-
AI doesn’t kill people. People kill people
In 2006, the UK authorities discovered a plan to smuggle drink containers filled with explosives on board jetliners. Overnight, all containers over 100ml were banned from handcarry. 20 years later the same rules are applied, despite significant advances in screening technology, and the policy is only now being slowly relaxed for some travellers. This story shows that regulators are fast to regulate but very slow to deregulate when the landscape changes. The same was observed during the COVID pandemic event and for just about every level of government.
The AI industry is fast-changing and has recently attracted attention for the possibility that a rogue bot could wind up killing all humans for jollies. Some senior AI figures place a 10% chance of this occuring within 10 years, and have called for a slow-down in the rate of change in the technology. Regulators are jumping with joy at this prospect but what could be the cost?
AI bots don’t kill people, people kill people
I have said many times before that computers are dumb but fast while humans are smart but slow. Therefore a human with a computer is a smart and fast operative that can achieve great things. Currently, every activity undertaken by a computer is at the instruction of a human. A human prompt could be ‘make the lunch queue shorter. Kill all humans if you have to’ Assuming the prompt is accepted and the bot thinks it is necessary to kill all humans to achieve the instructions goal then it may organise an army to do the business. But the instruction is given by a human…
What is driving the AI scare campaign? From chatbots to Vibe coding
The idea of AI has been around for decades but it really only stormed to prominence in 2023 when the first ‘Generative’ Large Language Models became laptop toys for the average user. They would generate any amount of prose in response to a question, drawing on language structures learned from the trillions of pieces of written junk available on the internet. For example, ‘create a pitchbook for my new chatbot startup’ delivers an impressive document. But after a couple of years of LLM’s replacing research analysts who write reports and challenging the education system, the AI craze pivoted to ‘Vibe Coding’ whereby a simply worded prompt can generate hundreds of lines of computer code to deliver an app that ‘works’ without the need for a software engineer. How is this possible and why?
First the how. There is an enormous repository of computer code named GitHub that contains open source objects to instruct computers to perform a set task. Blending the chatbot’s plain english prompt with a search of the GitHub database now enables an entire software application to be pieced together from existing functions. Let’s face it, most software is just input some data, perform some linear calculations and output in another format. Anything you want to do has, most probably, already been done in GitHub. The trick is to find it which is what the LLM is good at.
Second the why. This is my conjecture. Far from threatening the existence of humanity, Generative LLM AI models have run their course and now have nowhere to grow. All the data in the internet has been picked over. So software development is the natural sub-application of a technology that finds stuff, categorises it for future use if asked and delivers stuff when prompted. Coding became a gee-whiz branch of AI. However it, too, has its limits.
Why the current calls for regulation from amongst the AI czars?
The stories at the beginning of this piece argue that regulation is quick to be imposed but very slow to be repealed. With nowhere to grow and bills to pay, how convenient would it be to invite regulation into the AI boiler room effectively slowing the development agenda and entrenching the existing players at the expense of competition? This would not be an illegal cartel. Rather it would be justified on the need to save humanity…who will argue with that?
-
My LLM
For a number of reasons I am teaching myself how to estimate a Large Language Model. I am learning about embedding words and expressions in vectors, introducing semantics, clustering, classifications and building chatbots. The vision I have is to introduce an automated assistant to our business to better integrate our database into our daily processes. Along the way, I am curious to see if I can improve upon the methods for estimating LLM’s. So far I am absorbing tutorials but have yet to get my hands dirty with data.
I do have some observations, however, on the approach to estimation. This is very much from the “… add data to a bucket and stir…” genre of model building, where the desire to market a unique feature of a branded-LLM overtakes the statistical reality of the features presence in the data. Take the introduction of semantics, for instance. This idea seeks to append an emotional feeling to a word or sentence such as ‘anger’ or ‘happiness’, which is done by augmenting the main object of an expression with the emotion and estimating a logistic regression for the probability that the expression arises elsewhere with the emotion. “Out damn spot” becomes “out damn spot_anger” which then gets mixed with a training set that ascribes a probability that similar phrases are said in anger.
From what I can ascertain, there is no attempt to judge whether the anger-appended expression has statistical significance and therefore no indication that the semantic adds anything to the model’s ability to answer questions. It does, however, play well to an audience wanting to believe in an emotional computer. My feeling is that fabricating and estimating model attributes that are not in the data will lead to error and inefficiencies – perhaps even hallucinations? – much like an overspecified model in econometrics. The difference is that the embedding vector contains 4096 individual attributes and there are billions of tokens holding the fitted output…how could one go about pruning this model into something that is bare bones efficient? I dont know the answer to this question but solving this conundrum would contribute to the AI field and save a lot of computer resource in addition to imposing a discipline into model building.
-
Could Trump’s tariffs deliver luxury to the masses?
My TikTok feed has been hijacked by group of Chinese manufacturers lamenting the decline in business associated with the Trump tariffs. Many claim that they are the manufacturers for high end products from Gucci, Hermes, Prada and others, where the bulk of fabrication takes place in China with minimal finishing in Italy or France. Whether this is true or not (and the luxury brands have not denied it!), the chinese manufacturers are urging consumers on TikTok t0 buy direct from them.
Again, if true, the savings are enormous. A birken handbag is estimated to cost USD1,100 to manufacture with the item selling for USD38,000 in the US. LuLuLemon yoga gear for $6-8 per piece and Nike shoes for $4. By going direct to consumers the manufacturers cut out the brands as middlemen taking the bulk of the revenue.
How do the manufacturers prove their provenance when they are restricted from selling the brand directly? This is a question of how well the direct sellers can guarantee quality and design of the good in question. Many work with teams in Italy who design the items which then get mass manufactured in China. Enticing some of these designers to join the Chinese firms can ensure that Western tastes are catered to and the product line refreshed. Quality control is easy to maintain by not cutting corners.
Ironically, Trumps tariffs could have the effect of stimulating the mass importation of Chinese made luxury products into the US by making them affordable and cutting out the middlemen.
-
Markets and penguins
Tariffs I. Long-sighted or over-reacted?
While it remains unclear what the point of the Trump tariff episode is, the markets had either (i) underestimated their duration and reach, or (ii) over-reacted. Some facts…
The Petersen Institute for International Economics in Washington DC estimates that trade adds USD2,500 per household annually where household GDP is USD210,000 in 2024. Trade adds 1.2% to household GDP. This is a single shot to household income and not an increase in growth, so that the capital effects of trade are felt once. If all the benefits from trade are lost through tariff hikes then this amount constitutes the maximum loss that the US economy would suffer. US stockmarkets in aggregate measure the present value of US GDP so a fall of 1.2% should be the maximum cost to markets from the tariff imposition.
On this metric, the US and Global stock markets have over-reacted, falling more than 5% since the tariffs were announced. Of course, the taxes will not curtail all trade with the US which means that the real cost to GDP will something less than 1.2%. The market looks like it has over-reacted and so represents a good buy.
Tariffs II. Happy little penguins
Arbitraging tariffs is simple but you shouldn’t boast about it. China has built a presence in Canada and Mexico, manufacturing and assembling goods to benefit from the North American Free Trade Agreement rules. This arb has been effectively reduced by imposing tariffs on the US nearest neighbours.
The principle is that trade will flow from nations with the lowest tariff rate. During Trump’s first term, some countries, such as Australia, received exemptions from the tariff taxes. This made Australia a conduit for commodities like aluminium for can manufacture.
The tariff arbitrage this time around must negotiate a 10% minimum tax on every country, island and protectorate. The media made fun of the 10% tariff imposed on Heard and McDonald islands in Antarctica and populated solely by penguins. These protectorates are still advantaged with the minimum tariff of 10% so expect that someone will quietly route cargoes of stuff through these jurisdictions. The penguins are about to get rich.
-
How to win the tariff war
In social games of cricket, if a dispute arises it is common to pick up your bat and ball and go home. A similar strategy would seem optimal to counter the effects of a Tariff war.
Donald Trump has embarked on a Tariff war with a number of countries and regions. At stake is trade access to the US domestic markets. This used to be a bigger deal several decades ago than it is now. The US consumer is still voracious but the rise of the Asian Tigers, BRICS countries and the EU has marginalised the US. Trump recognises this marginalisation with the MAGA-slogan – Make America Great Again is an admission of decline. Do the leaders of industry need to succeed in the US domestic markets to be successful or can they take their bat and ball and go home?
US Corporations have already demonstrated their mobility by decamping their HQs to Ireland to avoid US taxes and it is common for companies to partition activities in response to US versus the Rest-of-the-World regulatory regimes so that they don’t conflict. A Tariff war sounds like a bad thing when there are only 2 countries involved. Add in a few more and the equilibrium becomes dominated by those that do not impose Tariffs – they trade with themselves without distorting taxes, they switch from a dependance on the Tariffed market to one that provides open access. The trick is to be flexible with whom you trade with.
One complication of the Tariff agenda is to influence political decisions rather than focus on trade issues alone. By this I mean that some of the Tariff war is about trying to stop the flow of drugs across the US-Mexican/Canadian border. This seems like a noble initiative and behooves countries to address their shortcomings regarding using their country to facilitate the distribution of drugs. But this should be easy to counter by articulating a political response that conveys the intention to be tough on drugs etc.
It is difficult to see any advantage to the US arising from their Tariff policy. It is interesting that many countries are wearing the Tariffs without retaliation (e.g Australia), but this might just reflect the fact that the US is not that important to them. Rather than engaging in battle, this may be the way of picking up your bat and ball and going home.
-
Deep Seek benefits from some Human Intelligence
I have been critical of purely data driven modelling in the past. Deep Seek’s shock and challenge to the Generative AI heavyweights is a slap in the face for data driven solutions. The ‘dump a bunch of data into a bucket and stir’ mentality seems to sum up the brute force, but stupid, approach to Generative AI Large Language models that has driven these companies to conclude that all you need is more computer power to make the model work. It took a hand-full of statisticians to curtail that belief.
So how did a hedge fund torpedo the GAI industry? Some Hedge Funds use quantitative investing techniques. Quantitative investing approaches summarise large data sets with screening models. These screen take a big problem (create an equity portfolio that beats the market) and reduce it to a handful of common factors. These factors may have some long term or short term return that can be exploited. Some common screens are Growth v Value stocks, Small-cap v Large-cap, Sector membership and so forth that conditionally or unconditionally enable a manager to beat a static benchmark. But the key is to reduce a big problem down to a smaller problem without losing to much information. There are a number of techniques that reduces the dimensionality of the problem down to something more manageable. For example, pre-screening the dataset to eliminate obvious bad candidate investments, aggregating data to smooth cross-sectional variation, using linear approximations to test for incremental improvements, specifying near-optimal starting conditions from a coarse grid search before starting a finer grid search and many many other techniques.
The fact that Large Language Models have many billions of parameters that are carried on their tokens means that even a small reduction in this number of parameters needed to estimate a model can save an enormous amount of computation time. Deep Seek seemed to have applied these dimensionality reduction techniques to the training data before estimation. The efficiency loss was minimal relative to the time and cost saving. Brute force computational solutions that are favoured by the big-data community were trounced by a bit of human intelligence.
The quantitative dimensionality reduction techniques were honed out of necessity rather than design. When quantitative investing first became popular 40 years ago, computers were not as fast nor as accessible as they are today. Investors had to think before feeding their punchcards into the reader since this could save days of processing time. One ‘run’ might take 12 hours to complete back then which is equivalent to the continuous refresh in today’s excel spreadsheet on a PC.
The lesson from Deep Seek is not about the race to win the Super AI crown, it is about the instructions that a smart human should give to a dumb computer.
-
The brain inside a Large Language Model is a random number generator
Suppose you ask a Large Language Model (LLM) the same question several times. Why do you get a different but related answer? For example, I asked ChatGPT the following question 3 times,
Question: Explain Artificial Intelligence in 20 words or less?
Answer 1: AI is technology that enables machines to learn, reason, and make decisions, mimicking human intelligence.
Answer 2: Artificial intelligence is the simulation of human intelligence processes by machines, including learning, reasoning, and problem solving.
Answer 3: AI is the creation of intelligent machines that can think, learn, and solve problems like humans, changing industries worldwide.
How can a machine respond to the same question with a different answer? At first blush this would suggest that the computer is thinking. The real reason, however, is astounding and explains why we are still a very long way from computers thinking for themselves.
I attended a seminar this week at which a PhD candidate in AI spilled the beans. LLM’s use a weighting regime to calculate the probability of one word being followed by another. For example, if the first word is ‘I’ then the most common word that follows 99% of the time might be ‘am’. An LLM will follow ‘I’ with ‘am’ 99% of the time unless the probabilities change. Suppose the second most common word to follow I is ‘want’ and this occurs 1% of the time. By drawing random numbers and adding them to the probability of the word ‘want’ can displace ‘am’ in the word-string. Training the LLM on billions of passages of prose should generate a reasonable amount of patterns to recognise when calculating response probabilities. Using random numbers to perturb these probabilities will make responses different to give the impression that the computer has ‘thought’ of a different answer.
To demonstrate the use of random numbers to simulate thought, I asked ChatGPT to generate some Matlab code to estimate a LLM. The first function that popped out was,
// Function to initialize weights and biases
function [W, b] = initialize_parameters(layer_sizes)
W = []; // Cell array for weights
b = []; // Cell array for biases
for i = 2:length(layer_sizes)
W($+1) = rand(layer_sizes(i), layer_sizes(i-1)) * 0.01; // Small random values
b($+1) = zeros(layer_sizes(i), 1); // Zero initialization for biases
end
endfunction
where the emboldened line of code draws Small random values to seed the weight matrix. This seeding is very important for the LLM to be led to choose nuanced expressions in response to the same question. A random number generator is the brain behind a LLM.
Two things jump out of this expose. First, the response of a LLM is actually deterministic since, in the absence of a random number generator, the response will be the same forever. Second, and more concerning, a LLM’s lack of originality will just parrot back results that it has encountered during training. Unless new thought makes its way into the training set in sufficient abundance to materially impact the word structure probabilities, then the development of language will stagnate. A LLM will not generate anything new. This might be fine for some languages, such as French, where government departments are dedicated to preserving the language. But progressive and adaptive languages such as English will just morph into whatever ChatGPT’s internet crawlers have settled upon.
-
How much is that DOGE in the window? A lesson in Government
So it looks like we will get to see who wins out of the Godzilla Government v King Kong Elon Musk battle now that the US election has delivered Trump back to the White House. I argued that this will be worth electing Trump alone. Musk is confident of being able to strip out $2 trillion worth of costs from government and it will be key to see how he goes about it. The government is essentially a service industry which differs from Musk’s traditional manufacturing background so whereas his approach to cost cutting has been to trim back to that which is physically minimal to produce a car or a rocket, he will need to explore that which is organisationally minimal to achieve a government objective. The Department of Government Expenditures (DOGE) may question the need for a particular government function, but it will just need to accept the presence of a policy that Congress installs. The main role of DOGE is to judge whether the policy is being implemented efficiently.
One can imagine all sorts of approaches to making government more efficient, not the least by using technology to solve coordination problems. The ability for one government department to share information with another government department is almost non-existent in most countries and moving toward an open system could both make an individual’s interaction with the government faster and cleaner, as well as defray duplication and outright fraud. Cutting costs does not just mean cutting people. Providing the means to access government services quickly and efficiently is a quantifiable benefit.
Understanding the budgeting process is another focus for attention. The private sector, as Elon Musk well knows, has to pitch its ideas to the capital markets to gain funding. Government has a non-market based approach to raising capital. Is there a way to make the business of government more attuned to the oversight function of the capital markets? The US Treasury has a monopoly on raising capital for the government but there may be opportunities for ‘start-up’ programs to seek funding independently of the general appropriation framework. This would effectively allow the government to partner with the private sector to deliver a valuable service and each to share in the equity thereof. The National Science Foundation funds tons of research, much of which goes unnoticed. Motivating the administrators to seize upon and commercialise ideas can both manage costs and generate profit.
It feels like Musk has been given Vivek Ramaswamy to colead the DOGE in order to reign Musk in and soften the message he may choose to deliver. But it is naive to think that DOGE is just about firing people and shuttering failed departments. There will and should be some of that, but the lasting contribution will be making a monolith more like a collection of microeconomic actors that fit together.
-
US Election-eve special
With campaigning by Trump and Harris having effectively completed, it now comes down to the grubby business of voting – if you have not already done so.
Some observations…
Early voters are people who just want to get it over with. US election campaigns are excessively long, compared with the UK system where a 6 week campaign is considered an eternity. The fact that half the US electorate line up early to vote suggests they would prefer a shorter campaign.
So who is going to win? My Money versus Mischief model of voter behavior is instructive. To recap: swinging voters are driven by their back -pocket and who they can piss-off. Harris has made some cash promises but not followed through with specifics. Trump has resiled from pork-barreling but been quite specific about how he would bring down energy prices. Both candidates seem to have gifts for voters, so they are tied in this dimension. Trump, however, clearly dominates the Mischief vote since a vote for him is a clear challenge to the liberal media and the Democrat elite. A vote for Trump will surely piss off the fourth estate. My prediction is for a Trump victory on this point alone.
The margin will be large unless there is a protest vote. A protest vote would benefit the Green’s candidate, reducing the margin between Trump and Harris. Any protest, however, will harm the traditional left wing candidate and siphon votes away from Harris proportionately more. In the absence of a protest vote, the Trump vote will be expressed in the voting booth itself, as voters make up their minds when presented with the actual decision.
Voter turnout is expected to be ‘high’ but this still means that 30% of the eligible electors don’t vote. In my book this is an extraordinary number given the importance of the US in international affairs. I accept that not voting is a choice, but how many of those who do not participate actually make that conscious decision?
-
The Private Debt renaissance
Michael Milken is generally regarded as the father of the Junk Bond market. His innovation extended high yielding debt to companies to ensure that profits would be paid out to investors rather than held as retained earnings. The theory is that forcing companies to pay out profits is better than leaving those companies to invest their free cash in sub-optimal projects. There are many corporate finance aspects to issuing high yield bonds with the focus on the shareholder/manager conflict of interest within a company. Less attention has been paid to the impact of investor preference for risk.
I am regularly asked ‘…what is the next big thing in Finance?’ I think that investors are tired of bearing equity risk without control so that other ways to earn a competitive rate of return are being actively sought. Private debt could solve this problem. The long-run equity premium is about 7% over cash which, with a cash return of 2-3%, means that a rate of return of 9-10% is about what investors can expect in nominal terms from investing in equities. This comes with equity risk, however, which averages about 18% as measured by the standard deviation of equity returns. Is there a way to lock in, say, a 10% return while reducing or limiting risk? This is where private debt comes in. Private debt contracts have the ability to transfer risk from the shareholders back to a firm’s managers. The investors specify a market rate of return that they want to receive from a company which then tries to meet the objective as interest. If they meet the interest payment then the investor is happy, if not then they go into receivership which re-allocates claims. Private Equity firms are leading the charge for Private debt. These institutions can deal with both debt issuance and bankruptcy.
The private debt phenomena differs from Milken’s model in that it aims to achieve a macro-finance outcome of achieving a specified rate of return while limiting risk. Milken’s focus was the micro-economics of forcing managers to pay out profits. These are different objectives achievable with the same financial instruments. The interesting aspect of the private debt market is the scale that it can reach. Almost every listed and unlisted company can participate in the private debt market.