pith. sign in

arxiv: 2606.18141 · v1 · pith:SOUGCWVRnew · submitted 2026-06-16 · 💻 cs.PL

Behind Python: The Languages That Power AI

Juan P. Licona-Luque , Beatriz A. Bosques-Palomo , Nezih Nieto-Guti\'errez , Gustavo de los R\'ios-Alatorre , Luis A. Mu\~noz-Ubando (Tecnol\'ogico de Monterrey , Monterrey , Mexico) This is my paper

Pith reviewed 2026-06-26 21:43 UTC · model grok-4.3

classification 💻 cs.PL
keywords programming languagesperformance benchmarkingAI algorithmsCC++RustGoJulia
0
0 comments X

The pith

C and C++ tie as fastest for AI algorithms implemented from scratch, with Python 315 times slower.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper implements five core AI algorithms from scratch in six languages to measure performance when libraries are unavailable. Shared random generators and bit-identical outputs ensure that measured differences arise from language features alone rather than from differing computations. Results place C and C++ at the top, Rust nine percent behind, Julia 3.3 times slower than C, Go five times slower, and Python 315 times slower on geometric mean. Memory footprints diverge sharply, with Julia carrying a large fixed overhead, and relative rankings shift across workloads such as k-means versus k-NN. The data supplies workload-specific guidance for selecting an implementation language in custom or constrained AI systems.

Core claim

When the same five algorithms (k-means, k-NN, MLP backpropagation, genetic algorithm, Mamdani inference) are written in Python, C, C++, Rust, Go, and Julia with identical logic and outputs, C and C++ achieve essentially identical runtimes, Rust trails by nine percent geometric mean, Julia runs 3.3 times slower than C, Go runs 5.0 times slower, and Python runs 315 times slower; Julia's JIT imposes a fixed 224 MiB memory cost while the others stay under 6 MiB, and Go's slowdown ranges from 2.6 times to 8.0 times depending on the algorithm.

What carries the argument

The six parallel implementations of each algorithm that share a common pseudo-random generator and produce bit-identical results, isolating language-level performance differences.

If this is right

  • C or C++ should be chosen when raw speed is the primary requirement for custom AI code.
  • Rust delivers near-C performance together with memory safety guarantees.
  • Go's ranking can move an entire tier depending on the specific algorithm, requiring per-workload testing.
  • Julia's fixed high memory footprint makes it unsuitable for memory-constrained targets even when its speed is acceptable.
  • Python remains impractical for direct implementation of these algorithms when performance matters.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Developers facing mixed workloads may need to profile each component separately rather than rely on a single language ranking.
  • The results suggest opportunities for language-specific backends under a common Python interface when both development speed and execution speed are required.
  • Memory-constrained embedded AI systems would likely exclude Julia regardless of its speed tier.

Load-bearing premise

The six versions of each algorithm perform exactly the same computation, so any speed or memory difference is caused only by the language.

What would settle it

Any pair of implementations that produce non-identical outputs on the same inputs, or that contain logically inequivalent steps, would invalidate the claim that observed differences are due solely to language choice.

Figures

Figures reproduced from arXiv: 2606.18141 by Beatriz A. Bosques-Palomo, Gustavo de los R\'ios-Alatorre, Juan P. Licona-Luque, Luis A. Mu\~noz-Ubando (Tecnol\'ogico de Monterrey, Mexico), Monterrey, Nezih Nieto-Guti\'errez.

Figure 1
Figure 1. Figure 1: Mean wall-clock time per benchmark and language (log scale). The three com￾piled systems languages (C, C++, Rust) are visually indistinguishable; Go and Julia occupy a middle band; Python sits two to three orders of magnitude above. Three tiers emerge. C and C++ are statistically indistinguishable: C++ wins four of five benchmarks by under 2% and C wins the fifth (fuzzy) by under 1%. Rust trails this pair … view at source ↗
Figure 2
Figure 2. Figure 2: Mean peak resident set size per benchmark and language (log scale). Julia’s runtime footprint is roughly two orders of magnitude above the systems languages and is essentially constant across workloads. 4.3 Developer-Cost Metrics (M3–M5) [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Normalized five-metric profile of all six languages (each axis scaled so that larger is better). The C and C++ outlines nearly coincide, confirming their tie; the systems languages dominate every runtime axis, while the interpreted and JIT languages recover ground only on developer cost (LOC) [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
read the original abstract

Python dominates AI development, yet the numerical work behind frameworks like PyTorch and NumPy is executed in C, C++, or Rust. When a developer must implement an algorithm without such libraries -- because none exists, the target is resource-constrained, or a new system is being built -- which language should they choose? This paper answers that question empirically. Five algorithms covering data mining (k-means), machine learning (k-NN), neural networks (MLP with backpropagation), computational intelligence (genetic algorithm), and fuzzy systems (Mamdani inference) are implemented from scratch in Python, C, C++, Rust, Go, and Julia. All implementations share a common pseudo-random generator, consume identical inputs, and produce bit-identical outputs, so every measured difference reflects the language rather than the computation. Three performance tiers emerge: C and C++ are effectively tied; Rust trails them by 9% (geometric mean); Julia runs 3.3x slower than C and Go 5.0x; Python sits at 315x. Memory tells a different story -- Julia's JIT runtime carries a fixed ~224 MiB footprint regardless of workload, while C, C++, and Rust stay below 6 MiB. Crucially, rankings are not stable: Go's slowdown swings from 2.6x on k-NN to 8.0x on k-means, showing that workload characteristics can shift a language's position by a full tier. The results provide concrete, per-workload guidance for choosing an implementation language in AI systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript presents an empirical performance comparison of six languages (Python, C, C++, Rust, Go, Julia) on five algorithms from scratch: k-means, k-NN, MLP with backpropagation, genetic algorithm, and Mamdani inference. All implementations share a common PRNG, identical inputs, and produce bit-identical outputs, allowing attribution of measured differences to language rather than computation. Results identify three tiers (C/C++ tied; Rust 9% slower by geometric mean; Julia 3.3x, Go 5.0x, Python 315x slower than C), note Julia's fixed ~224 MiB memory footprint versus <6 MiB for C/C++/Rust, and observe that rankings are workload-dependent (e.g., Go varies from 2.6x to 8.0x slowdown).

Significance. If the implementation equivalence holds, the work supplies concrete, per-workload guidance for developers implementing AI algorithms without library support in resource-constrained or novel systems. The controlled design and observation of unstable rankings are strengths that distinguish this from typical language benchmarks.

major comments (1)
  1. [Abstract (and implied Methods)] The central claim rests on logical equivalence of the six implementations per algorithm beyond language-level differences. The abstract states shared PRNG, identical inputs, and bit-identical outputs, but without the full methods section or code artifacts it is not possible to verify how numerical stability, floating-point semantics, and data-structure choices were aligned across languages with differing default precisions and memory models.
minor comments (2)
  1. Table or figure presenting per-algorithm raw times (not only geometric means) would allow readers to assess the stability claim directly.
  2. Clarify whether the reported memory figures include only heap or also stack/JIT overhead, and whether measurements were taken after warm-up for JIT languages.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the positive evaluation and recommendation of minor revision. The concern regarding verifiability of implementation equivalence is addressed below.

read point-by-point responses

  1. Referee: [Abstract (and implied Methods)] The central claim rests on logical equivalence of the six implementations per algorithm beyond language-level differences. The abstract states shared PRNG, identical inputs, and bit-identical outputs, but without the full methods section or code artifacts it is not possible to verify how numerical stability, floating-point semantics, and data-structure choices were aligned across languages with differing default precisions and memory models.

    Authors: We agree that greater detail on alignment would improve verifiability. Section 3 of the manuscript already specifies the common PRNG, identical inputs, and bit-identical output requirement, which directly constrains floating-point and numerical behavior because any divergence in semantics or stability would violate the bit-identity condition. Data structures were aligned by using equivalent array and record representations in each language (e.g., contiguous 64-bit float arrays for vectors and matrices). To make this fully transparent, the revised manuscript will add an explicit subsection in Methods describing the precision and structure choices per algorithm, and we will supply a public code repository link as a code artifact so readers can inspect the implementations directly. revision: yes

Circularity Check

0 steps flagged

No significant circularity; pure empirical measurement study

full rationale

The paper reports direct runtime and memory measurements from six language implementations of five algorithms. All differences are attributed to language after enforcing shared PRNG, identical inputs, and bit-identical outputs. No equations, fitted parameters, predictions, or self-citations appear in the derivation chain; the reported ratios and tiers follow immediately from the timed executions. The study is therefore self-contained against external benchmarks with no reduction of results to prior quantities by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that language-level differences alone explain the measured performance gaps once inputs and outputs are controlled; no free parameters or invented entities are introduced.

axioms (1)
  • domain assumption Implementations in each language are logically equivalent such that measured differences reflect only language characteristics.
    Invoked via the shared PRNG, identical inputs, and bit-identical outputs condition stated in the abstract.

pith-pipeline@v0.9.1-grok · 5856 in / 1415 out tokens · 31486 ms · 2026-06-26T21:43:35.960950+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 11 canonical work pages

  1. [1]

    Shah, and Alan Edelman

    Bezanson, J., Edelman, A., Karpinski, S., Shah, V.B.: Julia: a fresh approach to numerical computing. SIAM Review59(1), 65–98 (2017). https://doi.org/10.1137/141000671

    work page doi:10.1137/141000671 2017

  2. [2]

    Array programming with NumPy,

    Harris, C.R., Millman, K.J., van der Walt, S.J., et al.: Array programming with NumPy. Nature585, 357–362 (2020). https://doi.org/10.1038/s41586-020-2649-2

    work page doi:10.1038/s41586-020-2649-2 2020

  3. [3]

    In: Advances in Neural Information Processing Systems 32 (NeurIPS), pp

    Paszke, A., Gross, S., Massa, F., et al.: PyTorch: an imperative style, high- performance deep learning library. In: Advances in Neural Information Processing Systems 32 (NeurIPS), pp. 8024–8035 (2019)

    2019

  4. [4]

    arXiv preprint arXiv:2211.02740 (2022)

    Churavy, V., Godoy, W.F., Bauer, C., et al.: Bridging HPC communities through the Julia programming language. arXiv preprint arXiv:2211.02740 (2022)

    arXiv 2022

  5. [5]

    Computer 33(10), 23–29 (2000)

    Prechelt, L.: An empirical comparison of seven programming languages. Computer 33(10), 23–29 (2000). https://doi.org/10.1109/2.876288

    work page doi:10.1109/2.876288 2000

  6. [6]

    https://benchmarksgame- team.pages.debian.net/benchmarksgame/, last accessed 2026/06/09

    The Computer Language Benchmarks Game. https://benchmarksgame- team.pages.debian.net/benchmarksgame/, last accessed 2026/06/09

    2026

  7. [7]

    10th ACM SIGPLAN Int

    Pereira, R., Couto, M., Ribeiro, F., Rua, R., Cunha, J., Fernandes, J.P., Saraiva, J.: Energy efficiency across programming languages: how do energy, time, and memory relate? In: Proc. 10th ACM SIGPLAN Int. Conf. on Software Language Engineer- ing (SLE), pp. 256–267. ACM (2017). https://doi.org/10.1145/3136014.3136031

    work page doi:10.1145/3136014.3136031 2017

  8. [8]

    arXiv preprint arXiv:2206.05503 (2022)

    Bugden, W., Alahmar, A.: Rust: the programming language for safety and perfor- mance. arXiv preprint arXiv:2206.05503 (2022)

    arXiv 2022

  9. [9]

    In: Proc

    Nanz, S., Furia, C.A.: A comparative study of programming languages in Rosetta Code. In: Proc. 37th Int. Conf. on Software Engineering (ICSE), pp. 778–788. IEEE (2015). https://doi.org/10.1109/ICSE.2015.90

    work page doi:10.1109/icse.2015.90 2015

  10. [10]

    Lin, W.-C., McIntosh-Smith, S.: Comparing Julia to performance-portable par- allel programming models for HPC. In: Int. Workshop on Performance, Porta- bility and Productivity in HPC (P3HPC), pp. 94–105. IEEE/ACM (2021). https://doi.org/10.1109/P3HPC54578.2021.00010

    work page doi:10.1109/p3hpc54578.2021.00010 2021

  11. [11]

    In: Proc

    MacQueen, J.: Some methods for classification and analysis of multivariate obser- vations. In: Proc. 5th Berkeley Symp. on Mathematical Statistics and Probability, vol. 1, pp. 281–297. Univ. of California Press (1967)

    1967

  12. [12]

    Least squares quantization in PCM,

    Lloyd, S.P.: Least squares quantization in PCM. IEEE Trans. Inf. Theory28(2), 129–137 (1982). https://doi.org/10.1109/TIT.1982.1056489

    work page doi:10.1109/tit.1982.1056489 1982

  13. [13]

    Nearest neigh- bor pattern classification.IEEE Transactions on Information Theory, 13(1):21–27, 1967

    Cover, T.M., Hart, P.E.: Nearest neighbor pattern classification. IEEE Trans. Inf. Theory13(1), 21–27 (1967). https://doi.org/10.1109/TIT.1967.1053964

    work page doi:10.1109/tit.1967.1053964 1967

  14. [14]

    Rumelhart, Geoffrey E

    Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning represen- tations by back-propagating errors. Nature323, 533–536 (1986). https://doi.org/10.1038/323533a0

    work page doi:10.1038/323533a0 1986

  15. [15]

    MIT Press, Cam- bridge, MA (1992)

    Holland, J.H.: Adaptation in Natural and Artificial Systems. MIT Press, Cam- bridge, MA (1992)

    1992

  16. [16]

    Mamdani, E.H., Assilian, S.: An experiment in linguistic synthesis with a fuzzy logic controller. Int. J. Man-Machine Studies7(1), 1–13 (1975). https://doi.org/10.1016/S0020-7373(75)80002-2 Behind Python: The Languages That Power AI 17

    work page doi:10.1016/s0020-7373(75)80002-2 1975

  17. [17]

    Kamath, A

    IEEE: IEEE Standard for Floating-Point Arithmetic. IEEE Std 754-2019 (2019). https://doi.org/10.1109/IEEESTD.2019.8766229

    work page doi:10.1109/ieeestd.2019.8766229 2019