AI researchers spend their days debugging 'incomprehensible' training code, engineer says

gabrielchua · x · 2026-09-07

In a widely shared take, @thsottiaux contrasts AI research roles with engineering and product: researchers spend their days inside nearly incomprehensible Python codebases, requiring familiarity with a dozen papers and deep understanding of quirky systems that subtly affect training dynamics. Their work is mostly debugging failures without knowing whether the math or the system is wrong — while the engineering/product picture has looked very different for months.

Original post →

More from AGI Musings

AGI Musings channel →