Trusting AI, bitcoin and the future media

Glimpse into the near future

I curate a weekly newsletter Exponential View. Between emojis, I write about the delicate dance between exponential technologies and our society.

DeepMind’s new AlphaZero can teach itself to become super-human in a variety of different games in a matter of hours. Analysis on Chess.com is particularly good, pointing to a different quality to AlphaZeros’ tactical and positional play. From the original, and highly accessible, paper:

AlphaZero searches just 80 thousand positions per second in chess and 40 thousand in shogi, compared to 70 million for Stockfish and 35 million for Elmo… AlphaZero compensates for the lower number of evaluations by using its deep neural network to focus much more selectively on the most promising variations — arguably a more “human-like” approach to search.”

This more efficient method of searching the probability space seems to be important. Many real-world applications involve far too many degrees of freedom to be ‘brute forced’. Figuring out methods which focus on most promising course of actions is an important step forward. Of course, as Gary Marcus points out, AlphaZero still requires a large amount of human engineering to achieve this remarkable task.

(Separately, The Economist concludes that Google leads the world in artificial intelligence. This broadly aligns with my view, although I remain intrigued by the longer term prospects for the BATs, as I wrote here. Also, see great discussion on Baidu’s investments in AI.)

🤑💨 Might Bitcoin cost us our clean energy future? At the current rate of energy consumption, the network will consume as much electricity as the whole world uses today by Feb 2020. Bitcoin’s 32TWh of annual usage (growing 20–30% per month) is more than the total solar output of all but a handful of nations.

🤷‍ Umair Haque: How the way we think about the world failed. THOUGHT-PROVOKING ESSAY

💼 Four predictions of the future of work: more jobs, if we guide innovation; cities will compete for talent; an increase in freelancing; more flexible education, argues _EV_reader, Stephane Kasriel. For a robust data view of job trends, which backs Kasriel’s argument, LinkedIn’ Emerging Jobs research shows that tech is the engine of job growth, both complex engineering ones and those reliant on a harmony of soft skills. Mobility matters, as does education: many of the top jobs didn’t exist a few years ago. Brookings has a wonderful graph which illustrates the tech-driven ‘hollowing out’ of the US labour force. (Elsewhere, in the US gay men earn about 10% more than straight ones, on average. And the journey to be an independent creator is a long one. Fewer than 2% of users on Patreon make the minimum wage.)

🗝️ John Hagel on helping incumbent organizations manage the ‘big shift’, the fundamental transitions we’re witnessing today.

Every large institution has an immune system … never underestimate it.

📃 Trapped in Bork’s paradox: internet giants are eating their competition away, but the competition can’t do without them. Do anti-trust rules need to be rewritten? (See also: Google and Amazon are going against each other publicly, putting customers last.)

DEPT OF TRUSTWORTHY AI

We are making great progress in cataloguing some of the poor ethical and trust outcomes of AI systems, many of which we’ve discussed in previous issues of EV. We’re moving to the next stage of figuring out how to evaluate, discuss and fix such poor outcomes.

This paper coauthored by longtime EV reader, Adrian Weller, looks at procedural fairness in algorithmic decision making. Quite an interesting approach which introduces “three new measures of fairness that [can] account for individual’s moral sense” and uses human judgments about whether such features can be used in the decision making process. The conclusion is that procedural fairness, in that judgements are based on what people feel is fair, can be implemented in algorithmic decision making, but comes at a cost in accuracy.

Apple released a good discussion of how it implements differential privacy. This is a mathematically robust definition of privacy, which allows machine learning to apply to data which masked user-identifiable features. Naturally, this approach will come with accuracy trade-offs. But I’ve been reasonably pleased with how Apple has approached the private-data vs. machine-learning-performance conundrum. And, indeed, how they approach training of FaceID’s models which, if I recall correctly, involved recruiting large numbers of consenting, ethnically, gender and age balanced people to ‘lend their’ faces to the modelling. (Happy to be persuaded otherwise.)

💥 Kate Crawford’s talk at the NIPS conference: “bias is social issue first, technical issue second”. WORTH WATCHING

(See also: summary of the key work in ethical AI presented at the NIPS conference.)

Elsewhere:

• Robo-recruiters are hoped to eliminate bias in hiring, but some fear it might make it worse;
• Reinforcement learning appears to reduce the need for training data for AI systems, but here are five reasons why data remains key ingredient for most AI implementations;
• How marketers plan to use AI in the next few years. Infographic
• Autonomous drone races the best human drone pilot. Humans win! (20% faster.)
• Researchers at UC Berkeley teach a robot to predict consequences of its actions;
• Xiaomi to integrate AI in smartphones next year;
• Researchers predict voting pattern based on cars in the streets thanks to Google Street View;
• A new Tencent report says the industry will need millions of AI engineers;
• Will the US lead in AI chips keep America ahead of China in the AI race?
• The Hollywood Symphony Orchestra plays music composed through a machine learning system. Quite Morricone.

DEPT OF CRYPTOCURRENCY

Watching the bitcoin price on Thursday was absurdly fun. I checked in every three hours or so, with amazement. Beyond the obvious speculation, which leads to some kind of instability in the market dynamics of bitcoin, it is also clear that cryptocurrencies are being legitimated.

More people than ever before know about it and participate in it. Coinbase, a simple service to buy and sell popular cryptocurrencies, is a key facilitator of the growing trade. In fact, its app is now the most downloaded app in the Apple store. (It is also rumoured to be making $50m in EBITDA each month.) Early adopter bank, Revolut, has introduced bitcoin, litecoin and ether trading in its app, with a viral friend-get-friend mechanism built in.

🤓 A non-scientific poll I ran on Twitter, showed a significant bullishness on where the bitcoin price would end-up in 2018.

Bitcoin holdings unsurprisingly follow a power law. Of the 30m or so Bitcoin addresses (holders), at least 111 hold more than $100,000,000, 2,400 hold more than $10,000,000 and nearly two-thirds hold less than $100.

As Carlota Perez writes in “Technological Revolutions and Financial Capital”:

Frenzy is the later phase of the installation period. It is a time of new millionaires at one end and growing exclusion at the other. The paper economy decouples from the real economy… the regulatory framework [turns] impotent. It is also a time of speculation, corruption and unashamed love of wealth.

The turning point to making more wholesome use of the technology, she writes,

has to do with the balance between individual and social interests… giving greater attention to collective well being. [This] switch does not occur for ideological reasons, but as a result of the way in which the new installation of a new paradigm takes place.

After all, Pets.com, Webvan and Zap.com were all internet turkeys. Yet the Internet itself has proved to have lasting value.

Will Carlota’s framework hold true for distributed ledger technology, blockchains or Bitcoin? My sense is that the technologies need to prove their utility, more widely. Progress is happening at technical layers, and true to our exponential age, this is happening quickly. We’re also getting early user case studies of where blockchains may solve real problems. All positive signs of real value being demonstrated.

Elsewhere:

• Thought-provoking essay on token economics: why token velocity matters;
• Details of the city of Zug’s implementation of the first publicly verified blockchain ID;
• People will retweet anything for a bitcoin;
• The Bulgarian government is sitting on $3bn in bitcoin;
• Lightning, a bitcoin scaling layer, has passed a number of technical tests.

DEPT OF FUTURE MEDIA

🌨️ Winter is coming for the new breed of digital media houses like Buzzfeed. Did their war chests of venture funding encourage unhealthy rates of scaling?

The New York Times proves people will pay for content. More than 3.5m paying subscribers and growing.

The Economist aims to double its number of paid subscriptions and has a team of 16 focused on user retention.

eSports may be more popular amongst young American men than watching traditional sports. (Story from September.) The top video game, League of Legends, approached nearly 95m hours of viewing in October this year.

If you enjoyed reading, subscribe below to receive Exponential View every Sunday AM.