.svg)
THE BIG IDEA
How soon does it get scary?
Ezra Klein of the New York Times recently had Dario Amodei, the CEO of Anthropic, on his podcast and asked him how soon he thought the sci-fi future of AI would arrive.
Amodei led the team that created GPT-2 and GPT-3 for OpenAI. Then he left to co-found Anthropic with his sister, Daniela. In the last year, they’ve raised $7.3 billion. Claude-3, their latest model, performs as well or better than GPT-4.
Last year, Anthropic published its Responsible Scaling Policy, laying out a series of protocols to manage the risks of building mega-powerful AI systems. Within that framework lies our future: utopia, apocalypse, and nothingburger all rolled into one.
Currently, we’re at what’s called AI Safety Level 2 (ASL-2). Next up is ASL-3, which entails significantly higher risk and the potential for catastrophic misuse. Amodei seems most afraid of state actors using big AI models to develop new cyber and bioweapons—a risk which increases by 20% in ASL-3.
You don’t even want to know what ASL-4 is. In fact, Anthropic says it won’t attempt to define what ASL-4 is because it’s too far from present systems. Amodei’s best guess? “North Korea or China or Russia could greatly enhance their offensive capabilities in various military areas with A.I. in a way that would give them a substantial advantage at the geopolitical level.” Yikes!
Naturally, Klein then asked when we’ll reach ASL-3.
“I think ASL-3 could easily happen this year or next year,” Amodei said.
“Oh, Jesus Christ,” the normally even-keeled Klein blurted out.
What about ASL-4?
“Anywhere from 2025 to 2028.”
The future is closer than we think. Both the good and the bad. Which highlights more than ever the importance of safety research and greater oversight, both private and governmental.
Behind these predictions are the “scaling laws,” which aren’t laws so much as observable patterns that haven’t been disproven. The scaling laws suggest that growth in AI capability is exponential. It’s hard for our brains to comprehend how quickly this is moving.
As the physicist Albert Bartlett put it: “The greatest shortcoming of the human race is our inability to understand exponential growth.”
That said, these scaling laws rely on the constant input of vast amounts of new training data and compute. Some researchers think that we’ll run out of data in the next couple years.
Amodei is an optimist. His approach, in the absence of legislation from Washington, is to have Anthropic hold itself accountable to ensure that these catastrophic risks don’t come to pass. Even so, he said, “This may end up being the greatest geopolitical issue of our time.”
CHART OF THE WEEK
36% of workers think AI will replace their job
Stanford’s exhaustive 2024 AI Index Report gathered all the interesting studies around AI and compiled them in one place. They included a survey from Ipsos which found that 57% of respondents think AI is likely to change how they perform their current job within the next five years, and 36% fear AI may replace their job in the same time frame.
The fact that over half of respondents believe AI will change how they do their jobs shows just how far AI has crept into the public consciousness.
EXPERT CORNER
Understanding patient experience at scale using generative AI
Fannie McWatt, PhD, is a Strategy Director at healthcare advertising network IPG Health and one of the minds leading the design of their new AI tools for understanding patient experience. The tools process patient conversations about their symptoms, their disease, or their drug treatment to help health systems and pharmaceutical companies get a more empathic grasp on what people love and hate about their experience.
How do these tools work? And what are you hearing from patients?
We collect conversational data and use AI to help us understand people and their behaviors. And that can help guide our creatives with what kind of content they create. When they first get a diagnosis, how are they feeling? And how can we help them understand the options that they have for treatment?
It’s like a more responsive approach towards customer research.
Healthcare is often delayed in adopting new technology, but it's really catching up. We talk about Uber, Amazon, Spotify, and how they shifted the paradigm of how we purchase things and how we get our transportation. When we design patient support programs, we want to do the same thing. It's just a different type of consumer that needs a different type of experience. But all of the principles are the same.
It often feels like when we have feedback for our healthcare providers, they don’t really hear it.
Exactly. We're in a lot of patient forums where people can talk about specific categories, specific diseases that they are managing. And now all that information can be processed more easily at scale with AI. Hopefully we can move faster and more responsively.
What’s next? What would you build if you had a blank check?
There's such a huge opportunity in wellness and helping people take charge of their health. I would love to build something that helps me as a working mother in a fast-paced industry to work more efficiently, and to stay healthy and live better. That's something we all need. If I had a blank check, I would want to build something that could help the general population really leverage all of the information that's at their fingertips and put it to action. That will be the next frontier: Now that we know all of this information, how can we use AI to really change our behaviors?
WATERCOOLER
Wait, AI is good at emotions?
A new study from Harvard found that GPT-4 helped reappraise a difficult emotional situation better than 85% of humans.
Cognitive reappraisal, a technique from cognitive behavioral therapy, involves reshaping a person's emotional response by altering their interpretation of events. GPT-4 beat humans in 3 of four categories including effectiveness, novelty, and empathy.
Here’s an example from the study:
“A classmate sneers as you enter the room, reminding you of last week's mishap where you fell in the hallway. A typical human-generated reappraisal might emphasize personal resilience or the trivial nature of the incident, potentially abstracting away from the emotional nuance of the scenario. On the other hand, a GPT-4-generated reappraisal might focus on the ambiguity of the sneer and offer multiple interpretations, suggesting that the sneer could be unrelated to the observer or reflective of the classmate's own insecurities.”
This study comes at a time when many people grappling with mental health challenges are stuck looking for answers on Reddit. It’s interesting to consider the possibilities here, especially among populations with reduced access to sufficient psychological care.
EVENTS
A Human-First Approach to Healthcare Innovation
On May 21st, we’re bringing together an all-star group of healthcare operators, innovators, and technologists to tackle how to leverage AI without losing the human touch, strategies for improving the lives of healthcare workers instead of replacing them, and how we can use AI to tackle healthcare inefficiencies and failures — from improved scheduling to wearables that predict and mitigate burnout.
DISCOVERY ZONE
Ever wish you could smell your video game? GAMESCENT is an AI device that releases real-time scents to enhance your gaming experience — like an essential oil diffuser for gaming. SMELL-O-VISION is back!
MEME
.jpeg)
.jpg)
