The base model is the freshly pretrained network, which only continues text; the assistant is that same model tuned to respond as a helpful character in a conversation.
A base model is what comes out of pretraining: a network that has learned to predict the
next token over a huge corpus, and nothing else. It is also called pretrained or a
foundation model, a term popularized by Bommasani et al. (2021). The assistant
(an instruct or chat model) is the same network after instruction tuning and
preference optimization (Alammar & Grootendorst, 2024, pp. 355–356). The InstructGPT authors described
the base GPT-3 models as having "poorly characterized behavior" (Ouyang et al., 2022, sec. 3.5).
The difference is not one of architecture but of conduct. The base model continues any
document: a forum, a poem, a form. The assistant continues only one, the dialogue with a
helpful character. For janus, instruction following is merely "an exploitable subset" of the
patterns the model inherited from human language (janus, 2022) (see simulators).
Zhou et al. (2023) took this to the extreme with the superficial alignment hypothesis:
knowledge is acquired almost entirely in pretraining, and 1,000 well-chosen examples suffice
to teach the format. Lu et al. (2026) concur: the direction that marks "being the Assistant" in
assistants already exists in base models, where it promotes helpful human archetypes such as
consultants and coaches (2026, sec. 3.2.2).
Something is lost. The pretrained GPT-4 was well calibrated (its confidence matched its
probability of being right), and post-training made that calibration worse
(OpenAI, 2023, sec. 5). Variety is lost too (mode collapse).
Shanahan (2024) downplays the distinction: the tuned model still models the distribution of
words, only skewed toward what users and raters would approve (2024, p. 78). He
even imagines that, with endless raters and geological time, a base model could be obtained
entirely through RLHF; what counts is the product, not the process
(2024, pp. 78–79). If base and assistant differ in degree rather than in kind, who
is speaking when the assistant speaks: the character, the corpus, or the raters?
Sources
- Alammar, J., & Grootendorst, M. (2024). Hands-on large language models: Language understanding and generation. O’Reilly Media. §
- Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., Brynjolfsson, E., Buch, S., Card, D., Castellon, R., Chatterji, N., Chen, A., Creel, K., Davis, J. Q., Demszky, D., . . . Liang, P. (2021). On the opportunities and risks of foundation models (arXiv:2108.07258). arXiv. https://doi.org/10.48550/arXiv.2108.07258 §
- janus. (2022, September 2). Simulators. LessWrong. https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators §
- Lu, C., Gallagher, J., Michala, J., Fish, K., & Lindsey, J. (2026). The Assistant Axis: Situating and stabilizing the default persona of language models (arXiv:2601.10387). arXiv. https://doi.org/10.48550/arXiv.2601.10387 §
- OpenAI. (2023). GPT-4 technical report (arXiv:2303.08774). arXiv. https://doi.org/10.48550/arXiv.2303.08774 §
- Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P. F., Leike, J., & Lowe, R. (2022). Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems (Vol. 35, pp. 27730–27744). Curran Associates. https://doi.org/10.52202/068431-2011 §
- Shanahan, M. (2024). Talking about large language models. Communications of the ACM, 67(2), 68–79. https://doi.org/10.1145/3624724 §
- Zhou, C., Liu, P., Xu, P., Iyer, S., Sun, J., Mao, Y., Ma, X., Efrat, A., Yu, P., Yu, L., Zhang, S., Ghosh, G., Lewis, M., Zettlemoyer, L., & Levy, O. (2023). LIMA: Less is more for alignment. In Advances in Neural Information Processing Systems (Vol. 36, pp. 55006–55021). Curran Associates. https://doi.org/10.52202/075280-2400 §
