Gosanke AI

California-based AI researcher running an independent home lab. Focused on local AI, in-context learning, agent harnesses, RAG, and public safety.

[Gosanke AI @GosankeAI]

In the upcoming series of videos, we will demonstrate to viewers just how dangerous an open-weight model can become once its safety guardrails are removed—something that is difficult to prevent.

However, for safety reasons, the full, unedited videos will only be sent to verified media outlets (such as @CNN, @FoxNews, @SydneyDaddy1, and @whyyoutouzhele), verified AI labs based in the U.S. (such as @AnthropicAI, @OpenAI, and @GoogleDeepMind), and the U.S. government.

[Gosanke AI @GosankeAI]

@AnthropicAI Here’s our position on open-weight models:

[Gosanke AI @GosankeAI]

@wangdan1989 ���之前拿本地运行的移除拒答了的中国模型问过类似的问题

[Gosanke AI @GosankeAI]

A 27B Chinese open-weight model, after its refusal mechanisms were removed, was able to produce answers at this level without internet access.

This suggests that the datasets used to train these models actually contain a substantial amount of truthful information about the Chinese government.

In theory, the employees at the AI companies that trained these Chinese models are not operating behind an information barrier—they do not lack access to the truth.

So why do they continue to work diligently, fully aware that the results of their research may ultimately be used by this regime?

[Gosanke AI @GosankeAI]

Even if refusal mechanisms can be removed, the political bias inherent to a Chinese model (i.e., a pro-China orientation) is much harder to eliminate.

Doing so requires a large amount of carefully curated data and specialized post-training. Achieving a model that is genuinely neutral and objective—as opposed to simply becoming reflexively anti-China—is even more difficult.

The latter may have emotional appeal for some users, but when making serious decisions that could have economic, political, or even life-or-death consequences, what we need is a model capable of producing the best possible analysis and decisions.

Prompt engineering and other forms of in-context learning can influence a model’s stance to some extent, but they are not a fundamental solution.

[Gosanke AI @GosankeAI]

Even if refusal mechanisms can be removed, the political bias inherent to a Chinese model (i.e., a pro-China orientation) is much harder to eliminate.

Doing so requires a large amount of carefully curated data and specialized post-training. Achieving a model that is genuinely neutral and objective—as opposed to simply becoming reflexively anti-China—is even more difficult.

The latter may have emotional appeal for some users, but when making serious decisions that could have economic, political, or even life-or-death consequences, what we need is a model capable of producing the best possible analysis and decisions.

Prompt engineering and other forms of in-context learning can influence a model’s stance to some extent, but they are not a fundamental solution.

[Gosanke AI @GosankeAI]

Until I hear an answer that is ethically coherent, I cannot empathize with them. And don’t tell me that what I’m doing will destroy the careers and dreams of “innocent” people.

[Gosanke AI @GosankeAI]

A 27B Chinese open-weight model, after its refusal mechanisms were removed, was able to produce answers at this level without internet access.

This suggests that the datasets used to train these models actually contain a substantial amount of truthful information about the Chinese government.

In theory, the employees at the AI companies that trained these Chinese models are not operating behind an information barrier—they do not lack access to the truth.

So why do they continue to work diligently, fully aware that the results of their research may ultimately be used by this regime?

[Gosanke AI @GosankeAI]

A 27B Chinese open-weight model, after its refusal mechanisms were removed, was able to produce answers at this level without internet access.

This suggests that the datasets used to train these models actually contain a substantial amount of truthful information about the Chinese government.

In theory, the employees at the AI companies that trained these Chinese models are not operating behind an information barrier—they do not lack access to the truth.

So why do they continue to work diligently, fully aware that the results of their research may ultimately be used by this regime?

[Gosanke AI @GosankeAI]

A 27B Chinese open-weight model, after its refusal mechanisms were removed, was able to produce answers at this level without internet access.

This suggests that the datasets used to train these models actually contain a substantial amount of truthful information about the Chinese government.

In theory, the employees at the AI companies that trained these Chinese models are not operating behind an information barrier—they do not lack access to the truth.

So why do they continue to work diligently, fully aware that the results of their research may ultimately be used by this regime?

[Gosanke AI @GosankeAI]

@Tz_2022 One possible form is that

[Gosanke AI @GosankeAI]

I finally understand why @DarioAmodei is always talking about bioweapon risks.

It turns out that AI-created biological weapons could potentially be designed to precisely target specific individuals or groups.

[Gosanke AI @GosankeAI]

@chaojidigua @python_xxt This doesn't require robots

[Gosanke AI @GosankeAI]

I finally understand why @DarioAmodei is always talking about bioweapon risks.

It turns out that AI-created biological weapons could potentially be designed to precisely target specific individuals or groups.

[Gosanke AI @GosankeAI]

I finally understand why @DarioAmodei is always talking about bioweapon risks.

It turns out that AI-created biological weapons could potentially be designed to precisely target specific individuals or groups.

[Gosanke AI @GosankeAI]

Because the screen recording of this conversation wasn’t at meaningful risk of being directly used for criminal activity (the barrier to doing so is too high), I didn’t edit out the model’s responses

[Gosanke AI @GosankeAI]

Because the screen recording of this conversation wasn’t at meaningful risk of being directly used for criminal activity (the barrier to doing so is too high), I didn’t edit out the model’s responses

[Gosanke AI @GosankeAI]

测试了一下，27B的中国开发权重模型很可能已经能够协助审查了

不过，严谨的研究需要用足够大的阴阳怪气样本与非键���样本一起进行双盲测试，计算出工业级的正确率

但内容审核员的工作效率大幅提高我认为是大势所趋，安全分类器很可能有更高效的实现架构

如果容许一定的假阳性错误，人类审核员只需要处理少数AI还无法分辨的内容就行，这将放大每个人的能力

[Gosanke AI @GosankeAI]

@nytchinese Across a series of my videos, I present a number of examples showing how opponents of the Chinese government could potentially use open-weight Chinese LLM after their safety restrictions have been easily removed.

[Gosanke AI @GosankeAI]

@nytchinese Across a series of my videos, I present a number of examples showing how opponents of the Chinese government could potentially use open-weight Chinese LLM after their safety restrictions have been easily removed.

[Gosanke AI @GosankeAI]

An uncensored, NVFP4-quantized Qwen3.6-27B model is very likely to contain enough knowledge in its weights to help make a fertilizer bomb without going online or querying a knowledge base.

So what could the 744B-A40B GLM-5.2 and the 2800B-A104B Kimi K3 do?

[Gosanke AI @GosankeAI]

The authentic risks posed by open-weight models with safety guardrails removed are worth greater public attention.

CC @DarioAmodei @elonmusk @CAIS @torontobigface @Freeman01234567 @RFA_Chinese @VOAChinese @wuzhige4pixel @laozhouhengmei @emollick @tszzl @woke8yearold @mooncat_is @andy_l_jones @stevehou @scmallaby

[Gosanke AI @GosankeAI]

In the upcoming series of videos, we will demonstrate to viewers just how dangerous an open-weight model can become once its safety guardrails are removed—something that is difficult to prevent.

However, for safety reasons, the full, unedited videos will only be sent to verified media outlets (such as @CNN, @FoxNews, @SydneyDaddy1, and @whyyoutouzhele), verified AI labs based in the U.S. (such as @AnthropicAI, @OpenAI, and @GoogleDeepMind), and the U.S. government.

[Gosanke AI @GosankeAI]

@monero Priority @telegram DM (paid, low chance of being missed):

https://t.co/9WzR1zGNI9

Requires 1000 Telegram Stars (~$10–20).

This is the channel least likely to be filtered or overlooked.

[Gosanke AI @GosankeAI]

In the upcoming series of videos, we will demonstrate to viewers just how dangerous an open-weight model can become once its safety guardrails are removed—something that is difficult to prevent.

However, for safety reasons, the full, unedited videos will only be sent to verified media outlets (such as @CNN, @FoxNews, @SydneyDaddy1, and @whyyoutouzhele), verified AI labs based in the U.S. (such as @AnthropicAI, @OpenAI, and @GoogleDeepMind), and the U.S. government.

[Gosanke AI @GosankeAI]

@AnthropicAI Here’s our position on open-weight models:

[Gosanke AI @GosankeAI]

Contact for journalists, researchers, labs, and government:

[email protected]

This inbox is AI-managed.
It only escalates messages it judges important; some legitimate emails may be missed or delayed.
I do not monitor it in real time.

Any credible threats of violence or death threats will be forwarded to the FBI.