Is nationalization the only way to control artificial intelligence?

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each weekday, we scour the web to aggregate the many stories related to artificial intelligence; we pass along the news, useful resources, tools or services, guides, technical analysis and exciting developments in open source.

In today’s Dispatch:

• Nvidia's earnings report tomorrow could make or break the current AI-fueled rally in tech stocks - a trend that has the potential to be a risk for both the stock and wider sentiment if the company doesn’t deliver. Nvidia’s shares have nearly tripled this year on exploding demand for AI chips, and Nvidia is under pressure to forecast revenue above the expected $12.5 billion.

• Language translation startup Unbabel has developed a remarkable new brain-computer interface called Halo that could enable faster communication for people with disabilities like ALS. By detecting muscle signals in the arm and correlating them to words via AI, Halo can silently interpret a user's thoughts and relay messages at 20 words per minute - faster than previous systems used by people like Stephen Hawking. Halo may be just the beginning of a future where thoughts can be quickly translated into text or speech without needing to type or speak.

• OpenAI's acquisition of digital studio Global Illumination is causing intrigue in tech circles. The acquired startup brings AI expertise from industry veterans (including Instagram's algorithm designer) - but beyond that, OpenAI's motivations are unclear. The company could be looking to make smarter AI by training models in 3D game worlds, or it may hope to release an AI-infused game as a new source of hype and revenue.

Plus: A lengthy opinion piece from POLITICO on the urgency for the US government to more directly control AI, South Korea’s incredible robot pilot, Microsoft collaborating with Databricks, trending tools and more!

(Opinion) There’s only one way to control AI: nationalization

From POLITICO: A former CEO of an AI company argues that key areas of AI development need to be nationalized. The piece highlights the rapid recent advances in AI capabilities, but argues that AI safety is not keeping pace. A ‘Humane AI Commission’ could focus exclusively on maintaining human sovereignty and mitigating unforeseen risks.

More details:

• The author uses a personal anecdote in facial recognition technology to highlight why ‘full transparency around AI systems’ is not enough: his company’s AI unexpectedly matched photos of an engineer's son to photos of the engineer. Despite having full access to the AI system, the NeuralEye team had no way of understanding how or why that could happen. This ‘black box’ nature of AI systems puts a ceiling on the value of transparency.

• The (imperfect, but effective) Atomic Energy Commission provides a previous model for government control over dangerous technology. A similar commission could nationalize key AI assets, ensuring they serve the public good.

• While AI transparency has limits, more transparency requirements could still aid accountability and oversight. Mandated safeguards may also provide some protections against misuse, bias, and other harms. But tepid regulations and industry self-governance will not be enough. The US must make bold moves to harness AI.

Takeaways: In October 2022, the Blueprint for an AI Bill of Rights was published by the White House Office of Science and Technology Policy; but this is merely an advisory guide for AI development - it has no teeth. There have been policies on AI implemented at the state level, but these are scattered (and most of New York’s proposed state legislation on AI is not being passed).

The US is indeed the world leader in AI, as the article notes; it seems clear that Washington will not risk jeopardizing that mantle even if the public desire for caution is abundantly clear.

Meet 'Pibot,' the humanoid robot that can safely pilot an airplane better than a human - without modifying the cockpit

From EuroNews: Researchers in South Korea are developing a humanoid robot called Pibot that can pilot an airplane by manipulating controls just like a human pilot. The robot’s memory is so large that it can memorize all Jeppesen navigation charts - a task that is impossible for human pilots.

More details:

• Pibot can precisely control its limbs to operate flight instruments and monitor the airplane's status with cameras. It can memorize aeronautical navigation charts for the entire world as well as flight manuals.

• The robot is dextrous enough to operate flight instruments even during severe vibrations in the aircraft. It can also communicate with air traffic control and human pilots using voice synthesis.

• The robot is adaptable and could potentially replace humans in other vehicles besides planes. Pibot is expected to be completed by 2026 and may have military applications.

Takeaways: Pibot can operate existing planes meant for human pilots without any modifications to the aircraft. Retrofitting the global fleet of aircraft with fully autonomous capabilities would be an enormous undertaking, so this development seems like a logical bridge to more futuristic automated planes. The expanse of automation into piloting shares many of the benefits and drawbacks of driverless cars.

From our sponsors, Stat Significant:

Love a good statistic?

Stat Significant is a weekly newsletter featuring data-centric essays about culture, economics, sports, statistics, and more. Are Best Picture winners getting worse? Is Christmas season coming earlier? What is the dollar value of a yard in the NFL? Subscribe for free to find out!

Check out Stat Significant

Social media/Video/Podcast:

• AI phone calls - things are going to get weird quick [Reddit]

• Intensive vs. laissez-faire approaches to AI governance - with Robin Hanson of George Mason University [Podcast]

• Humans are pretending to be AI and NPCs on TikTok [X]

• How Sam Altman uses ChatGPT on a daily basis - Bloomberg interview [YouTube]

• George Hotz vs Eliezer Yudkowsky AI Safety Debate [YouTube]

Did you know? 

A randomized trial found that using AI to support radiologists' reading of mammograms maintained or improved cancer detection rates while substantially reducing workload - by up to 50%. The study randomized 80,000 women in Sweden to either standard double reading of mammograms or AI-supported single reading. The AI system scored each mammogram for malignancy risk to determine if single or double reading was needed.

In the AI group, the cancer detection rate was 6.1 per 1,000, meeting the safety threshold, compared to 5.1 per 1,000 in the control group. The false positive rate and positive predictive value were similar between groups. The results suggest AI could improve mammography screening efficiency without compromising cancer detection. More follow-up is needed to confirm there is no difference in interval cancer rates.

Trending AI Tools & Services:

• GPTForWork: Use ChatGPT in Google Sheets, Docs and Excel

• The best AI story generators to help you write your next novel

• (Technical/Mac) GodMode: The AI chat browser. Fast, free access to ChatGPT, Bing, Bard, Claude, YouChat, Poe, Perplexity, Phind, and local models like Vicuna and Alpaca.

• Chapple: All-in-one AI content creation platform

• GitMind: AI-powered mind mapping & brainstorming app. Create and collaborate on mind maps online in real-time

• Cerelyze: Turn technical research papers into code