This year, AI will get better and cheaper, and maybe even gain independence

Building up interest.co.nz’s technology coverage was very rewarding in 2024, but there was a massive elephant in the room. That’s artificial intelligence or AI of course. The plan was always to cover AI developments, but it was hard to foresee just how dominant the topic would grow last year.

So much so that at times, it feels like AI has sucked the oxygen out of other tech topics. With AI everywhere, from the devices in your hands to personal computers and in data centres, there’s no avoiding the technology.

That’s particularly true for traditional tech companies, which have to have an AI strategy that sounds convincing to investors. If not, said investors will punish you like they did with the venerable chip giant Intel, which is now in dire straits and has dumped chief executive Pat Gelsinger in a panic move.

Samsung also got lost on the road to AI riches. The South Korean behemoth saw its electronics division profits drop with the share price following suit, and then had to apologise publicly for missing the AI bus.

For the record, Intel and Samsung have AI parts to sell; they’re just not the right ones that command Nvidia-style premiums.

We’ve also heard how electricity hungry the technology is, to the point that tech companies are serious about using nuclear power to feed AI and are missing their decarbonisation goals. The enormous power requirements of AI are literally destabilising the mains grid and could cause widespread damage, an analytical piece from the end December in Bloomberg suggest. 

Related to this, as 2024 was winding down, in December the company that started the current frenzy, OpenAI, announced its o3 model which will become available January onwards. 

It provides yet another leap in capability, as o3 can expend more computing power to deliberate and reason so as to produce better results. The model isn’t generally available yet, but while the consensus seems to be that o3 isn’t an Artificial General Intelligence (AGI) that can rival human capabilities, parts of it could have reached that level of reasoning ability.

Unfortunately, that technological leap forward comes at a staggering environmental cost, the AI sustainability lead at customer relationship management company Salesforce, Boris Gamazaychikov suggests.

Gamazaychikov calculated that the energy footprint of testing o3 on an admittedly complex and large benchmark was 1785 kilowatt-hours, or roughly the same as an average household in the United States uses in two months.

That’s just for a single task on the benchmark, and the amount of CO₂ is 684 kg, or five tankfulls of petrol, he estimated.

It’s not cheap either: toting up the cost of power and infrastructure, Gamazaychikov arrived at US$3400 per o3 task in the benchmark testing.

On an already overheated planet, that’s just a bad message in every respect. Furthermore, that kind of pricing puts o3 out of most users’ reach. 

There’s another way to look at it however. Models are becoming cheaper to train and to use, as their efficiency goes up. The chief executive of OpenAI, Sam Altman, pointed out that the smaller o3-mini model not only outperforms but is cheaper than the existing o1. 

Then there’s China’s DeepSeek-V3 which also popped up in December. That 671 billion parameter model is said to be competitive with GPT-4o and Anthropic’s Claude 3.5 Sonnet, but cost only US$5.6 million to train. 

Which is still a good chunk’o’change, but the AI industry was floored by what it considers to be a very low cost for model training, which can be 10 to 15 times more expensive than what DeepSeek forked out.

High training and infrastructure costs are bugbears for Western AI companies which will be looking at how DeepSeek got to where it is for a relatively small amount of money.

Nobody seems to care that DeepSeek isn’t very useful if your perspective of history doesn’t align with the official view in China though but them’s the knocks.

Meanwhile, I can run a biggish model like Meta’s Llama 3.3 with 70 billion parameters on an admittedly powerful MacBook M4 Pro Max, a feat which AI Svengali Simon Willison thought would require a big data centre with more than 40,000 graphics cards. 

That’s because Simon says Llama 3.3:70b is GPT-4 class, or state of the art. I haven’t been able to get Llama 3.3 to admit to that yet, as it insists it's merely a GPT-3 class model. As a related aside, parameters decide how AI processes data with more being better for complex, nuanced tasks. There’s no free lunch however and the more parameters a model has, the greater the computational power is required.

Anyway, that’s where we’re headed in 2025, it seems: cheaper and more capable models, which more people can afford to use, particularly for app development, and which can take much more data input than earlier variants. Perhaps GPT-5 will finally pop up in OpenAI's repository as well?

Cheap-to-run AIs that consume less energy despite being able to ingest more data could result in more and more people using the tech. The AI crowd is already talking about such a Jevons Paradox scenario, one which nullifies the lessened environmental impact from the AI efficiency improvements as usage increases. 

Lower costs for AI also drops the threshold of the technology being used for bad purposes. It’s not just the obviously abusive cases, like impersonation of others for deception and data collection of intellectual property and sensitive personal information, but unthinking use of the technology.

One great example of that involves Meta which as we know has jumped into AI with both feet. You can have MetaAI everywhere, like in WhatsApp, to chat with a virtual Snoop Dogg whose responses are toe-curlingly cringe-making and completely pointless.

Meta’s main property, Facebook, is already a bad experience thanks to scam ads and fraudsters impersonating people you know. Now there’s much “AI slop” on FB as well, nonsensical images and other such inauthentic content.

Not content with degrading the user experience in the above manner, Meta intends to roll out AI generated characters on its sites, with profile bios and pictures, and the ability to create and share content. The idea is to drive engagement with billions of users.

It’s not a new idea though. Around the New Year, people found some existing AI generated Instagram profiles that were created in 2023, and which bombed completely and lay unused until Meta recently deleted them. It takes a special kind of person wanting to get social with an AI bot, and most of us aren’t interested in doing that (and please may that continue to be so).

In just 8 years and with billions of dollars we have been able to recreate Microsoft's Tay, a chatbot which is instantly broken the moment it comes into contact with users

[image or embed]

— Brad Westness (@brad.westness.cc) January 4, 2025 at 4:58 AM

Think about that for a second: on social networks, you, the user, are the product sold to advertisers. Why would advertisers pay Meta to access an AI generated “audience”?

Maybe “agentic” (horrible word) AI could work in that context, acting as a virtual audience? 

Claude 3.5 Sonnet describes Agentic AI like so:

Agentic AI refers to artificial intelligence systems that can act independently to achieve specific goals, make autonomous decisions, and take actions in their environment without direct human input. These AI agents can understand their objectives, plan accordingly, and modify their behaviour based on feedback and changing circumstances. The concept of AI agency is about the system's ability to act on behalf of itself or its users with some degree of independence and purpose.

Unlike simple reactive AI systems, agentic AI demonstrates more sophisticated behaviour through goal-oriented actions and decision-making capabilities. However, there is ongoing debate about the extent and nature of AI agency, particularly regarding consciousness and true autonomy.

I’ve found a great app called Chorus from https://melty.sh which lets you talk to several public AIs at the same time, and locally hosted ones. They all give different answers to queries, sometimes miles apart from each other. With that in mind, I can’t wait for agentic AI to get interactive with it, making decisions without direct human input.