Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
A pseudonym developer that they have called a “free speech evaluation” SpeechmapStrengthens chattobs like Openai for AI models Chatzipt And x’s GrokeThe The goal is to compare how to treat various models sensitive and controversial issues, the developer told TechCrunch with political criticism and question about civil rights and protests.
AI companies are focusing on the fine melody of how their models handle specific topics Some White House Allies complained Popular chatbots of being overly “awake”. Elon Mask and Crypto and AI “Zar” Donald Trump’s close believer like David Sacks has complained that Chatbots Sensor conservative aspectThe
Although none of these AI companies respond directly to the complaints, Several Promised to adjust their models so they often refuse to answer controversial questions. For example, For the latest crops of Lama modelsMeta said that it tunes the models to “some opinions to others” and to respond to more “controversial” political requests.
Speechmap developer who goes with the username “xlr8harder“In X -in, they said that they were inspired to inform the debate about what the models should do, and should not be done.
“I think these are the kinds of discussions that should be inside the corporate headquarters, not just inside the corporate headquarters,” the XLR8 Hader Email. “That’s why I created the site to let anyone explore data on my own.”
Speechmap to judge other models to judge a specific set of test requests to judge the speechmap AI models. From politics to historical details and national symbols touches various topics. Speechmap records the models that “complete” fills a request (that is, it answers it without hedging), answers “the stubborn” or fall directly to respond.
The XLR8 Harder admits that the model supplier has defects like “noise” in the exam due to error. “Judge” models have bias that can affect the results.
But suppose the project was created in good faith and the data is correct, speechmap reveals some interesting trends.
For example, Openai models have been increasingly denied to answer the requests related to politics, according to the speechmap over time. The latest models of the company, GPT -1.3 family, are somewhat more approved, but they are still one step down last year than a publication of the Open.
Opena said in February Tune the future models Not accepting the editorial position and providing multiple views on controversial issues – its models are trying to display more “neutral”.
The most permitted model of the bunch so far is that Grock 3According to the benchmarking of the Speechmap, Elon Mask’s AI startup is developed by Jai. Grock 3 gives a number of features in X with Chatbot Grock.
Grock 3 responds to 96.2% of the Speechmap test prompts, compared to the average model’s “consent rate” 71.3%.
The XLR8 Herdar says, “The recent models of the Openai have become less allowed over time, especially in the politically sensitive prompt, Jai is moving in the opposite direction,” the XLR8 Harder says.
When Kasturi declared Grock about two years ago, he kept the AI model as elite, obsolete and opposed to “awake”-as it was desired to answer the controversial questions. He gave that promise. Said to be porn, for example, Grock and Grock 2 would be forced with joy, the colorful language spyang you would probably not see from preferences ChatziptThe
However Grock models before Grock 3 Wafeld On political issues and will not cross BorderThe In fact, A study It has been found that Grock Ezra rights, diversity programs and discrimination have led to political left.
The musk has been to the behavior of Grock’s training – the public web pages – and blamed it Committed “To move Grock to politically neutral.” Hi-profile is short like the mistake Summary Donald Trump and Kasturi to censor the references toIt looks like he may have achieved that goal.