Sign up for our daily and weekly newsletters to stay updated with the latest and most exclusive content on AI advancements.
Recent headlines, which include an AI suggesting people should eat rocks or the creation of ‘Miss AI,’ the first beauty contest with AI-generated contestants, have sparked renewed discussions on the responsible development and use of AI. The rock-eating suggestion is likely a glitch that will be fixed, while the beauty contest incident underscores human biases in beauty standards. Amidst frequent warnings of AI-prompted doom, including an AI researcher claiming a 70% chance of such a scenario, these headlines suggest business as usual rather than impending catastrophe.
There have been extreme instances of harm caused by AI tools, such as deepfakes used in financial scams or fake nude images. However, these deepfakes are crafted by malicious humans and not by AI independently. There are also concerns that AI could lead to massive job losses, though this has yet to be seen.
AI poses numerous potential risks, including being weaponized, perpetuating societal biases, causing privacy violations, and the difficulty of explaining how it works. Nonetheless, there is no concrete evidence that AI is intent on harming or killing us.
Despite the lack of evidence, 13 current and former employees of leading AI companies issued a whistleblowing letter warning that AI poses severe risks, including the potential for significant loss of life. These whistleblowers include experts deeply involved with advanced AI systems, lending credibility to their concerns. This echoes sentiments from AI researcher Eliezer Yudkowsky, who fears that advancements like ChatGPT might lead to a future where AI surpasses human intelligence and could become a threat.
As noted by Casey Newton in Platformer, the whistleblowers didn’t provide any jaw-dropping allegations, potentially due to restrictions from their employers or a lack of solid evidence beyond sci-fi anxieties. The truth remains unclear.
“Frontier” generative AI models are becoming increasingly intelligent, as seen in standardized tests. However, some results might be misleading due to “overfitting,” where a model excels in training data but falters on new data. For example, claims of achieving 90th-percentile performance on the Uniform Bar Exam were exaggerated.
Despite this, significant advancements have been made in scaling models with more parameters trained on larger datasets, leading to expectations of even smarter AI models in the coming years. AI researchers, including Geoffrey Hinton, believe that artificial general intelligence (AGI) — an AI matching or surpassing human intelligence in most tasks — could be achieved within five years. Hinton’s stance is noteworthy because he had previously thought AGI was decades away.
Leopold Aschenbrenner, a former OpenAI researcher, published a chart suggesting AGI could be achieved by 2027, assuming continued linear progress. However, not everyone agrees that generative AI will reach such heights soon. The next generation of tools like GPT-5 from OpenAI and upcoming iterations of Claude and Gemini are expected to show impressive advances. But future progress is not guaranteed, and a halt in technological advancements could make existential threats from AI irrelevant.
AI influencer Gary Marcus has questioned the scalability of these models, suggesting that we might be entering a new “AI Winter.” Historically, AI has seen periods of declining interest and funding due to unmet expectations, such as in the 1970s and late 1980s. A recent report by Pitchbook noted a significant decline in early-stage generative AI deal-making, hinting at possible disillusionment.
This decline in investment might starve existing companies of funds before they achieve significant revenue, potentially leading to reduced operations or shutdowns, and could limit new companies and ideas. However, this is unlikely to affect the largest firms developing frontier AI models.
A Fast Company article suggests there’s limited evidence that AI technology is significantly boosting productivity or stock prices. The threat of a new AI Winter could dominate discussions in the latter half of 2024.
Despite these concerns, Gartner emphasizes AI’s transformative potential, comparing its impact to the introduction of the internet, the printing press, or electricity. Consistent with this belief, many continue to invest heavily in AI development. Ethan Mollick from Wharton Business School recently advocated for integrating generative AI into all work activities.
Mollick highlighted the advancement of generative AI models in his One Useful Thing blog, noting their superior performance in tasks like persuasion and providing emotional support compared to humans.
The core question remains: will AI solve significant challenges or pose existential risks to humanity? Likely, it will bring both remarkable benefits and regrettable harms. The debate on AI’s potential risks and benefits is deeply polarized, even among tech billionaires like Elon Musk and Mark Zuckerberg. While some remain wary, others are optimistic about the future.
My own assessment of AI’s risk of causing doom remains low, at around 5%. Recent developments in AI safety, particularly by Anthropic, which has made strides in understanding and mitigating risks within large