Author: Conor O'Higgins

Can AI make cocktails? We found out.

Posted on by Conor O'Higgins

Christmas party season is here, a time for congregating with friends indoors for hot whiskeys, mulled wine, brandy, and maybe something a little more singularitarian.

I had a burning question: “Can AI make new, original cocktails that taste good?” My AI-sceptic friend said, “I prefer recipes written by people with taste buds”. Large Language Models simply combine strings of words and don’t really know what anything tastes like – can they act like they do?

There are several major LLMs on the market – which LLM will make the best recipes?

Most importantly, will there be any delicious recipes here, ones worth keeping? Can I produce something that will please my friends at Christmas parties? And can I repeat that feat later for different themes?)

I gave the same prompt to five AI models: I asked them all “Invent some new Christmas-themed cocktails.”

  • ChatGPT 4o generated ten cocktails
  • Gemini (formerly Bard) generated five cocktails
  • Mistral generated six cocktails
  • Claude generated five cocktails
  • Llama (Llama 3.1 405B Instruct to be precise) generated five cocktails

You can read the raw outputs here in the interest of transparency. That adds up to 31 cocktails. This had to be trimmed down for two reasons: firstly I can’t drink 31 cocktails, and secondly I can’t be expected to get the likes of ‘chestnut liqueur’ or ‘lavender bitters’ for this experiment.

We ended up making and tasting 11 Christmassy cocktails:

  • Cranberry Moscow Mule (ChatGPT)
  • Santa’s Spiced Mule (ChatGPT)
  • The Holly Jolly Julep (ChatGPT)
  • The Yule Tide Toddy (ChatGPT)
  • Jingle Bell Julep: (Gemini)
  • Merry Mistletoe Mojito: (Gemini)
  • Mistletoe Mule (Mistral)
  • Eggnog Espresso Martini (Mistral)
  • Spiced Apple Cider Sangria (Mistral)
  • Gingerbread Martini Wonderland (Claude)
  • Reindeer’s Kiss (Llama)

Which AI’s recipes are hard to make?

60% of ChatGPT’s cocktails were eliminated for being too hard to make, 60% of Gemini’s, 50% of Mistral’s and 80% of Llama’s. Llama seems pretty bad on this metric, but it’s too little data to draw a conclusion.

AIs in general seem to have decided (see the raw output) that white chocolate liqueur + peppermint schnapps + vanilla vodka is the taste of Christmas. Those are obscure ingredients to source. Should we make those cocktails in a Part Two of this article? Let me know in the comments.

This is as much my fault as the AIs’ – I could have altered the prompt to exclude obscure ingredients. However, I wanted to keep the prompt simple for the purposes of this experiment, and to give all the models the same prompt for a level playing-field. If my aim had been only to make great cocktails (and not to test the models), a more specific prompt would be better.

The Yule Tide Toddy, by ChatGPT

ChatGPT says –

A warm, comforting cocktail with bourbon, honey, and herbs. 
**Ingredients:**        
- 2 oz bourbon        
- 1 tbsp honey        
- 1/2 oz fresh lemon juice        
- 4 oz hot water        
- Fresh thyme sprig (for garnish)   

**Instructions:**        
1. In a mug, combine bourbon, honey, and lemon juice.        
2. Pour in hot water and stir until honey dissolves.
3. Garnish with a sprig of fresh thyme.

Verdict: this is low on the creativity; the AI just took a normal hot whiskey and threw thyme at it to make it Christmassy. But doesn’t hot whiskey usually have cloves? Cloves are arguably more Christmassy than thyme “except for the fact that thyme looks like a Christmas tree”, according to one reviewer.

Two points worth commending: adding squeezed lemon juice is an improvement over the usual hot whiskey procedure of adding a slice of lemon, and the quantities of lemon juice and of honey are basically spot-on.

Overall, our group felt this was a middling cocktail.

Santa’s Spiced Mule, by ChatGPT

A holiday twist on the classic Moscow Mule with rich spices and cranberry.

**Ingredients:**
2 oz vodka
1 oz cranberry juice
1/2 oz ginger syrup (or fresh ginger juice)
1/2 oz fresh lime juice
4 oz ginger beer
Cinnamon stick and fresh cranberries (for garnish)

**Instructions:**
1. In a mule mug or glass, combine vodka, cranberry juice, ginger syrup, and lime juice.
2. Add ice and top with ginger beer.
3. Stir gently, and garnish with a cinnamon stick and cranberries.

Verdict: overall, this is a good drink. The smell of gingerbread syrup is nice, and goes well with the cinnamon and ginger beer – especially with the ginger beer because you see normally when you drink ginger beer it only has a mild ginger flavour, and adding a stronger ginger flavour cleverly fixes that.

The drink does taste Christmassy. “I think it’s too sweet personally,” said one reviewer, and it suffers from the same weakness as the Cranberry Moscow Mule below: too much sweetness, and arguably should have doubled the lime juice.

Cranberry Moscow Mule, by ChatGPT

A festive take on the Moscow Mule with fresh cranberry and lime.
**Ingredients:**
2 oz vodka
1 oz cranberry juice
1/2 oz lime juice
4 oz ginger beer
Fresh cranberries and lime wedges (for garnish)

**Instructions:**        
1. In a mule mug, combine vodka, cranberry juice, and lime juice.
2. Add ice and top with ginger beer.
3. Stir gently and garnish with cranberries and lime wedges.

This was overall one of the worst. It was sweet and not much could be tasted beyond ginger beer. There is quite a lot of mixer to vodka, and that took away the alcoholic bite. Cranberry juice isn’t assertive enough to redeem it, and lime juice could be, but not a half-measure. We commented on this and the other ChatGPT mule (above) that doubling the lime juice would be an improvement. And that’s exactly what the French model did –

Mistletoe Mule, by Mistral

- **Ingredients:**        
- 2 oz vodka        
- 1 oz cranberry juice        
- 1 oz lime juice        
- Ginger beer        
- Fresh cranberries and a sprig of rosemary for garnish

**Instructions:**       
- Fill a copper mug with ice.        
- Add vodka, cranberry juice, and lime juice.        
- Top off with ginger beer.        
- Stir gently and garnish with fresh cranberries and a sprig of rosemary.

In the mule head-to-head, Mistral is the clear winner. The ChatGPT mules were both too sweet and had too much ginger beer, and too little bite and lime. Mistral fixed this, doing just what the doctor ordered by doubling the lime juice. Is this to do with the way Americans like things sweet, and Mistral is French tech?

The other difference is the garnish was rosemary in place of lime wedges; this is probably also an improvement. “I think the rosemary gives something to it,” said one of our reviewers.

Merry Mistletoe Mojito, by Gemini

2 oz white rum            
1 oz lime juice            
1/2 oz simple syrup            
A handful of fresh mint leaves            
A few cranberries        

Muddle the mint leaves, lime juice, and simple syrup in a highball glass. Add the rum and fill the glass with crushed ice. Top with a splash of club soda and garnish with cranberries and a sprig of mint.

Verdict: all the flavours here are weak. We could taste the mint a little, and the lime a little, but it’s largely just rum and sugar. A bland cocktail and among the worst on the list.

Jingle Bell Julep, by Gemini

2 oz bourbon            
1 oz simple syrup           
A handful of fresh mint leaves   
A splash of club soda

Muddle the mint leaves and simple syrup in a julep cup. Add the bourbon and fill the cup with crushed ice. Top with a splash of club soda and garnish with a sprig of mint.

This was even worst than Gemini’s Merry Mistletoe Mojito; that mojito already was too sweet, and this one doubled the simple syrup. Jack Daniel’s (the bourbon we used), when paired with sugar and club soda winds up tasting something like off-brand cola.

The Holly Jolly Julep, by ChatGPT

A julep with a holiday twist using mint, pomegranate, and bourbon.        
**Ingredients:**
- 2 oz bourbon
- 1 oz pomegranate juice
- 1/2 oz simple syrup
- Fresh mint leaves
- Crushed ice

**Instructions:**        
1. Muddle mint leaves and simple syrup in a julep cup.        
2. Fill the cup with crushed ice and add bourbon and pomegranate juice.        
3. Stir well and garnish with a sprig of mint.

ChatGPT has been defeated in the mule head-to-head by Mistral. Can it redeem itself in the julep head-to-head? Yes it can; this is an interesting twist on the julep, because the combination of pomegranate juice and bourbon is creative and they do compliment each other. The pomegranate deepens the flavour of the Jack Daniel’s. ChatGPT avoided Gemini’s error of using too much simple syrup, and it had a stroke of real creativity with that combination.

Eggnog Espresso Martini, by Mistral

- **Ingredients:**  
- 2 oz vodka     
- 1 oz espresso  
- 1 oz eggnog        
- 1 oz coffee liqueur        
- Grated nutmeg for garnish        

**Instructions:**        
- In a cocktail shaker, combine vodka, espresso, eggnog, and coffee liqueur with ice.   
- Shake well and strain into a chilled martini glass.        
- Garnish with grated nutmeg.

This was a good-tasting cocktail – people were talking about giving it 8 or 9 out of 10 for taste. The nutmeg gives it a nose, which I liked as it added complexity, but it did draw some criticism from my friends who just dislike nutmeg. The balance of vodka and coffee was pretty good, but maybe if there was slightly less vodka it would be better. First you smell the nutmeg, then you taste the coffee, then you taste the booze – an interesting and complex sensory experience.

Spiced Apple Cider Sangria, by Mistral

**Ingredients:**
- 1 bottle of red wine
- 2 cups apple cider        
- 1 cup brandy        
- 1/2 cup orange juice        
- 1/4 cup lemon juice        
- 1/4 cup simple syrup        
- 1 apple, sliced        
- 1 orange, sliced        
- 1 lemon, sliced        
- Cinnamon sticks and star anise for garnish

**Instructions:**        
- In a large pitcher, combine red wine, apple cider, brandy, orange juice, lemon juice, and simple syrup. 
- Add sliced apple, orange, and lemon.        
- Stir well and refrigerate for at least 2 hours to let the flavors meld.        
- Serve over ice and garnish with cinnamon sticks and star anise.

Our reviewers said they could barely taste the alcohol, but I assured them it was there: a bottle of wine and a cup of brandy is not a weak brew. Sangria tends to hide its teeth. The fruit tastes boozier, and when I distributed forks and teaspoons, the fruit drew coos of approval.

Kiki suggested that mulled wine is a spiced Christmas drink, and sangria is a summery drink – maybe this was the AI’s attempt to hybridise the two? It’s an interesting concept that would be worth exploring more; for instance, you could prompt a LLM ‘Create a recipe for a Christmassy sangria inspired by mulled wine’.

The mulled wine notes came mostly from the cinnamon and the star anise. If I had doubled or tripled the amount, or the time spent brewing, they would have been stronger. (I put in 2-3 cinnamon sticks and 4-5 stars for about four hours: a reasonable amount. I would advise readers to be less reasonable and throw in fistfuls.)

While making it, I thought that ¼ cup lemon juice plus a lemon is a lot of lemon. It didn’t taste exactly sour and lemony, but dialling it back a little probably would be an improvement, we all agreed.

Overall a good drink. Nothing is terribly out of balance, though more spice and less lemon were suggested as improvements. Then I heated it up on the stove and served it warm. It smelled amazing warmed up, and we all preferred it.

Gingerbread Martini Wonderland, by Claude

A creamy, spiced cocktail featuring vodka, gingerbread syrup, and a graham cracker rim       
Ingredients:        
* 2 oz vodka        
* 1 oz gingerbread syrup        
* 1 oz heavy cream        
* Dash of cinnamon        
* Graham cracker crumb rim        
- Garnish with a tiny gingerbread man cookie and a sprinkle of nutmeg

A website told me to replace ‘graham crackers’ (which aren’t available to me) with “a mixture of digestve and rich tea biscuits”, so I did that and it worked well.

One reviewer said, “Tastes like a White Russian with a bit of biscuitty gingeriness; it’s like eating a ginger biscuit and drinking a White Russian at the same time”. This cocktail was well-liked in our group, and people commended its creativity, though it did receive some complaints for having too much vodka.

Reindeer’s Kiss, by Llama

* 2 oz Whiskey        
* 1 oz Spiced apple cider        
* 1/2 oz Maple syrup       
* 1/2 oz Lemon juice        
* Dash of Cinnamon powder        
Mix whiskey, spiced apple cider, maple syrup, and lemon juice in a shaker filled with ice. Strain into a rocks glass and sprinkle with cinnamon powder. Garnish with a cinnamon stick and a cherry, if desired.

I presume what they mean by ‘cider’ here (and also in the Spiced Apple Cider Sangria) is the American non-alcoholic kind. This ambiguity is annoying, but happens in human-written recipes as well. It wasn’t easy to get that ‘cider’ in Ireland, but Llewellyn’s Orchard make it in Lusk.

This was Llama’s only cocktail that got past the qualifying stages, four of its five suggestions having been eliminated as impractical. How did it do? Spectacularly. It was the best cocktail in the game with no close second. My coworkers are making it and I am going to the market to get the hard-to-source cider before Christmas Day. The whiskey (I used Jameson), the apple, the spice, are all present in perfect proportion, and it’s just delicious in a way I can’t put into words, so you know what to do.

Best and worst cocktails

Winner: Reindeer’s Kiss (Llama)

Honourable mention: Gingerbread Martini Wonderland (Claude), Eggnog Espresso Martini (Mistral)

Worst: Cranberry Moscow Mule (ChatGPT), Merry Mistletoe Mojito (Gemini), Jingle Bell Julep (Gemini)

How did the AI models compare?

Poor Gemini, it’s not looking good for you. The two cocktails that Gemini generated were bad. ChatGPT was hit-and-miss but pretty fine overall, and Claude and Mistral similarly held their own.

Llama did the worst in the qualifying round; I mean that it suggested things that I can’t really make. But it more than redeemed itself by suggesting the clear stand-out of the tournament. With a more specific prompt, it could perhaps improve on practicality. Going on this little data it is hard to draw solid conclusions, but either way we have a superb new recipe to tell our friends and readers about.

Let us know your thoughts! Sign up for a Mindplex account now, join our Telegram, or follow us on Twitter. 

Flying cars: Faster, Higher, Stronger

Posted on by Conor O'Higgins

In a video released in April 2022, German startup Volocopter stated their goal was to get their flying car zipping around the skies over the 2024 Summer Olympics.

Volocopter is one of hundreds of start-ups pursuing advanced air mobility – essentially flying cars – and their vehicle is called the Volocity. Publicly announcing timelines in technology is usually considered a bad idea, and as late as February 2024, Politico rubbished their bold claim.

Volocopter can now gloat at the doubters. Last week they announced Olympic success. The technology and legal permissions will be ready for the Paris Olympics, the opening ceremony of which is today, 26 July. People will indeed be able to use flying taxis to hop around Paris during the upcoming Olympics, and President Macron is invited to go for a spin.

It is an interesting moment for electric aviation. Battery technology is boldly tearing forward with sodium-ion, solid-state, silicon anode, and other improvements. Cars, boats, and trains are switching to electric motors. “The electrification of aircraft is leading to many opportunities to develop fundamentally new configurations and to take advantage of distributed and mechanically disconnected propulsion”, one paper says. At the same time: drones have established themselves as a major industry, thanks largely to improvements in computer technology that can control many rotors at once. Most consumer and military drones are electric but unmanned, but it’s time to revisit the question of 1950s futurists: When can I get my flying car?

Credit: Conor O’Higgins using Stable Diffusion

eVTOL and eSTOL

There are two approaches to building flying cars: eVTOL and eSTOL. These stand for ‘electric vertical takeoff and landing’ and ‘electric short takeoff and landing’.

Look at a map of a city and you’ll see that in most cities, airports are the largest plots of land dedicated to any single purpose. There are two reasons for this: space and noise. Planes need long takeoff and landing distances (1.5km or 2km), and they are noisy, so they have to be set apart from homes and schools and stuff. These limitations often force airports outside of cities, and that adds an hour to your trip to Spain.

Dublin-airport-Mindplex
Dublin airport (outlined in orange) compared to Dublin. Note that the airport is huge, maybe 5% the size of the city herself, and is set apart from the city. (Credit: OpenStreetMap)

eVTOLs promise a quieter flying machine that can take off and land vertically (say on a roof). eSTOLs also aiming to run with quiet electric motors, and to take off and land a very short runway. Whereas commercial planes need 1.5-2km runways, eSTOLs aim to need “between 45 and 90 meters of runway” according to the COO of Electra.aero.

This video from Electra beautifully illustrates what an eSTOL should be: it’s simply a plane, but a battery-powered one optimised for urban deployment:

Flying with rotors, flying with wings, a bit of both

Why would anyone build eSTOLs for short runways when eVTOLs with zero runway are an option? It’s because the engineering is significantly simpler: an eSTOL uses familiar airplane technology: it drives forward to create lift under its wings and take to the sky. It is an adapted plane. This allows engineers to use well-understood principles and try to win the race-to-market.

eVTOLs are a little more complex. Most of them (Volocopter’s Volocity is one of the exceptions) require engineers to think about two phases: one where rotors or ducts pick it up and set it down for the vertical takeoff/landing, and another fixed-wing mode for cruising around. This second plane-like flight is quieter and more energy-efficient than the drone-like one.

Little planes or big drones?

Therefore eSTOLs should theoretically be simpler and cheaper than eVTOLs, but in this bubbling, surging industry, design philosophies abound. The Volocity by Volocopter that will share the skies with the Olympic polevaulters is an eVTOL with a single propulsion-type. It takes off and lands vertically like a drone, and then – unlike its competitors Joby, Beta Technologies, Lilium etc. – it continues to fly around in the manner of a drone. The simplicity of this design philosophy got the Volocopter Volocity to market before their competitors. Whereas eSTOLs are small planes, Volocopter’s machines are big drones.

A good Forbes article said, “It’s been estimated that around 300 different companies are trying to build new “flying car” electric VTOL aircraft for the anticipated revolution, and there are almost as many different design philosophies. Most are opting for hybrid designs that feature rotors for vertical takeoff and landing, but regular fixed wings for horizontal flight. There’s a good reason for that — fixed wing flight is much more efficient, and for electric aircraft, battery weight is the key issue, and that makes efficiency really important. In spite of this, one of the companies furthest along in actual deployment is using a much more basic electric multirotor design with no fixed wings. It’s effectively a human sized drone”.

So Volocopter’s approach appears to have won on simplicity and quick deployment. The tradeoffs? Range and noise. Flying with rotors is less energy-efficient than fixed-wing flight, so you drain your battery quicker. The Volocity has a range of 35km.

Is Volocity’s 35km range a problem? Not really. You only need your vehicle’s range to be as long as your trip. Hundreds of kilometers is for city-to-city. A range of 35km is grand for hopping around one city.

For comparison, a video released by Joby, another of a many eVTOL startups, in 2021, says they flew 154.6 miles (248.8km) in a test. That was their battery-powered version; for longer ranges, they feel hydrogen is the way to go, and last month (June 2024) flew their hydrogen eVTOL for 523 miles.

An invaluable resource for tracking and comparing these projects is the Advanced Air Mobility Reality Index. Tech news is so full of hype and self-awarded “breakthroughs” that it’s handy to have someone independent keeping track of the projects and their technology readiness levels. It provides a league table (Volocopter is currently winning) that usefully lists the intended use-cases (flying taxi being the most common) and whether the bird is piloted or autonomous.

The Volocity aircraft, made by Volocopter, flying at the Paris Air Show in 2023. (Credit: Photo by user Ibex73 from Wikimedia Commons)

Noise

I used to write about breaking tech that would change the world. Now I write about breaking tech that won’t make such a racket. I must be getting old.

The noise made by helicopters is a major nuisance and has limited them to very infrequent flying. And if you’ve ever been near a drone, you know that their noise is not a detail.

Noise is the hardest engineering problem in eVTOL/eSTOL, and, in my opinion, the biggest shortcoming of Volocopter’s Olympic success. Electric cars run much quieter than combustion ones, but no such luck with aircraft. The noise from helicopters, drones, and airplanes doesn’t mostly come from the engine, but mostly from the interaction of the craft and the air (Likewise, electric cars are actually just as noisy as combustion cars at high speed, because wheel-noise is more important than engine-noise).

A valuable paper on low-noise electric aircraft puts it this way: “Although these aircraft use quiet electric motors instead of noisier combustion engines, this is not likely to have a significant effect on the overall noise radiation of the vehicle, because the noise of rotor and propeller driven aircraft is generally dominated by the aerodynamically-generated noise of the rotating blades. Instead, the main acoustic impacts of electrification are a result of the new freedoms of electric propulsion, especially distributed electric propulsion, offered to the aircraft designer”.

The fluid dynamics of why flying machines are noisy is extremely complex. eVTOLs spin their rotors at slower speeds than helicopters do to avoid the loud thwap-thwap noise that helicopters make. Yet turbulent flows bumping the body of the bird cannot be fully avoided, and complex multirotor designs send vortexes all over the place, including into collusion with their neighbouring rotors, creating noise.

Progress is being made on these problems (for example, a March 2023 paper found that using six blades instead of four lowered noise by 5 to 8 decibels, losing only a 3.5% thrust in the process. Volocopter say they are using “the lowest disc loading currently on the market… and a low RPM (revolutions per minute) rate” to reduce noise (‘Disc loading’ is the ratio of the bird’s weight to the area of its rotors; Volocopter positively bristles with rotors.)

This detailed fluid dynamics work can chip away at the noise problem – sensible work for centibel gains – but the reality of the engineering says don’t expect a masterstroke that will suddenly make eVTOLs 100× quieter. The only technology that may be an exception is the Lilium Jet. Instead of rotors, it uses 36 small ducted fans (so that the ducts trap sound), and is an undeniably beautiful-looking product:

Lilium sits at a mediocre 11th on the ‘Reality Index’, and until independent tests have certified the noise-levels of each experimental eVTOL, we have to be cautious.

A loud note of caution

I agree with the need for electric urban mobility. But let me question the need for flight.

‘Energy-efficient flight’ is, to some extent, a contradiction in terms. Under 0.25% of international freight is transported by plane (most goes on ships), and there’s a reason for that: energy costs. When you have to expend energy to do move a load from A to B, why also expend energy fighting gravity? You need a justification. Traffic might be a justification, but there are simpler ways to solve that (such as bicycle lanes). Emergency vehicles such as ambulances should fly: they have an excellent justification.

The noise issue is the hardest engineering problem around flying cars. Safety can be solved. Autonomy can be solved. eVTOL companies are quick to make impressive claims about their quiet birds, so we need an independent agency inspecting them for sound and publishing data. As this data is missing, and as I haven’t had the chance to get up close, we have to use some guesswork to figure out how loud they are.

I lean towards skepticism, and suspect that Volocopter have not made much progress on the sound problem. The approval from the City of Paris means they have achieved good safety standards, but it comes with limitations: they aren’t just allowed zip anywhere at any time. The authorities want them flying a limited number of flights at constrained times of the day. Volocopter’s website says they will “map out routes inside the city that ensure the Volocopter aircraft do not generate a cacophony that exceeds the city’s permitted noise levels. Part of its approach will involve flying at specific times of day.” This strongly implies there is quite a lot of noise. The most skeptical interpretation is to say they’ve simply built a small helicopter: those are also approved to fly low flight volumes at major events like the Olympics.

Let us know your thoughts! Sign up for a Mindplex account now, join our Telegram, or follow us on Twitter. 

Paul Cockshott on the Economics of Automation

Posted on by Conor O'Higgins

Paul Cockshott is a computer scientist and economist, with an interest in how advanced computerised planning can supplant our existing economic order. I spoke with him about how Artificial Intelligence will automate jobs that people currently do. The discussion focuses on the economic costs of training AI models, how they weigh up against labour costs, and the economic circumstances under which human jobs will be automated. Over to Paul Cockshott:

* * * * * * *

The use of AI requires a great deal of processing power. It needs it in two distinct ways. The first is in training, and the second is in application. 

Let’s look at the training aspect. This has become feasible because of two developments over the last 15 years, in data and in power. 

Data

The build-up of large collections of images and text on the internet that can be used as training data for neural networks. I recall back in the 90s when a team I worked with was developing neural network video encoding, one of our problems, pre internet, was just getting and collections of image data to train with. We resorted to capturing TV broadcasts and training neural nets on those. Now, of course, due to Android camera phones, Google has almost unbounded collections of images from around the world on which neural nets can be trained for vision purposes. In addition, there are vast quantities of indexed images on the web, with dubious ownership, that smaller groups like Stability.AI can use. The same applies to text. It is the ready availability of a vast corpus of academic papers and books that makes systems like ChatGPT and Bard able to answer questions, if not like an expert, at least like a 3rd year student. 

Power 

Actual nervous systems work by electrochemical means to aggregate multiple discrete impulses to produce a discrete response. The Church–Turing–Deutsch principle states that any physical system can be emulated to an arbitrary degree of accuracy by a universal computing machine. This includes the semi-analogue, semi-digital processes that occur in nervous systems. Whilst this has been theoretically known at least since the 1980s and in informal terms since the 1950s, it was, until recently, impractical to apply on a large scale.

To emulate the analogue aspects of synaptic responses requires a great deal of floating-point arithmetic, more specifically it requires a lot of matrix to vector multiplication. A lot of work from the 1960s went into developing supercomputers for matrix mathematics, since these techniques turn out to be of very general applicability in the physical sciences. By the end of the last century, this had produced machines that were well able to handle tasks like climate simulation and weather forecasts. 

But the scale of maths required by artificial intelligence was considerably greater. The human brain contains tens of billions of neurons, and each neuron would have to be represented by a vector of synaptic weights. If each neuron has or the order of 10,000 synaptic weights and can fire about 10 times a second, we would require a vector processor of from 10¹⁵ to 10¹⁶ operations per second to emulate the brain: that is to say it would have to reach the petaflop range. 

The first machines in this range became available about 12 years ago. Last year, Tesla launched its Dojo supercomputer complex with a processing power of 10¹⁸ operations a second. That makes it equal to around 100 human brains in processing rate. The downside is the power usage – in the region of 1-2 megawatts. In comparison, the metabolic energy consumption of 100 human brains would be of the order of 1.5 kilowatts, so the Dojo system is about 1,000 times as energy intensive. 

The machine is built of 120 individual ‘training tiles’ as shown below. 

 

Credit: Paul Cockshott

However, at this point we are just comparing operations per second, not information storage. A brain with 80 billion neurons each with 15,000 connections would have 1.2 quadrillion weights. Tesla stores its weights in cfloat8 format, so that each of their training trays can store about 11 billion weights or about 1/100,000 of a human brain. 

So the current best Tesla technology is 5 orders of magnitude behind the human brain in storage, and 3 orders of magnitude behind in energy-efficiency: overall about 8 orders of magnitude away from the storage x power efficiency of the human brain. 

The consequence is that whilst it is possible, by consuming megawatts of power, to train a system on a specialist skill like driving, it is not yet possible to incorporate a human level of intelligence and knowledge into the car itself.  

A human can be taught to drive with a few tens of hours of driving instruction, and they can still do other jobs after they have driven to work. Tesla must spend years of processing time at a huge power bill to obtain the set of neural weights that a person needs to drive. 

Credit: Tesfu Assefa

Of course, the Tesla business plan is to train once and then replicate the information in all their cars. But the size and power hungry nature of the chips at present prevents them being put in each car. 

It will take some time, one or two decades, before the energy × storage efficiency of chips reaches a point where mobile robot devices with general intelligence comparable to humans are likely to be available. So, to harness general AI, a lot of effort must go into improving power consumption and memory capacity of chips. Until that point, it will be only available as remote online services running on big data centers. 

These in turn will increase demand for electricity at a time when, due to environmental considerations and political instability, energy is becoming scarce and expensive. The implication is that an ever-increasing share of GDP is going to have to be directed to producing non-fossil fuel energy infrastructure. 

Let us know your thoughts! Sign up for a Mindplex account now, join our Telegram, or follow us on Twitter.