Người sáng lập Reka - startup AI của Singapore: cần nhiều người làm AI trực tiếp hơn chỉ nói suông

- Reka, một startup AI với 4/5 đồng sáng lập đến từ Google Brain và DeepMind, ra mắt vào tháng 7/2023 và nhanh chóng tung ra các mô hình ngôn ngữ đa phương thức có khả năng cạnh tranh với các sản phẩm tương tự từ OpenAI, Google và Anthropic.

- Công ty hiện có giá trị 300 triệu USD và chỉ với đội ngũ 22 nhân sự. Tuy nhiên, Yi Tay cho rằng quy mô nhỏ gọn lại chính là lợi thế cạnh tranh của Reka, giúp tập trung vào chất lượng sản phẩm và mối quan hệ với khách hàng.

- Yi Tay, đồng sáng lập kiêm nhà khoa học trưởng của Reka, cho rằng ngành AI ở Singapore cần có nhiều người thực sự hiểu và trực tiếp làm AI hơn là chỉ nói suông về lĩnh vực này.

- Ông nhận định việc các quan chức cấp cao trong chính phủ không hiểu rằng trong AI, những cá nhân trực tiếp đóng góp mới là người tạo ra tác động lớn nhất, chứ không phải các nhà quản lý chỉ tham gia các cuộc họp.

- Tay cũng chỉ ra rằng việc Singapore mời các chuyên gia chính sách đến nói về an toàn AI thay vì những người thực sự am hiểu sâu về công nghệ này là một vấn đề cần thay đổi nếu muốn trở thành trung tâm AI toàn cầu.

- Trước khi đồng sáng lập Reka, Tay từng làm việc tại Google Brain trong 3,5 năm và là một trong những người đóng góp quan trọng cho các mô hình PaLM và PaLM 2 - tiền thân của mô hình Gemini.

- Tay cho rằng trí tuệ nhân tạo tổng quát (AGI) và khả năng lập luận của AI vẫn còn nhiều điểm mơ hồ, khó xác định rõ ràng về phạm vi cũng như đánh giá tiến độ phát triển, và cần nhiều nỗ lực hơn nữa từ cộng đồng AI.

📌 Reka, startup 22 người của các cựu kỹ sư Google Brain và DeepMind, đang thách thức các ông lớn trong cuộc đua phát triển AI với các mô hình ngôn ngữ đa phương thức tiên tiến. Đồng sáng lập Yi Tay nhận định Singapore cần nhiều người thực sự làm AI hơn là chỉ nói suông, đồng thời cho rằng AGI và khả năng lập luận của AI vẫn còn nhiều điểm mơ hồ cần làm rõ. Ông cũng chỉ ra tầm quan trọng của việc các cá nhân trực tiếp đóng góp trong việc tạo ra những đột phá trong lĩnh vực AI.

https://www.techinasia.com/singapores-ai-scene-doers-talkers-reka-founder

Singapore’s AI scene needs more doers and less talkers, says Reka founder

As AI fever takes hold everywhere from Silicon Valley to Shenzhen, one early-stage startup – with Singaporean and Indonesian co-founders – is taking the fight directly to the big guns.

Reka, whose large language models (LLMs) can be used for the likes of online customer support and caption generation, emerged out of stealth mode in July 2023. Less than a year later, the company launched multimodal language models that are “competitive” with similar offerings from OpenAI, Google, and Anthropic.

Reka co-founder and chief scientist Yi Tay / Photo credit: Tech in Asia

Valued at US$300 million during its 2023 fundraise, Reka’s newcomer status didn’t stop data cloud giant Snowflake – one of the startup’s customers as well as an investor – from pursuing a rumored US$1 billion acquisition. The talks reportedly ended without a deal, and chief scientist Yi Tay declined to comment when asked by Tech in Asia.

Reka’s quick trajectory is perhaps less surprising once you know the team’s caliber: four out of five co-founders came from Google’s Brain and DeepMind teams.

That includes Tay, who hails from and is based in Singapore. At a meetup for Tech in Asia’s paying subscribers, he talked about the startup’s beginnings, how staying small has been a competitive advantage, as well as the AI trends to look out for – including where Singapore stands in its quest to become a global AI hub.

More coding, less meetings

Reka’s rise has put Tay squarely in the middle of not just AI’s increasing importance globally, but also Singapore’s own ambitions in the field.

The city-state has launched a revised national AI strategy and invested over S$1 billion (US$742 million) in the industry, while inviting the likes of Nvidia and AWS to make AI-related investments there.

But for Tay, Singapore’s path would require a “paradigm shift” – at least when it comes to the government. While not unique to the city-state, Tay finds that senior officials in any government may not understand that, in AI, individual contributors are the ones making the most impact.

In other words, “the people making impact are the people who are on the ground,” he said.

That is the case not just at Reka, but also at the likes of Google DeepMind, OpenAI, and other so-called “frontier labs” – a term referring to companies working on highly capable, general purpose AI models like ChatGPT or Gemini.

In this sense, AI is different from – and “a little bit harder” than – software engineering when it comes to the level of difficulty in making impact and breakthroughs, said Tay. Here, it’s about getting very senior people who are hands-on and have a lot of experience, not “management-style people” that “think they know what they’re doing, but they actually don’t know,” he noted.

“So it’s no longer about having 10 interns, 20 interns, 100 interns” who do all the base work while the senior person “just takes meetings,” he pointed out. “The senior person writes code, everyone writes code … Nobody should not write code.”

Such mindsets may have brought about Singapore’s other hurdles in its quest to be a global hub for the sector. For instance, if the country wants to be an AI hub, “you don’t invite policy people to come and talk about AI safety. You invite people who actually know this stuff, right?” said Tay.

But the “people who can really do it” are not necessarily present in droves in Singapore, he pointed out.

To be fair, that perhaps applies to anywhere in the world outside of innovation hotspots like Palo Alto or Shenzhen.

Beginnings at Google Brain

After getting his Ph.D. from Singapore’s Nanyang Technological University in 2019, Tay joined Google Brain, where he worked for three and a half years. At Google, he was part of a team that did research on transformers, which refers to neural networks that track relationships between sequential data – including text, speech, and even DNA – to glean context and meaning.

He was one of the contributors on Google’s PaLM model and was a co-lead of modeling for PaLM 2, the precursor to Gemini.

“That was during the era where only Google and OpenAI were working on LLMs,” Tay said.

At the time, such models were more or less unknown outside of tech or even AI circles. Then ChatGPT happened, whose public response Tay found to be “perplexing,” as AI scientists had been working with such technologies for some time.

Google, for instance, had launched its Meena chatbot in January 2020, which was followed by LaMDA in 2021. In other words, Google was already working on these technologies for four to five years by the time ChatGPT launched publicly.

The mainstream audience’s response to ChatGPT was also a contrast to the AI community’s – which to Tay made it even more interesting. Tay recalled that OpenAI launched ChatGPT at AI industry conference NeurIPS in 2022, and for practitioners, the product wasn’t that novel compared to what other companies have been developing.

Staying small

Still, the AI wave that ChatGPT brought forth to the mainstream eventually spurred Tay and his co-founders to start Reka, which launched just months after the NeurIPS conference.

At the time, Google was transitioning from PaLM 2 to Gemini, and like any big company, its entire efforts were focused on its LLM. Tay, meanwhile, was itching to experience training AI models outside the walls of big tech.

“I did identify as a scientist and an engineer more than an entrepreneur,” he explained. “So it was the challenge of how about we train these models ourselves? Then we have full control over what we train and what models we build and stuff like that.”

Reka’s achievements have come on the back of US$60 million in venture funding, a comparatively smaller sum to what other frontier labs have raised. It also has a team of just 22 people. According to Tay, however, staying small has been a competitive advantage.

“When people want to work with us, they’re not only getting our models,” he said. “They also care about relationships and [having people] to spend some time thinking about the problem.”

Fundamentally, the firm sees itself as an AI research and product company. In other words, rather than, say, brand name or reach among consumers, Reka is defined partly by its products but also by its team.

Why AGI is still fuzzy

The LLM race aside, AI conversations nowadays often point to artificial general intelligence (AGI), which refers to AI that can emulate human intelligence and, in turn, teach itself. Another is reasoning – Tay called this AGI’s “younger brother” – which refers to AI that can make logical deductions the way a human brain does.

From a technical point of view, Tay finds AGI to be a point of singularity where “AI self-improves without human intervention.” He pointed out that “AGI is a very abstract thing, but I view it more like a technical breakthrough” that can lead to, say, a certain percentage of jobs being replaced by AI.

On both AGI and reasoning, however, Tay finds defining the meaning and scope itself is difficult – especially with multiple AI experts likely having different takes on methodology or evaluation. It’s therefore difficult to make progress or define a timeline – outside of relatively early iterations like grade-school mathematics or booking flights.

“Right now, the process of improving reasoning is getting human annotators to pick out a bunch of data to teach reasoning,” Tay said. “A lot of things still have to come into play.”