When OpenAI released ChatGPT, that month was pure chaos. Large language models had been circulating internally for a while, but everyone was waiting for someone to make the first move. OpenAI jumped, and suddenly the giants couldn't hold back—Google rushed out its Bard model within weeks.
In Bard's launch demo, someone asked: "How do I explain a James Webb discovery to my 9-year-old?" Bard replied: "The Webb telescope took the first photo of a planet outside our solar system."
Sounds plausible. But if Bard had been reading Sin Chew Daily, it would know this is wrong. Yes, Webb studies exoplanets. Yes, it has incredible imaging capabilities. But the first exoplanet image was actually captured twenty years ago, when adaptive optics became standard on large ground-based telescopes.
When this error went viral, markets lost confidence in Google's AI ambitions. Their market cap dropped by hundreds of billions.
How These Models Actually Learn
Why did Bard make this "hallucination" error—confidently spouting nonsense as fact?
To understand, you need to know how these models train. Unlike traditional programming, where we give machines explicit instructions, machine learning lets machines discover patterns from data on their own.
The way machines learn language is remarkably similar to fill-in-the-blank exercises from elementary school. Take massive amounts of text from the internet, delete some words, and try to guess what's missing. "(Someone) is Malaysia's tenth Prime Minister." "(Something) is a traditional Malaysian food." Keywords get hidden; the model's job is to predict them. Sometimes multiple answers fit, so machines don't just guess—they assign probabilities.
Simple in concept, but models learn an enormous amount this way. Just like fill-in-the-blank exercises in different subjects taught different things—language for grammar, history for facts, math for logic—machines study patterns across massive datasets, cramming for an endless exam.
Large language models are like slot machines at Genting Highlands. Every pull produces different output. Training with big data is like rigging the machine so its output better matches expectations.
The Real Power of Open Source
Now you can see why Bard short-circuited. Webb has many "firsts"—it does study exoplanets, it does have advanced imaging. But combining these facts incorrectly is like answering "What should I eat in Malaysia?" with "Roti canai stuffed with durian, served with laksa."
Current AI still makes rookie mistakes. But there's more to large language models than ChatGPT. ChatGPT is one model among many. And while OpenAI is no longer "open," plenty of models—including Meta's LLaMA series—remain open source. We can fine-tune them and unlock capabilities the original developers never imagined.
Think of it this way: you have a group of excellent university graduates. They don't have specialized domain knowledge—say, astrophysics—but if you put them through a focused graduate program, they can accomplish far more. This is exactly what my research team has been exploring.
Teaching AI to Read 300,000 Astronomy Papers
Recently, we fine-tuned open-source models on roughly 300,000 astronomy papers published over the past thirty years. Same fill-in-the-blank training method. We also added some tricks—like having two language models compete against each other, one playing a student answering astronomy questions, the other playing a professor grading the answers.
Honestly, we weren't sure it would work. We figured we'd just try.
The results shocked us. We used the AI to generate PhD thesis research directions and had astrophysicists rate them. The AI-generated topics scored higher than what typical astronomy PhD students propose. Some colleagues joked: "You really need to stop—we're all going to be unemployed."
To be clear: our goal isn't replacing humans. But we need to accept that AI-generated answers have become the new baseline. Only work that exceeds AI is considered valuable. Sounds brutal, but it's not necessarily bad—AI has raised everyone's floor. We all have individual strengths that still exceed current models, but models can cover our weaknesses elsewhere.
Take me. I always felt my research was solid, but writing journal articles in English was painful—a real disadvantage for non-native speakers. Current models won't produce Dickens or Wilde, but I can now write with something approaching native fluency.
Toward a More Level Playing Field
From a big-picture perspective, I'm optimistic. These AI models can promote equality and create opportunities for people who've been left behind.
We've long relied on citation counts to evaluate contributions. But academia isn't perfectly fair. Sometimes mediocre papers go viral while valuable work gets buried. Metrics don't always reflect true worth. One direction we're exploring: using our fine-tuned model to read all papers, ignore subjective human evaluations, and identify true pioneers objectively.
Looking ahead, AI should make research more accessible. Astrophysics is a high-investment field. Most talented people, lacking resources or connections, get stuck in dead-end positions. But fine-tuned language models can hold their own—their capabilities rival domain PhDs. They can become research partners for disadvantaged researchers, ensuring circumstance doesn't limit achievement. This is our primary motivation. We're gradually publishing AI-generated astrophysics thesis directions online.
Remember IBM's Deep Blue?
Some say AI spells humanity's doom. This narrative has never been absent from history—and has never come true.
When Deep Blue beat Kasparov at chess, some declared the game was over. Twenty years later, people play and learn with AI anywhere, anytime. Chess is more popular than ever, flourishing even in countries without traditional chess cultures.