URSABLOG: Deeply Seeking

There has been so much written about AI this week that I am tempted not to add to the mountains of uninformed opinion already created. But as the deluge of tariffs recently imposed have yet to make any sense to me, and have yet to be reacted to by their targets, I think AI is a more suitable subject, if only because it has made me think more, and discover more about it, and, well, stop ignoring it as a load of hype-driven nonsense.

In the aftermath of the US presidential election, where the tech bros were out in force, and with NASDAQ and other indices being driven ever upwards by the promise of AI, and Nvidia was churning out the chips to drive them, a small and hitherto unknown Chinese company called DeepSeek released a new version of AI that apparently is just as good as the US developed models, but at a fraction of the cost, the computing power and energy that the US ones have required.

In fact DeepSeek v3 has been out for almost a month now, but was not really noticed by the competition (see Drop Site’s essay and summary of the story of all you need to know about how we got here, and the likely repercussions) it took a while for the news to hit the market – maybe everybody was distracted by wondering what to wear for the inaugural balls. But the news hit – big – this week: Nvidia stocks were down almost 15% pre-market on Monday, losing approximately US$420 billion from its market capitalization – the largest ever fall by a single company, ever – triggering a bloodbath across semiconductor stocks. Later in the week other stocks that were riding on the Nvidia wave – electricity generators, hardware suppliers, battery makers – were dragged down too.

One of the intriguing things about all of this is that DeepSeek is open source – meaning that its source code is made freely available to the public, which can then be modified and further distributed, as opposed to the US developed Large Language models, which are not. There are ironies here: the US, land of the free, champions of liberty have pursued a policy where their products – ChatGPT  and Gemini to name but two – are closed and can only be used via payment, powered by the latest and most powerful graphics processing units (GPUs) supplied by Nvidia, the very ones that the CHIPS Act (enacted by the Biden Administration) banned for export to China for national security reasons. Instead DeepSeek developed a model using less, and less powerful, GPUs which need less energy to run and developed a leaner, more efficient model that – despite the protestations of the Meta, OpenAI and Google – appears to be just as good as anything the US can produce.

The money – private and public (in the form of huge federal subsidies) – that has been thrown at developing the US models seems have been wasted if DeepSeek v3 is able to do the same at a fraction of the cost and energy required..

There are comparisons here to be made with the recent bipartisan SHIPS Act put before Congress before the inauguration. In it the US accuses the Chinese of using massive subsidies and abundant supplies of labour (surely this is not China’s fault entirely) and thereby decimating the US shipbuilding industry. This is despite the fact that the US was not interested – privately or publicly – in building ships for the global market at the time, and the ships it did produce were entirely unsuitable for that market anyway, being too expensive and mostly for domestic use only, the costs distorted by the Jones Act. The Chinese meanwhile – by fair means or foul – just got on with it, developing their shipbuilding industry for the global market that is now not only good enough to compete on similar standards but also now has the majority of world commercial shipbuilding capacity. They wanted to do it, the US didn’t, so they just got on with it.

To cry foul now, when the US has been out of this market now for all intents and purposes for the last fifty years – well before Chinese shipbuilding got its act together let alone started to build for foreign clients – seems a little tardy, to put it mildly.

Likewise, the tech bros are crying foul because DeepSeek copied/stole/learnt (delete where appropriate) from US LLMs and then modified them – i.e. innovated – cheaply. The US CHIPS Act offered subsidies of US$ 50 billion by the way.

Sure, it is well known that President Xi had targeted AI development as a crucial area for China – he certainly didn’t make a secret of it, and subsidised it heavily too – although the irony of a country run by the ever controlling and watchful CCP has maybe got ahead in the race towards Artificial General Intelligence (AGI) by releasing an open source model, together with detailed guidance of its training methods, into the market shows that we live in strange times.

I have little sympathy with the big beasts of Silicon Valley. Some – like Sir Demis Hassabis, co-founder of Google DeepMind and joint recipient of the Nobel Prize for chemistry were scientifically and academically respectful towards the new kid on the block. He, for one, will appreciate the fact that DeepSeek is open source.

Others were less respectful. Yesterday I watched an interview with Dario Amodei, founder and CEO of Anthropic who went on at some length to try and prove that those who have the precious new GPUs in abundance will succeed in the end through sheer processing power. I thought to myself: “size isn’t everything you know, Dario.” Further in the interview he repeats the apparently common thought that when we achieve AGI – the hypothetical intelligence of a machine that possesses the ability to understand or learn any intellectual task that a human being can – we will be able not only to create drugs that will allow us to live longer, but also enable us to predict – his word not mine – accurately and consistently.

Well, Mr Amodei’s own powers of prediction aren’t so great if he failed to see more than a trillion US dollars wiped off the stock markets in a couple of days, even though the DeepSeek model had been released a month earlier. I wonder if his AGI or China’s AGI – if and when it comes – will do any better. Because no one – no human, no machine – can predict the future. If we could, we may as well all go home and die quietly, because there would be no market, no surprises, no fun. But this will not happen, ever. It is – I believe – a fundamental philosophical error to believe that we could – or could ever – control the future by prediction. We cannot even guess at the latest wheeze that will come out of Donald Trump’s mind. I can imagine that any AGI model would be similarly flummoxed.

Yet in shipping – and especially in shipbroking – people are earnestly wondering whether AI or some version of it will replace brokers and even charterers. There is cause for concern, but I remain sceptical, mainly because the concerns tend to focus on the cargo matching or ship finding task of a shipbroker – which is important – rather than the persuasion, negotiation and execution part, let alone the important role of human relationships.

One of my colleagues came back from a business trip to Dubai this week, and told us he had been asked by one client on our opinion of the Mitsui 66 design for ultramax bulk carriers, a ship that has a wider beam than standard ultramaxes, developed before the Panama Canal locks were widened to tap into potential demand for wider ships, a demand that has never really materialised. Our research, chartering and ship sale and purchase teams all put their views in, and we replied accordingly. We concluded that it was not really a design worth pursuing.

However in the process, one of our chartering colleagues put the enquiry into an AI platform – Regenerate I think – and without realising what it was (I am far slower than many humans let alone LLMs) I started critiquing it for is bland and meaningless waffle. It was trying – perhaps because the question asked for the advantages – to promote the ship design in ways that were almost embarrassing.

Here are a few of the choicer comments:

Increased Cargo Capacity: The wider beam allows for a greater deadweight tonnage, enabling the vessel to carry more cargo compared to traditional Panamax vessels. This is particularly beneficial for bulk carriers, maximizing revenue per trip.

Cost Efficiency: The ability to load and unload cargo in shallower ports can reduce waiting times and demurrage costs associated with loading and unloading, leading to overall cost savings for charterers.

Enhanced Operational Flexibility: The wide beam increases stability, making these vessels suitable for a variety of cargo types, including bulk commodities, logs, aggregates, and more, enhancing their employability across different markets.

Reduced Port Fees: Frequently, ports may charge lower fees for handling vessels with shallower drafts or for less time spent manoeuvring in port, further enhancing cost-effectiveness.

Adaptability to Changing Market Demands: The broad design can accommodate fluctuations in cargo demand, making it easier to adapt to cargo types that are currently in demand, thus increasing chartering opportunities.

In summary, wide-beam/Post-Panamax shallow-draught Ultramaxes provide significant advantages in terms of capacity, versatility, operational efficiency, and adaptability, making them a strong choice in the current chartering market.

Expect they aren’t. My apologies to those not active in the dry bulk carrier market, but those that are will find the above less than convincing, puzzling and in some case just plain wrong.

As I was pondering this over a glass of wine later, I smiled at myself as I realised that this was probably how I used to sound as a young broker, without much experience or knowledge, but eager to please and waffle in order to get a deal done, which I certainly had the desire and hunger to do. I probably even convinced a few to consider my points of view.

But then I stopped smiling to myself. If I don’t do now what an earlier and naïve version of my did twenty-five to thirty years ago, it is because I have learnt not to. I am – hopefully – wiser and more knowledgeable. But thirty years, even ten years, is a very long time these days.

In his interview at Davos, Sir Demis Hassabis said: “We folded all proteins known to science, all 200 million . . . [T]he rule of thumb is it takes a PhD student their entire PhD to find the structure of one protein. So 200 million would have taken a billion years of PhD time.” This, as he said, is “science at digital speed”.

Forget about science, what about shipbroking? If a PhD student takes a minimum of three years – probably more – to map the structure of one protein, no wonder it took a humble ship sale and purchase broker so much longer to learn his craft. Change is coming, and fast.

But there is still hope for us. As the philosopher Michael Oakeshott said:

“In every ability, there is an ingredient which cannot be resolved into information, and in some skills this may be the greater part of the knowledge required.”

I’m hanging on to this thought for now, for comfort and for hope. And thinking of downloading DeepSeek at the same time, just in case.

Simon Ward