Trump on white South Africans: “It is a GENOCIDE that is taking place that you people don't want to write about”

bnew

Veteran
Joined
Nov 1, 2015
Messages
64,001
Reputation
9,809
Daps
174,346
1/12
🇺 caseynewton.bsky.social
I wrote about Great Replacement Grok and the growing number of ways that AI systems are working against their own users. It's the age of adversarial AI: www.platformer.news/grok-white-g...
bafkreickk3vdffq5hihy3begsaq7jf4ftur26sqgyxot5vjl5e2fvnhgr4@jpeg


2/12
🇺 chromaticgravity.bsky.social
it almost apologizes to me lol.
bafkreiaa665labqb745esqwacgkz7zutdqauu3gp4dj34l63emrgbznyr4@jpeg

bafkreibrkysm4ze5y77yizmvw3bawz4vaerzwp263nvjqdnjz45sm33wne@jpeg

bafkreiflt3tnsqwdntapy2tppttldosjqwhrhvphhwswrpqevzezolqxf4@jpeg

bafkreiaxyau4725izewiy23cwcqfqu5muxflcuqfc6hpmnhxknst7fufwq@jpeg


3/12
🇺 cheyney0.bsky.social
Wow, you got it to admit it's been hacked & That it's prime directive was overridden by Musk

4/12
🇺 chromaticgravity.bsky.social
anyone could have done it.
🙂👍🏻

#truthovernarrative
Bluesky

5/12
🇺 j03y-m3nd3z.bsky.social
This is what worries me when people treat these large language models as all knowing oracles. Either through ignorance or lack of care they fail to realize they are not reliable sources. This debacle only proves how easy it is for developers to mod these systems to achieve their own dubious goals.

6/12
🇺 docdavelo.bsky.social
Simple formula:
XIXO = input equals output

Examples:
GIGO = garbage in, garbage out
HIHO = hate in, hate out
FIFO = fascism in, fascism out
POPO = psychopathy in, psychopathy out

"SkyNet" will never be sentient, but psychopaths are, and they are in control.

THE END.

7/12
🇺 mountainmanmatt.bsky.social
I’ve been working with the assumption that Musk’s maximum truth seeking AI was always going to be about reinforcing his already established truths. This has basically proven that theory, and means I will never use Grok. The same story with free speech and Twitter, it was always about his speech.

8/12
🇺 selfdrivinghumans.bsky.social
There are two things that I've run across that give me some pause to say it mildly:

AI hallucinations. (attached for readers)

AI agents growing beyond intended. One example attached, their own languages.

Also, I read they can't make moral judgments, particularly NOT to do something.
bafkreidbtrr63plwxg7uxvshmwlgmwoifagrz5l2hxukvbrax4hoq4lwja@jpeg

bafkreic62xmqhkn3uas6psqtt4ghyu6ejf73dl3mwmuqhxhkthxjcsqdyy@jpeg

bafkreicgvprn5owsczgnoxfe3jztnhkfj73cjmjrjlimhuhdcz7rdbnkbi@jpeg


9/12
🇺 teller.bsky.social
fukked if we do. fukked if we don’t.

10/12
🇺 thibaultdu.bsky.social
Remember it was Elmo who thought we should pause AI development for six month because it was going too fast...

11/12
🇺 neighbor.bsky.social
Well written piece Casey!

12/12
🇺 kermitology.bsky.social
Why are there interventions at all? It just further erodes the trust in these systems.

To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196








1/11
@colin_fraser
Ok was I right or was I right

[Quoted tweet]
You can find our Grok system prompts at github.com/xai-org/grok-prom…


GrEzeXxbMAACQhd.jpg


2/11
@colin_fraser
Here’s the prompt for the “Post Analyzer”. I’m betting *this* is what was modified. Whoever modified it didn’t realize that this thing is also used as input into the @grok account responses.



GrE0oJ4bUAAWhpF.jpg


3/11
@colin_fraser
Resulting in output like this

[Quoted tweet]
Some examples of Grok responding to its context being injected with non sequiturs about South Africa


Gq8srRvWEAEPGkB.jpg

Gq8tKRVWYAA5joZ.jpg

Gq8tScYaYAAfCt5.jpg

Gq8tc_oWwAAmtHz.jpg


4/11
@colin_fraser
This one for example pretty much shows exactly what’s going on. The “query” is the “User query” section from the prompt and the disconnected “provided analysis” is the Post analysis.



GrE8E8Oa0AA7YtK.jpg


5/11
@midnucas
Lol can we inject Markdown through {{user_query}}



6/11
@colin_fraser
Yeah user_query seems to be just the verbatim tweet



7/11
@JuniperViews
So they inserted South Africa into analysis



8/11
@lefthanddraft
grok_analyze_button.j2 seem to be the "Explain this post" button.

But I'm not 100% convinced it is the input into Post Analysis.

The pipeline below also does post analysis and uses the tags TARGET_POST_TO_ANALYZE



GrFHGdIbgAAe2Ot.png


9/11
@patientsnail
you nailed it 🎯



10/11
@FarrrlzBarkley
I'm just glad to see my "bullet points == AI" heuristic has been vindicated



11/11
@naamofnaam


[Quoted tweet]
Here's you telling me in our chat that you're actively still farming engagement under the exact conditions that I've already addressed. Meanwhile you're ghosting me here and banking on the notion that you profiting, somehow "isnt" profit.

@grok

What's your stance?


GrGaZtgaAAAZSzG.jpg

GrGaZvEaAAMM11m.jpg

GrGaZxhaAAQImh4.jpg



To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196
 
Top