It all started with a snarky tweet that stung the researchers’ pride. “What are those people at Athena Research Center doing? Aren’t they going to make a Greek language model already?”

Those dismissive words from an influential voice in the AI world struck a nerve with the multidisciplinary team of linguists, computer engineers, and scientists in Athens. After years of research into computational linguistics and language technologies focused on Greek, they had been put on the spot. Inspired by the viral success of OpenAI’s ChatGPT, the team took the gibe as a gauntlet challenge. If not us, then who would pioneer this cutting-edge technology for the Greek language?

Nasos Katsabanis, chief researcher and deputy director of the Institute for Language and Speech Processing, vividly remembers the galvanizing moment. “We were meeting with colleagues from America who asked what we were doing for the Greek language. Then that tweet landed like a grenade. I called up Vasilis Katsouras right afterward and said, ‘We have to do this.’ There was an opportunity to access powerful cloud computing infrastructure, so we immediately emailed 10 colleagues to see who was interested in joining the project.”

What ensued over the next year were many long nights, countless cups of Greek coffee, and no shortage of exasperated expletives as the newly-formed team pushed to overcome one daunting obstacle after another. Collecting and cleaning a massive dataset of 30 billion Greek words from approved sources like dissertations, books, laws, and educational materials was just the start.

“Avoiding social media data was crucial,” researcher Dimitris Pappas explained in an interview with Kathimerini newspaper. “We didn’t want our model learning to ‘speak’ like the toxicity you can find on platforms like Twitter.”

For Giorgos Paraskevopoulos, the project’s unique difficulty was driven home by the fact that Greek’s distinct alphabet and linguistic roots lack overlap with other languages. “A Portuguese person can understand a Spaniard, so you can adapt Portuguese to a Spanish model or Bulgarian to a Russian one. But that’s not possible for Greek. We’re talking about a language spoken by only around 15 million people globally – less than 0.5% of the world population.”

There were make-or-break moments when the slightest delay could prove disastrously expensive. “One three-day mistake could cost us over $7,000,” Paraskevopoulos recalled. “The cloud computing resources cost $100 per hour, and we had reserved them for a set window. I’ll never forget the anxiety I felt when we realized we had to fix an error quickly to avoid hemorrhaging money.”

Technical glitches, suddenly obsolete approaches due to rivals’ latest research breakthroughs, and strategic pivots added recurring drama. “I remember saying we’d pull an all-nighter, only to wake up having dozed off at the keyboard,” Leonidas Voukoutis recounted with a laugh to Kathimerini.

“You can’t create something new without experimenting, adjusting course, and even making mistakes,” team leader Vasilis Katsouras philosophized. “So it’s crucial for a research team to have the flexibility and wiggle room built into the process.”

Yet for each headache-inducing frustration, proud breakthrough moments spurred the team on with potent motivation. Bringing their “own ChatGPT” for the Greek language into existence from scratch ignited an unmistakable sense of achievement.

The culmination is Meltemi – a pioneering language model that uniquely combines cutting-edge AI capabilities with preservation of Hellenic linguistic identity. Potential applications seem limitless, spanning sectors like healthcare, education, tourism, culture, legal services, journalism and more.

In healthcare, a Meltemi-powered medical digital assistant could instantly decipher doctors’ handwritten notes or complex diagnoses and then converse with patients to explain treatment plans in clear language. An AI tutor for schoolkids could engage in dialogue to identify knowledge gaps, then provide customized lessons drawing from curricula, simplifying complex concepts – even generating practice exercises overnight from phrases like “Create 10 math word problems involving fractions.”

For Greek cultural institutions, Meltemi could power AI tour guides that could discuss ancient artifacts in rich historical context, or chat engagingly about classic works of literature. Lawyers could use it as a research assistant to scour massive case law databases and statutes, summarizing key insights in plain language. Journalists too could harness Meltemi for sophisticated research endeavors.

As Katsouras told Kathimerini, “We’re here for the digital survival of the Greek language.” After all, what’s the point of breakneck technological progress if it leaves humanity’s rich ancestral languages and cultural context behind as mere artifacts?

“Language and communication are what makes us human. They’re how we pass on knowledge and make sense of our world,” added researcher Stelios Piperidis. “But as our understanding of the world keeps evolving, languages must move in parallel – and therefore, language models must keep evolving too. If not, in a hundred years, we risk having our linguistic heritage devolve into just fossils studied by scientists.”

Meltemi represents Greece’s hard-won foothold in a future where artificial intelligence is poised to pervade and reshape industries. But the work has only begun. The team views their pioneering Greek language model as just the inaugural step in building entire AI ecosystems that can fully capture the nuances of Greece’s vibrant linguistic history – the regional dialects, evolving modern vernaculars, and even Ancient Greek classics.

“In 2000, a couple of guys in a Silicon Valley garage could develop world-changing tech,” Piperidis said. “But modern artificial intelligence requires extensive capital, computing power, data, and entire cross-disciplinary teams working in concert. This is civilization-scale work.”

So, while Meltemi is currently only available for research institutions and commercial partners, the Athens team continues fine-tuning, updating, and expanding the knowledge base in preparation for an eventual open-release interface akin to ChatGPT and other AI assistants. The ambitious long-term vision is to give everyone, from schoolkids to philosophers, the ability to naturally converse with an AI assistant steeped in all facets of the Greek language across its entire historical evolution – Doric, Ionic, Attic, Koine, Medieval, Modern, and beyond.

From an initial goading tweet to this pioneering reality, the Athena Research Center team reminds us that big breakthroughs often start with resolve spurred by derision and doubt. By merging linguistics, computer science, and the latest neural network wizardry, Meltemi is ensuring the indomitable Greek language doesn’t become an obsolete fossil in our AI age, but rather evolves to both reflect and shape humanity’s next minds.

“We truly made something new and original for Greece that hasn’t been done before anywhere else,” researcher Dimitris Roussis told Kathimerini. For this dedicated group of Greek academics and technologists, that’s a source of immense pride – but just the motivational spark for even bolder AI achievements yet to come.

Source: “Meltemi: Και εγένετο το πρώτο ‘ελληνικό ChatGPT’” by Sofia Christou, published in Kathimerini newspaper on June 4, 2024