Products · Product
TensorFlow
Raised as a usability problem by 81 people on Hacker News over 11 years (2015–2025), in 66 threads.
No issue was raised by at least 3 people in 3 threads this year.
No issue was raised by at least 3 people in 3 threads in the last 5 years.
Every issue, all time, most people first
Building it makes the whole computer unusableUnreviewed
20152026- TensorFlow is unusable on the author's AMD GPU setup2024-11-29
- Original TensorFlow was unintuitive and slow to program2022-10-21
- Building TensorFlow is so slow the whole computer becomes unusable2021-01-16
Unreviewed issues were grouped by machine: Claude Haiku labelled each comment as a usability complaint, named what it's about and summed up the problem in a line, and a small model grouped similar problems. Nobody has checked them by hand. Expect mistakes.
The newest 50 complaints
- Steep learning curve and broken docs after TF2 changed APIs2025-11-07
- Can't print debug tensor dimensions, making it hard for a beginner to use2025-11-07
- Low-level API is powerful but incomprehensible, poorly documented, and ever changing2025-11-07
- Hard for beginners to use, with sharp edges throughout2025-11-07
- Graph building and session running too complex, with global state and scoping confusion2025-10-25
- Docs gave no indication of required inputs, forcing trial-and-error to learn argument types2025-09-05
- Lazy evaluation makes debugging very difficult, so people chose PyTorch instead2025-05-24
- Fails on Python 3.12 with non-obvious error messages and no documentation2024-11-29
- TensorFlow is unusable on the author's AMD GPU setup2024-11-29
- Searching for a linear layer lands on unrelated monotonicity docs; the right layer is named Dense2024-10-01
- Symbolic model is less intuitive to work with and its errors are harder to debug2024-08-16
- Build error messages and debug output were inscrutable, impossible to understand or solve2024-06-30
- Grossly complex for common tasks, with many dependencies, making it a ridiculous pain to use2023-05-26
- Manually building computation graphs in Python in its first version was infuriatingly difficult to use2023-04-26
- Backwards-incompatible TF2 changes left outdated docs and posts confusing newcomers2023-04-26
- Feels clunky and slow to make progress compared with PyTorch2023-01-16
- Became too sprawling to reason about after trying to serve every audience2022-10-21
- Early releases were impossible to debug and tutorials were broken for beginners2022-10-21
- Original TensorFlow was unintuitive and slow to program2022-10-21
- Developer API is awful to use and hard to work with2022-07-16
- Install instructions from five pages never consistently work, leaving the user stuck2022-03-09
- Some operations take a long time with no clear reason why2021-12-14
- Google's large, confusing ecosystem of overlapping tools is hard for newcomers to navigate2021-12-14
- API behaviour can't be intuited, so users must rely on heavy documentation navigation2021-12-14
- Too many inconsistent ways to save a model; clunky APIs that don't work together2021-12-14
- Graph-mode TensorFlow makes it hard to debug or inspect tensor values interactively2021-11-06
- Pip install fails with an obscure numpy error, no clear cause2021-06-09
- Building TensorFlow is so slow the whole computer becomes unusable2021-01-16
- Setup took five days; a library file had moved and docs were wrong2020-12-31
- Hard to understand, with unnecessary abstractions; user spent 8 hours unable to extract gradients2020-12-16
- No way to figure out why something is slow or uses huge memory; a black box2020-12-15
- API was cumbersome and hard to work with compared to PyTorch2020-10-26
- TensorFlow 2 runs noticeably slower than TensorFlow 1 for the same work2020-10-17
- Linking to the TensorFlow C++ library is painful and lacks SavedModel support2020-10-10
- Text gives impression FP16 is slower than FP32, confusing without the table2020-09-29
- Opaque high-level libraries hide what code does, making subtle bugs hard to debug2020-09-20
- Control flow and array indexing require awkward, hard-to-use patterns, frustrating users2020-09-20
- Decorator-based module placement makes functions hard to locate, breaks IDE autocompletion, and hides global state2020-08-23
- Heavy abstractions and dependency injection make the framework hard to understand2020-08-05
- Had to trawl the repo and Python wrappers to learn how to load SavedModels2020-07-17
- API design is unintuitive, so it is hard to use correctly2020-07-17
- Obscure bugs and error messages that are hard to interpret2020-06-18
- Duplicate APIs and growing complexity make the library confusing and opaque2020-02-17
- Graph fails at execution with an error stack trace that does not reveal the cause2020-02-17
- Hard-to-follow error messages, unintuitive APIs, multiple ways to do the same thing, difficult debugging2020-02-17
- Developer experience is seen as worse than PyTorch's2020-02-17
- Developer finds the TensorFlow Haskell bindings difficult to work with2020-02-13
- Hard to debug because sess.run() hides what the network does2020-01-31
- Tensor shape wrangling and indexing ops like tile, gather, scatter_nd are a real headache2020-01-31
- Hard to use from languages other than Python2019-12-20