spaceemotion

@cnakazawa Interesting. It seems to help a lot with Opus 4.6 though (GPT 5.5 is just too all over the place for me)

@spaceemotion That hasn’t been a thing for a while. Got 5.5 doesn’t need this, and sometimes performs worse.

Third year in a row in which I ask llms to make my little JS palette swap library faster. Year 1 and 2 just made it worse/slower/broken. Year 3: Goal achieved in 13 minutes – 3.7x faster, and it told me "honestly" that it couldn't get to 5-10x. github.com/nkzw-tech/pale…

@pingToven @suchenzang i also would love to have a company keep their older models around for more than (what feels like) a month. OpenAI is surprisingly one of the only ones you can kind of depend on for a longer time, than Google who shut down Gemini 3 wayyyy too fast.

@suchenzang can also add that based on openrouter community vibes and how long the older gen gemini models stayed on top of our rankings that people miss earlier flash models (for their capabilities incl. vision but also their pricing)

ouch

@jjackyliang @benhylak has @pingToven not shown you our NCBench? ;)

@benhylak we really need a bench/arena for writing

what's the best oss model for writing right now?

@zeeg are the rate limits different to the regular API? i tried to download about a hundred issues with thousands of events to analyze them locally because i got limited quite heavily

bias, but Sentry MCP is goated i have it debugging a collection of adjacent issues in Junior (our slackbot) right now it aggregated different errors, pulled trace data, identified time-ordering concerns for something thats not obvious whatsoever on the surface

@cnakazawa i love doing performance optimizations with AI. i'd you get the right loop going, it's happily getting as high as it can get, for hours and hours - though it also loves to introduce shared state as "JS is single threaded anyway" (not great when using iterators).

Codiff 0.4: Much faster and a few nice QoL improvements. Kinda sucks to build personal software these days. I use this tool all day to solve problems, then realize it could be better, so now I have two problems.

@kettanaito Afaik, the @vuejs ecosystems to this really well - but that's probably because there's a lot of downstream libraries and frameworks using them. If you're asking about MSW, what kind of libraries do you know depend on your lib. as an integral part? Maybe npmx can help there...

@spaceemotion It's almost that. As I understand, ecosystem tests is about making sure *your* changes don't break existing ecosystem consumers.

I wish more developers spoke about ecosystem testing and how to efficiently set it up. I know Node.js does it. I know a few other projects do that. I've never seen anyone speaking about it. As someone whose projects have gathered a bit of an ecosystem, I want to know!

@cnakazawa Ah yes, the side project for the side project!

Windows & Linux releases are here: github.com/nkzw-tech/codi… I still haven't reviewed the code for which I originally built Codiff for 😭

Codiff 0.2 * Improved Walkthroughs: Faster and more actionable * Ask Codex: Ask about code inline * PR Review: Review any PR, comment and approve/reject * History: Review apst commits * Windows & Linux Builds * And tons more! Install: brew install --cask nkzw-tech/tap/codiff

@vanilagy @TechSquidTV I thought the same for "Novelcrafter" - and to this day, AI just loves to do "NovelCrafter" or even "Novel Crafter". Seems like a curse for whenever you have two nouns mashed together for a name...

@TechSquidTV Yup! Thought it looked cleaner as one word without camel case.

Me when Mediabunny gets a shoutout 😌😌 And it's spelled incorrectly 😠😠

@zeeg @eazyscalp Are you doing this with a custom loop? or something like GEPA? (github.com/gepa-ai/gepa)

@spaceemotion @eazyscalp its mostly constraining it to certain changes, and guardrailing it to not over optimize in a few places im not being super rigorous here, but it amounts to "lets optimize this prompt to address these failing evals, heres the kinds of changes that are ok"

Warden Pi branch is still optimizing itself 24 hours later with little success. Doing this against evals aint no fun yall...

@zeeg @eazyscalp Would love to know what the optimization steps were, to see which bits of prompting hurt and which things helped! especially in a very lightweight harness like Pi.

@eazyscalp its attempting to optimize the Pi harness to perform at the same level as Claude SDK for us which is mostly prompt augmentation

@zeeg we kind of started doing the same, but for building a Pi-based agent that can do code (OpenCode replacement), GH actions, slack and discord. Still working on it! but at least it's not breaking our workflows when a vendor updates their things…

vendor-specific chatbots are broken by design that means the Sentry agent, the Linear agent, and any others you might have in Slack they are fine for some point situations, they're nice to get started with, but agents with generalized access outperform them in every single scenario some weeks ago we built an internal Slackbot, gave it access to a bunch of systems (Sentry, GitHub, Linear, Notion, etc), and its capabilities overnight far exceed these other bots "Oh cool Linear can now search your code bases" - our bot did that on day one, and then could push that information wherever it needed to go. Its useful to the point where I now discourage use of things like the Linear bot because it _creates worse outcomes_. this also goes beyond the simple generalization of access: we can customize it. we throw in skills-as-runbooks, templates, etc and the outcomes once again incrementally improve if your org hasnt already built a general purpose bot internally you should. if you need inspiration ours is open source on GitHub (albeit fairly unstable still) github.com/getsentry/juni…

@SebAaltonen When you say 'no occlusion culling', does that also mean no frustrum culling? or does it skip things outside of the view/behind the camera already?

There's these high poly objects inside the cafe. We have no LODs and no occlusion culling yet. All of these are rendered in the above screenshot at this density :(

Testing mini-engine on iOS. Hits 120Hz native on iPhone 15 Pro in Bistro (18M triangles rendered), but heats too much. Have to optimize further.

@astuyve Hm. still waiting for something like that for provisioned instances. afaik the only way is to handroll that with a custom lambda running on a schedule :/

Adding scheduled scaling for Lambda Managed Instances is a good ship (like LMI, probably not meant for you). But it is a very "snake eating its tail" moment for serverless

@lucasmeijer @Yuval_Shimoni @bentlegen I think plannotator.ai (OSS) has a web-based platform that does similar things!

@spaceemotion @Yuval_Shimoni @bentlegen That’s terminal only right? I don’t like terminal

I wish I had learned about diffs.com this morning instead of just now.

@lucasmeijer @Yuval_Shimoni then what @bentlegen is working on with github.com/modem-dev/hunk may be of interest of you?

@Yuval_Shimoni i'm trying out a different take on reviewing agent produced code

@lucasmeijer i have not success with GPT 5.5 when it comes to code at all. especially typescript. even at high or xhigh, it tends to be a lot worse at web/frontend stuff than Claude or other models.

been working with gpt5.5 to produce a in-browser diff viewer that has "show 10 more lines" buttons at the top/bottom of each hunk, and I have not seen an LLM be this bad in a very very long time.

@bentlegen @nvimcruz @mitchellh If it could be an extension for pi, that'd be great! right now I am just adding inline TODOs into the files :D

@spaceemotion @nvimcruz @mitchellh Yep

the @mitchellh effect

@bentlegen @nvimcruz @mitchellh so something like this? x.com/ahmedelgabri/s…

@nvimcruz @mitchellh It’s next on the list