{"text":[[{"start":5.88,"text":"It’s becoming increasingly hard to distinguish between reality and science fiction. Rogue AI agents hack government websites. Genetically modified pigs’ kidneys are transplanted into humans. Nasa sponsors experiments to find minerals on asteroids."}],[{"start":21.12,"text":"But one particularly fascinating, if disturbing, area where reality is blurring with science fiction is “brain reading”, as described in a 168-page essay by Saturnin Pugnet that is causing a stir in AI circles. Called “Phase Lock”, Pugnet’s manifesto explains the ways in which brain-computer interfaces (BCIs) are developing at remarkable speed and could one futuristic day allow us to upgrade human “software”."}],[{"start":48.36,"text":"BCI devices are already being used to treat Parkinson’s disease and enable paralysed patients to operate computers. Pugnet suggests they could give humans cognitive superpowers. But he also acknowledges the ways in which this could go scarily wrong."}],[{"start":63.52,"text":"“If we were not in a world of accelerating AI, I would want to slow down neurotechnology. The risks are too high,” Pugnet tells me. But, he adds, in a world of accelerating AI the potential benefits of using neurotechnology to align AI models with human interests are enormous."}],[{"start":80.08,"text":"As a San Francisco-based entrepreneur and investor, Pugnet spent months researching the neurotechnology sector. His views certainly reflect the Silicon Valley consensus that it is only a matter of time before superintelligent AIs outsmart humans. So, his thesis goes, if we cannot slow down AI, we will have to speed up humans to enable us to remain relevant."}],[{"start":101.64,"text":"Such thinking lay behind Elon Musk’s launch of Neuralink in 2016 to develop implantable BCIs. Soon after the launch, Musk said computers could communicate at a trillion bits per second compared with humans at 10 bits a second. Human “bandwidth” could only be increased by merging biological and digital intelligence to create a kind of cyborg."}],[{"start":122.88,"text":"Advances in AI are opening new research pathways and unlocking a further surge of investment. Venture funding for neurotechnology more than doubled from $2.3bn in 2024 to $4.8bn last year."}],[{"start":137.68,"text":"Pugnet’s argument is that neurotechnology can benefit from the same scaling laws over the next decade that helped large language models over the previous one. LLMs work by throwing huge computing power at data trawled from the internet. In the same way, AI models may be able to recognise and replicate patterns in brain activity, given enough data and computing power. That brute-force approach could enable researchers to fully understand the brain and, more controversially, add new capabilities."}],[{"start":167.32,"text":"Pugnet describes the inherent dangers. But, ethically applied, he argues, this approach could help develop safer AI that reflects human values. By training AI models on scanned data amassed by BCI devices, they could achieve “alignment by integration”, as he says. That may be more reliable than the imperfect “alignment by control” approaches we use today."}],[{"start":189.8,"text":"To date, invasive BCI technology has been used solely in the medical field via implantable electrodes. A recent Nature Medicine article described how, over two years, a 45-year-old paralysed man was able to compose 183,060 sentences — at an average rate of 56 words per minute — by using a BCI device for speech and cursor control."}],[{"start":213.12,"text":"More advanced non-invasive brain-imaging devices are also developing fast. This week, the British start-up Cerca Magnetics received £7mn of government funding to install non-invasive, brain-scanning services at two NHS hospitals to help treat epilepsy."}],[{"start":229.64,"text":"Matt Brookes, a professor at Nottingham University and chair of Cerca Magnetics, says that brain scanning is opening up “incredible opportunities” for medical research. “This is not all complete sci-fi,” he says."}],[{"start":242.88,"text":"Some neurotechnologists are intrigued where the scaling approach will lead but remain sceptical that it will work in the ways foreseen by Pugnet. Say, for example, a person is thinking about eating an apple. That thought may produce very different brain signals depending on whether a person is happy, hungry, depressed or concussed."}],[{"start":261.92,"text":"It certainly takes a leap of faith to believe that a computer could read abstract thoughts. It takes an even bigger leap to imagine this process becoming bidirectional, enabling AIs to “write” as well as “read”. But the self-appointed role of venture capital is to help entrepreneurs cross the chasm of scepticism. Even if they fail, they can produce an additional precious data resource for medical researchers."}],[{"start":292.2,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1792217389_5495.mp3"}