For example, we are still talking about what ChatGPT’s return on investment is, after two years of it being around in the public domain and public eye. Yes, they have gone from zero to 300 million users in these two years and they have started to generate huge amounts of revenue. But they are not profitable just yet. 

But that is the magic of Silicon Valley, isn’t it? One minute, we have companies that are deeply unprofitable for years and decades, and then go on to become the biggest companies in the world. Exhibit A: Uber.

As we leave 2024 behind, and dive head-first into 2025, we will take a look at some of the hits and misses of AI technology, in the hopes that the tech masters at the top take note and deliver us better advancements this year.

HITS

1. Bigger Models

While the progression of large language models (LLMs), say from GPT-3 to GPT-4, didn’t exactly light up the sky, we also can’t brush it under ‘progress for the sake of progress’. 

The ‘bigger is better’ mantra of scaling laws also hit a plateau, sparking whispers of ‘where do we go from here’ and ‘where do we get more data from’. OpenAI’s 01 surely stepped in, teaching AI to think just a bit longer for better answers.

The takeaway? AI is a costly and high-stakes gamble. Someone, somewhere, is going to crack the right formula and cash in big on these LLMs. It wasn’t the flashiest year, but the groundwork is being laid for something game-changing.

2. AI Agents

AI agents are the latest hit, with OpenAI, Microsoft, Apple, Google and Amazon working on putting out their respective versions sometime this year (Anthropic has already released theirs). These digital do-it-alls — from managing your inbox to drafting emails — promise convenience, but they do come with their own set of hiccups. 

Despite flashy demos from tech giants, AI agents currently feel more like interns learning the ropes. The computer is still figuring out how to use a mouse or how to navigate a web browser. 

Critics argue that the hype might outpace utility, as the use cases aren’t catching up with users, and the tools frequently ‘hallucinate’ errors. Still, for those willing to work with their inaccuracies, AI agents hint at a future where automation could save a lot more time.

3. Promise of AGI

The thing that OpenAI’s Sam Altman is most excited about in 2025 is artificial general intelligence (AGI). AGI is a level of intelligence where the lines separating human intelligence and the artificial one get blurred, so to speak. It is when a machine reaches the ability to understand, learn and apply knowledge across a wide range of tasks at a level equal to or exceeding human capabilities.

Joining Altman in this prediction is Anthropic CEO Dario Amodei, who put the realisation of this advancement at a safe 2026. But the real problem in assessing whether we have achieved AGI or not is the different definitions companies will use to ascertain this.

There is no universally accepted standard for what constitutes AGI, nor are there established criteria for verifying its existence, making it difficult to reach a consensus on when we’ve truly reached that milestone.

Will we reach AGI in 2025? We may be able to reach some level of it, but we will have to be careful not to confuse an extremely intelligent artificial system with AGI.

MISSES

1. Apple Intelligence, Vision Pro

Apple's AI or Apple Intelligence, as they call it, has been criticised for being too "two years behind" the competition, offering basic functionalities that have existed in competitor tools for much of the last year and the year before. For example, Apple’s AI struggles with basic tasks, like pulling flight details from emails and summarising text messages in a way that removes personality.

On the other hand, Apple’s Vision Pro has been considered a letdown. It was expected, though it doesn’t have the first mover advantage (Hello, Meta!) that Apple would bring in a compelling daily use case. The Vision Pro's US$ 3,500 price tag wasn’t the issue for Apple’s affluent customer base, but rather the fact that there was little to do with it beyond playing games.

Looking ahead, Apple plans smaller, cheaper versions of the Vision Pro, but the key challenge will be offering more practical applications. 

2. Models are still pretty stupid

AI is still pretty dumb. Take Google's recent AI product releases: They might do well in benchmark tests, but in real-world usage, they’ve been disappointing. Google introduced AI Overview in 2024, which told people to eat rocks and put glue on their pizza. Remember how ChatGPT couldn’t tell how many Rs there were in the word Strawberry?

What such instances point at is that there’s a clear gap between technical training and actual user-friendly applications. This is a symptom of AI models being trained to ace certain tests, but failing when asked to deliver results in practical and real-word scenarios.

3. Ambiguity over data usage

Data privacy still remains a roadblock in the development and deployment of AI models, particularly as these models grow in both capability and scope. A significant challenge lies in the fact that we don’t know how our data is collected, processed and stored. 

AI systems, such as ChatGPT, bank upon huge amounts of data to train their models, but how they get their hands on such a vast trove of data remains questionable. The process is not as straightforward as it may seem.

A New York Times report said that OpenAI developed a tool to transcribe YouTube videos to amass conversational text for AI development, which is illegal.

In some cases, AI models have allegedly violated copyright infringement laws across different countries. By generating content based on data derived from various sources — often without explicit permission — these models risk breaching intellectual property rights. 

The challenge becomes complex when AI systems operate across borders, where local laws may vary significantly. Without standardised regulations, companies behind AI technologies may inadvertently skirt legal boundaries.