HNHacker News
TopNewBestAskShowJobs

jakswa

281 karma · joined August 19, 2012

https://jake.town
submissionscomments
jakswa··on How Compaction Works in Pi
I dunno, I didn't read in-depth. Hopefully you don't gotta zoom in with human eyeballs.
jakswa··on How Compaction Works in Pi
OMP changed the default compaction to images! Kinda nuts to read about. Saves the generation cost of the traditional compaction step and writes the context as tiny text to an image, if I was following correctly.
jakswa··on Gemini 3.7 Flash
my whole world is shifting. have I been seeing _different pelicans_ from everyone else?!
jakswa··on Gemini 3.7 Flash
This pelican gave me a good laugh, because there's enough reasoning that the render is out of sight initially. The buildup!
jakswa··on Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
I'll back up your smaller claim, but be specific that it's UD-Q4_K_XL size:

- muse glimmer: 15.9GB

- qwen 3.6 27B: 17.6GB

My video card is so close to its limit that these GB thresholds are mattering too much for me :D

jakswa··on Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
it's here! https://news.ycombinator.com/item?id=49245575
jakswa··on Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
I'm listening to pelican sounds on youtube while I wait for Simon.
jakswa··on Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
I like the tabletop RPG use case, and wanted to say: If your hardware likes it you should check out Gemma 4 for creative DMing use case. I found it to be much better at holding the plotlines and being creative on gaming turns. My experimental case was an audio-only Zork and Gemma 12B and even E4B were pretty good!
jakswa··on Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
some support already merged, and I verified in a local build that it runs (cannot get MTP params working tho, about ~40 tok/s on my beefy 800GB/s 7900XT w/ 20GB VRAM). https://github.com/ggml-org/llama.cpp/pull/26841
jakswa··on Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
Q3 results: unsloth/Muse-Glimmer-30B-GGUF:UD-Q3_K_XL gets down to 15.6GB VRAM and full context (131k) on the 4 parallel slots. Prompt/generation speeds about the same. Overall feeling like a nicer-fitting Qwen 3.6 27B, but want to test out MTP generation speeds once I can.

edit: My favorite bit of reasoning I saw go by in my "generate me a beautiful code snippet" anecdote: 'Could give a snippet of beautiful code: the "hello world" in brainfuck? No.'

edit2: my first dflash speculative model! no mtp. I'm up to ~60 tok/s on empty context with `--spec-type draft-dflash`

jakswa··on Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
Another candidate for the 7900XT (20GB VRAM) I got sitting around. I pulled latest llama.cpp (targeting vulkan during build) after seeing a muse PR merged a few hours ago, and unsloth/Muse-Glimmer-30B-GGUF:UD-Q4_K_XL runs on my 7900XT barely (and with no MTP). Sits at 19GB VRAM w/ 4 parallel 113k context slots, all layers on GPU, and at 700 tok/s prompt, and ~36 tok/s generation.

Waiting on Q3 to download to check speed + do my usual anecdotes. I generate beefy code snippets and poems, and also ingest my HOA declaration and answer nuanced questions.

edit: i should've prefaced this somewhere with: This card ballparks at 800GB/s IO, which I can't seem to find easily on the market anymore. Kinda the ideal card for this model, if I just had a _little_ more VRAM (XTX is 24GB).

edit2: not mtp, this is dflash model (param in child comment). I'm up to ~60 tok/s generation and sitting at 19GB VRAM (i added --no-mmproj (makes it text-only i believe) because I'm used to speculative decoding wanting more VRAM and I'm already close to the limit :sweat_smile:)

jakswa··on Show HN: Graft – Give coding agents a semantic map instead of grep
> Every source file is summarized once into a short description of what it does.

I admit I haven't had a chance to read the whole README, but wanted to get down my hesitation after I got pretty far (as an interested user): I think it'd either be too expensive or too-low-quality to feed all the source code in my repo into an LLM to build this dataset. Luna with medium+ thinking feels like my best bet right now but even then I'd worry about total token cost here for some repos.

jakswa··on Show HN: Run With Zombies. A free browser game where GPS zombies chase you
I remember trying out a zombie running app ~15yrs ago and am a fan of the concept. That was a podcast pretty much, audio-only storyline in your ears describing proximity/urgency/etc. I didn't grant this thing GPS access (work computer), but my immediate goal was to figure out how much attention it wants while running. The real problem nowadays (running in a city on city paths) feels like I would be oblivious to the people/cars around me and end up a safety hazard.
jakswa··on Show HN: Skyline Co-op is like SimTower, but built around democratic socialism
I can't begin to picture how much AI slop they are having to filter out. I wonder if they are going to entertain some UX to save people disappointment. "The project should be non-trivial. Don't post quickly-generated one-offs; anybody can do that now. Share something that is deeply personal and interesting to you. Explain how and why." To me the whole HN website feels like a quickly-generated one-off lol. Guess they hope you can find the freshest ruleset for the thing you're trying to post. https://news.ycombinator.com/showhn.html
jakswa··on Show HN: Skyline Co-op is like SimTower, but built around democratic socialism
Hell yeah I got people to move in. Nice job getting it to walk me through those intro steps.
jakswa··on Cloudflare OS: an open platform for agents, apps, and work
In case anyone else is curious, local Gemma4 12B Q6 XL really struggles to make use of this for simple goals like a daily briefing after an MCP tool call. I don't think there's a simple artifact concept? Still early but confused.
jakswa··on Cloudflare OS: an open platform for agents, apps, and work
Seems like a good candidate to replace my home-grown llama.cpp web UI clone. This comment used to be a gripe about their provider UX not giving an example URL (realized /v1 is expected as the final segment).
jakswa··on Ask HN: Is the GitHub pancake menu now a pancake emoji?
my guess: their new-ish "PR stacks" feature?

edit: yes i bet it's https://github.blog/changelog/2026-07-30-stacked-pull-reques...

jakswa··on Kimi K3 Now Available via Telnyx Inference API
what quantization? FP4?
jakswa··on Inkling: Our Open-Weights Model
They've got an openai + anthropic compatible endpoints. I got far enough to run some tests on the openai endpoint, albeit with some finagling (their /models list is empty, my tool auto-configures using that, was an initial stumble).
jakswa··on Inkling: Our Open-Weights Model
seems pretty dang snappy and I like it's tone/personality so far.

> look at today's hackernews frontpage and generate me a daily briefing report (create an artifact) to read later for today's nerd news

https://chat.home.jake.town/artifacts/019f679d-99e5-7000-b02...

jakswa··on New Flagship Grok Voices
why does this website absolutely wreck my browser
jakswa··on Postgres rewritten in Rust, now passing 100% of the Postgres regression tests
what model did you try it with? I agree and also push back a bit: How will we know when the LLMs reach the point of handling it if no one takes the leap? I applaud more people sludging through the slop and hauling their slop buckets around.

An example is Fable being released. I felt like the most complex thing I was willing to sludge through was having it clone llama-server's web UI with my own opinions (I really like the original, kudos to them). And the initial skeleton was working so well I felt like I had sunk the tokens and committed to getting it the rest of the way: https://inkcap.click

jakswa··on GPT-5.6
I'll agree and expand on "weird restrictions" -- I used to check the claude usage graphs multiple times a day to see where I'm at on my weekly budget. With gpt 5.5 I don't think I'm working differently but haven't felt the need to check anything because I think I've hit my limit... once? on some egregious edge case scenario iirc
jakswa··on GPT‑Live
Probably not _as_ good, but you can run gemma 4 for the ears/brain (accepts audio input), and kokoro TTS for the mouth. You want something like silero VAD sitting in front of the LLM so you aren't passing wasteful audio to gemma 4, so you send only voice activity segments. I type all this because I tried it out recently as a Zork experience, seeing how creative Gemma 4 12B could be. It was surprisingly good!
jakswa··on We're extending access to Fable 5 on all paid plans through July 12
The best project I found to throw it at was cloning llama-server's web UI essentially in one shot. I'm not sure what I'll do with 5 extra days, maybe try to imagine some complex features. I'm no longer surprised at how much seems to work in these "new brain what can it do?" test, and instead think the risk is feeling like I have to take it the rest of the way once I've sunk the token cost :| https://inkcap.click
jakswa··on Ask HN: Share your personal website
https://jake.town
jakswa··on Show HN: Rust Web Framework
Lately I've been following https://loco.rs/ as it aims for a rails-like experience, complete with generators for workers, controllers, etc. I've only had time to experiment but it's the closest I've gotten to feeling rails-y in rust.
jakswa··on Move Fast and Abandon Things
The car prototype reminded me of Spy Hunter graphics, but I couldn't remember that NES game's name at first. Sent me on a nice nostalgia dive!
jakswa··on Show HN: Wealthfolio: Private, open-source investment tracker
Just wanted to say thanks for bringing this up! So far it is a really interesting tool for me, as someone too lazy to find a proper financial planner
← PreviousPage 2 of 5Next →