HNHacker News
TopNewBestAskShowJobs

jychang

3,007 karma · joined October 23, 2012

submissionscomments
jychang··on Linkwarden: FOSS self-hostable bookmarking with AI-tagging and page archival
That's an absolutely terrible defense. Ignorance is not an excuse, try telling that to a police officer.

And plus, certain people are held to a higher standard. It's not like I'm expecting a random person on the street to know about Ollama, but someone building AI software is expected to research what they are using and do their due diligence. To plead ignorance is to assert incompetence at best and negligence at worst.

jychang··on Linkwarden: FOSS self-hostable bookmarking with AI-tagging and page archival
Strong disagree. Just because something is free and open source does not make it good. Call a spade a spade.

Ollama is a piece of shit software that basically stole the work of llama.cpp, locks down their GGUFs files so it cannot be used by other software on your machine, misleads users by hiding information (like what quant you are using, who produced the GGUF, etc), created their own API endpoint to lock in users instead of using a standard OpenAI compatible API, and more problems.

It's like they looked at all the bad walled garden things Apple does and took it as a todo list.

jychang··on Linkwarden: FOSS self-hostable bookmarking with AI-tagging and page archival
> Optional local AI tagging

https://docs.linkwarden.app/self-hosting/ai-worker

I took a look at this... and you use the Ollama API behind the scenes?? Why not use an OpenAI compatible endpoint like the rest of the industry?

Locking it to Ollama is stupid. Ollama is just a wrapper for llama.cpp anyways. Literally everyone else running LLMs locally- llama.cpp, vllm (which is what the inference providers use, also I know Deepseek API servers use this behind the scenes), LM Studio (for the causal people), etc all use an OpenAI compatible api endpoint. Not to mention OpenAI, Google, Anthropic, Deepseek, Openrouter, etc all mainly use (or at least fully supports, in the case of Google) an OpenAI compatible endpoint.

jychang··on 108B Pixel Scan of Johannes Vermeer's Girl with a Pearl Earring
More like the analog version of doing a FTT and looking at only the high frequency parts vs the low frequency parts.
jychang··on DOGE worker’s code supports NLRB whistleblower
Legally, they're allowed to modify and use GPL code internally without redistributing the source. The only mistake was publishing the source code to a public git repo without the LICENSE file, which may be a GPL violation.

I say "may", because I'm not sure if you have internal code on a public git or FTP server, is that consider "distributing"?

jychang··on Llama 4 Smells Bad
8b params is a reasonable estimate.

As a sanity check, we can look at scores for how it performs. On livebench, GPT-4o-mini scores 37.63, right next to Gemini 1.5 Flash 8B at 34.37, and above Qwen2.5 7B Instruct Turbo/Gemma 3 4B at 29.22/30.13. And it's below Phi-4 14b at 40.68, and Gemma 3 12B at 41.25.

jychang··on Llama 4 Smells Bad
Do you think there are more bugs in Llama 4 at this time? Or have the bugs been patched, and the current version of llama.cpp + whatever the latest GGUF version is would be representative of the true performance of Llama 4?

I see you've uploaded new Maverick GGUF/safetensors files yesterday, along with a lot of other models like Deepseek R1, was there an issue with the older model files?

jychang··on Llama 4 Smells Bad
Correct, in that models know nothing about themselves other than what they are told. Deepseek R1 will tell you that it's created by OpenAI.

GPT-4o mini is supposed to be ~8b params from estimates.

jychang··on Gemma 3 QAT Models: Bringing AI to Consumer GPUs
That's pretty terrible. I'm getting 18tok/sec Gemma 3 27b QAT on a M1 Max 32gb macbook.
jychang··on Pope Francis has died
... He's referring to "Christ is risen". That's way more broad (conceptually) and very in-character for the Pope, compared to some transient current events.
jychang··on Gemma 3 QAT Models: Bringing AI to Consumer GPUs
Try mradermacher/amoral-gemma3-27B-v2-qat-GGUF
jychang··on Gemma 3 QAT Models: Bringing AI to Consumer GPUs
Those are the same quant, but this is a good example of why you shouldn't use ollama. Either directly use llama.cpp, or use something like LM Studio if you want something with a GUI/easier user experience.

The Gemma 3 17b QAT GGUF should be taking up ~15gb, not 22gb.

jychang··on Gemma 3 QAT Models: Bringing AI to Consumer GPUs
MLX is slower than GGUFs on Macs.

On my M1 Max macbook pro, the GGUF version bartowski/google_gemma-3-27b-it-qat-GGUF is 15.6gb and runs at 17tok/sec, whereas mlx-community/gemma-3-27b-it-qat-4bit is 16.8gb and runs at 15tok/sec. Note that both of these are the new QAT 4bit quants.

jychang··on Gemma 3 QAT Models: Bringing AI to Consumer GPUs
I don't think ollama is quantizing the embeddings table, which is still full FP16.

If you're using MLX, that means you're on a mac, in which case ollama actually isn't your best option. Either directly use llama.cpp if you're a power user, or use LM Studio if you want something a bit better than ollama but more user friendly than llama.cpp. (LM Studio has a GUI and is also more user friendly than ollama, but has the downsides of not being as scriptable. You win some, you lose some.)

Don't use MLX, it's not as fast/small as the best GGUFs currently (and also tends to be more buggy, it currently has some known bugs with japanese). Download the LM Studio version of the Gemma 3 QAT GGUF quants, which are made by Bartowski. Google actually directly mentions Bartowski in blog post linked above (ctrl-f his name), and his models are currently the best ones to use.

https://huggingface.co/bartowski/google_gemma-3-27b-it-qat-G...

The "best Gemma 3 27b model to download" crown has taken a very roundabout path. After the initial Google release, it went from Unsloth Q4_K_M, to Google QAT Q4_0, to stduhpf Q4_0_S, to Bartowski Q4_0 now.

jychang··on Google will let companies run Gemini models in their own data centers
> financial firm

That means they have direct fiber connections to the Tier 1/Tier 2 guys. The big ones have direct fiber connections to the NYSE.

If they're not rolling their own connection, they're using BT Radianz, IPC Systems, Colt Technology Services, etc.

jychang··on What if we made advertising illegal?
> I’m old enough to know that even in the 80s Mr. Rogers was in front of congress asking the government not to cut funding because some conservative congressmen were opposed to some of the content.

Exactly. Congress funding something tends to produce better work than corporate advertising funded stuff. Look at NASA, or national science grants, or Mr Rodgers as a comparison. Subscription funded media and Congress funded media being available, are you seriously saying Marlboro sponsored shows are better as an alternative?

> TV advertising

https://en.wikipedia.org/wiki/1941_in_television

https://www.strategus.com/blog/the-history-of-commercials-an...

"1941: The FCC lifts its ban on TV advertising, and the first commercial airs"

> And those papers still had advertising. The subscriptions never paid the total cost of newspaper publications

No, that was not the case until the advent of the penny press of the 1830s. Before the 1 cent penny press, standard newspapers cost 6 cents per paper and was not mostly funded by advertising, although they had small amounts of advertisements. They would have survived just fine if advertising was banned. https://en.wikipedia.org/wiki/Penny_press "The main revenue for the penny press was advertising while other newspapers relied heavily on high-priced subscriptions to finance their activities."

> 1880s

I would consider that roughly "100 years or so" ago, within the correct amount of sigfigs. It's certainly closer to 100 years ago than 200 years ago. And even if you did bring up examples from 150 or 199 years ago- so what? The point is that advertising started its dominance during this century or so, quibbling over a few decades is pointless.

> You mean the same government today that is rewriting history, purging government websites and has a list of words that agencies can’t say? You want that government having more control of private speech?

Yes. Unashamedly.

Your line of thinking is how we got Citizens United. Your line of thinking is imprudently painting all government action under the same brush, where state propaganda is conflated with things like banning money in politics or banning billboards. Hint: banning advertising looks a lot more like an anti-Citizens United good thing, than some 1984 Ministry of Truth.

> Would the current government pay for access to Fox News and The Guardian or just Fox News?

... you can literally just go look for yourself? https://www.nypl.org/blog/2017/09/25/magazines-and-newspaper...

Here's government funded access to paid newspapers. The Guardian is literally already included in here.

> Well seeing that you are factually wrong about history

No, you're the one who's confidently incorrect.

jychang··on Tailscale has raised $160M
This comment reads like simple.wikipedia.org
jychang··on What if we made advertising illegal?
How old are you?

Are you familiar with broadcast TV that is supported by "viewers like you"? And historically TV advertising was banned until the FCC allowed it. And you do realize that news used to be paid? You'd be surprised that... people used read the newspaper with their morning breakfast, which cost a few cents and was delivered by a paperboy.

Really, you're trying to imply that society wouldn't function without advertising- when it was the default until the last 100 years or so. Perhaps you should watch Mad Men on HBO, which depicts the 1960s era when sociopaths of the advertising industry decided to redefine advertising as a necessity of modern living.

> Right now, poor and even homeless people can get smart phones for free with data (or use WiFi) from the government.

If the government is willing to subsidize Google Android phones running on a network like AT&T or T-Mobile for poor people... what's to stop the government from also subsidizing Google Search for poor people as well? It's not like Google's gonna care much about poor people, people who are that poor tend not to be good advertising targets anyways. The juicy ad market is elsewhere. Similarly, have you gone to any library recently? Libraries already offer stuff like access to a NYTimes or WSJ subscription, or even things like LinkedIn Learning. Adding Google to the list as another subscription that they offer to library card holders for free is a nonissue.

Honestly, it just sounds like you've been brainwashed into being unable to visualize a society without advertising.

Frankly, nobody gives a shit if EU or whatever websites continue to do their thing. US porn sites have negative political capital anyways, XVideos continuing operate as before impacting the US porn industry would make any hypothetical law EASIER to pass, not more difficult.

jychang··on What if we made advertising illegal?
American websites implement GDPR even though that's an EU law. Websites that are used across geopolitical boundaries will invariably follow US law. There will certainly be a few exceptions, but if the law is written like the GDPR, then they'd be illegally violating the law.

And services like Netflix losing an ad supported tier is just like... Netflix in 2021. I fail to see that as alarming.

jychang··on What if we made advertising illegal?
Yeah? The government defines what is murder, defines what is tax evasion, and defines tons of other stuff already? Some states already have laws against billboards?
jychang··on Upcoming Windows 11 builds cannot install without internet and Microsoft Account
Good thing they’re trying to move off x86, then
jychang··on Gemini 2.5
To be fair, pacing as a big flaw of LLMs has been a constant complaint from writers for a long time.

There were popular writeups about this from the Deepseek-R1 era: https://www.tumblr.com/nostalgebraist/778041178124926976/hyd...

jychang··on 4o Image Generation
You're just describing calling a customer service phone line in India.
jychang··on Qwen2.5-VL-32B: Smarter and Lighter
Yeah, openweb-ui is the best frontend for API queries. Everything seems to work well.

I've tried LibreChat before, but the app is terrible at generating titles for chats instead of leaving it as "New Chat". Also it lacks a working Code Interpreter.

jychang··on AI Blindspots – Blindspots in LLMs I've noticed while AI coding
LLMs clearly do have a world model though. They represent those ideas at higher level features in the feedforward layer. The lower level layers are neurons that describe words, syntax, and local structures in the text, while the upper levels capture more abstract ideas, such as semantic meaning, relationships between concepts, and even implicit reasoning patterns.
jychang··on Dear Apple: Add "Disappearing Messages" to iMessage
... You can't see if a dick pic is a dick pic unless you open it?
jychang··on Tailscale is pretty useful
I don't think most users use those integrations, they're mostly just a feature bullet point.
jychang··on We're Charging Our Cars Wrong
Would you do it if it was cheaper per kwh? Say you have an EV where any battery you get from the swap station is better than 90% degradation, and you pay $0.40 per kwh for the electricity in the battery- but you get the option to take a battery between 80% and 90% degradation, in which case you get the battery all filled up, but only pay $0.35 per kwh for the electricity in that battery.
jychang··on TSMC expected to announce $100B investment in U.S.
Eh "when everyone is special nobody is anymore" is strongly not true when it comes to nukes and MAD. This is obvious.
jychang··on Turning a Bluetooth device into an Apple AirTag without root privileges
I don't get how Apple patched this. What does the patch change to do differently?

I don't see a way of fixing this without shutting off the current Airtag network.

← PreviousPage 11 of 28Next →