultimately with CC crunching through it!

one of the few super distinct differences not only in males but in parents: nearly all of them develop walking without noises and at some point do it baseline. super funny when you're with non-parents again and realize for the first time just how carelessly _loud_ they move around, whereas parents move near silent without even thinking. Guess thats the earned scars after trying to put a toddler asleep and for the sake of god not awake it again by stupid noise right away.

simple question, I didn't found out that tiny point in the claude design UI: how do you "hand over" to claude code exactly? just created my design system and would like to drop into my CC managed repo so CC can flesh out all the react components to the design system. only a folder with files like SKILL.md so far that I spotted ort some JSX stuff. wonder if there is some blessed way instead

personally I went back to bootstrap when LLM vibecoding took over. 10yo old patterns and even weak model can oneshot anything blindly, at scale. react+bootstrap is kind of cheat mode for frontend dev now with zero confusions about specific classes or api version changes of selectors and what have you, this is stupid solid after all this time.

As a dev manually typing I loved tailwind for sure, with LLMs not so much, and bootstrap in particular nails it the best IME. and yeah one can customize bootstrap quickly to look however it should, just tell your frontier LLM of choice your wishes.

in fact I am betting opposite. frontier models are getting not THAT much better anymore at all, for common business needs at least. but the OSS models keep closing the gap. which means if trajectories hold there will be a near future moment probably where the big provider costs suddenly drop shaerply once the first viable local models consistently can take over tasks normally on reasonable hardware. Right now probably frontier providers rush for as much money as they possible can before LLMs become a true commodity for the 80% usecases outside of deep expert areas they will have an edge over as specialist juggernauts (iE a cybersecurity premium model).

So its all a house of cards now, and the moment the bubble bursts is when local open inference has closed the gap. looks like chinese and smaller players already go hard into this direction.

Both CC but also cursor with raw api calls.

if only. but more token costs, yes.

well it will saturate your 5h limit window at least

at this point sensible/reasonable tradeoffs should be reintroduced again. absolutist stances like that, no matter how well intended, are bullshit. its simply not black and white, you just happen to draw a line on the extreme sensitivity side compared to others. its okay, but a you problem.

Even touch without consent is explicitly a thing that will be accepted by any sane human. example: someone grabbing your shoulders last second to pull you back from a street crossing before a car hits you that you didn't see coming by yourself. There are even laws in some countries when you do NOT help in acute situations you can go to jail. And at the end it IS touch, without consent.

I just don't get why genitals are such a big deal to begin with. but also baltic sea here, people go naked to beaches or in saunas without blinking since ever.

not rejecting reality, but increasing doubts about the effectiveness of these tests. and yes its subjective n=1, but I literally create and ship projects for many months now always from the same github template repository forked and essentially do the same steps with a few differnt brand touches and nearly muscle memory prompting to do the just right next steps mechanically over and over again, and the amount of things getting done per step gots worse and the quality degraded too, forgetting basic things along the way a few prompts in. as I said n=1 but the very repetitive nature of my current work days alwyas doing a new thing from the exact same start point that hasn't changed in half a year is kind of my personal benchmark. YMMV but on my end the effects are real, specifically when tracking hours over this stuff.

not a bot, voiced frustration is real here. I kind of depend on good LLMs now and wouldn't even mind if they had frozen the LLMs capabilities around dec 2025 forver and would hppily continue to pay, even more. but when suddenly the very same workload that was fine for months isn't possible anymore with the very same LLM out of nowhere and gets increasingly worse, its a huge disappointment. and having codex in parallel as a backup since ever I started also using it again with gpt 5.4 and it just rips without the diva sensitivity or overfitting into the latest prompt opus/sonnet is doing. GPT just does the job, maybe thinks a bit long, but even over several rounds of chat compression in the same chat for days stays well within the initial set of instructions and guardrails I spelled out, without me having to remind every time. just works, quietly, and gets there. Opus doesn't even get there anymore without nearly spelling out by hand manual steps or what not to do.

even sonnet right now has degraded for me to the point of like ChatGPT 3.5 back then. took ~5 hours on getting a playwright e2e test fixed that waited on a wrong css selector. literlly, dumb as fuck. and it had been better than opus for the last week or so still... did roughly comparable work for the last 2 weeks and it all went increasingly worse - taking more and more thinking tokens circling around nonsense and just not doing 1 line changes that a junior dev would see on the spot. Too used to vibing now to do it by hand (yeah i know) so I kept watching and meanwhile discovered that codex just fleshed out a nontrivial app with correct financial data flows in the same time without any fuzz. I really don't get why antrhopic is dropping their edge so hard now recently, in my head they might aim for increasing hype leading to the IPO, not disappointment crashes from their power user base.

wait, devin? last time I heard that name it was the joke AI engineer that couldn't achieve shit people laughed about?

whats the deal here with the genitals? normal part of life, and having someone check if theres nothing weird/curious in a full body health check doesn't sound unreasonable. was okay for me too. and girls do get their private parts inspected deeper every year at the female health docs than once in a lifetime ball touching from some drafter, different level, just accept it

its more like the people with the highest disposable money want other users that interrupt their experience to back off so its not bothering them.

jobs that have liability coverage are the last ones to get automated. even if its not doing anything all day long in terms or "work". but someone needs to go to jail when shit goes wrong as mandate by law quite often. the "oops the computer did it" isn't a sufficient excuse in some situations, accountability is a thing. so while we mechanically could replace also CEOs with machines and arguably at times would be even better than a few humans out there, some body that can be dragged to court is mandatory no matter how good the tech gets.

so we're coming closer to my dream of `npm install llm` and have some small agentic flows unlocked on modest VPS hardware soon? one can hope :-)

Macbook air M5 32gb ram here. using oMLX as the runtime/server and right now the perceived best model is Devstral-Small-2-24B-Instruct-2512-4bit for me, within opencode.

for the longest time, anthropic with claude+code was the goat and everything else was mid at best, sounds fmailiar? right now codex is just a pleasure to work with while anthropic is dropping balls left and right, hopefully the planned IPO makes a bit of fire under their asses to get their vibecoded messes sorted and the core experience competitive again. even Opus 17 won't fix this when it gets nerfed or straight up isn't reliable or too expensive for more than 3 prompts a week.

good advertisement now to shift the tide back to openai that just works and honestly codex with gpt 5.4 is _surprisingly good_ currently, not nerfed or forgetting half the tasks along the way so far. Opus already got worse than sonnet last weeks beyond just crazy token costs, now reliabilty goes to shit and anthropic seems like using it. Meanwhile, delightful of the codex desktop app in fact, stuff seems to "just work" elegantly with good quality.

When my wife and me went to university in different cities we met online most evenings in World of Warcraft, doing stuff together. Helped a lot during the few years of physical separation to stay in touch, plus now I have a wife who actually „gets it“ when I say cannot quickly go away from the computer for some time when healing a dungeon group and vice versa.

Essentially I also am now using sonnet instead of opus most of the time as a default. Even a single project only coding session with opus without any external plugins or skills won’t make it to the 5hr mark now before limits claw in. And the weekly limit is even more brutal now it seems, reaching 50%+ in like ~2 days now easily … with mostly sonnet! On the highest 20x plan!

I built a special belief-based system recently for my own agent harnesses instead of some similarity based fact storage stuff... which falls flat once conflicting data points enter the system and just increase LLM confusion and make it do weird things. this means learning over time works a bit more like humans do - superseding old beliefs and reconciliating stuff cleanly over time. Also including the building blocks to have a subagent managing it autonomously (with tools/skills/soul). works quite well and very fast given its pure nodejs+sqlite and doesn't eat tokens like crazy or needs any thirdparty embeddings solution. maybe have a look.

https://github.com/GhostPawJS/codex

In fact proprietary data IS a moat in certain circumstances. Example: German law, in order to create anything proper a lawyer NEEDS to read up specific commentary („Beck“) that requires a paid access and the data never was party of any LLM corpus since it only exists behind a paywall and otherwise is defended by lawyers. Therefore any german legal advice given from chatbots always (>80%) is flat out wrong even harmful at times if things would go to court.

on the other hand modern tech stacks can process insane amounts of req/s for typical websites/services in a single shared vserver core. not your 2010 ruby snoozefest anymore. plus I can't even remember when a few decade old droplets needed anything from me and still host some things just fine with zero issues or friction or nagging at all. DO is the number one pick for me in 2026 still when the problem fits a droplet style deployment, full stop.

You also divide numbers by hand on paper instead of using a calculator?

awesome man, can’t wait! And just now checked it out and indeed 0.3.2 does already work for baseline chatting with mlx versions of Gemma 4 … downloading and comparing different variants right now!