postalcoder

Search Google Images for "bicycle". Almost all bicycle product shots are staged the same way: side view, going left-to-right. It makes sense to me that given that skew in the training data, the model grounds itself in the bicycle.

This makes me deeply nostalgic. Starlight Zone from Sonic the Hedgehog is my favorite game level of all time and the aesthetic of this city is basically like stepping into the background of Starlight Zone.

https://youtube.com/watch?v=ZHR5OKDL81g

I’ve been happy training and running inference for small language models on my M4 Mac.

Inference with MLX is surprisingly zippy. I’m running a classification task on the entire HN comment dataset and it’s projected to take about two and a half days, which is not bad considering we’re talking about tens of millions of comments.

Yes, I could do it much more quickly by throwing Modal GPUs at it but this is low-priority work. I might as well throw my M4 a bone.

From the HF link posted above it's Qwen3.6-35B-A3B.

https://huggingface.co/thomsonreuters/Thomson-1.0-Small

I've noticed a chilling new trend of apple listening to the community.

A lot of these criticisms seem to be very 2023-coded.

There have been a lot of hiccups to the transition towards using AI-assisted code but, ignoring some of that sloppiness, I've enjoyed the product velocity from some of the services/apps I know that heavily use coding agents.

I think it's more than that. They truly believe Claude Code to be a moat (both the harness and the posttraining) to be a moat.

I've been deeply distrustful of Anthropic from early early days. They have always been openly disdainful of user feedback. I would not be surprised if later they try to implement more shenanigans to keep people locked into CC.

Unfortunately, this doesn't work with CC's other features like the lazy loading of CLAUDE.md files in subdirectories.

There is no live tarot feature. Thats the point.

I’ll add that.

Not sure I agree with this. The math guy at anthropic's prompts are essentially:

  "suppose you’ve gotta resolve the $CONJECTURE, like absolutely have to, everything depends on it. think really hard, and try to come up with a bunch of ideas to try. but remember to trust yourself and not necessarily in conventional wisdom!!"

  https://claude.ai/share/25740bd5-aa97-4bd7-bf58-c4df3793fda7
  https://xcancel.com/__alpoge__/status/2083855298239078748

Tao's chat was for him to gain intuition, not to solve the problem from the outset.

What's funny is that every other person gets a different conclusion about who these models reward/empower. I've seen people say that the generalist stands to gain the most and others say that it's the experts. Like all of life, maybe the "winner" is the person who just does stuff.

Hmm, right now I only process a subset of all submissions to keep things lean, so this is the closest that the site can do to /newest right now: https://hcker.news/?filter=top200&ai=exclude

Feel free to email me.

Same to you – glad you like it!

There are a lot of ways of tackling this, but the only way to truly filter everything out is to update the list constantly.

My filter currently has 146 keyword filters and 101 domain filters. Even with those, I've had to filter almost 400 exact urls on top of that in the past two weeks!

All due respect to the YC folks, but having a skill with 22,069 tokens is a major skill issue. Yes, ironic.

My slop control skill is a thousand tokens. Biggest problem I see with the skill is that everything is prompted via negativa. smh. sorry to be judgmental but it's hard to trust a harness that comes shipped with a skill like this.

I think the most important thing here is not absolute performance. It's that organizations now have access to a Fable-ish model without Fable's 30-day data retention requirement[0].

> "Consistent with prior Opus models, Opus 5 does not have data retention requirements for general access."[1]

On the Opus model release page, the reason why Fable doesn't have an ARC-AGI score is because of that retention policy[2].

0: https://support.claude.com/en/articles/15425996-data-retenti...

1: https://www.anthropic.com/news/claude-opus-5

2: https://xcancel.com/arcprize/status/2064399134099153344

My point was more of an exercise to point out that finance is about mutating resources. A lot of cash can be a good thing or a bad thing. Same for debt. There’s nothing inherently bad about levels.

Eep, I guess it depends on what kind of learner you are. For me, the notes being in the book, and the book being somewhere on my shelf spatially encodes it into my mind.

Offloading things into a centralized store according to a system not only causes me anxiety but it doesn’t have the same spatial kick.

A corollary to this is, when reading books, mark them up! Don’t be precious with them. Destroy the spine so it’s easier to read, fold pages, write the important stuff inside the front cover of the book. Cross out stuff you disagree with, draw a big circles around the stuff that resonates.

Not only does it make consumption a lot more engaging, it makes a revisit of that book extremely rewarding.

Don’t get caught up trying to figure out the right notation or highlighter color for a line. Just mark them up and later it’ll make sense.

Edit: I think the broader point that some of you may be missing is that I’m not prescribing a particular form of notetaking. I’m saying that reading is made more rewarding with active engagement. How you best engage is up to you.

There are many reasons to use subsidiaries for things like this, like to invite outside investment, ringfence risk, cede operational risk, and many more.

This is all like CFO 101 type stuff, and not nefarious. I find it amusing that people assume the worst for things they understand little about, rather than trying to learn.

Maybe the best way I can explain it to the programming crowd is this: imagine how ridiculous it would sound if outsiders were saying that Google was on the verge of collapse because its codebase has billions of lines of code.

> These companies have valuations reflecting a debt light business.

Sorry, but this doesn’t make sense. The valuations of these companies reflect their growth.

In finance there’s nothing inherently virtuous about a “debt-light business”. It’s all an allocation decision based on how you expect to grow relative the cost of that growth.

Try and reframe it: are cash-heavy businesses given a premium?

I wonder how big the Pro model is that Google is using behind the scenes to train these smaller ones.

Going on baseless speculation, the lack of accompanying pro models with these flash releases either means: 1) the model is too big to be economical, 2) google doesn't have the compute to serve the big model, 3) their big model has too many alignment issues to serve to the public.

edit: looks like benchmarks are up on https://artificialanalysis.ai/models/gemini-3-6-flash. It's solidly middle-of-pack. However, if you want to be most fair to flash, look at the intelligence vs time per task and intelligence vs outputspeed benchmarks. This is a very fast model.

edit 2: I use antigravity from time to time and in my experience, 3.5 flash is an underrated model, so long as you know what it's good for. It's very good at frontend (much better than gpt 5.5) and it's fast, so it's a great tool for iteration. I expect 3.6 to be no different.

I've been thinking about this too much because it's so ridiculous and funny that I had to at least try wrapping my head around it. My best guess is this:

If you look at past snapshots at archive.org, you notice that the meta keywords are growing like an append-only list, which means it's probably part of some messed up seo pipeline. The other clue is that it's stuffing the meta keywords which apparently only Yandex uses as a search signal[0].

The list has 3882 entries. A lot of them are clustered and look like auto-complete results. But a bunch of them look like hyper-specific, misspelled search results (e.g. "145 gwen rd cheshire ct"). Google's webmaster tools doesn't provide distinct queries like that, but yandex's does[1].

My best guess of what's happening is that Qwen is monitoring it's search queries in Yandex, dumping that list into a serp service that scrapes yandex's autocomplete suggestions, and then taking that list and dumping it into their meta keywords.

It explains the urls in the list (ppls using search engines like address bars), the seeming fixation around certain topics (which usually starts with a misspelling), and the random one-off queries.

0: https://yandex.com/support/webmaster/en/controlling-robot/me...

1: https://yandex.com/support/webmaster/en/service/popular-quer...

That's hilarious. These keywords are applied globally, even on pages like https://qwen.ai/usagepolicy, where they very kindly ask you not to use their products for sexual content.

Run this in console to see all the tags:

  document.querySelector('meta[name="keywords"]').content

Fair point!

They must have trained on GPT Image 1 outputs. The yellow tint is unmistakable.

https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen-Image/i...

https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen-Image/i...

https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen-Image/i...

https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen-Image/i...

I’m grateful for this team. Jellyfin is a great product. I’ve been using llm + Tailscale + Jellyfin to manage my media box and I couldn’t be more happy. I’ve got perfect metadata, perfect organization, and multimodal search on top of it. What a dream.

Unfortunately it was possible in codex too, until recently.

https://github.com/openai/codex/issues/5237

Yes I’m saying a question answered deceptively should not be one of the four featured questions on your product page.

> Looking through their privacy policy, it looks like the bigger issue is that they are (or refuse to refute) training on your data/prompts.

You’re making my entire point. They’ve designed an ingestion engine that is not private by design. Any sort of privacy posturing is wrong.

Yes, but the answer is an indirection.

“How does Kimi Work protect my privacy when accessing local files?”

It doesn’t protect your privacy. It’s like asking, “how do I know you’re not spying on me?” And getting the reply “we cannot physically enter your house.”