The difference between this approach and its predecessors is that DeepMind hopes to use “dialogue in the long term for safety,” says Geoffrey Irving, a safety researcher at DeepMind.
“That means we don’t expect that the problems that we face in these models—either misinformation or stereotypes or whatever—are obvious at first glance, and we want to talk through them in detail. And that means between machines and humans as well,” he says.
DeepMind’s idea of using human preferences to optimize how an AI model learns is not new, says Sara Hooker, who leads Cohere for AI, a nonprofit AI research lab.
“But the improvements are convincing and show clear benefits to human-guided optimization of dialogue agents in a large-language-model setting,” says Hooker.
Douwe Kiela, a researcher at AI startup Hugging Face, says Sparrow is “a nice next step that follows a general trend in AI, where we are more seriously trying to improve the safety aspects of large-language-model deployments.”
But there is much work to be done before these conversational AI models can be deployed in the wild.
Sparrow still makes mistakes. The model sometimes goes off topic or makes up random answers. Determined participants were also able to make the model break rules 8% of the time. (This is still an improvement over older models: DeepMind’s previous models broke rules three times more often than Sparrow.)
“For areas where human harm can be high if an agent answers, such as providing medical and financial advice, this may still feel to many like an unacceptably high failure rate,” Hooker says.The work is also built around an English-language model, “whereas we live in a world where technology has to safely and responsibly serve many different languages,” she adds.
And Kiela points out another problem: “Relying on Google for information-seeking leads to unknown biases that are hard to uncover, given that everything is closed source.”
The EU wants to put companies on the hook for harmful AI
The new bill, called the AI Liability Directive, will add teeth to the EU’s AI Act, which is set to become EU law around the same time. The AI Act would require extra checks for “high risk” uses of AI that have the most potential to harm people, including systems for policing, recruitment, or health care.
The new liability bill would give people and companies the right to sue for damages after being harmed by an AI system. The goal is to hold developers, producers, and users of the technologies accountable, and require them to explain how their AI systems were built and trained. Tech companies that fail to follow the rules risk EU-wide class actions.
For example, job seekers who can prove that an AI system for screening résumés discriminated against them can ask a court to force the AI company to grant them access to information about the system so they can identify those responsible and find out what went wrong. Armed with this information, they can sue.
The proposal still needs to snake its way through the EU’s legislative process, which will take a couple of years at least. It will be amended by members of the European Parliament and EU governments and will likely face intense lobbying from tech companies, which claim that such rules could have a “chilling” effect on innovation.
In particular, the bill could have an adverse impact on software development, says Mathilde Adjutor, Europe’s policy manager for the tech lobbying group CCIA, which represents companies including Google, Amazon, and Uber.
Under the new rules, “developers not only risk becoming liable for software bugs, but also for software’s potential impact on the mental health of users,” she says.
Imogen Parker, associate director of policy at the Ada Lovelace Institute, an AI research institute, says the bill will shift power away from companies and back toward consumers—a correction she sees as particularly important given AI’s potential to discriminate. And the bill will ensure that when an AI system does cause harm, there’s a common way to seek compensation across the EU, says Thomas Boué, head of European policy for tech lobby BSA, whose members include Microsoft and IBM.
However, some consumer rights organizations and activists say the proposals don’t go far enough and will set the bar too high for consumers who want to bring claims.
China is betting big on another gas engine alternative: methanol cars
Today, the leading company making methanol from carbon dioxide is Carbon Recycling International, an Icelandic company. Geely invested in CRI in 2015, and they have partnered to build the world’s largest CO2-to-fuel factory in China. When it’s running, it could recycle 160,000 tons of CO2 emissions from steel plants every year.
The potential for clean production is what makes methanol desirable as a fuel. It’s not just a more efficient way to use energy, but also a way to remove existing CO2 from the air. To reach carbon neutrality by 2060, as China has promised, the country can’t put all its eggs in one basket, like EVs. Popularizing the use of methanol fuel and the clean production of methanol may enable China to hit its target sooner.
Can methanol move beyond its dirty roots?
But the future is not all bright and green. Currently, the majority of methanol in China is still made by burning coal. In fact, the ability to power cars with coal instead of oil, which China doesn’t have much of, was a major reason the country pursued methanol in the first place. Today, the Chinese provinces that lead in methanol-car experiments are also the ones that have abundant coal resources.
But as Bromberg says, unlike gas and diesel, at least methanol has the potential to be green. The production of methanol may still have a high carbon footprint today, just as most EVs in China are still powered by electricity generated from coal. But there is a path to transition from coal-produced methanol to renewables-produced methanol.
“If that is not an intention—if people are not going to pursue low-carbon methanol—you really don’t want to implement methanol at all,” Bromberg says.
Methanol fuel also has other potential drawbacks. It has a lower energy density than gasoline or diesel, requiring bigger, heavier fuel tanks—or drivers may need to refuel more often. This also effectively prevents methanol from being used as an airplane fuel.
What’s more, methanol is severely toxic when ingested and moderately so when inhaled or when people are exposed to it in large amounts. The potential harm was a big concern during the pilot program, though the researchers concluded that methanol proved no more toxic to participants than gas.
Beyond China, some other countries, like Germany and Denmark, are also exploring the potential of methanol fuels. China, though, is at least one step ahead of the rest—even if it remains a big question whether it will replicate its success in developing EVs or follow the path of another country with a major auto industry.
In 1982, California offered subsidies for car manufacturers to make over 900 methanol cars in a pilot program. The Reagan administration even pushed for the Alternative Motor Fuels Act to promote the use of methanol. But a lack of advocacy and the falling price of gasoline prevented further research of methanol fuel, and pilot drivers, while generally satisfied with their cars’ performance, complained about the availability of methanol fuel and the smaller range compared with gas cars. California officially ended the use of methanol cars in 2005, and there’s been no such experimentation in the US since.
Can we find ways to live beyond 100? Millionaires are betting on it.
But to test the same treatments in people, we’d need to run clinical trials for decades, which would be very difficult and extremely expensive. So the hunt is on for chemical clues in the blood or cells that might reveal how quickly a person is aging. Quite a few “aging clocks,” which purport to give a person’s biological age rather than their chronological age, have been developed. But none are reliable enough to test anti-aging drugs—yet.
As I leave to head back to my own slightly less posh but still beautiful hotel, I’m handed a gift bag. It’s loaded up with anti-aging supplements, a box with a note saying it contains an AI longevity assistant, and even a regenerative toothpaste. At first glance, I have absolutely no idea if any of them are based on solid science. They might be nothing more than placebos.
Ultimately, of all the supplements, drugs and various treatments being promoted here, the workout is the one that’s most likely to work, judging from the evidence we have so far. It’s obvious, but regular exercise is key to gaining healthy years of life. Workouts designed to strengthen our muscles seem to be particularly beneficial for keeping us healthy, especially in later life. They can even help keep our brains young.
I’ll be penning a proper write up of the conference when I’m back home, so if your curiosity has been piqued, keep an eye out for that next week! In the meantime, here’s some related reading:
- I wrote about what aging clocks can and can’t tell us about our biological age earlier this year.
- Anti-aging drugs are being tested as a way to treat covid. The idea is that, by rejuvenating the immune system, we might be able to protect vulnerable older people from severe disease.
- Longevity scientists are working to extend the lifespan of pet dogs. There’ll be benefits for the animals and their owners, but the eventual goal is to extend human lifespan, as I wrote in August.
- The Saudi royal family could become one of the most significant investors in anti-aging research, according to this piece by my colleague Antonio Regalado. The family’s Hevolution Foundation plans to spend a billion dollars a year on understanding how aging works, and how to extend healthy lifespan.
- While we’re on the subject of funding, most of the investment in the field has been poured into Altos Labs—a company focusing on ways to tackle aging by reprogramming cells to a more youthful state. The company has received financial backing from some of the wealthiest people in the world, including Jeff Bezos and Yuri Milner, Antonio explains.
From around the web
An experimental Alzheimer’s drug appears to slow cognitive decline. It’s huge news, given the decades of failed attempts to treat the disease. But the full details of the study have not yet been published, and it is difficult to know how much of an impact the drug might have on the lives of people with the disease. (STAT)
Bionic pancreases could successfully treat type 1 diabetes, according to the results of a clinical trial. The credit card-sized device, worn on the abdomen, can constantly monitor a person’s blood sugar levels, and deliver insulin when needed. (MIT Technology Review)
We’re headed for a dementia epidemic in US prisons. There’s a growing number of older inmates, and the US penal system doesn’t have the resources to look after them. (Scientific American)
Unvaccinated people are 14 times more likely to develop monkeypox disease than those who receive the Jynneos vaccine are, according to the US Centers for Disease Control and Prevention. But the organization doesn’t yet know how the vaccine affects the severity of disease in those who do become unwell, or if there is any difference in protection for people who are given fractional doses. (The New York Times $)
Don’t call them minibrains! In last week’s Checkup, I covered organoids—tiny clumps of cells meant to mimic full-grown organs. They’ve mainly been used for research, but we’ve started to implant them into animals to treat disease, and humans are next. Arguably the best-known organoids are those made from brain cells, which have been referred to as minibrains. A group of leading scientists in the field say this wrongly implies that the cells are capable of complex mental functions, like the ability to think or feel pain. They ask that we use the less-catchy but more accurate term “neural organoid” instead. (Nature)
That’s it for this week. Thanks for reading!