nateb2022

The author's AI applied a thin layer of varnish to what is more or less handwaving. I think more people need to take an undergraduate level statistics course; the conclusions made are completely unsupported by the random trials.

Weighing six rocks a hundred times doesn't tell you the average weight of a mountain.

> ds pro and flash are fast but not something id actually use for important things, unlike sol, fable and maybe 3.7 here and there

As someone who's used Gemini 3.7 Flash (Google sub mostly for the storage) and DS4 Flash a lot (~6B tokens), I'd actually place DS4 Flash (even pre-0713) above Gemini 3.7 Flash. Gemini has a tendency to leave some things unimplemented; perhaps it's agy which frankly leaves a bit to be desired as a harness.

Although I will praise DS4 Flash any day, it no longer makes sense for me after the price increase (GPT 5.6 Luna is a much better price point) and I have completely migrated my high volume workflows to Muse Spark 1.2 Contributor (which I find to perform better than DS4 Flash 0713, happily).

Well OpenRouter went from ~$5M in mid 2025 to ~$140M about a year later, so although 50x might be compared to current revenue, they probably factored in some growth models to approximate its value over time.

To bring people back to earth for a bit, two observations.

1. Current battery technology wouldn't enable continuous recording in such a form factor for longer than ~1hr.

2. Were continuous recording to be enabled, the earbud case would easily reach temperatures in excess of 115 F within minutes, which would be pretty noticeable.

On macOS (zig 0.16.0) a compile with zig build -Doptimize=ReleaseSmall yields a 5.8M binary for me.

Can Sprite replace Pepsi?

There's some potential research suggesting it has positive long term cardiovascular benefits, including decreased risk of heart attack, stroke, and overall cardiovascular mortality.

Anecdotally, a lot of people use it off-label to improve performance in strength training. I would be dismayed if blindness is confirmed to be a long term effect.

OpenRouter already takes in around $140M in yearly revenue. How would paying 5% of yearly revenue make any sense for an acquisition??

browser-use is a lot more capable and generally more efficient, and a lot easier to embed in other programs than Claude's driver. I haven't used Vercel's agent browser enough to speak to it.

https://github.com/browser-use/browser-use has been great so far

official site seems to be https://deepseek.com/harness/en/

Which quant would cut it closest on a 16GB VRAM card?

Currently using https://github.com/placemark/check-geojson for this usecase right now.

Noticed your code includes a custom JSON parser, which seems to silently overwrite duplicate keys, which is absolutely something a linter should warn about.

And instead of using fast-glob, I'd recommend https://superchupu.dev/tinyglobby or better yet, use node:fs and node:fs/promises which are built in to modern node already.

And I noticed your package.json includes some bizarre fields,

  - "types": "./dist/index.d.ts" is unnecessary since you're an ESM module and require Node >22
  - "./package.json": "./package.json" is a Node 12-14 ESM workaround

The bench is not a black box, if you scroll down it gives an example of what the test cases look for

[dupe] https://news.ycombinator.com/item?id=49299605 (3 hr ago, 217 comments)

On the contrary; it seems Luna (max) is 2x as capable at 1/5 the cost: https://cognition.com/frontiercode

This was the older Gemini 3.7 submission and linked to the blog post; the submission you marked this as a dupe of occurred 20 min after this one, and linked to the much more barren https://ai.google.dev/gemini-api/docs/models/gemini-3.7-flas... URL.

Or a 3.7 Flash-Lite

[dupe] https://news.ycombinator.com/item?id=49288847 (35 points, 8 comments)

Codex CLI should have been usable for a while now, and you could also sign in to your ChatGPT subscription using Opencode which runs perfectly fine on a Raspberry Pi.

That's my bet, I doubt they'd have the resources to have trained anything substantially larger

Yup. I posted about the huggingface countdown: https://news.ycombinator.com/item?id=49272534 but seems like both the huggingface page, and the modelscope.cn URL that commenters found, were quickly taken down. Not sure if the submission being flagged is also the doing of whoever took down the huggingface/modelscope.cn pages.

Edit: for anyone curious, I saw the huggingface countdown page for Kimi K3, which gave me a hunch that led me to the countdown page for Qwen3.8-27B. As of ~11am EST today, the countdown page said 2 days and 2 hours to release.

This modelscope.cn url, like my huggingface url, seems to have been taken down.

URL was taken down; seems like the modelscope.cn link that other commenters posted also got taken down.

It used to feature a countdown, at least when I posted. 2 days and 2 hours from Wednesday 12 August at ~11am. Huggingface or Qwen took down the URL it seems.

Edit: looks like the modelscope.cn URL that other commenters posted also got taken down.

Think of an LLM as a thesaurus, but for entire trains of thought rather than words. Your initial query yields something pertinent to the task at hand. But let it endlessly recurse and... you end up with something completely useless.

People would do well to acquire at least a modest familiarity with what an LLM actually is. NLP is fascinating. So is entropy.

Seems to be entirely AI written. Full of the stereotypical rule of 3, and "not X, it's Y," in fact I think each paragraph features at least an instance of one of them if not both.

Some well written devblogs are also available:

  - https://lwss.github.io/Duty-Of-Kisak/
  - https://lwss.github.io/Kisak-Black/

I use Luna a lot (over 1T tokens since it came out) and I'd rank Luna (high/xhigh) on par with Sonnet 5, without hesitation.

> tiny (3MB)

I think small is a fair description. Say tiny and I'd expect under 1MB.