top of page
Search

How an AI Engineer Builds the Future Model

In November 2024, a consortium of Indonesian corporations and academic institutions unveiled an initiative on a stage in Jakarta that would have previously seemed like a concept from science fiction. They commenced the development of a substantial language model tailored to comprehend Bahasa Indonesia and various regional dialects, all situated within Indonesia.

By the middle of 2025, that model expanded to encompass 70 billion parameters and was responding to inquiries for millions of users via a mobile application. The individuals responsible for this achievement included more than just executives and ministers.

At the core of the project resided a reserved and progressively in-demand specialist: the AI engineer. The term "the future model" can signify various interpretations.

It may signify the forthcoming frontier system that expands the boundaries of machine capabilities. It may refer to the particular model a corporation will depend on in the upcoming year to operate its offerings. For an AI engineer, it signifies both options, and the journey to either one requires the same rigorous discipline.

Comprehending the mechanics of that craft is the optimal approach to grasp why this position has emerged as one of the most esteemed occupations in technology and why a student in Jakarta is particularly well-positioned to pursue it.


How does the nation's AI engineer build the future model?

An AI engineer builds the future model using a systematic pipeline that includes defining a specific problem, collecting and refining extensive datasets, selecting a model architecture like the transformer, training the model on robust hardware, aligning it with human objectives, and then assessing, deploying, and continuously improving it in a production environment. The project resides at the intersection of software engineering, mathematics, and product development.

In contrast to a pure researcher, who investigates novel possibilities, an AI engineer transforms those possibilities into systems that actual users can use safely and dependably. Constructing a future model is consequently less an isolated stroke of brilliance and more a sequence of deliberate choices, each quantifiable and each subject to enhancement.


The AI engineer, and how the role differs from a researcher

Clarity regarding the task is beneficial. An AI engineer conceptualizes, constructs, implements, and expands systems driven by machine learning and, progressively, by extensive language models. A scholar typically emphasizes theoretical progress and innovative techniques.

An AI engineer utilizes those techniques, in conjunction with available tools and pretrained models, to address specific business challenges. The differentiation is significant as the majority of the value in the present AI surge is derived from engineering rather than mere discovery: integrating models with data pipelines, encapsulating them in dependable software, and ensuring their functionality when actual users engage.

The fundamental toolkit embodies these principles. It typically encompasses robust programming in Python, deep learning frameworks like PyTorch and TensorFlow, operational tools for production systems such as Docker and Kubernetes, and an expanding repertoire of skills related to large language models, retrieval, and assessment.


Step one: framing the problem and gathering the data

Each forthcoming model originates not from programming but from an inquiry. What functions should this system perform, and how can one ascertain its success?

Once that is established, the engineer focuses on data, as a model's efficacy is solely dependent on its learning sources. This phase is unremarkable yet pivotal.

The engineer gathers textual, visual, auditory, or organized data pertinent to the assignment, subsequently refining that unprocessed material, as actual data is rife with redundancies, inaccuracies, and interference. Frequently, the data requires annotation to provide the model with accurate examples for replication.

The magnitude at play is difficult to exaggerate. Contemporary language models acquire knowledge from datasets comprising trillions of words, and the caliber of this compilation influences all subsequent outcomes.

Acquiring accurate data is often where a seasoned engineer dedicates most of their efforts.

Step two: choosing an architecture

Armed with data, the engineer determines the construction of the model. The prevailing solution for language and an expanding array of tasks since 2017 has been the transformer, which was presented by Google researchers in the paper "Attention Is All You Need" (Vaswani et al., 2017). The core concept, known as "self-attention," enables a model to evaluate the relationships between all components of an input simultaneously, rather than processing them in a linear fashion.

This resulted in models that were significantly more parallel, considerably quicker to train, and markedly superior at maintaining extended context.

Nearly every significant model from recent years, including the bidirectional comprehension model BERT (Devlin et al., 2018) and the frameworks powering widely used chat assistants, as well as Indonesia's national model, originates from that architecture.

An AI engineer typically adapts existing architectures instead of building entirely new ones. The expertise involves choosing the appropriate option, sizing it judiciously, and tailoring it to the specific issue and available funds.


Step three: training at scale, and the mathematics of bigger

Training is the phase in which the model acquires knowledge and where expenses escalate significantly. The engineer inputs the data into the selected architecture over numerous iterations, modifying billions of internal parameters to enhance the model's predictions.

The training process operates on dedicated hardware, typically arrays of graphics processing units, and the computational demands can be substantial. Training a cutting-edge scale model may require thousands of petaflop days of computational power and incur costs in the millions of dollars (Ouyang et al., 2022).

Due to the high costs involved, engineers depend on established principles for effectively allocating a computing budget. Preliminary research indicated that increasing model size was the primary mechanism (Kaplan et al., 2020).

In 2022, a research study conducted by DeepMind altered the prevailing consensus. The authors demonstrated through the training of over 400 models that many contemporary large models were significantly undertrained, indicating that for a constant computational budget, model size and training data should increase in tandem, approximately 20 words of data for each parameter (Hoffmann et al., 2022).

The Chinchilla model, comprising 70 billion parameters and trained on 1.4 trillion words, surpassed competitors that were significantly larger, with subsequent independent research reaffirming this conclusion (Besiroglu et al., 2024).

For an AI engineer, these scaling principles serve as pragmatic direction: they transform an abstract aspiration to create something formidable into specific decisions regarding the model's size and the volume of data to utilize.

Step four: teaching the model to be helpful

A newly trained language model may accurately forecast the subsequent word, yet its performance does not equate to being beneficial, truthful, or secure. Transforming a rudimentary model to align with human intention constitutes a distinct phase, and it represents one of the paramount responsibilities of an AI engineer.

The approach that revolutionized the domain is reinforcement learning derived from human feedback, as detailed in the InstructGPT study by OpenAI (Ouyang et al., 2022). It comprises three phases. Initially, human authors present the model with exemplary responses, and the model is honed to replicate them.

Secondly, individuals evaluate various model responses from most favorable to least favorable, and these evaluations inform a distinct reward model that encapsulates human preference. Third, the primary model is calibrated to perform favorably in relation to that reward model. The outcome was remarkable: a significantly smaller aligned model, approximately 1.3 billion parameters, yielded responses favored by human evaluators compared to a base model nearly one hundred times its size (OpenAI, 2022).

Two teachings resonated within the field. Individuals remain engaged continuously, and meticulous coordination can yield greater benefits than merely constructing a larger entity.

An engineer's discernment regarding individuals, language, and culture is equally crucial as their proficiency in mathematics.

Step five: evaluation, deployment, and the loop that never ends

A model that appears remarkable during a demonstration may still falter discreetly in practical applications; thus, assessment is imperative. The engineer assesses the model using reserved examples with defined metrics, constructs evaluation frameworks to identify deficiencies, and subsequently proceeds to deployment.

Deploying a model into production and ensuring its reliability constitutes a distinct discipline, commonly referred to as machine learning operations, which integrates data science with infrastructure. The efforts persist even post-launch.

Linguistic patterns and user conduct are continually evolving, necessitating the oversight, retraining, and enhancement of models. Numerous teams additionally link their model to new, proprietary data via a method referred to as retrieval augmented generation, enabling it to respond using a company's specific documents instead of relying solely on memory.

Proficient AI engineers excel at assessing quality and maintaining system reliability, which is crucial since a team that cannot evaluate a model cannot securely deploy it.


The Indonesian future model: sovereign AI and Sahabat AI

The concept of "constructing the future model" is most evident and relevant in Indonesia's initiative to create one.

Sahabat AI, translating to "close friend," is a collection of open-source large language models developed by the telecommunications firm Indosat Ooredoo Hutchison in collaboration with the technology consortium GoTo, with assistance from partners such as NVIDIA and AI Singapore (GoTo, 2024).

Initially introduced in late 2024 and enhanced to 70 billion parameters by mid-2025, it comprehends Bahasa Indonesia as well as Javanese, Sundanese, Balinese, and Batak (GoTo, 2025). Two design decisions render it an exemplary case for budding engineers.

Initially, it was optimized for efficiency to operate on merely two H100 graphics processing units, significantly fewer than what many comparable models necessitate, thereby reducing expenses and enhancing accessibility (Dealstreet Asia, 2025).

Secondly, it is developed and operated on Indonesian infrastructure, a national cloud referred to as GPU Merdeka, ensuring that data remains within the country's boundaries (NVIDIA, 2024). The construction was based on a nationwide partnership involving the University of Indonesia, Gadjah Mada University, and Bandung Institute of Technology, and by the end of 2025, the models had been downloaded by developers tens of thousands of times (Sea Lion, 2025).

This development is significant as Indonesia ranks as the second most linguistically diverse nation globally, boasting over 700 spoken languages across approximately 17,000 islands (Aji et al., 2022).

Global models predominantly trained in English inadequately support other languages. A forthcoming model, created by Indonesians for Indonesians, bridges this divide, safeguards linguistic heritage, and promotes what the nation refers to as digital sovereignty, which is integral to the broader Golden Indonesia 2045 initiative (Fortune, 2025).

For an engineer, it serves as a tangible exhibition that premier model construction is no longer restricted to a select few laboratories overseas.

Why this is one of the most valuable careers in technology

The economic principles align with the level of difficulty. Projections for the average remuneration of AI engineers significantly exceed that of general software developers, with one substantial dataset indicating an average of approximately USD 145,070 annually, while senior professionals earn notably higher (Glassdoor, 2026; Coursera, 2026).

The United States Bureau of Labor Statistics indicates that computer and information research scientists, encompassing a significant portion of this field, received a median salary of approximately USD 140,910 in 2024, with employment anticipated to increase by around 20 percent over the next ten years, significantly outpacing the average profession (U.S. Bureau of Labor Statistics, 2025).

The demand for AI positions has exceeded supply to such an extent that job listings for these roles skyrocketed in 2025 and 2026, with positions requiring AI expertise offering a salary premium of approximately 25% compared to similar jobs (365 Data Science, 2026).

The potential reward is significant: a widely cited projection suggests that AI could contribute up to USD 15.7 trillion to the global economy by 2030 (Coursera, 2026). Indonesia is well placed to take advantage of this opportunity.

The nation boasts approximately 229 million internet users, a digital economy projected to reach about USD 180 billion in worth by 2030, and generative AI utilization among knowledge workers estimated at nearly 92 percent (Digital in Asia, 2026; Business Indonesia, 2025). The national AI market is anticipated to reach approximately USD 10.88 billion by the year 2030 (Introl, 2025).

A nation developing its own models requires engineers capable of constructing them, and that necessity is just commencing.

Building the future model from Jakarta

For Indonesian students, the reassuring reality is that model construction yields greater benefits than mere coding, highlighting the significant advantage of a school grounded in design and business.

At Raffles Jakarta, located on Jalan M.H. Thamrin in the heart of the capital, the artificial intelligence curriculum provides students with foundational knowledge in machine learning and model creation, complemented by the relevant disciplines that enhance a model's significance.


Identifying the appropriate issue and linking a model to economic worth relies on the business administration and one-year English language MBA programs.

Instructing a model to assist via human feedback cultivates an appreciation of human behavior and language that the psychology curriculum fosters. Transforming a proficient model into an enjoyable user experience requires applying interface techniques learned in digital media design and visual communication design.

Studying in Jakarta immerses students in the local market, characterized by its diverse languages and swift digital advancement, where the nation's future paradigms will be developed.

The engineer responsible for constructing those models will not have to journey far to locate them. The task is already present.

Frequently Asked Questions

What does an AI engineer actually do? An AI engineer designs, builds, deploys, and scales systems powered by machine learning and large language models. The work runs across a full pipeline: framing the problem, collecting and cleaning data, choosing and training a model, aligning it to human intent, and then evaluating, deploying, and continuously improving it in production. The focus is on building reliable, useful systems rather than on pure research.

How is an AI engineer different from a machine learning researcher? An AI engineer is more applied, while a researcher is more exploratory. A researcher explores new methods and often publishes papers. An AI engineer takes existing methods, tools, and pretrained models and turns them into working products that solve concrete business problems. Both are valuable, but the engineer's success depends on systems that ship and keep running.

What are the main steps to build a large AI model? The main steps are framing the problem, gathering and cleaning data, choosing an architecture such as the transformer, training the model on powerful hardware, aligning it to human intent through methods like reinforcement learning from human feedback, and then evaluating, deploying, and monitoring it. Each step is measurable, and the process repeats as the model is improved over time.

What are scaling laws in AI? Scaling laws describe how a model's performance improves as its size, its training data, and its compute grow. A landmark 2022 study found that for a fixed compute budget, model size and training data should be increased together, at roughly 20 words of data per parameter, and that many large models had been undertrained (Hoffmann et al., 2022). Engineers use these laws to decide how big a model to train and how much data to use.

Does Indonesia build its own AI models? Yes. Indonesia has built Sahabat AI, a family of open-source large language models led by Indosat Ooredoo Hutchison and GoTo, which was upgraded to 70 billion parameters in 2025 and understands Bahasa Indonesia and regional languages such as Javanese, Sundanese, Balinese, and Batak. It is trained and hosted on Indonesian infrastructure to support national digital sovereignty (GoTo, 2025).

Where can I study to become an AI engineer in Jakarta? You can study artificial intelligence in Jakarta at Raffles Jakarta, located on Jalan M.H. Thamrin in Central Jakarta. Its artificial intelligence program combines machine learning and model development with the school's business, psychology, and design disciplines, and its location gives students direct access to the Indonesian market where local AI models are being built.





References

Aji, A. F., Winata, G. I., Koto, F., Cahyawijaya, S., Romadhony, A., Mahendra, R., Kurniawan, K., Moeljadi, D., Prasojo, R. E., Baldwin, T., Lau, J. H., & Ruder, S. (2022). One country, 700+ languages: NLP challenges for underrepresented languages and dialects in Indonesia. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (pp. 7226 to 7249). Association for Computational Linguistics. https://aclanthology.org/2022.acl-long.500/

Business Indonesia. (2025). AI to shape Indonesia's digital economy as it moves toward USD 180 billion by 2030. https://business-indonesia.org/news/ai-to-shape-indonesia-s-digital-economy-as-it-moves-toward-usd-180-billion-2030

Besiroglu, T., Erdil, E., Barnett, M., & You, J. (2024). Chinchilla scaling: A replication attempt. arXiv. https://arxiv.org/abs/2404.10102

Coursera. (2026). How much do AI engineers make? 2026 salary guide. https://www.coursera.org/articles/ai-engineer-salary

Dealstreet Asia. (2025). Indonesia's GoTo and Indosat launch an updated version of the AI model Sahabat AI. https://www.dealstreetasia.com/stories/goto-indosat-sahabat-ai-444237

Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). BERT: Pretraining of deep bidirectional transformers for language understanding. arXiv. https://arxiv.org/abs/1810.04805

Digital in Asia. (2026). What is the state of Indonesia's digital economy in 2026? A comprehensive market overview. https://digitalinasia.com/indonesia-digital-market-overview-2026/

Fortune. (2025). Indonesia's Indosat and GoTo unveil a new sovereign AI that can chat in the country's most used languages. https://fortune.com/asia/2025/06/02/indosat-goto-sovereign-ai-sahabat-bahasa-javanese

Glassdoor. (2026). AI engineer: Average salary and pay trends 2026. https://www.glassdoor.com/Salaries/ai-engineer-salary-SRCH_KO0,11.htm

GoTo. (2024). Indosat Ooredoo Hutchison and GoTo launch Sahabat AI: Indonesia's open source LLM for empowering digital sovereignty. https://www.gotocompany.com/en/news/press/indosat-ooredoo-hutchison-and-goto-launch-sahabat-ai-indonesias-open-source-llm-for-empowering-digital-sovereignty

GoTo. (2025). Sahabat AI gets smarter: Indosat and GoTo launch a new 70 billion parameter model with a multilingual chat service. https://www.gotocompany.com/en/news/press/sahabat-ai-gets-smarter-indosat-and-goto-launch-new-70-billion-parameter-model-with-multilingual-chat-service

Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., de Las Casas, D., Hendricks, L. A., Welbl, J., Clark, A., Hennigan, T., Noland, E., Millican, K., van den Driessche, G., Damoc, B., Guy, A., Osindero, S., Simonyan, K., Elsen, E., Rae, J. W., Vinyals, O., & Sifre, L. (2022). Training compute-optimal large language models. arXiv. https://arxiv.org/abs/2203.15556

Introl. (2025). Indonesia AI: 92% adoption, USD 10.88B market by 2030. https://introl.com/blog/indonesia-ai-revolution-infrastructure-investment-2025

365 Data Science. (2026). AI engineer job outlook 2026: Trends, salaries, and skills. https://365datascience.com/career-advice/career-guides/ai-engineer-job-outlook-2025/

Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., & Amodei, D. (2020). Scaling laws for neural language models. arXiv. https://arxiv.org/abs/2001.08361

NVIDIA. (2024). Indonesia tech leaders team with NVIDIA and partners to launch Indonesia'snation's AI. https://blogs.nvidia.com/blog/indonesia-tech-leaders-sovereign-ai

OpenAI. (2022). Aligning language models to follow instructions. https://openai.com/index/instruction-following/

Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., 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., Leike, J., & Lowe, R. (2022). Training language models to follow instructions with human feedback. arXiv. https://arxiv.org/abs/2203.02155

Sea Lion. (2025). Sahabat AI case study. AI Singapore. https://sea-lion.ai/case-study/sahabat-ai/

U.S. Bureau of Labor Statistics. (2025). Computer and information research scientists. Occupational Outlook Handbook. https://www.bls.gov/ooh/computer-and-information-technology/computer-and-information-research-scientists.htm

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017). Attention is all you need. arXiv. https://arxiv.org/abs/1706.03762

bottom of page