That Space Cadet Glow #119
Issue 119 - 27th May 2021
There are alternatives to the tech giants' platforms, even ones that can make money without exploiting your data - I look at one of the best examples. Plus, using AI to help paralysed people write again, discovering secrets from 100,00-year-old DNA, and an AI that can crack cryptic crosswords. The masthead image is from Quebecois artist Karen Cantú. You can (as I have done) buy prints of her images here.
If it looks like a Google, takes your data like a Google, then...
It is usually really difficult to avoid the Faustian bargain that comes with using the tech giants' platforms; you get to use their stuff for free and they take all of your data (to make buckets of money through ads). If you don't want to 'share' your data (i.e. give unknown amounts of it away for exploitation by unknown companies across the internet) then your only practical option is not to use their services at all, but then you will miss out on discovering what your friends had for lunch. One internet service, though, does not depend on network effects and is, therefore, easier to switch to lesser-known providers - that of 'search'. Google is obviously the market leader here in usage (with 92% of the market), but also way up the list in data gathering and revenue ($162 billion in 2019). But, as this Wired article succinctly points out, there are ways of making money out of search engines that don't require you to be a creepy company. They spotlight DuckDuckGo, a privacy-focused search engine that has been profitable since 2014. DuckDuckGo does collect your data when you use their search engine so that it can show you ads, but, crucially, it doesn't store or share your data. It doesn't even retain your search terms - every time you load the search engine, you’re a stranger. It will take a lot to reverse the trend of surveillance capitalism, and legislation may be required, but it is good to know that regular capitalism is at least an option and has, as in the case of DuckDuckGo, been proven to work.
In the last issue I wrote about Neuralink's amazing achievements in (apparently) getting a monkey to play a game of computer pong with just its mind. I also talked about the (Elon Musk-owned) company's penchant for hype, which only seems to sow doubt in people's minds. An equally amazing, yet much less hyped, success was reported in New Scientist earlier this month. Just as Neuralink did, a team at Stanford University embedded neural sensors into a brain in order to read the electrical signals and then use AI to interpret those signals into something useful. With the Stanford work, though, the subject was a 65-year-old man who has a spinal cord injury that left him paralysed below the neck since 2007. The system was able to generate characters on a computer as the man thought about writing them with a pen. He was able to achieve 90 characters per minute and with an accuracy of 94.1% (which is much better than my own typing!). The AI model was trained using data from the subject's brain activity but, because of the huge amounts of data required, was supplemented with synthetic data that had 'noise' added. Because of the uniqueness of each of our brains, and the complexity of positioning the sensors, the model developed here can only be used on that one person. But, because of the relatively small training requirements, it can be replicated easily for other people once the sensors are installed. One clear advantage of this system over Neuralink's is that it is able to pick up very fine motor signals, such as for hand-writing, which contrasts with the much larger signals from arm movements, as with the joystick in pong. The ability to communicate with others is a fundamental human need, and being able to give that capability back to those who have lost it is surely a worthwhile pursuit (just don't over-hype it).
It's amazing what you can find out from dust. By isolating tiny specimens of 100,00 year-old blood and excrement from the dust found on a cave floor, scientists have been able to identify major changes in Europe’s Neanderthal populations. As this article in the Guardian explains, this new technique allows researchers to study DNA recovered from cave sediments, without the need for any fossils or stone tools. The technique was developed a few years ago, but the first meaningful results have been generated recently by the Max Planck Institute for Evolutionary Anthropology. The Galería de las Estatuas cave system in Spain has been studied very carefully by palaeontologists and archaeologists so that each layer of cave sediment has been analysed and dated precisely, which means they could put an exact date to the samples of DNA that they found in each layer. The team found that, about 100,000 years ago, the population who had been living in the cave for millennia were replaced by a completely different group of Neanderthal people. This is an unexpected discovery and demonstrates the power of the technique. Other cave systems are being planned for analysis, with the hope of new insights regarding the Denisovans and the homo floresiensis being discovered. God knows what scientists will discover when they start analysing our own dust and dirt in 100,000 years time...
1 across: The first and ninth letters of the alphabet (2)
We have seen AI master chess, Go, space invaders and sudoku. The latest success, as told by Slate, comes over crosswords, and that journey to success, by winning the American Crossword Puzzle Tournament, is an interesting AI story. The program, called Dr. Fill, was written by Matt Ginsberg, a bit of a polymath, and originally came 141st in 2012 (everyone who beat it got a badge!). It was built using what most AI people call "good ol' fashioned AI" which means it is more of a brute force engine than an intelligent system (Deep Blue, when it beat Kasporov at chess, was of a similar design). The program got better and better over the years, peaking at 11th place in 2017. But just before this years (online) tournament, a team from the Berkeley Natural Language Processing Group got in touch and offered to collaborate. The key thing about the Berkely AI is that it is based on neural networks, so it is much better at understanding the subtleties of the language in the clues. By combining the two approaches, and literally with only days to spare, the super-charged system managed to win the competition (it completed the final puzzle in 49 seconds). The combination of an expert system and machine learning is not really exploited as much as it should be, mainly because AI folk tend to fall into one camp or the other, so perhaps Dr. Fill's victory will be a useful example for the AI community to follow.
Hypnotic Brass Ensemble - Soon It Will Be Fire
There is an interesting, but niche, genre of music that can only be described as 'brass bands'. But these are not your traditional Brighouse & Rastrick sort - most hail from New Orleans and have been generating beautiful tunes and covers, many of which fuse jazz and hip-hop, over the past 10 years or so. Of especial note are the Hot 8 Brass Band, the Youngblood Brass Band and, the subject of this month's pick, the Hypnotic Brass Ensemble. They hail from Chicago and the band is made up of the seven sons of the jazz great Phil Cohran. They've played with Prince, Mos Def , Mick Jones and Damon Albarn in their time. Their latest LP is a reworking of Richard Youngs’ 1998 album Sapphie, with the first track, a mesmerising and beautiful tune that features Moses Sumney on vocals, available now. The full album, called This is a Mindfulness Drill, will be released on 25th June on Jagjaguwar Records.
Last month I wrote about the dangers of using AI to identify emotions. Spotify obviously didn't get the memo.
How did AI benefit the world during the pandemic? This paper looks in depth at that question and gives it a B+.
Last night the BBC investigative program Panorama focused on the risks of AI. Good to see that all of the issues they raised have been discussed in this newsletter at some point.
Andrew Burgess is the founder of Greenhouse Intelligence, a strategic AI advisory firm.