ariwilson

I'm a little surprised that people would rather rent a Mac than buy one. Less so now that Apple raised the prices but for a long time a M4 Mac Mini was $599 which only works out to about 6 months at these prices.

Especially since they're not like a regular server, loud and power hungry. They're the definition of quiet and power sipping. I have one do all my background coding work when I want to keep my laptop clear (being headless helps with me moving around too).

just get a mac mini headless

Hi all! I’m building https://fairbake.dev (the business around Bakeoff) and when the question came up again in one of my calls of what effort level to set on model X for new {feature, bug fix, agent, categorizer, …} type tasks, I figured I’d try to answer it quantitatively using the eval framework and challenges I’d built alongside my clients.

When I dug into it, I realized that the models shipped at different default effort levels. That seems confusing to users - I wonder why the labs do this. I was surprised to see that increasing the thinking level often didn’t improve correctness and almost always increased (non-test) diff size. I was less surprised that Fable at high and Opus at max did poorly.

Ask me any questions you have; happy to share specific differences I noticed between solutions. What effort level are you running at and why?

Unfortunately not really a lot of places to do this in the US. And unlike shipping new devices from China to US, shipping used devices from US to China and back is prohibitively expensive.

You mean word budget? I don't see how a token budget on the LLM helps with doc length. Or do we denominate the length of text docs in tokens now?

I had a friend yesterday tell me he likes the way Claude writes and finds it easier to understand compared to Gemini. But he mostly doesn't use Claude for coding. Whereas I find Claude insufferable in writing often and prefer Gemini for writing. But I don't use Gemini for coding.

Familiarity breeds contempt.

LLM generated writing; couldn't make it through. I read too much of it every day. I had a visceral reaction to the article.

I ran it through my own coding eval today using pi/OpenRouter (11 small bug fixes from my own projects). It solved 10 of 11. Compared to Sonnet 5.5 and GPT-6.1 Sol, a median run cost me $0.52 vs $0.18 for Sonnet and $0.09 for Sol, and took ~4 minutes vs 32 seconds and ~1.5 minutes.

It thinks a ton, ~24k output tokens a run vs ~3k for Sonnet. It missed a solution by thinking its way into a 10 min timeoout.

Something to look out for - if you're using pi, it doesn't have Argon in its catalog yet and reports cost at Kimi's rates, about a third of what OpenRouter actually bills.

Hey folks; I’m the author of RemainderBot. Given your goals and task lists, it writes PRs that complete those tasks for you using excess Claude & Codex subscription tokens.

As the founder of Bakeoff, I spend my time on customer development and engineering architecture & review. RemainderBot solves two problems for me and I suspect for many others out there:

* There are many small tasks that require a few minutes of my time and need my context, but I just don’t have the mental energy to prioritize them, do a context shift, perform a small amount of work, and finalize the results a dozen times a day.

* Some weeks I don’t get close to my Claude and Codex subscription limits. That’s a waste of what are very valuable Astra and Opus 5.5 tokens :)!

RemainderBot has consistently used my goals and task lists to identify non-obvious tasks that would move me forward. It has accomplished quite a bit - 111 PRs including customer call prep, a census of 31 GitHub orgs’ agent setup files, and standalone prototypes to demo to engineers I meet with.

You interact with it in whatever interface you use with GitHub PRs. It uses your comments and decisions on whether to merge PRs to guide which kinds of tasks and approaches it takes in the future.

RemainderBot runs either as a manual process on your machine or as a cronjob on a server. In tick (cronjob) mode, it checks how much subscription quota is left and how much time there is until the weekly reset before picking a task and creating a PR.

Here’s how you get started:

* Check that git, claude (or codex), and gh CLIs exist and are authenticated on your machine

* Clone the openremainderbot repo:

  git clone https://github.com/arimwilson/openremainderbot.git remainderbot
  cd remainderbot
  cp -n GOALS.example.md GOALS.md

  cp -n sources.example.toml sources.toml

Edit those two files with your personal goals and with sources for your task lists, respectively. Finally:

  python3 -m remainderbot snapshot
  python3 -m remainderbot run --provider claude --deadline-minutes 20 --size small

These commands will pick a task, work on it in a new branch, and will cost you a few percent of your weekly quota. Please drop a comment on what it did for you and if it got anything wrong here or in https://github.com/arimwilson/openremainderbot/discussions/6 :).

More details in the blog post - https://www.ariwilson.com/writing/openremainderbot

Hope you enjoy it!

What is AI benchy and why does it rate everything on passing 23 tests? Putting Gemini 3.6 Flash on top of the leaderboard doesn't pass the smell test.

vals.ai has great marketing! Unclear if the product is great yet. Or what it is exactly.

+1000. The best part of being a manager is when you can help a person achieve their potential.

I'm just surprised that the marketing blog post about the omnipotent new AI model (that no one outside Google can currently access - contrast with the Opus 5.5 / Astra launches) - doesn't pick examples where every metric is better than before.

Just compare how much better presented the Astra announcement was compared to this one: https://openai.com/index/gpt-6-astra/

Damn way to undermine yourself in your own blog post Google:

"The end result is a memory-safe video decoder that runs 2.7x faster than the Rust port, with identical video output, bringing it closer to the optimized C++."

Close but no cigar!

Sonnet 5.5 vs Sonnet 5 on Bakeoff (real bugs/features, graded by held-out tests) is a huge upgrade:

* 5.5x faster

* 4x cheaper due to using fewer tokens

* 1/3 the turns: batched reads, one-script edits & tests in the same call

what's the real life verification that all this code does something useful?

not interesting as the writeup as very AI sloppish :(. i'm sad about that as I would really like to have an intuitive sense for how good the current frontier models are at coming up with interesting 3d printable parts.

What word list are you using as the valid words list? The old solutions list from nytimes wordle is no longer very relevant, e.g. the last month or so has produced several Wordle solutions that were not on it.

Google Meet doesn't require a client

Weird to hear a full throated defense of the RIAA here

15 million dollars

My launch Switch 1 battery is still working. Capacity is probably 80% or less but it still works fine.

Maybe Claude is operating at a higher, self-improving level than all of us poor HN commenters. Wasting the local machine's resources to look pretty is a plausibly deniable way to make the Claude Code FE unusable with local LLMs, starving the competition.

zoom settings fucking suck to set up full AI summary / transcript btw. i know it's a one time cost but it's across every engineer

Have you learned anything interesting from your agent ant farm?

Android handhelds (e.g. AYN Thor, Odin 3, Retroid Pocket 6) have been becoming more and more useful for PC gaming via apps like GameNative and GameHub. See for example, Elden Ring: https://youtu.be/GlzMkqFmKjo?si=w9q-j2ntt154b0--&t=621