On January 6, according to The Inside of Business, the data label company Invisible Technologies CEOs Matt Fitzpatrick IN THE PODCAST 20VC, IT'S ABOUT GENERATION AI TrainingA core misunderstanding is gradually being corrected:
The role of human beings in the data production and feedback chain will not be replaced by synthetic data in the coming decades。
Fitzpatrick points out that there is a long-standing view within the industry that, as synthetic data capacity increasesAI ModelsReliance on human feedback will be removed within two to three years。
IN HIS VIEW, HOWEVER, THIS JUDGEMENT IS NOT JUSTIFIED BY THE RATIONALE. THE REASON FOR THIS IS THAT IN THE REAL WORLD THERE ARE A LARGE NUMBER OF DIVERSE AND COMPLEX TASKS, AND AI UNDERSTANDING AT THE LINGUISTIC, CULTURAL AND PROFESSIONAL LEVELS STILL DEPENDS ON HUMAN CALIBRATION。
HE CITED THE LEGAL PROFESSION AS AN EXAMPLE WHERE THERE WAS A GREAT DEAL OF NON-PUBLIC INFORMATION, WHERE MODELS WERE NOT DIRECTLY ACCESSIBLE AND HAD TO RELY ON MANUAL PROCESSING AND LABELLING. IN ADDITION, THE TRAINING OF GENERATING AI STILL REQUIRES HUMANS TO SCREEN, SEQUENCE AND CORRECT MODEL OUTPUTS TO ENSURE QUALITY AND SAFETY。
Invisible Technologies completed $100 million in financing last September, valued at $2 billion. Companies compete with Scale AI, Surge AI and other data labeling enterprises to provide high-quality training data for technology giants。
As modelling capacity increases, the industry has shifted from relying on generic labelers to recruiting specialists in mathematics, science and other fields to meet higher-precision data requirements。
Several peer companies, CEOs, have expressed similar views. Mercor CEO Brendan Foody said last September that data quality remains at the core of business, and that this depends on "good people who are well treated."。
Handshanke CEO Garrett Lord noted that as the model absorbed almost all of the Internet's open language, industry's requirements for labelers were shifting from “general talent” to “experts”。