MANILA, July 13 — The technology behind chatbots and digital assistants that many families across the region interact with daily is undergoing a quiet but significant transformation. Large language models, or LLMs, represent a class of artificial intelligence trained on vast collections of text to produce human-like writing, and their development is accelerating rapidly. At the heart of this shift is the transformer architecture, introduced by researchers including Vaswani and colleagues in a landmark 2017 paper.
This design underpins nearly every major LLM in use today. OpenAI released GPT-3 in 2020, a model with 175 billion parameters that cost an estimated $4.6 million to train.
Its successor, GPT-4, arrived in 2023; the company has not disclosed the training cost, though analysts believe it was significantly higher. Google’s PaLM, launched in 2022, uses 540 billion parameters. Meta’s LLaMA family, also from 2023, reaches up to 65 billion parameters.
How these models learn
The training process relies on next-token prediction — the model learns to guess the next word in a sequence by processing enormous datasets. One of the largest is Common Crawl, a repository containing petabytes of web text. Training such models requires specialised hardware; deployment and inference typically run on graphics processing units such as Nvidia’s A100 and H100 chips.
But the costs extend beyond money. Training GPT-3 alone emitted an estimated 552 metric tons of carbon dioxide, a figure that highlights the environmental toll of building ever-larger models.
Current limitations also include hallucination, where the model fabricates facts with confidence, and a fundamental lack of genuine reasoning ability. These are not abstract problems for researchers — they affect how trust is built with users across Asia who may rely on LLM-powered tools for education, business, or everyday information.
Key players shaping the field
Several organisations dominate the landscape. OpenAI, founded in 2015 and led by CEO Sam Altman, remains a central force. Google DeepMind, formed through a merger in 2023 and headed by CEO Demis Hassabis, combines two of the world’s leading AI research teams.
Anthropic, founded in 2021 by CEO Dario Amodei, focuses on safety and alignment. Meta, under CEO Mark Zuckerberg, has released its LLaMA models as open-weight systems, a move that has spurred wider experimentation.
The global AI market is projected to reach $1.8 trillion by 2030, according to a 2023 report from Grand View Research. That figure underscores the economic stakes for companies and governments alike. For communities in the Philippines and across Southeast Asia, the question is how these tools will be adapted to local languages, contexts, and needs — not just imported from Silicon Valley.
What comes next
Future directions point toward multimodal models that can process not only text but images and other data types. Examples include OpenAI’s GPT-4V and Google’s Gemini family. Researchers are also pursuing improved efficiency through techniques such as sparse models and quantization, which reduce the computational load.
Alignment methods, including reinforcement learning from human feedback and constitutional AI, aim to make models safer and more reliable. These developments matter closer to home.
As LLMs become embedded in services used by Filipino families — from customer support to educational platforms — the reliability, cost, and cultural fit of these systems will shape their real-world impact. The technology is still young, and its trajectory is far from settled.
Sources
{“@context”: “https://schema.org”, “@type”: “NewsArticle”, “headline”: “Large Language Models Grow Smarter and More Costly in AI Race”, “description”: “MANILA, July 13 — The technology behind chatbots and digital assistants that many families across the region interact with daily is undergoing a quiet but…”, “datePublished”: “2026-07-13T09:30:00+00:00”, “author”: {“@type”: “Person”, “name”: “Merry Gel Sigui”}, “publisher”: {“@type”: “Organization”, “name”: “InfoPulse Today”, “url”: “https://newsfeed.infopulsetoday.com/”}, “citation”: [{“@type”: “CreativeWork”, “url”: “https://en.wikipedia.org/wiki/Large_language_model”}]}
