AI state of the union H1 2024

9 minutes estimated reading time

The AI state of the union H1 2024 post came about as we had a number of trends starting to come into view. To paraphrase Charles Dickens the AI state of the union H1 2024 represented both the best of times and the worst of times in generative AI.

Is the current state of AI analogous to the dot com boom?

In this respect, discussions around a dot com type boom around generative AI are less helpful. The dot com boom didn’t have the same naysayers at the time, aside from what would be now called edge lords worried about money. Like with all economic cycles they would eventually be proved right, but not until we had broadband and shopped at Amazon.

Culturally, we are in a very different, darker place and aren’t riding an economic boom. I have tried to lay out some of the nuances in the themes currently driving AI discussions.

I have broken this down into three themes and a focus on Japan.

Vendor knife fight
Trough of disillusionment
Synthetic data
Japan focus

Right let’s get into The AI state of the union H1 2024.

Vendor knife fight

The AI state of the union H1 2024 sees generative AI being added to products and business plans, in a similar way to being web-enabled in the late 1990s. Spring Apps estimated that there almost 58,000 AI companies.

Microsoft, Alphabet, Amazon and Meta increased their investment in generative AI to $106 billion during the first six months of the year. In addition, Meta is looking to leverage the open source model of software development to drive progress in its Llama model. This echoes, how the open source model and software like Linux, MySQL and PHP were used during the dot com boom and the move to web 2.0 to provide greater efficiencies and possibilities.

Some applications such as Adobe Creative Suite has become much more powerful thanks to adding generative AI. Alphabet has finally had generative AI companies parked on its front lawn. Open AI joined Perplexity in providing a search product.

Apple came up with Apple Intelligence that provides a platform and front end for a mix of generative AI models. Microsoft has Co-Pilot that it has been selling to enterprises.

Financial institutions have led the charge to try and get productivity gains such as JPMorgan’s IndexGPT and the continued automation of back office processes.

Meanwhile in China, Baidu’s ERNIE has attracted almost a million developers looking to use the generative AI platform in their projects.

Trough of disillusionment

The enthusiasm for generative AI hasn’t managed to drown out dissonant voices. The number of objections are diverse.

Business issues

Morgan Stanley in a research report quoted a large pharmaceutical company CTO who abandoned the use of Microsoft Copilot in their organisation. The crux of the argument was that they weren’t seeing the value. Presentation creation was described as middle school level. Pharma companies tend to use PowerPoint as a publishing platform rather than a ‘presentation tool’ with data rich busy slides, so I can understand why Copilot became unstuck.

It isn’t only ‘high-end’ knowledge work conducted by large corporates that is underwhelming. McDonald’s is withdrawing generative AI systems deployed at its drive-thru restaurants due it not working as planned. A Gartner survey of IT leaders indicated that McDonald’s wouldn’t be alone with nearly one in three generative AI projects to be scrapped in 2024.

That might not be such a bad thing as businesses are currently in a process of experimentation, so long as the lessons learned are captured and internalised.

Goldman Sachs have pivoted over a matter of months from being bullish about generative AI, to being concerned that the return on investment for generative AI may take far too long.

Societal impact

What if speed isn’t the goal? The process of reading isn’t only about the ability to parse information quickly but also affects other aspects of human thinking and behaviour. There are clear benefits for certain groups of people including neurodivergent and second-language learners. But it also poses a risk to close reading skills which impacts developing or improving existing skills. Secondly, the generative AI can miss key facts from a document, given up as speed is prioritised over nuance and accuracy.
Knowledge collapse – By mediating access through AI tools moving forwards, due to the model’s focus on the centre of of the distribution of its data set. Restricting access to the edges is likely to cause harm to future innovation, human understanding and cultural development outlined in Peterson’s paper AI and the Problem of Knowledge Collapse.

Technology-specific issues

Environmental impact – like web 3.0 and the crypto economy, generative AI requires a lot of energy to run high performance data centres. This means that Open AI is losing money hand-over-fist paying for computing capacity from Microsoft at a significant discount.
Model collapse – the relative lack of human-made data and the rise of synthetic data used in training generative AI systems is likely to lead to a rapid degradation in those models, indicating a ceiling on amount of progress that generative AI based on LLMs is likely to make.

Synthetic data

Synthetic data is probably one of the most difficult subjects to write for AI state of the union H1 2024. On one hand, you have Mark Ritson’s endorsement of synthetic data based on what we saw from B2B marketing generative AI startup Expenza AI. Ipsos have also got some credible interesting offerings that seem to be based on the provision of synthetic data.

Is it any good? A lot depends on how the LLM is trained and the way it’s being used in terms of what you want to achieve. As with any tool, it can be useful for the right jobs. The MRS Delphi Group gave a range of feedback on the way it should be used, some of which seemed to contradict each other. We don’t know how accurate a picture the LLM is creating, what is being called algorithmic fidelity.

Until concerns about algorithmic fidelity is addressed sufficiently well; marketers would be wise to exercise a degree of caution.

Japan focus

I have included Japan in my AI state of the union H1 2024 post for a few reasons.

Prior to the current exuberance about generative AI; Japan was doing really interesting things using different parts of AI including fuzzy logic and software agents. The Panasonic rice cooker that cooks rice that’s perfect for your preference. Error correction for video and audio playback, from CDs to Blu-Rays. Complex camera programme algorithms including image stabilisation. Sophisticated non-playable character behaviour in computer games. There has even been synthetic singers like Hatsune Miku and virtual influencers.
If Star Trek influenced the flip phone, the smartphone ( think the tri-corder, especially when used with the likes of Oxford Nanopore‘s products) and tablets (on Star Trek’s next generation), then cyberpunk and Japanese anime have influenced AI in a similar manner. Elon Musk and Sam Altman would fit right in as villains in the Ghost In The Shell series.
Finally, even though Japan influenced cyberpunk based on William Gibson’s experience meeting Japanese students, it has a paradoxical relationship with technology. For instance, the Japanese government recently stopped using 3.5″ floppy disks. Ancient crafts are still highly prized and Japanese brands like The Real McCoy’s and Grand Seiko who provide premium manufactured goods to artisanal standards.

Matt Alt has put together a good overview of the policy and cultural context for generative AI in Japan. It’s less of a clear cut issue than the Japanese body politic seems to believe. The Japanese government believes that its population is likely to view AI positively because of anime plot lines. While Atom Boy is a positive example there are lots of negative examples in the Ghost In The Shell franchise alone. There is also a tension between government aspirations for international exports of increased amounts of media content.