Perspective on Risk - Dec. 9, 2022 (Technology)
AI & ML; Quantum Computing; Blockchain; Energy, Batteries & Electric Vehicles; Robots Need Jobs Too
I’ve written in the past about my three-driver long-term framework, and I’ve done a long piece on Globalization. Back when at AIG I produced a newsletter highlighting aspects of technological change. This Perspectives is in that vein. It provides some information on trends in computing and energy that will be important; I don’t focus on things like CRISPR and other areas that might be transformational. I also don’t dwell much on the implications in this post; those will come out in subsequent posts. But the TL;DR is that the world in 5 years may be very different from today.
WARNING: this Perspectives is quite long1; if it doesn’t all come through on email you may have to jump over to the Substack site.
Perhaps the two best things Peter Hancock did for me in his time as AIG CEO was 1) to help frame a future driven by non-linear change, and 2) to help recognize that human judgment is just an unobservable internal algorithm or heuristic that is likely inferior to an unbiased model. So I clearly appreciated this piece of research.
Algorithm appreciation: People prefer algorithmic to human judgment (Science Direct)
Even though computational algorithms often outperform human judgment, received wisdom suggests that people may be skeptical of relying on them (Dawes, 1979). Counter to this notion, results from six experiments show that lay people adhere more to advice when they think it comes from an algorithm than from a person.
Paradoxically, experienced professionals, who make forecasts on a regular basis, relied less on algorithmic advice than lay people did, which hurt their accuracy. These results shed light on the important question of when people rely on algorithmic advice over advice from people and have implications for the use of “big data” and algorithmic advice it generates.
Artificial Intelligence & Machine Learning
“the best approximation to what we know is that we know almost nothing about how neural networks actually work and what a really insightful theory would be,” said Boris Hanin, a mathematician at Texas A&M University and a visiting scientist at Facebook AI Research who studies neural networks.2
The Impact of AI
Our World in Data has written a few articles recently on AI.3 We are at the tipping point, where machines are overtaking humans in many areas. This is only likely to accelerate during our lifetimes.
As of 2020, AI systems out-perform humans on handwriting, speech and image recognition, reading comprehension, and language understanding tests.
Generative AI is rapidly progressing (images, text).
While the early systems focused on generating images of faces, these newer models broadened their capabilities to text-to-image generation based on almost any prompt. The image in the bottom right shows that even the most challenging prompts – such as “A Pomeranian is sitting on the King’s throne wearing a crown. Two tiger soldiers are standing next to the throne” – are turned into photorealistic images within seconds
When will we have “human-level”4 AI?
The futurist Ray Kurzweil has long been making accurate predictions about the pace with which computers will surpass human capabilities. In 2017, he formally stated that “By 2029, computers will have human-level intelligence … [and] I have also set the date 2045 for singularity — which is when humans will multiply our effective intelligence a billion fold, by merging with the intelligence we have created.”
When Our World in Data asked AI experts, “half of the experts gave a date before 2061, and 90% gave a date within the next 100 years.”
Similarly, when they asked a group of professional forecasters, “the forecasters believe that there is a 50/50-chance for an ‘Artificial General Intelligence’ to be ‘devised, tested, and publicly announced’ by the year 2040”
GPT-4 & ChatGPT
GPT-3, GPT-4 and ChatGPT are called ‘large language models (LLM); LLM are generative AI models; LLM are ‘artificial intelligence tools that can read, summarize and translate texts and predict future words in a sentence letting them generate sentences similar to how humans talk and write.’ GPT-3 is the current state of the art, but its successor is on the immediate horizon.
GPT-4
OpenAI’s GPT-4: The Much-Anticipated Follow-Up to GPT-3 (Medium)
GPT-4 is a diffusion model, which means that it uses machine learning algorithms to quickly generate answers to questions. This upgrade from GPT-3 is said to give the model real capabilities, allowing it to answer complex queries and perform better than its predecessor. OpenAI has also created a demo of WebGPT, a chatbot powered by GPT-4, which can answer questions with a level of factuality that was not possible with GPT-3.
GPT-4 is an improvement from its predecessor in several ways. Its larger size and more complex diffusion models make it better equipped to answer questions more accurately than GPT-3. Additionally, GPT-4 has significantly more real capabilities than GPT-3, including the ability to generate long pieces of text that are more accurate and coherent.
ChatGPT
OpenAI “trained a model called ChatGPT which interacts in a conversational way. The dialogue format makes it possible for ChatGPT to answer followup questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests.” This is essentially a better front-end for GPT-3; it is answering based on the collective knowledge of what is in its training data; it is not reasoning, but reflects the sequence of thinking in its dataset. You can try it yourself by clicking on the above link.
ChatGPT just crossed 1 million users; it's been 5 days since launch.
Here ChatGPT answered a quantum physics question correctly:
Ben Thompson of Stratergy (a great newsletter) summarizes these large language models as essentially a “conventional wisdom generator.” However he notes that “making conventional wisdom available is very valuable.”
AI is making it very cheap to produce mid-level writing, so the supply of mid-level writing will soar and the compensation for mid-level writing will plummet. AI is making it cheaper to provide answers; the compensation for providing answers will fall, and the compensation for asking how to ask the right question, and when to ask the question, will rise.
How ‘intelligent’ is ChatGPT?
Well, of course, someone had ChatGPT take an SAT test. Scored at the 52nd percentile, despite having problems with graph and picture questions. Sergey Ivanov gave chatGPT an IQ test; its came in with an IQ of 83 (low average). @emollick gave GTP-3 standard human creativity tests and it basically maxes out the two most common: the RAT remote association word test, and the AUT alternative uses test. @debarghya_das gave ChatGPT the Myers-Briggs personality test and it is an ISTJ (Logistician). David Rozado determined that ChatGPT has a liberal libertarian political bent.
Some examples of what ChatGPT has been able to do:
Hold a 20 minute conversation about the history of modern physics.
Mimic a 1970s dial-in test-based game based on Harry Potter
Write a script for a Seinfeld episode where Jerry needs to learn the bubble sort algorythm.
ChatGPT seems particularly good at debugging computer code
For some/many questions in provides more usable answers than Google.
It answers Reddit’s favorite question: “In a fight between one horse-sized duck and a hundred duck sized horses, who would win?”
It even seems to handle the Monty Hall problem
However, ChatGPT confidently presents its information, even when it is wrong.
For instance, it explained why an abacus is faster than a GPU for machine learning, sometimes it doesn’t understand that kids are indivisible (though other times it does), and the work of Thomas Hobbes.
This is completely wrong.
Anyway, I’ve already added ChatGPT to my phone. You can easily do it too.
Perplexity.ai integrates ChatGPT with the Bing search engine. Quite good. I’m using it now before I choose to Google.
For your power-user staff, ChatGPT is particularly good at quickly giving you needed Google Sheets formulas. You can also
Image Models (Dalle-e, Midjourney) are Generative Models
As we discussed in earlier posts, generative AI now is considered to pass the Lovelace Test, which is the Turing Test equivalence for creativity.
ChatGPT meets Midjourney
The logical evolution is to start to ‘wire’ several of these advanced AIs together. Here, Twitter user @arbedout asks ChatGPT for a prompt that can be used in the generative art AI Midjourney. Generative text creates prompts for generative art.
What Other New Things Can Artificial Intelligence Do Better Than Humans?
AI Discovers a Faster Way to Solve a Fundamental Math Problem
DeepMind’s game-playing AI has beaten a 50-year-old record in computer science (MIT Technology Review)
WHAT: DeepMind has used its board-game playing AI AlphaZero to discover a faster way to solve a fundamental math problem in computer science, beating a record that has stood for more than 50 years.
WHY IS IT IMPORTANT: there are more ways to multiply two matrices together than there are atoms in the universe (10 to the power of 33, for some of the cases the researchers looked at). “The number of possible actions is almost infinite,” says Thomas Hubert, an engineer at DeepMind.
HOW: The trick was to turn the problem into a kind of three-dimensional board game, called TensorGame. The board represents the multiplication problem to be solved, and each move represents the next step in solving that problem.
Physicists Are Using AI to Find Hidden Relationships in Data
Powerful ‘Machine Scientists’ Distill the Laws of Physics From Raw Data (Quanta)
Symbolic regression similarly identifies relationships in complicated data sets, but it reports the findings in a format human researchers can understand: a short equation. … [I]n recent years the algorithms have grown mature enough to ferret out undiscovered relationships in real data — from how turbulence affects the atmosphere to how dark matter clusters.
The goal of symbolic regression is to speed up such Keplerian trial and error, scanning the countless ways of linking variables with basic mathematical operations to find the equation that most accurately predicts a system’s behavior.
A Bayesian machine scientist to aid in the solution of challenging scientific problems
Here, we introduce a Bayesian machine scientist, which establishes the plausibility of models using explicit approximations to the exact marginal posterior over models and establishes its prior expectations about models by learning from a large empirical corpus of mathematical expressions. It explores the space of models using Markov chain Monte Carlo. We show that this approach uncovers accurate models for synthetic and real data and provides out-of-sample predictions that are more accurate than those of existing approaches and of other nonparametric methods.
AI Learns To Negotiate
AI now is superior to humans at quite a few games (chess, Jeopardy!, Go, Starcraft 2 and other video games). Now, Meta has created an AI, CICERO, that has bested humans at Diplomacy. This is a big deal because players (typically human) must collaborate, form alliances, communicate strategy, coordinate moves with other players and/or strategicly lie in real-time.
Yann LeCun, the Chief AI scientist at Meta, calls this a pretty big deal.
Human-level play in the game of Diplomacy by combining language models with strategic reasoning (Science)
We introduce Cicero, the first AI agent to achieve human-level performance in Diplomacy, a strategy game involving both cooperation and competition that emphasizes natural language negotiation and tactical coordination between seven players. Cicero integrates a language model with planning and reinforcement learning algorithms by inferring players’ beliefs and intentions from its conversations and generating dialogue in pursuit of its plans. Across 40 games of an anonymous online Diplomacy league, Cicero achieved more than double the average score of the human players and ranked in the top 10% of participants who played more than one game.
AI is Being Used in Earthquake Prediction
Artificial Intelligence Takes On Earthquake Prediction (Quanta)
In a new preprint, A Silent Build-up in Seismic Energy Precedes Slow Slip Failure in the Cascadia Subduction Zone (arxiv), researchers claim that their machine learning algorithm can predict the start of a slow slip earthquake to “within a few days — and possibly better.”
Driverless Cars Are Beginning To Arrive
Uber riders can now hail an autonomous ride in Las Vegas (Yahoo)
Here's how it works – in the Uber app, customers can select UberX or Uber Comfort Electric for the opportunity to be matched with a robotaxi. Then, customers will have to opt in if they're matched to an autonomous vehicle and, to get into the car, there's also a button they will have to press in the app.
Next, Uber plans to launch its public robotaxi service in Los Angeles.
AI Maturity by Industry
Quantum Computing
This article, What Makes Quantum Computing So Hard to Explain? (Quanta), explains what makes quantum computing potentially so powerful and important. Quantum computers will differ in structure from typical computers, and are particularly well-suited for certain applications. Specifically, quantum computers
achieve better “scaling behavior,” or running time as a function of n, the number of bits of input data. This could mean taking a problem where the best classical algorithm needs a number of steps that grows exponentially with n, and solving it using a number of steps that grows only as n2.
They are particularly hard to create due to
decoherence, which means unwanted interaction between a quantum computer and its environment — nearby electric fields, warm objects, and other things that can record information about the qubits. This can result in premature “measurement” of the qubits, which collapses them down to classical bits that are either definitely 0 or definitely 1.
There continues to be systematic progress in creating real-world quantum computers.
IBM pushes qubit count over 400 with new processor (ArsTechnica)
With more than three times the qubit count of its previous-generation Eagle processor, Osprey is the first to offer more than 400 qubits, which indicates the company remains on track to release the first 1,000-qubit processor next year.
From Forbes The State Of Quantum Computing: Future, Present, Past, the first practical quantum computers are about 5 years away.
Doug Finke, Managing Editor, Quantum Computing Report: The next five years of quantum computing will be the era of the NISQ [Noisy Intermediate-Scale Quantum] machine and we will see increasingly powerful NISQ machines being introduced. Although there might be a few applications that will use these to achieve a quantum advantage, most potential quantum applications still won’t find these NISQ machines powerful enough to outperform classical computing-based solutions. However, by the end of the five-year period we will start seeing the emergence of error-corrected fault-tolerant quantum processors and this will be the inflexion point for large-scale quantum computing adoption in real world applications.
Blockchain
In AWS & Blockchain, Tim Bray recounts conversations with Andy Jassy (AWS) and numerous Wall Street firms back in 2016 when there was a huge moment where every database was going to be moved onto a Blockchain. It really is a great read.
We really only had two questions, both for the big-finance players and for the startups. “What is it you want to do?” and “How does blockchain help?” ¶
The answers, to our disappointment
“Ledgers are useful, cryptography tech is useful, blockchains aren’t, the field is full of grifters, but we could build distributed-ledger infrastructure and then these cool services on top of it.”
AWS decided [in 2016] not to make a strategic investment in blockchain
In the WSJ, David Solomon of Goldman recently stated:
I think bitcoin is really not the key thing. The key thing is how can blockchain or other technologies that are not developed yet accelerate the pace of the digitization of the way financial services are delivered.
However, some of the earliest blockchain projects have recently pulled the plug.
ASX Blockchain Project Fails
ASX is the Australian Securities Exchange, and they have been working on a project to replace their settlement and clearing platform with a blockchain-based system. However, they have recently paused the project after reviews and reassessments. ASX is a long-running exchange with a history of world-firsts and new technologies, and their blockchain project was intended to provide opportunities for their people to feel valued and undertake meaningful work. Unfortunately, the project has been marred by delays and concerns from users.5
ASX was one of the first high-profile attempts to adopt a distributed ledger based on the blockchain for a major commercial purpose. Many of you may have heard of ASX as its technology partner was Digital Asset Holdings, Blythe Masters (formerly of JPMC and credit derivative fame) firm.
You can read the Accenure report, ASX CHESS Replacement Application Delivery Review, for details. The techies among you will love it.
From the Australian Financial Review ASX grip on clearing shaken by blockchain disaster:
ASX pulled the plug on a seven-year project to replace the ageing CHESS system – which transfers ownership and manages payment for equities – with distributed ledger technology known as blockchain.
It will write off $245 million to $255 million and may have to compensate trading firms that spent at least $100 million on their own upgrades for the experimental clearing and settlement project.
This came after a devastating report from Accenture identified multiple problems with the beleaguered project including uncertain timelines, communication issues with technology vendor Digital Asset and excessive complexity.
TradeLens Blockchain Project Fails
IBM's Tradelens blockchain project was introduced in 2018 as an open and neutral supply chain platform underpinned by blockchain technology, with the goal of improving global trade. It quickly gained nearly 100 users but has since been discontinued due to a lack of technology to deliver on what was promised.6
IBM and Maersk Abandon Ship on TradeLens Logistics Blockchain (Coindesk)
Maersk and IBM will wind down their shipping blockchain TradeLens by early 2023, ending the pair’s five year project to improve global trade by connecting supply chains on a permissioned blockchain.
TradeLens emerged during the “enterprise blockchain” era of 2018 as a high-flying effort to make inter-corporate trade more efficient. Open to shipping and freight operators, its members could validate the transaction of goods as recorded on a transparent digital ledger.
The idea was to save its member-shipping companies money by connecting their world. But the network was only as strong as its participants; despite some early wins, TradeLens ultimately failed to catch on with a critical mass of its target industry.
Central Bank Digital Currencies
So with the implosion of the crypto-currency industry, attention has refocused on CBDCs.
A Central Bank Digital Currency (CBDC) is a digital form of central bank money that is widely available to the general public. It is an electronic form of fiat currency with potential wide use by households and businesses to store value and make payments, and it is a claim on the central bank instead of printed money.7
Recently, the NY Fed’s New York Innovation Center has started a pilot, Facilitating Wholesale Digital Asset Settlement, that will test how banks using digital dollar tokens in a common database can help speed up payments.
In the 12-week project—the Regulated Liability Network U.S. Proof of Concept—the NYIC will experiment with the concept of a regulated liability network (RLN). RLN is a concept for a financial market infrastructure (FMI) facilitating digital asset transactions that connect deposits held at regulated financial institutions using distributed ledger technology.
This theoretical FMI provides a multi-asset, always-on, programmable infrastructure containing digital representations of central bank, commercial bank, and regulated non-bank issuer liabilities, denominated in U.S. dollars.
Participants include BNY Mellon, Citi, HSBC, Mastercard, PNC Bank, TD Bank, Truist, U.S. Bank and Wells Fargo. Swift, the global financial messaging service provider, is supporting interoperability across the international financial ecosystem.
In addition, the BIS provided an informative report on Central bank digital currencies in Africa.
The interest of African central banks in CBDCs has shot up in recent times. While all of those surveyed are analysing CBDCs, only few have projects at advanced stages (pilot or live). Some countries, in particular in East and West Africa, stand out as promoting fast payment systems through mobile money, but half of the surveyed central banks think that CBDCs can provide a superior solution.
Energy, Batteries & Electric Vehicles
We are at, or rapidly approaching, a number of tipping points. The IEA thinks solar will exceed coal in about three years.
Renewable capacity expansion in the next five years will be much faster than what was expected just a year ago. Over 2022-2027, renewables are seen growing by almost 2 400 GW in our main forecast, equal to the entire installed power capacity of China today. That’s an 85% acceleration from the previous five years, and almost 30% higher than what was forecast in last year’s report, making it our largest ever upward revision. Renewables are set to account for over 90% of global electricity capacity expansion over the forecast period. The upward revision is mainly driven by China, the European Union, the United States and India, which are all implementing existing policies and regulatory and market reforms, while also introducing new ones more quickly than expected in reaction to the energy crisis.
Policy efforts are turning hydrogen production from wind and solar PV into a new growth area
This power will facilitate a transition away from combustion automobiles. Here, the US is behind, but is also approaching a tipping point.
In the first quarter of 2020, just 3.5% of Chinese car sales were electric vehicles. In the first quarter of 2022, it was 26%. EVs represent 25-30% of new car sales in Europe and China. Norway has the highest percentage of electric vehicles in new car sales, with EVs making up 86% of all car sales in the country.
The share of electric vehicle models in new vehicle sales in the US has been increasing, with it being 2.4% in 2020 and projected to reach 25.4% by 2028.
No American car company is yet in the top-5 in EV sales; the top five slots belong to BYD, Tesla, SAIC, Volkswagen and Geely-Volvo.
Three firms currently make 70% of the batteries in EVs; CATL, LG Energy Solution and Panasonic.
EV batteries currently require large quantities of the minerals Lithium, cobalt, nickel, copper and several rare-earth minerals.8
The green transition will make some economies rich overnight. But most "electrostates" are ill-equipped to manage windfalls. The majority of the world’s 96 commodity-linked SWFs are backed by sales of fossil fuels; only 7 metals exporters have established rainy-day funds.
Robots
Amazon confirms its latest warehouse robot uses AI to handle millions of items
Sparrow is Amazon's latest warehouse robot that is designed to manage specific inventory. According to the company, Sparrow is its very first robotic system in our warehouse that "can detect, select, and handle individual products in our inventory." Amazon explains that Sparrow leverages an artificial intelligence system that has been fed millions of items giving it the ability to recognize these items, pick them up, and place them into the desired location.
Robots Need Jobs Too!
One reader suggested I wrote too many Perspectives; let’s see how he likes dealing with a single Perspective of such prodigious length.
Foundations Built for a General Theory of Neural Networks. Quanta Magazine
A human-level AI would be a machine, or a network of machines, capable of carrying out the same range of tasks that we humans are capable of. It would be a machine that is “able to learn to do anything that a human can do”, as Norvig and Russell put it in their textbook on AI. Human-level AI was defined as unaided machines being able to accomplish every task better and more cheaply than human workers.
This paragraph was written by Perplexity.ai in response to the query “Provide me with some background on the Australian ASX project.”
Ibid.
Ibid.
Japanese Tesla Supplier to Take on Chinese Rivals to Tap EV Boom (Bloomberg)
Japan’s Sumitomo Metal Mining Co. wants to start making materials for a cheaper type of electric-vehicle battery -- containing no nickel -- that’s poised for global popularity after storming the world’s top EV market, China.