HN user
Thanks Bonsai team.
Now open weight LLMs/VLMs/LMMs are becoming even larger to the extent that consumer-grade hardware are no longer able to run these models. In contrast, quantization and pruning make the model better at the size-performance pareto and provide people with strictly more possibilities.
Other details for the official (Yang Youlin) in this news.
---
Whistleblower Yang Hai already reported Yang Youlin for his economic misconduct in July 2008. The whistleblower was detained because of the report at Nov 21st, 2008.
---
People's comment on this matter around March, 2009:
哇塞！终于有人敢动杨友林啦，杨海好样的。杨友林此人早该除无奈碍于他的势力。除掉杨友林大快人心。Wow! Finally, someone dares to take action against Yang Youlin. Good job, Yang Hai. That guy should have been dealt with long ago, but we were stuck with his influence. Getting rid of him is incredibly satisfying.
不杀此贪官，难平民愤。If you don't execute this corrupt official, you won't appease the public's anger.
江宁有一个传说，谁也动不了杨永林！There's a legend in Jiangning: nobody can touch Yang Youlin!
他的保护伞是谁？Who is protecting him?
希望引起中央的重视！I hope this gets the attention of the central government!
现在社会怎么啦？好多天了根本没人关注这件事？是上层没有看到？还是视而不见？还是怕牵连自己？What's going on with society lately? It's been days and no one is paying attention to this! Did the higher-ups miss it? Are they turning a blind eye? Or are they just afraid of getting dragged into it?
---
Since the report, there were several pieces of news about "the investigation about Yang Youlin is ongoing", but no real progress until 2023.
The criminal complaint missing in this article: https://www.justice.gov/usao-ndil/media/1450651/dl
Basically, Microsoft logs things from windows users, including and not limiting to, the machine GDID, IPs that come with the GDID, and when and what exact URLs accessed. So for windows users, privacy, an important part of information security, is totally destroyed by these logs enforced on windows. Another important part of information security is bug fixing, and microsoft did make at least one security researcher angry [1].
And a simple solution to that problem is moving to linux. You save yourself a lot of time and energy for leaving the adversarial information security condition imposed by windows and microsoft.
P.S. You may consider debloating windows for a more information security friendly environment. However, that is nearly impossible, as long as you realize that windows is an OS composed of thousands of closed source softwares, and doing security audits on all of these will be costly.
Yeah, searxng instance reliability problem goes deeper than simple uptime. The google response time function on public instance page [1], a good measure of the search engine availability of one instance, is broken for some time. Without real testing, people really cannot get the full view of the service status of a single instance.
I've been using searxng for several years now. I don't run my own instances because the inhumane network censorship imposed by GFW, and proxy detection enforced by search engines. Instead, I rely on public instances on the list [1] and libredirect [2]. Note that service from a single instance is not guaranteed, but you can always switch to other available instances with little cost within a minute.
I won't say searxng can help you degoogle because metasearch engine calls other search engines, e.g., google, to collect results. However, if you try searxng, you can at least get rid of things like ai reviews in no time.
In the end, thank you people after searxng project and public instances.
Android developer verification program, together with recent reCAPTCHA push [1], and Manifest v2 force depreciation on chrome [2], make one thing crystal clear. When companies like GOOGLE talks about things in the name of "your security", it's a sign that they want you to sacrifice your own things, e.g., privacy, freedom, etc., for their own security. And if you trust them and show your consent by doing nothing, you pay the price.
Glad to see more open models. However, where are the 31b models?
I've been using FUTO voice input for several month. It's not perfect as I have to manually do the correction like always. However, it definitely saves me some time and effort. It also helps me to degoogle, starting from gboard.
Apparently this new swipe function is tied to FUTO keyboard. I don't use the FUTO keyboard as there is no support for the language I use intensively. Nevertheless, this line of work deserves more recognition.
We paid for your things, AMD.
If you want to strip some features from things we bought after the purchasing, you must ask me and every other customers for consents explicitly, with a reasonable explanation, and before the strip happens. If one of us show no consent, you cannot do that.
-------------
See the github issue [1]. @benkilpatrick found out the problem in April. There was absolutely no consent asking information transparency at all. There was inefficient to no information even for people willing to spend THEIR OWN TIME to solve the problem. After about two months of back and forth with motherboard manufacturer, @benkilpatrick found out the problem stems from some components inside the bios, and the components came from AMD. Another ~three weeks passed and no problem resolution at all. It was after things blow up AMD PR came out and said something about "valuable feedback".
Wait, what if there's no enough pushback? What if this github issue as well as the problem it raised is ignored by all? Just see this thread, that thread [2] and whatnot. Is your customers going to screw themselves and being stripped silently for being your customers and believing that new bios will solve their problems without causing shenanigans?
-------------
I won't upgrade bios without future third-party bios integrity checks showing the problem is solved properly.
To people who silently downvoted this: please explain why you did that, you're doing things like AMD.
The github issue that never made it in this news: https://github.com/AMDESE/AMDSEV/issues/292
Silent enshittification in the name of updates is getting out of hand. There are several evidences that downgrading BIOS/AGESA to below 1.2.7.0 to 1.2.0.3 brought back TSME for their AMD cpus.
I downgrade my bios as a price for my blind trust on AMD, and yes TSME is back.
You lost my trust AMD. The lesson learned is that if your PC with AMD cpu is stable, don't do any bios upgrade, as AGESA in the bios is adversarial to you, the users of AMD cpu.
Seems like the affected version of chrome is 150 and 151, and the timewindow is near.
Because of ungoogled-chromium, I'm used to manually update the browser. This time I won't update until MV2 problem is solved properly. I've already lost my extensions once by updating to 138 without checking google's adversarial move, and got the forbidden knowledge that I cannot manually downgrade chromium because it won't start.
Please do not claim you trained a new model, only to got caught red-handed by others. There are already several people or groups did that, got caught, and vanished in no time.
Check how the "authors" of "this model" react to this problem [1]. See how they deal with this problem by first changing their affiliation from https://iplanrio.rio.rj.gov.br to https://iplanrio.prefeitura.rio [2], then saying that they are sorry for being caught [3], then just remove all their affiliations once for all [4].
I think the "authors" of "this model" [5] should be held accountable until they upload new checkpoints, and the performance of the new model is verified by third-parties.
P.S. To people who downvoted me, show me why you're doing this.
[1] https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B/comm...
[2] https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B/comm...
[3] https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B/comm...
[4] https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B/comm...
I don't know how open source AI wins. The description is too vague for serious discussions. What I do know is that, once closed source AI groups become anti-you, you should punish them, or help open source groups, or both.
If you really want specific open source {LLM, LMM, research, harness, whatever} groups to win over closed source counterparts, you may show your care by trying open source solutions first when solving problems. And if they're really capable, award them with contributions or something.
Hilarious read. I laugh out loud multiple times during the read. In the end, I think amd should pay the author simply for the will of debugging for a broken software written by amd, as well as the sheer amounts of loose ends this exploration leads to.
Thanks gemma team for this release.
Compared to autoregressive decoding, diffusion is huge for local MoE inference because of the improved token generation efficiency, especially for normal GPU + ram offload setting.
However, there are models which are better positioned on the performance vs memory pareto front, i.e. dense models, so I'll just wait.
P.S. QAT is really something as it reduces the performance fluctuations compared to the normal one. Thanks again.
More rants about local inference, consider yourself warned.
Together with bf16 related deliberate hardward degrades on consumer-level nvidia gpus, i.e., gtx 10, rtx 20, 30, 40, 50 series, things gets sour really quickly.
From the perspective of a local llm user, I think the qat doesn't solve the major problem of the gemma models.
Gemma family (gen 1 to gen 4) is consistent with extreme range of activations, i.e., 600000, essentially forcing people to use bf16 kv cache and accept a short context window, e.g., 31b, iq4_xs quantization, 100k context window on 32gb memory. Or, people use q8 kv cache, 200k context window, and accept a large performance penalty.
In contrast, for qwen 3.5 family, the largest activation is below 2000, making q8 or even lower-precision kv cache essentially free estates. Together with linear attention, which doesn't require kv cache, full 262k context window can be easily reached.
Qat training with w4a16 target, while improving performance on inference with low-precision weighs, doesn't solve kv cache problem at all.
In the end, a qat is a qat, and there are unseen efforts behind qat checkpoints. Thank you gemma team for releasing qat checkpoints.
A small dense multimodal model with audio support, interesting.
Wait, *Excluding Chinese language.
This is ... curious.
P.S. Where is gemma 4 124b?
This website brings me some good chuckles. Now I really know how powerful an on-demand bullsh*t generator is.
Like the recent copilot silent signing incident, the without consent part is blatant foul move.
If you don't like be treated like anything but human, you should seriously consider replacing chrome with ungoogled chromium or other browsers.
Yeah, this is part of the reason why vscodium exists.
Wow. Just like using ungoogled-chromium instead of chrome, lineage os instead of oem android, using vscodium instead of vscode is again justified. These decisions really are the ones that I'll never regret.
In addition, using the word microslop instead of microsoft is again justified, too.
One thing that makes me wonder is that there are 4 security issues raised and all of them were automatically commented and closed by some bot called `pl-ghost` [1][2][3][4]. In the end, only this one [4] properly handled, and all bot comments are deleted. You can see the bot comments in another report [5], which is more informative than the OP one.
[1] https://github.com/Lightning-AI/pytorch-lightning/issues/216...
[2] https://github.com/Lightning-AI/pytorch-lightning/issues/216...
[3] https://github.com/Lightning-AI/pytorch-lightning/issues/216...
[4] https://github.com/Lightning-AI/pytorch-lightning/issues/216...
[5] https://socket.dev/blog/lightning-pypi-package-compromised
Although the performance claim of 8b dense matching 32b moe is somewhat questionable, thank you granite team for releasing small dense LLMs.
Thank you Qwen team. Small DENSE LLMs shapes the future of local LLM users.
When Qwen 3.5 27b released, I didn't really understand why linear attention is used instead of full attention because of the performance degradation and problems introduced with extra (linear) operators. After doing some tests, I found that with llama.cpp and IQ4_XS quant, the model and BF16 cache of the whole 262k context just fit on 32GB vram, which is impossible with full attention. In contrast, with gemma 4 31b IQ4_XS quant I have to use Q8_0 cache to fit 262k context on the vram, which is a little annoying (no offenses, thank you gemma team, too).
From benchmarks, 3.5->3.6 upgrade is about agent things. I hope future upgrades fix some problems I found, e.g., output repetitiveness in long conversations and knowledge broadness.
Check out Fig. 6 in this paper, it shows the comparison between the proposed method and pytorch native FSDP offload method.
No open weights.
Besides, I'm old enough to recall that META has trained a version of LLAMA 4 specifically for LM arena elo benchmaxxing and PR things, and proceeded to release a different version of LLAMA 4.
Thank you Gemma team for releasing small dense VLM(s).
The elo ranking [1] is too good to be true. I don't know why gemma-4-26b-a4b performs better than gemma-4-31b.
Also waiting for more bugfixes in llama.cpp, sglang and vllm to do proper evaluations.
[1] https://arena.ai/leaderboard/text/expert?license=open-source
Please no.
If you want to install APKs directly on Android phones selling in China, you'll face even more draconian restrictions imposed by both Chinese OEMs and Chinese government, e.g., cannot install telegram [1], cannot install VPNs [2], called by local police station after installing VPNs [3], and so on. And you do not have the freedom to even talk about these restrictions freely without getting sued or censored.
[1] https://xcancel.com/whyyoutouzhele/status/168915238841261670...
[2] https://xcancel.com/whyyoutouzhele/status/197843066556268971...
[3] https://xcancel.com/whyyoutouzhele/status/170299205759627676...