Travis Oliphant: NumPy, SciPy, Anaconda, Python & Scientific Programming | Lex Fridman Podcast #224

Travis Oliphant tells the story of how a poor graduate student scratching his own itch ended up creating the libraries that power modern scientific Python. He walks through writing SciPy in 1998-2001, unifying the fractured array community by writing NumPy in four months, and later founding Anaconda and conda to solve Python's packaging problem. Throughout, he wrestles with a question that has driven him for 25 years: how to reconcile the cooperative ethos of open source with the need to make a living, leading him deep into economics and entrepreneurship. He closes with hard-won advice on programming, hiring, marketing, and building rather than destroying.

• Oliphant wrote the first version of NumPy in four months only because two students dropped his MRI class, leaving him with empty teaching time.
  00:54:43
• He took on unifying the split array community partly out of duty, knowing his electrical-engineering department would not appreciate the open-source work for tenure.
  00:55:44
• His 2006-07 tenure bid split the department all the way to the university president; they told him to reapply in two years, and he left academia instead.
  01:07:36
• To fund his lab he self-published 'Guide to NumPy' with a market-determined price, selling 30,000 copies and using the money to pay his grad student.
  01:11:44
• Microsoft offered to buy Anaconda in 2014, but the offer came from two levels below Satya Nadella while Microsoft was doubling down on R, so it went nowhere.
  01:27:46
• Oliphant calls TensorFlow's original Python API so bad he publicly told the team to never show it at a PyData conference again.
  01:24:39
• His business relationship with M-thot co-founder Eric Jones ended badly over Oliphant starting Anaconda, and the two are no longer friends.
  01:15:51
• He admits NumPy's array community split exists largely because he could never raise enough money to work on NumPy full-time.
  00:40:13

• Oliphant believes language shapes thought, citing how Russian's history of suffering is encoded in its sad, hopeful songs and untranslatable poetry.
  00:05:06
• APL pioneered array-based programming in the 1960s but needed an entirely new keyboard to type its dense glyphs, which doomed adoption.
  00:18:30
• His book's deal was self-imposed: once he earned $250,000 or five years passed, it would become free; he stopped early and opened it up.
  01:13:18
• He considers Python 3.3 the real Python 3.0, arguing earlier 3.x releases changed too much without adding enough worth switching for.
  01:38:15
• Andrej Karpathy got a real speedup by switching from np.sqrt to math.sqrt, and plain Python x**0.5 can be even faster on scalars.
  01:41:21
• Anaconda built the very first CUDA JIT compiler in 2013, letting a Python function run on GPUs with up to 1000x speedups.
  02:04:34
• Oliphant notes the U.S. spent over six trillion dollars in the Middle East after 9/11, contrasting it with how little funds the programmers who build the world's tools.
  01:54:43
• He once deleted a tweet saying Emacs is better than Vim after getting overwhelmed by the backlash.
  02:53:40
• Oliphant pitches a repeatable model: every venture fund should have an associated open-source research lab funded by its carried interest.
  02:30:22
• Oliphant's recurring hope: 'nations that code together don't go to war together.'
  01:54:43