omneity

RAG is infinitely more accessible and cheaper than finetuning. But it is true that finetuning is getting severely overlooked in situations where it would outperform alternatives like RAG.

128GB but it's not using much.

I'm running Q4 and it's taking 17.94 GB VRAM with 4k context window, 20GB with 32k tokens.

I'm using Qwen3-30B-A3B locally and it's very impressive. Feels like the GPT-4 killer we were waiting for for two years. I'm getting 70 tok/s on an M3 Max, which is pushing it into the "very usable" quadrant.

What was even more impressive is the 0.6B model which made the sub 1B actually useful for non-trivial tasks.

Overall very impressed. I am evaluating how it can integrate with my current setup and will probably report somewhere about that.

Looking at the Persian IPA table[0] for the letters you wrote, we get `/ʒ/` for `ژ` and `/tʃʰ/` for `چ`

In Arabic[1], there are two close phonemes: `/dʒ/` for `ج` and `/ʃ/` for `ش`

The difference in both phonemes is minimal and are practically affricates[2] of each other (where `d` or `t` can precede a `ʒ` or a `ʃ`), so it seems these sounds are present in both Arabic and Persian.

These variations are also within the dialectal distribution of either languages. For example `ج` is pronounced `/dʒ/` in Algeria and `/ʒ/` in Morocco.

0: https://en.wikipedia.org/wiki/Help:IPA/Persian

1: https://en.wikipedia.org/wiki/Help:IPA/Arabic

2: https://en.wikipedia.org/wiki/Affricate

Arabic is not Aramaic. Please correct your sources.

I’m also quite curious about the sounds of “ch” and “zh” which exist in Arabic as ش and ج, or did you mean something else?

It doesn’t have to be contiguous AI-written chunks. This could also mean X% accepted AI suggestions on Y% codebase where X*Y < 30%.

AI suggestions can also be as simple as autocomplete and still be counted for the sake of engagement metrics.

Oh and in enterprise settings and especially MS shops Github Copilot is being pushed everywhere, (forceful) adoption rates are much higher than the market average.

You are right, it doesn’t make sense. If that was a real reason then why not augment their current staff with AI tools to “scale content to more learners” and still have a sense of quality and human touch.

Encrappification at play.

I already found Duolingo quite bad, struggling to make it past the repetitive and often not-so-sensical question/answers, but this finally proves it’s not for me.

I would like to challenge the fundamental premise of the article. Just because you can generate 50 PRs doesn’t mean you should. In fact the same bottleneck they’re describing is present if you have 50 coders in your team.

The problem therefore is not how to scale PR review and rather how to select meaningful work to perform, which brings me to the second point made by TFA: humans making judgment calls on whhich PR should be prioritized being a uniquely defining human feature.

I beg to differ here as well. All the problems described in the article are high context decisions, you need to take a lot in consideration (user request, product strategy, market dynamics, cost/benefit, rou..) to decide which feature should be prioritized in the next release. What prevents LLMs from being able to help with that is the sheer amount of information to ingest which is a still a limitation despite the long context windows we see nowadays.

tl;dr: this is a problem of prioritization and product strategy and nothing specific to AI. Scaling so-called judgment is a red herring and better focus and scope management should be aimed for instead.

Excellent release by the Qwen team as always. Pretty much the best open-weights model line so far.

In my early tests however, several of the advertised languages are not really well supported and the model is outputting something that only barely resembles them.

Probably a dataset quality issue for low-resource languages that they cannot personally check for, despite the “119 languages and dialects” claim.

For the uninitiated, what this comment means is that LEO-orbiting payloads have a naturally limited lifespan as the orbit decays over time on its own causing the LEO satellites to get closer to earth and burn down to smithereens due to friction with the atmosphere.

This is why Starlink keeps launching new satellites on the regular. I believe theirs have a six years lifespan but don’t quote me on that.

This somehow reminds me of Agent-3 from [0].

0: https://ai-2027.com

You probably know this already but I’ll leave it for posterity.

https://mermaid.js.org/syntax/entityRelationshipDiagram.html

Yeah those are pretty cool and on my radar! And thanks for sharing the tip :)

Just checked your work on covary and it's pretty rad! What's your backend like?

> So, our question was: what happens to brain fingerprints when we lose consciousness, such as during the artificial sleep induced by general anesthesia?

The most important question here, and yet the article keeps talking about "fingerprints" of anecdotal but limited insightfulness.

Maybe by doing it enough times o3-mini will end up reimplementing itself?

A long time ago, I believed I wanted to implement a custom interpreter in Ruby. And then I discovered method_missing, and later on the rabbit hole that is Metaprogramming in Ruby, and that worked wonders to get the interpreter thought of my system, but not before implementing at least 3 DSLs of dubious value.

While it worked great to gently expand my knowledge on what's possible in programming (I would have burned myself on a lisp, not for the lack of trying as I did with Haskell), it also taught me about the meaning of shooting oneself in the foot, and in a sense do not regret the absence of such powers in JS and Python, even if Proxy objects are dangerously close!

If I was the paper author I would have done it and include the results as an appendix or a repo.

Or your model wasn't trained so well (weights are too spiky)

This is similar to the halting problem. You can only write an effective policy if you can predict all the side effects and their ramifications.

Of course you could do like deno and other such systems and just deny internet or filesystem access outright, but then you limit the usefulness of the AI system significantly. Tricky problem to be honest.

I guess the "lucky 10000" effect is extremely strong today!

This is like an OS developer who has never heard of Linux.

Block-based editors maybe?

This would be such an interesting experiment to perform on other social platforms as well alongside some rough semantic analysis to understand which topics are being silenced.

I already got quite a lot of the data pipeline setup for this, so if anyone wants to collab hit me up!

Exactly. There was a certain simplicity that might be lost. But yeah I can imagine it might work out differently for others as well. Glad to hear it is for you!

Have you considered other options? Curious if you came across anything particularly interesting from the simplicity or DX angle.

Put simply, Svelte and React were at two ends of a spectrum. React gives you almost complete control over every aspect of the lifecycle, but you have to be explicit about most of the behavior you are seeking to achieve. Building an app with React feels about 80% on the JS and 20% on the HTML side.

Svelte on the other hand felt like a breeze. Most of my app is actually plain simple HTML, and I am able to sprinkle as little JS as I need to achieve my desired behaviors. Sure, Svelte <=4 has undefined behaviors, or maybe even too many magic capabilities. But that was part of the package, and it was an option for those of us who preferred this end of the trade-off.

Svelte 5 intends to give that precise level of control and is trying to compete with React on its turf (the other end of that spectrum), introducing a lot of non-standard syntax along the way.

It's neither rigorous Javascript like React where you can benefit from all the standard tooling developed over the years, including stuff that wasn't designed for React in particular, nor a lightweight frontend framework, which was the initial niche that Svelte happily occupied, which I find sadly quite empty now (htmx and alpinejs are elegant conceptually but too limiting in practice _for my taste_).

For me it's a strange "worst of both worlds" kind of situation that is simply not worth it. Quite heartbreaking to be honest.

Disclaimer: OT and pretty ranty.

I don't know if that's what the GP hinted at, but as a Svelte developer and big advocate for more than 6 years (single handedly training and evangelizing 20+ developers on it), I found so many concerns with Svelte 5 that it simply made me use React again.

It's a temporary choice and I'm desperately evaluating other ecosystems (Looking at you SolidJS).

It’s actually not too far off. Yes MCP is designed for LLM interactions, but we observed that it’s an invocation API that’s pretty generic. So we built a package format that encapsulates computations and makes them accessible from any of MCP, REST, JSON-RPC over WS (the generic cousin of MCP)..

We build logic once and make it automatically accessible from any of these consumption methods, in a standardized way to our clients, and I am indeed piping some of these directly in the CLI to jq and others for analysis.

There’s a modern version of it on Steam with multiplayer which is decent (could be much worse given most modern remakes of classic IPs)

I can actually see the emoji for number 4 on iOS Safari.

It’s in the 3rd line, the one before the website url (penultimate).

I thought hackernews strips out emojis, interesting.