I chose Labellerr as the labeling/curation tool because it promises high‑quality labelled datasets essential for effective model fine‑tuning.
I decided not to go live because “Can we go live with this? --> No”.
I noted that the secret to training a GPT‑3‑scale LLM in just 30 minutes on a single machine is using a high‑quality labeled dataset (https://lnkd.in/gsi6czAa).
The high quality labeled dataset enables rapid training.
Training time reported: 30 minutes.
Hardware used: 1 machine.
I observed that more than 99 % of businesses cannot build foundation models themselves, which drives a thousandfold increase in AI adoption across businesses.
Current generative‑AI reliability is about 50 % and the target reliability is 90 % according to Mustafa Suleyman.
He stresses that closing this gap mainly requires "data curation and labeling" with humans in the loop, even if the process is imperfect.
I observed on 2025-09-19 that Meta’s AI models need more real human quality data and a better WiFi, as believed by #mark. 😀