Francois Chollet is leaving Google

Original Article ↗ Hacker News Discussion ↗

Departure and career plans

  • The Keras creator is leaving Google to start a new company with a friend; no move to another major lab.
  • They remain US‑based for now, but are positive about the AI scene in Paris.
  • Some see Google’s blog farewell as unusual and possibly a “soft launch” for the new venture.

Keras, TensorFlow, PyTorch, JAX

  • Many recall early Keras (on Theano/TF1) as transformative: easy, Pythonic, and critical to deep learning’s “takeoff,” especially vs Theano, Caffe, Torch7.
  • Common criticism: the abstraction was “too easy for basics, too hard for custom work” (custom losses, RNN variants, bespoke training loops), pushing researchers toward raw TensorFlow and then PyTorch.
  • PyTorch is widely seen as the current default: better flexibility, LLM tooling, multi‑GPU support, performance, ecosystem, and community momentum.
  • JAX is praised as powerful yet under‑appreciated; there are claims Google uses it heavily internally and that TensorFlow is losing ground.

Multi‑backend Keras and production use

  • Several see the 2018–2019 folding of Keras into TensorFlow as the moment Keras “died,” and believe it accelerated PyTorch adoption.
  • The Keras author clarifies they did not decide that merger; it was a higher‑level TensorFlow leadership decision and, in hindsight, likely a mistake.
  • Keras is now standalone and multi‑backend again (TF, JAX, PyTorch), with explicit emphasis on backward compatibility and “progressive disclosure of complexity.”
  • Users report Keras still running reliably in production (often since ~2018–2019) for vision and recommendation workloads; others report heavy technical debt and migration to PyTorch.
  • The author lists many large companies using Keras; skeptics counter that this reflects legacy rather than current research leadership.

ARC‑AGI benchmark and AI progress

  • Extensive debate around the ARC benchmark and a recent $1M prize:
    • Some describe strong results (e.g., systems using GPT‑4‑generated synthetic tasks) as “gaming” a benchmark that was supposed to resist brute‑force and big‑data memorization.
    • Others argue this is legitimate progress in problem‑solving and test‑time fine‑tuning, not a hack.
  • Concerns are raised that human baselines measured via Mechanical Turk underestimate motivated human performance.
  • The organizer plans ARC 2 with tasks that are harder to brute‑force yet similar human difficulty, and sees ARC as a high‑leverage path toward AGI‑relevant research.
  • They expect ARC to be solved within a few years, see that solution as a stepping stone (not AGI itself), and maintain skepticism about an “intelligence explosion,” citing diminishing returns and the need to separate intelligence from autonomy.

Google culture, hierarchy, and AI startups

  • Multiple commenters portray Google as comfortable but bureaucratic, where higher‑level decisions (e.g., around TensorFlow/Keras) can override project creators and dampen ambition.
  • Others defend large hierarchies as necessary for 100k+‑employee firms and note that senior engineering levels typically have broad systems experience.
  • There is broader discussion of AI startups: some claim top researchers can easily raise ~$100M; others argue that’s insufficient to sustain a competitive foundation‑model company, so future startups will likely focus on specialization and post‑training rather than training new general LLMs from scratch.