Bah, Humbug! A Christmas quiz for GenAI sceptics

The approaching year is sure to come with its own well-choreographed drama. Like the falling snow in Finnish December, markets are certain to descend from lofty heights at some point next year. And when they do, the pundits—GenAI sceptics—will be emboldened by every dip to parade their arguments adjusted for developments they may not have expected.

These proclamations, which will echo in the halls of digital media during the bear phases, will revolve around the theme that GenAI is nothing but a bubble about to burst. Fundamentally, they will revolve around perceived capability shortfalls, insufficient use cases, overzealous investment, stagnating model improvements, or perhaps the ever-controversial valuations.

While such articles are as inevitable as the holiday season, those who harbour doubts might find the quiz below useful. It might just offer a moment of reflection. During future moments of dread, these questions may help settle at least one doubt: GenAI is not a hoax. It has impressive use cases and staggering utility.

Here goes the quiz. Each question addresses recent advances in large language models (LLMs) and their cousins—transformers, generative agents—across a multitude of fields, from weather forecasting to medical diagnostics, robotics to drug discovery, and well beyond.

At present (late 2024), approximately how many individuals worldwide have used a Large Language Model (LLM) at least once?

A. Around 5 million

B. Around 50 million

C. Nearing a hundred million

D. Nearing or over a billion

Answer: D. Nearing or over a billion.

In case one wonders, major LLMs have entered countless apps and platforms. The number of users has reached a level that makes even estimates difficult.

Transformer-based models have effectively interpreted which range of ‘languages’?

A. Strictly modern English only

B. All human languages, including long-dead tongues

C. All human languages plus animal communication and major programming languages

D. All of the above plus patterns in DNA, atmospheric data, and quantum-level signals.

Answer: D. All of the above plus patterns in DNA, atmospheric data, and quantum-level signals.

These models decode patterns wherever data forms a language-like structure, proving that ‘linguistics’ can mean much more than human speech.

When DeepMind trained its LLMs on raw environmental data, the model’s weather forecasting capabilities compared to best-in-class environmental prediction models were:

A. Significantly worse, justifying all sceptical headlines

B. Marginally worse, confirming at least some fears

C. On par, making doomsayers shuffle their notes

D. Clearly better, prompting a rethinking of traditional modeling approaches

Answer: D. Clearly better.

With no fundamental equations explicitly programmed, the models inferred complex atmospheric relationships, outperforming long-cherished models. It is remarkable how these models, without any embedded scientific laws, deduced superior forecasting methods simply by training on historical data—an approach strikingly analogous to how chatbots are trained to decode human languages.

According to recent surveys, who achieves the highest diagnostic rate in medical diagnostics?

A. Doctors alone

B. Doctors with assistance from LLMs like ChatGPT

C. LLMs like ChatGPT alone

Answer: C. LLMs like ChatGPT alone.

When viewed by failure rates, doctors alone have a failure rate of 26%. With the use of LLMs, they improved to 24%. However, without any doctor interventions, LLMs alone had a failure rate of 10% as per a survey published in the New York Times.

With useful text data already used in the training of models, how much more can they still learn?

A. Not much more—the models are peaking as predicted by ‘Scaling Laws’

B. They have some way to go with synthetic data but are approaching stagnation

C. There is a vast pool of untapped visual and multimodal data to analyze

D. LLMs can expand their learning by integrating data from all domains, including DNA sequences, molecular structures, weather patterns, and even senses beyond the audio-visual spectrum

Answer: D. LLMs can expand their learning by integrating data from all domains.

The next frontier for LLMs is multimodal learning, which involves synthesizing information from diverse data types. This approach promises breakthroughs in interdisciplinary fields, blending biology, climate science, physics, and beyond.

In Singapore hospitals in 2025, where is one most likely to see LLMs in action?

A. Patients engaging with chatbots for medical queries while recovering

B. Doctors using LLMs as diagnostic assistants

C. Robot nurses employing LLMs for logical decision-making and patient care

D. All of the above

Answer: D. All of the above.

In Singapore hospitals, LLMs have seamlessly integrated into multiple facets of healthcare. From patient interaction to diagnosis and robotic care, they exemplify the future of AI-driven medical services.

When sceptics abound and doubt clouds your holiday cheer, which webpage offers a comprehensive update on the rapid progress and spread of GenAI?

A. GenInnov’s collection of articles at https://www.geninnov.ai/blog

B. GenInnov’s Fund information at https://www.geninnov.ai/fund

C. GenInnov’s curated media content at https://www.geninnov.ai/media

D. A vast repository of external articles tracking LLM innovations at https://www.geninnov.ai/our-innovation-age

Answer: D. A vast repository of external articles tracking LLM innovations at https://www.geninnov.ai/our-innovation-age.

The selected webpage is a treasure trove of curated external articles documenting transformative developments in genomics, mobility, material sciences, urban planning and beyond—all propelled by GenAI.

Bonus question

Which of the following is yet to happen or has no evidence on the GenInnov website?

A. Virtual labs powered by ‘AI scientists’

B. AI-boosted scientists discovering 44% more new materials and filing 37% more patents

C. New ‘Hulk’ or wearable Robots reducing the muscle strain of workers by 60%

D. AI-powered communication for people suffering from ALS

E. LLMs surpassing human neuroscientists in predictions

F. AI-generated video games

G. Robots watching surgeons and learning to perform surgeries

H. AI-Robot’s painting fetching more than a million dollars

I. Fragrance innovation with AI-powered molecular discovery

Answer: None. This is a trick question, honouring the traditions set by some of the toughest professors this writer has encountered. All of the above have happened, and these are from the articles we compiled only in the last few weeks.

Finally, as sceptics abound and technology continues its march forward, may this holiday season spark curiosity, optimism and innovation. Let us look ahead to 2025 as a year illuminated not just by technological brilliance, but by the boundless potential of human ingenuity to harness it for a brighter future.

The author is a Singapore-based innovation investor for LC GenInnov Fund.