armanj

the moment debate becomes philosophical it's over

Bonsai 8B and 1.7B were on Qwen3.5 the benchmark is from a few months ago. However I'll add Qwen3.6 to the benchmark too.

Bonsai vs Qwen (quick) Benchmark: https://github.com/ArmanJR/PrismML-Bonsai-vs-Qwen3.5-Benchma...

can someone tldr me why choose apple container (and its ui) over docker (and orbstack)

one main marketing leverage of 23andMe, AncestryDNA, etc are fulfilling the curiosity of people who want to know which part of the world their genes are from. I guess that dataset should be preparatory.

funny how we may have to wait even longer for llms to pick up this update in their pre-training

useless fact: there is no mention to "gpt" in this article. the ai is referred to as "An internal OpenAI model".

I'm in Toronto and I can confirm.

i assumed you're connected to internet and using codex/claude code

for vibe coding stuff, especially when you're outside touching grass, I believe MacBook Neo is perfect. it fills the gap between the phone remote control (which is too painful for chatting with ai cli) and, well, not having any dev device.

seems like `karpathy/autoresearch` on steroids

buy

good luck

How reliable is this uptime? and why it's sooo different from gh's official status numbers?

i've been a zed user for almost 6 months. i've encountered maaany bugs which i reported, or that were already reported. they're still there. meanwhile, every single update shipped a feature or bugfix for "ai agents".

not sure how 1.0 ships with that massive pile of bugs. but ai agents are the first-class citizen in this editor, and developer experience is not a priority.

funny thing is i uninstalled zed right before the 1.0 release. kinda relieved i didn't miss anything.

I have a few lightweight apps using deepseek api, and funny how the initial credit I topped up for using r1 is still left. Nothing makes the user happier than getting more for less. cc: anthropics with its fancy token-wasting claude code "features"

hn is this true

I did a quick benchmark & compared it with Qwen3.5: https://github.com/ArmanJR/PrismML-Bonsai-vs-Qwen3.5-Benchma...

in my results, accuracy-wise Ternary-Bonsai-8B is on par with Qwen3.5-4B. But in accuracy-per-byte, bonsai is the clear winner:

=> Ternary-Bonsai-1.7B achieved 65.1% from 462 MiB, beating Qwen3.5-0.8B by 12 points while being ~5% smaller on disk. => Ternary-Bonsai-4B is the accuracy-per-byte winner above 1 GiB. 83.0% from only 1.1 GiB, within 2 points of Qwen3.5-4B at 40% of the weight size.

they show strong promise on edge devices and where disk space is limited. I think this lab is worth watching.

while it seems even with 4.7 we will never see the quality of early 4.6 days, some dude is posting 'agi arrived!!!' on instagram and linkedIn.

I recall a Qwen exec posted a public poll on Twitter, asking which model from Qwen3.6 you want to see open-sourced; and the 27b variant was by far the most popular choice. Not sure why they ignored it lol.

kinda ironic you can clearly see signs of Claude, as it shows misaligning table walls in the readme doc

This project is early and experimental. Core concepts are settled, but expect rough edges. Local mode: relatively stable - Hub-based workflows: ~80% verified - Kubernetes runtime: early with known rough edges

i guess gastown is a better choice for now? idk i don't feel good about "relatively stable"

People in the comments are, in my opinion, overcomplicating this and making it more philosophical than it needs to be. The reason for their decision is dead simple: there aren’t enough GPUs, so they have to cut access somewhere, and they’re starting with claw.

It’s really that straightforward. If tomorrow they decide GPUs are better allocated to enterprise use, they could start removing the $20 plan just as quickly overnight, the same way they did tonight.

Anti-distillation: injecting fake tools to poison copycats

Does this mean `huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled` is unusable? Had anyone seen fake tool calls working with this model?

claude code started as an experimental project by boris cherny. when you’re experimenting, you naturally use the language you’re most comfortable with. as the project grew, more people got involved and it evolved from there. codex, on the other hand, was built from the start specifically to compete with claude code. they chose rust early on because they knew it was going to be big.

A well defined doc evolves over time. it gets sharper with real-world scenarios, incidents, and experiments. Before Voyager 1, we didn’t have that kind of experience. You can’t predict everything upfront.

Theory only takes you so far

they're already there ?? https://openrouter.ai/qwen/qwen3.5-27b

“remove this section entirely, we don’t need caching here” — rejecting a proposed approach

I wonder why you don't remove it yourself. Aren't you already editing the plan?

https://arman.pw

I referenced 3B1B for the name: youtube.com/watch?v=8idr1WZ1A7Q

Years ago, I often struggled to choose between Amazon products with high ratings from a few reviews and those with slightly lower ratings but a large volume of reviews. I used the Laplace Rule of Succession to code a browser extension to calculate Laplacian scores for products, helping to make better decisions by balancing high ratings with low review counts. https://greasyfork.org/en/scripts/443773-amazon-ranking-lapl...