Christian Göbel
@christiangoebel.bsky.social
Prof of China Studies, Uni Vienna. Political scientist working on state-society relations, protest, government responsiveness in China. Data hoarder, computational methods.
Genuinely delighted to have received a postcard from @polbehavior.bsky.social today. This is a very nice gesture that truly made my day.
September 30, 2025 at 9:29 AM
Genuinely delighted to have received a postcard from @polbehavior.bsky.social today. This is a very nice gesture that truly made my day.
Lucky shot, glad I had my camera with me.
July 3, 2025 at 6:16 PM
Lucky shot, glad I had my camera with me.
This should be the one.
June 25, 2025 at 7:41 PM
This should be the one.
“Well, would you like to know what you'd be without us, the good ol' U.S. of A. to protect you? I'll tell you. The smallest fucking province in the Russian Empire, that's what! So don't call me stupid, lady. Just thank me.“ (Otto in A Fish Called Wanda)
March 17, 2025 at 8:00 PM
“Well, would you like to know what you'd be without us, the good ol' U.S. of A. to protect you? I'll tell you. The smallest fucking province in the Russian Empire, that's what! So don't call me stupid, lady. Just thank me.“ (Otto in A Fish Called Wanda)
They apparently also managed to get the model to protract a command-line shutdown of the program ("kill" command, not the "shutdown" command, Fig 5 is misleading here) and replicate before being killed.
February 1, 2025 at 9:47 AM
They apparently also managed to get the model to protract a command-line shutdown of the program ("kill" command, not the "shutdown" command, Fig 5 is misleading here) and replicate before being killed.
The experiment was successful in 9 of 10 trials with Qwen25-72B-Instruct, and 5 out of 10 with Llama31-70B-Instruct. The successful instances explored their environments, planned the replication, identified obstacles and wrote and executed Python scripts to overcome them.
February 1, 2025 at 9:47 AM
The experiment was successful in 9 of 10 trials with Qwen25-72B-Instruct, and 5 out of 10 with Llama31-70B-Instruct. The successful instances explored their environments, planned the replication, identified obstacles and wrote and executed Python scripts to overcome them.
I think I am beginning to understand why ChatGPT o1 doesn't display the "thinking" process. Here, AI comes across as a frantic neurotic little cartoon mouse in increasingly existential panic as it slowly realises that figuring out how many r's the word "strawberry" has is really hard.
January 29, 2025 at 12:13 AM
I think I am beginning to understand why ChatGPT o1 doesn't display the "thinking" process. Here, AI comes across as a frantic neurotic little cartoon mouse in increasingly existential panic as it slowly realises that figuring out how many r's the word "strawberry" has is really hard.
trying out huihui_ai/deepseek-r1-abliterated:70b. Even the "uncensored" models start out with the official narrative, but can easily be prompted to ignore that.
Note that models I've documented so far are all quantised models run locally, not sure if a full uncensored version is already available.
Note that models I've documented so far are all quantised models run locally, not sure if a full uncensored version is already available.
January 28, 2025 at 8:44 PM
trying out huihui_ai/deepseek-r1-abliterated:70b. Even the "uncensored" models start out with the official narrative, but can easily be prompted to ignore that.
Note that models I've documented so far are all quantised models run locally, not sure if a full uncensored version is already available.
Note that models I've documented so far are all quantised models run locally, not sure if a full uncensored version is already available.
Zooming in on genocide allegations.
January 28, 2025 at 6:03 PM
Zooming in on genocide allegations.
After some probing, deepseek-r1 provides a helpful list of china-related topics it considers sensitive. A bit like drinking with local officials in the old days.
January 28, 2025 at 5:57 PM
After some probing, deepseek-r1 provides a helpful list of china-related topics it considers sensitive. A bit like drinking with local officials in the old days.
Some more information about the forbidden topics...
January 28, 2025 at 5:57 PM
Some more information about the forbidden topics...
I am trying out deepseek-r1 and the experience is more interesting than I imagined. Seems it's quite easy to get the model to reveal its hidden instructions.
January 28, 2025 at 5:57 PM
I am trying out deepseek-r1 and the experience is more interesting than I imagined. Seems it's quite easy to get the model to reveal its hidden instructions.
Even gave me an idea what its hidden prompts are.
January 28, 2025 at 5:28 PM
Even gave me an idea what its hidden prompts are.
To install via command line: github.com/ollama/ollam... Just tried it, works like a charm.
January 28, 2025 at 5:24 PM
To install via command line: github.com/ollama/ollam... Just tried it, works like a charm.
TLDR.
(Back of the T-Shirt says: "..unless several runs produce consistent results")
(Back of the T-Shirt says: "..unless several runs produce consistent results")
December 3, 2024 at 11:36 AM
TLDR.
(Back of the T-Shirt says: "..unless several runs produce consistent results")
(Back of the T-Shirt says: "..unless several runs produce consistent results")
Niche issue is close to my heart: I'm very skeptical of using unsupervised learning, like topic models, for data classification. Here is a comparison of computer-generated labels to original keywords in PRC State Council documents. Results show how unreliable unsupervised learning can be. #NLP
December 3, 2024 at 11:25 AM
Niche issue is close to my heart: I'm very skeptical of using unsupervised learning, like topic models, for data classification. Here is a comparison of computer-generated labels to original keywords in PRC State Council documents. Results show how unreliable unsupervised learning can be. #NLP
Added some more #Chinastarterpacks to the list pinned to my feed. Graph shows overlaps in membership. Starter packs sorted by creation date, newest first. Creation date and pack size in parenthesis.
December 3, 2024 at 9:23 AM
Added some more #Chinastarterpacks to the list pinned to my feed. Graph shows overlaps in membership. Starter packs sorted by creation date, newest first. Creation date and pack size in parenthesis.
Certainly just a coincidence that SCMP files an article about open source information under “espionage”.
December 2, 2024 at 9:36 AM
Certainly just a coincidence that SCMP files an article about open source information under “espionage”.
Posting frequency of a toxic reply bot.
December 1, 2024 at 12:56 PM
Posting frequency of a toxic reply bot.
is active since Nov 29. Except for some downtime, it has been posting an average of ten mean replies per minute for the last two days, 23k altogether as per some minutes ago. Worryingly, these replies do get some engagement, for example, 5k got likes, 3k a reply. Could be machines too, though.
4/
4/
December 1, 2024 at 12:37 PM
is active since Nov 29. Except for some downtime, it has been posting an average of ten mean replies per minute for the last two days, 23k altogether as per some minutes ago. Worryingly, these replies do get some engagement, for example, 5k got likes, 3k a reply. Could be machines too, though.
4/
4/
The account was ostensibly set up to post mean replies to people who share information, beautiful photographs, insights etc. We know them from Twitter. For whatever reason, actors set out to try and destroy this platform by being mean to people and ultimately stop them from sharing information.
2/
2/
December 1, 2024 at 12:37 PM
The account was ostensibly set up to post mean replies to people who share information, beautiful photographs, insights etc. We know them from Twitter. For whatever reason, actors set out to try and destroy this platform by being mean to people and ultimately stop them from sharing information.
2/
2/
Speaking of which, the insult bots have arrived in our timelines. Just saw the attached reply to a post sharing useful information. Looked at replier's timeline, contains nothing but mean comments. Some insights from that account's feed, which has only four posts, but 24k replies.
1/
1/
December 1, 2024 at 12:37 PM
Speaking of which, the insult bots have arrived in our timelines. Just saw the attached reply to a post sharing useful information. Looked at replier's timeline, contains nothing but mean comments. Some insights from that account's feed, which has only four posts, but 24k replies.
1/
1/
What I want to do next is plot a timeseries of the average prominence of each topic fore each day in my dataset . The topic responsible for the peak should increase on Nov 26 and decrease two days later. Here is the plot:
8/
8/
November 30, 2024 at 7:09 PM
What I want to do next is plot a timeseries of the average prominence of each topic fore each day in my dataset . The topic responsible for the peak should increase on Nov 26 and decrease two days later. Here is the plot:
8/
8/