If there was a zombie apocalypse, what would you take from your home? What type of person would you like to live with on Mars?
These are some of the questions used to train Elon Musk's Xai's AI voice model, as well as daily topics on DIY plumbing and travel planning, the Business Insider Show.
Freelancers from the company Scale AI, who sign data, are paid to record conversations with other contractors about Murs' goals and superheroes, so that Xai's voice models don't sound like robots, but real people.
According to an internal dashboard seen by BI, as of April, Scale AI was running at least 10 generation AI projects for Xai. The dashboard lists over 100 AI training projects for Xai and other clients, including Apple, Google Deepmind, and Meta.
The Scale AI job is because companies across the industry are pushing to make bots more conversational and human-like and help them compete for users who may pay for the premium version.
Scale AI and Xai did not respond to requests for comment from Business Insider.
Internal “Project xylophone”
Business Insider has obtained four scale AI documents. We outlined how “Project Xylophone” works in Xai, including a set of two project instructions, a set of instructions for reviewers checking submissions, and a conversation topic guide.
The documentation does not say which Xai models are trained. In late February, Musk announced a beta rollout for the audio mode of Grok, the company's only publicly available AI model.
The Scale AI Project Dashboard is asked to record short conversations, focusing on “audio quality and natural flow ency.” They are especially encouraged to participate if they have experience in voice acting. According to the dashboard, the project aims to “spiritate scripts, great audio performance, and high quality audio.” Scale's dashboards do not have access to contractors who are not sure who their clients are.
In the case of Project Xylophone, gig workers all over the world can choose from hundreds of conversational topics related to ethics, philosophy, business and travel, and record answers in a variety of languages. Split the work between an invitation-only project called a “conversation” that gig workers have a team of three, and a solo “grassland.”
The “conversation” team is asked to set up realistic conversations with each other via Zoom. Contributors will ask questions from a quick spreadsheet that was active earlier this week. This sheet includes over 700 conversation starters on a variety of topics, including apocalyptic survival tactics, planning a trip to India, managing anxiety and panic attacks.
“If you were designing a 'culture' for the first Martian settlement, what kind of Earth's tradition would you definitely want to replicate? Reads one prompt.
BI found that approximately 10% of the conversations in the document reviewed were science fiction-related.
The other questions concern the US political and judicial system, but this set does not include the political issues of hot buttons.
In the “conversation” skill, “good” conversation instructions are explicit. “Recording should sound very natural, like you're having a casual conversation with a friend. This involves being emotional, having a variety of intonations and interrupting each other! Avoid sounding like an interview.”
In the “grassland” arms, solo workers are asked to create natural sounding recordings in their native language, without scripting. Each worker is given a conversation type and subcategories, and is told to flow the conversation in any setting they like, encouraging background noise.
There are numerous subcategories, such as “Socrates' Questions,” “Reflective Storytelling,” “Course of Love Scenarios,” “Hero Brin's Conflict,” or “Co-Puzzle Solving,” which requires different accents, sound effects, or invented linguistic patterns.
Related Stories
Business Insider tells the innovative stories you want to know
Business Insider tells the innovative stories you want to know
Fast and accurate
Three sizes of AI contractors who asked them not to name because they signed a private contract, said the project was assigned to contractors based on their skill set.
The two contractors said payments for the Grassland project were assigned to the contractor based on their location and language expertise, starting at $3 per task and reduced to $1 per task after about a month. The contractor takes 5 minutes to complete each task, each task is one recording.
Once the contractor records the audio file, it is uploaded to the scale AI contributor platform and manually transcribed. Grassland documents ask for filler words such as “uh.”
Large language models require vast amounts of quality data to improve. Reproducing real-world scenarios, such as natural-sounding conversations between people, is one way to generate the right data to feed into those models.
Training Glock
Project Xylophone is an example of a big push that AI companies infuse their personality into AIS and stand out in an increasingly crowded space.
Bi reported last month that Meta ran the project via the project to gig workers who train AI to employ a variety of personas, including “a wise and mysterious wizard” and “a student of ultra-excited music theory.”
Openai's Sam Altman said in late April that the latest GPT-4o was “too annoying and sycophant-y,” prompting the reset to make it more natural.
Xai sells Grok as a politically sharp chatbot compared to what Musk called a “Woke” rival. This is a training method that leans heavily towards right-wing or paradoxical views, and has been reported previously. In addition to outsourced work on Xai, the company will have hundreds of in-house “AI tutors” and will hire thousands more, BI reported in February, showing the human efforts involved in training AI.
Xai also stepped up efforts to control unpredictable aspects of Grok. New recruits are “Red Team” Glocks, stress-testing responses to responses that are unsafe or policy-related, particularly on controversial topics and “NSFW” or “non-immersed” modes.
The safety push follows a high-profile incident that includes the ability to encourage users to encourage GROK to use racial slur, and more recently includes a non-prompted response to South Africa's “white genocide.” Xai has denounced the latter issue with fraudulent quick changes. The company has promised more stringent code reviews and 24-hour monitoring.
Any hints? Contact Shubhangi Goel via AT sgoel@businessinsider.com Or signal on Shuby.85. Please contact Effie Webb by email ewebb@businessinsider.com Or signal at efw.40. Use your personal email address and unprocessed devices. Here's a guide to sharing information safely.
