-
DeepSeek vs: Winnie-The-Pooh: the winner is Chinese Censorship — unless you know how to circumvent it
Prompted by a blog comment from Ben Tasker, I tried to repro what Ben had done; I asked three questions (illustrated) and it responded to the third at considerable length before something else kicked-in and redacted the answer which had been printed up to that point, replacing it with a “beyond my current scope” phrase.
Fediverse reactions
-
Suffice it to say that DeepSeek’s LLM doesn’t have problems commenting upon political massacres so long as they didn’t happen in China
See the above. I wonder what can be reverse-engineered out of the model to find out what other things it does not want us to know?
Fediverse reactions
-
Dan Brickley on Bluesky: “Testing DeepSeek R1 censorship UI (mobile web), since it will get heavy use worldwide” | …thread with multiple screenshots
Screenshots: session one, snapshot of the ui rewriting. Session two, numbering messages to see which stay in history. Finally, its response to uploading 1st screenshot (initial engagement switching to refusal)
Fediverse reactions
-
“deepseek r1 has an existential crisis” | …more fuel for the upcoming “Open Source LLMs on Raspberry Pi’s are a National Security issue” crisis
Further to my posting from yesterday: I’m very much in favour of people publishing AI models & individuals being free to use them at home, but if the attached sort of thing happens, it’s clear there will be an enormous brouhaha about “are models free speech or should they be controlled for [reasons]?” I look
-
Ofcom Session for Small, Low-Risk Community Sites | Wednesday 12 February
12:30 – 14:00 GMTIn light of the recent proposed shutdown of several hundred online communities incapable of resource for meeting Ofcom’s bar, this should be interesting: …a lunchtime session with Ali Hall & Andrew Park from Ofcom to discuss the new Online Safety responsibilities for small, low-risk community sites. As well as an overview… there will be time
Fediverse reactions
-
If what @BrianRoemmele claims is delivered, running a Chinese LLM on a Raspberry Pi at home should become trivial
We should be able to hear the US National Security wonks & the European AI regulators screaming, any moment now. More on DeepSeek-R1 here: https://arxiv.org/abs/2501.12948 and here: https://api-docs.deepseek.com/news/news250120
Fediverse reactions
-
Software > Patents: open-source Python plugin for slicer software implements staggered layers, improves strength, violates dodgy patents
Via Reddit, the Bricklayers plugin for PrusaSlicer and OrcaSlicer is a dinky little python script that implements staggered-layer 3D-printed walls that improve print strength commonsensically… but open a can of patent worms
Fediverse reactions
-
Europol Director: “The European People are Criminal Scum and cannot be trusted with robust online privacy” | …or words to that effect
Technology giants must do more to co-operate with law enforcement on encryption or they risk threatening European democracy, according to the head of Europol … companies had a “social responsibility” to give the police access to encrypted messages … “Anonymity is not a fundamental right” … “You will not be able to enforce democracy [without
Fediverse reactions
-
Shallow Deepfakes circumvent legal/licensing issues streaming Australian Tennis … as done by the *organisers*
We should call this what it is, and it’s fascinating: using “live deepfake” technology is used to re-skin actual tennis players as Wii animations, so it’s not really “their” tennis-match being streamed to YouTube. Goodbye, licensing issues. Where are the anti-AI, anti-Deepfake activists? Tennis Australia has transformed the world’s best tennis players into animated versions
Fediverse reactions
-
One massive, glorious upside of the supposed “end of factchecking” is that Americans are finally learning how to Satire without a safety net
Too long the only US exponent of satire has been The Onion, and that seems only to survive on the basis of brand; but now people are saying things which are pointedly, gloriously and incredibly (nb: in-credible, viz: not credible on their face) false. Please, continue! Let’s hope that the populace now swim in enough
Fediverse reactions