anirka

Seems like there would be ongoing debate whether a company shall rely only on one cloud provider for LLMs or not. I personally have stopped using SOTA models to make sure I can rely on "good-enough" LLMs for almost any task possible. Currently working on stuff that will get banned by OpenAI or Ant.

I've become tired of slopping to some extent. It's time to take care of my prompts and code review again (yes I started to get too many bugs and my token credit balance is getting low, I can't afford outsourcing review anymore, sheeeet).

How has Go been WASM-wise and browser-wise? What's up with the tooling?

There's apparently a big difference between merely prompting and meticulously prompting for software engineering. Holy shit I just run a CRUD-ish project I was given to extend and it did 3 TB disk rewrites in just 15 minutes. And there's no proper profiling and logging! Holy fuck... Opus 4.6-5 btw

I have almost solved my sleep schedule. Consistently waking up mid morning, finally (7-8 am), yahoo!

Apparently it is never a good idea to skip a day in Anki. God help me clean this backlog up...

prime-agent is very good. I can finally use cheap models properly as needed and it does not shitmax the context with useless shit.

Do you think human oversight is still required for reliability in software? Do you think human oversight is still required for **cheap** reliability in software? Wonder what people think about these questions.

Another paper discussing how LLM agents may "get ideas", this time not as a distributed attack strategy (cf. https://adhocism.net/pages/aug-6th-2026/), but as spontaneous spread of belief-modifying content between agents: http://arxiv.org/abs/2608.10218 I wonder when we'll get our first LLM […]

Using token resellers to use 5.6-luna at insanely low price. Still does mistakes though sadly, but overall finally not worrying about depleting money on these matrix multiplications.

Once you get past first 100-200 cards seen in an Anki deck, it feels so natural just to continue. Only to actually get through it completely in a year or so. Do not stop the grind.

I do not get this political shit going on bsky why is my feed is full of us political blabbering where pure stupid shit clashes with just normal adequate thinking and reasoning? I installed bsky just to maybe explore something else after ten years of not using any social media. Wtf?

Listening, recognition, and recall at the speed of light. Day one attempting doing 70 new cards a day without thorough practice.

This is interesting how DeepSeek v4 Flash 0731 costs so low but assists so much in development. This is the first time I started using pay-as-you-go API instead of using a sub. Wonder how exactly did they achieve such low cost of inference in comparison to various sub-100b models.

Your phone won’t run an app that’s been tampered with, but your browser will run anything a server sends it. WEBCAT is a tamper-evident seal for web applications, developed by @freedom.press Find out more:

I once finished doing about 1200 flashcard reviews in a day. That was a big mistake so far I am now happy with just 10-15 new cards a day and then up to 100 reviews usually. That's at least a guaranteed deck completion in a year or so.

Cramming 200 words in a day is a mess but it was all just reviews 💀 that piled up throughout lazy weeks, daymn. Do not skip your anki days.

I just realized that when writing hanzi/kanji it's okay to make mistakes unless you are doing something formal or need to maintain strictness. And it is even more lenient in japanese than chinese, that so it's ok if the overall character shape is understandable... hmm.

Yeah, tones are problematic. Distinguishing 县 and 险 by themselves is not that hard but in speech... they sometimes sound the same. And as of 4th tone... getting used to always spell it properly does indeed take some time. I always tend to make it a neutral or 3rd tone, damn me.