Subscriber OnlyInnovation

We are at last creating AI approaching human intellect but the underpinnings are insubstantial

Power consumption required for GenAI could be heading towards entirely impractical levels

The first British regular public newspaper is generally attributed to Nathaniel Butter in 1622, a London printer. Unlike formal government gazettes that announced official diktats from the leadership of the day going back as far as Julius Caesar’s Acta Diurna, Butter’s Certain News of the Present Week encouraged a more informed society including on international current affairs.

Butter, however, became prone to filtering stories or even making up entirely fictitious news to boost his business, which in turn had a tendency to become amplified by other newspapers and in public discourse. The playwright Ben Jonson became particularly dismissive towards Butter and the developing news sector, frequently parodying his newspaper.

Fake news is now familiar in our lexicon. Asserting that various news items are concocted has led people to question truthful news and societal norms, affecting opinions on critical issues and topics. Fake news has mutated facts and beliefs, raising doubts about information credibility and quality.

The American comedian WC Fields quipped: “You can fool some of the people some of the time – and that’s enough to make a decent living!” With social networks and partisan media swamping established news channels, some pundits and politicians have built substantial personal wealth that would have been impossible when the news industry had a business model of rigorous fact-checking that in turn attracted advertisers seeking credibility and authority.

READ MORE

Generative artificial intelligence (GenAI) is further pushing the boundary between reality and illusion. Articles, photographs, images, videos, audio, scripts, clips, music – frankly all content – can be algorithmically spawned from the original seeds of appropriate training data. The resulting content alchemy can be challenging to distinguish from reality.

One of the core roots of GenAI is “generative adversarial networks”, invented in 2014. Two algorithms are set to compete with one another. The first generates new data extrapolated from given training data. The second attempts to distinguish the generated fake data from the original training data. The procedure iterates until the second process is largely unable to distinguish between real and generated results, thus indicating that the first has learned to produce convincing but artificial data.

Livelihoods are now being impacted by GenAI. In the same way that a typography font enables whole books to be printed in a particular visual style, a “voice font” capturing just a few spoken syllables can be used to generate arbitrary audio speech in a particular accent and tone.

Voice-over artists face an end to their profession. An “expression font” can be used to credibly mimic the dynamics of an individual’s face. A “body font” of a few movements can generate a digital twin of a human. Actors, celebrities and film extras face a challenging future.

The video gaming industry is building illusions of the real world, with its technology also now becoming widely adopted in film production, architectural planning and training, including by security forces. Privately held Epic Games from North Carolina is the industry leader, well known for its Fortnite online game.

Tim Sweeney, the founder and CEO, believes that simulations will imminently become completely indistinguishable from reality.

Some observers caution against the ebullience of GenAI, such as a recent syndicated Financial Times column in this newspaper (see here).

Nevertheless, the technology is rapidly improving. For example, assistants such as perplexity.ai supplement their generated responses with citations, allowing you to easily fact-check and detect any tendency for inaccurate answers or fake news.

Some GenAI start-ups are already faltering, such as San Francisco-based Tome and its GenAI business presentation tool which, having raised $81 million (€76 million) since last year, announced lay-offs last week.

Nevertheless, the investment momentum remains significant. Microsoft and Open AI have announced a collaborative plan to build a $100 billion data centre, called Stargate, for their next technology iteration of GenAI, at a time when the construction cost of a typical data centre is by comparison just a few hundred million dollars.

Criticism of the enormous power consumption required for GenAI remains. The chief marketing officer of Arm Holdings, a UK semiconductor design company, last week asserted that although GenAI consumes about 2 per cent of electricity today it could require a quarter of all electricity in the United States by 2030. This is entirely impractical, particularly when the human brain consumes only about the power of a 10-watt light bulb in comparison and yet still outperforms the intelligence of the current generation of GenAI.

Since biological systems considerably outperform in terms of intelligence for power consumed, I believe that the foundations of GenAI are profoundly incorrect. While we are at last creating machines approaching human intellect, the underpinnings are insubstantial.

The IDA was correct to observe the opportunities in the Irish economy for GenAI in a report last week. However, there is also considerable scope for researchers and theoreticians to disrupt the current iteration of GenAI to produce a more viable artificial intelligence.