If we are serious about p(doom), we should also discuss p(human); inside China's plan for the future of AI; Silicon Valley freaks out over California AI bill; is the cloud ready for its AI moment?

Sam Altman's investment empire explained; the SPV market goes wild for AI startups; AI to hold center stage Apple's WWDC event; how Humane's Ai Pin flopped; OpenAI is giving Facebook vibes

Jun 07, 2024

∙ Paid

Shortly after appearing on Dwarkesh Patel’s podcast this week, Leopold Aschenbrenner, a former OpenAI safety researcher, posted a chart on X arguing that, given the current pace of progress with AI, we will have AGI by 2027 and maybe even “superintelligence” beyond 2030 thanks to millions of GPUs concentrated in 10 GW data centers.

On his website, Mr Aschenbrenner has written an essay describing how we’re on the path to training clusters costing hundreds of billions by 2028 and requiring power equivalent to a medium-sized US state

He’s not the only one to make such claims. The launch of ChatGPT has created a cottage industry of people whose main focus appears to be stoking both excitement and fear about AI on social media. The rot starts at a top, with many executives racing to make predictions of AI's imminent rise to surpass human intelligence and why only their company is best positioned to save us from the impending p(doom). However, amidst the hype, a critical perspective is often overlooked: the possibility that AI may never achieve human-level intelligence. This perspective is not just a contrarian view but is supported by respectable AI researchers and scientists who have highlighted the intrinsic limitations and challenges faced by AI.

There were two phases in recent history which have led us to this moment: first, the rise of deep learning and the ImageNet breakthroughs achieved from 2012 until 2015. Although this phase started just before the decade in question, the impact was felt strongly in the following years. In 2012, a team led by Geoffrey Hinton won the ImageNet competition using a deep convolutional neural network called AlexNet, significantly outperforming other methods. This marked the beginning of the deep learning revolution. By 2015, deep learning models were surpassing human performance on certain image recognition tasks. This breakthrough led to massive investments in AI research and applications across industries.

In 2015, deep learning models became just as good as humans at image classification

Then, in 2017, Google introduced the transformer architecture in the paper Attention Is All You Need. The real impact of this achievement was only felt years later, when OpenAI amassed enough data and compute to train GPT-3 and then launched ChatGPT in late 2022, bringing conversational AI to the mainstream and sparking widespread public interest and debates about AI's potential impact on society.

Yet, these successes often mask a fundamental truth: these systems excel in controlled environments with clear rules and abundant data, but they struggle with the nuanced, flexible thinking that characterizes human intelligence.

I hate to go all Gary Marcus on you in this newsletter but sometimes you have to give the devil his due: current AI systems lack the deep understanding and adaptability of human cognition, because they are simple mathematical models that make probabilistic calculations based on large data sets and access to compute.

Let’s unpack the sentence above: while it is true that the human brain is wired in a neural network similar to the artificial ones powering deep neural networks, human intelligence is not just about processing information quickly or recognizing patterns. It encompasses a broad spectrum of cognitive abilities, including abstract thinking, emotional understanding, and creative problem-solving. These capabilities are deeply rooted in our biology and experiences, shaped by millions of years of evolution.

Secondly, the data which these large models have been trained on is produced by people, and therefore limited by our ability to produce it. It is more likely that the best chart to describe artificial intelligence is not a linear pattern but more of an asymptotic curve that is always tangent to the direction of human intelligence. The other assumption made by Mr Aschenbrenner is that human intelligence is a constant on a chart while AI is permanently evolving. That is factually not true. As humans make progress in fields such as technology and science, we acquire more knowledge and get smarter.

Finally, if we want to move closer to human-level AI, 10 GW compute clusters alone are not going to get us there. Instead, we need a new architecture that pivots away from auto-regressive models and towards causal reasoning. Without this shift, AI will remain a powerful tool for specific applications rather than a rival to human intellect.

I want to close on one important point which has received little attention in my view. Unlike some people who mock the concept of p(doom), I actually believe there is an existential risk when it comes to frontier AI. There are several companies today building advanced AI systems for military applications, including self-guided missiles or drones. As these machines get more capable, we might be lulled into a false sense of security and let them operate autonomously. Sooner or later, they will make a mistake which could trigger a global conflict. So the way I look at p(doom) is not from the lens of AI reaching superhuman cognition and wiping out humanity, Skynet-style. Instead, I can see a scenario in which p(doom) goes over 50% when we rush to take the human out of the loop in military applications by thinking we have reached AGI when in reality we have not.

Nevertheless, investing in AI research should continue, but with a balanced perspective that considers its capabilities and its boundaries. AI can be a transformative tool in areas like healthcare, education, and environmental management, where its ability to process vast amounts of data and identify patterns can lead to significant advancements. However, the goal should be to develop AI systems that augment human decision-making and creativity, rather than aiming for—or scaring people with—hypothetical scenarios of an elusive human-like intelligence.

And now, here are this week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

Business Insider: China's new plan to dominate the future of tech will reshape the world
FT: Silicon Valley in uproar over Californian AI safety bill
WSJ: The Opaque Investment Empire Making OpenAI’s Sam Altman Rich
TechCrunch: VCs are selling shares of hot AI companies like Anthropic and xAI to small investors in a wild SPV market
Bloomberg: Here’s Everything Apple Plans to Show at Its AI-Focused WWDC Event
Wired: How to Lead an Army of Digital Sleuths in the Age of AI
The Economist: Robots are suddenly getting cleverer. What’s changed?
The Atlantic: OpenAI Is Just Facebook Now
CNBC: The quiet Apple executive behind Apple’s AI strategy
WSJ: Will Cloud Software Be Ready for Its AI Moment?
Bloomberg: Sam Altman Was Bending the World to His Will Long Before OpenAI
Wired: OpenAI Offers a Peek Inside the Guts of ChatGPT
New York Times: ‘This Is Going to Be Painful’: How a Bold A.I. Device Flopped
WSJ: How Apple Fell Behind in the AI Arms Race

⚙️Computer does

AI in the wild: how artificial intelligence is used across industry, from the internet, social media, and retail to transportation, healthcare, banking, and more

MIT Technology Review: This AI-powered “black box” could make surgery safer
Business Insider: AI in the classroom has some people worried. Teachers aren't.
BBC: Could AI put an end to animal testing?
TechCrunch: Google looks to AI to help save the coral reefs
BBC: Scientists enlist AI to interpret meaning of barks
Fortune: At this gym, customers can choose an AI best friend or drill sergeant
Axios: New AI system hunts for satellites behaving oddly in space
TechCrunch: Wix’s new tool taps AI to generate smartphone apps
Business Insider: How one agency uses AI to track and manage thousands of campaign assets — building its own library of 'collective knowledge'
TechCrunch: eBay debuts AI-powered background tool to enhance product images
FT: St James’s Place uses AI to spot and help ‘vulnerable’ customers
The Guardian: AI used to predict potential new antibiotics in groundbreaking study
Business Insider: Mastercard's AI system is helping banks keep fraudsters in check — and it could save millions of dollars
Android Authority: Even Spotify could soon get its own Gemini Extension
BBC: Researchers use AI to analyse cosmic explosions
Business Insider: 'What are your clothes made of?' is a deceptively difficult question. AI can help answer it.
BBC: New AI tech developed to detect heart failure earlier
Business Insider: How Daily Harvest used AI to optimize product packaging and improve customer service
The Verge: Amazon’s Project PI AI looks for product defects before they ship

🧑‍🎓Computer learns

Interesting trends and developments from various AI fields, companies and people

Keep reading with a 7-day free trial

Subscribe to Computerspeak by Alexandru Voica to keep reading this post and get 7 days of free access to the full post archives.