Spectrum’s Top AI Stories of 2023

2023 may go down in history as one of the wildest and most dramatic years in the history of artificial intelligence. When OpenAI released its latest large-scale language model, GPT-4, in March 2023, people were still struggling to understand the power of OpenAI's ChatGPT, introduced in late 2022 (LLM is essentially is the brains behind consumer applications).

Throughout the spring of 2023, important and trusted people were worried about the potential negative effects of ever-improving AI, ranging from the mildly alarming to the existentially negative. First an open letter calling for a moratorium on the development of advanced models, followed by a statement on existential risks, the first international summit on AI safety, and groundbreaking developments in the form of a US executive order and EU AI law. regulations have been announced.

Below are Spectrum's top 10 AI articles of 2023, based on the amount of time readers spent on the articles. Get a taste of AI as 2023 could be a historic year…unless 2024 is even crazier.

10. AI art generators can be tricked into creating NSFW images:

Paisi Lee/Getty Images

With text-to-image generators like Dall-E 2 or Stable Diffusion, the user enters a prompt describing the image they want to generate, and the model does the rest. And while safeguards are in place to prevent users from producing violent, pornographic, or otherwise unacceptable images, both AI researchers and hackers have found it difficult to find ways to circumvent such safeguards. I feel joy in it. Jailbreak is a new hobby for white hats and black hats.

9. OpenAI Moonshot: Solving the AI ​​Tuning Problem:

Daniel Zehnder

This Q&A with OpenAI's Jan Leike takes a deep dive into the problem of AI coordination. It is a concern that we will build superintelligent AI systems that are inconsistent with human goals and could potentially lead to the extinction of our species. This is obviously an important question, and OpenAI is devoting significant resources to finding ways to empirically study problems that don't yet exist (because superintelligent AI systems don't yet exist).

8. Nvidia's secret to success:

Chen Yihua/Bloomberg/Getty Images

Nvidia has had a great year with its AI-accelerating GPU, the H-100, becoming arguably the hottest hardware in tech. At the conference, Bill Daly, the company's chief scientist, reflected on his four elements that propelled NVIDIA into the stratosphere. “Moore's Law was a surprisingly small part of NVIDIA's magic, and new number formats were a very large part,” he writes. IEEE spectrum Senior Editor Sam Moore.

7. ChatGPT illusions can hinder your success:

Zuma/Alamy

One of the problems that plagues LLMs is their habit of making things up and unexpectedly lying in a very confident tone. This habit is especially problematic when you try to use it for something really important, like drafting legal briefs. While OpenAI believes this is a solvable problem, some outside experts, such as Meta AI's Yann LeCun, disagree.

6. 10 key insights into the state of AI in 2023 (charts):

It's a list within a list! Every year, spectrum The editors sift through the massive AI index published by the Stanford Human-Centered Artificial Intelligence Institute and distill the report into a few charts that show the most important trends. Highlights for 2023 include the cost and energy required to train large-scale models, and industry's advantage over academia in hiring PhDs and building models.

5. The creepy new digital afterlife industry:

harry campbell

Below is an excerpt from a wonderful book called. us data, by Wendy H. Wong. This excerpt details the services that are emerging as part of the new digital afterlife industry. Some companies offer to send messages on your behalf after your death, while others allow you to ask questions and record a story that others can replay later. question. And there are already several examples of digital replicas being built based on data left behind by deceased loved ones.

4. AI Apocalypse: Scorecard:

IEEE spectrum

This project was born as spectrum The editors note how surprising it is that really smart AI practitioners, people who have worked in the field for decades, have such opposing views on two important issues. We discussed. So, does his LLM today indicate that AI will soon achieve superhuman intelligence, and will such superintelligent AI systems spell doom for humanity? homo sapiens. We created a scorecard to help our readers understand the different opinions.

3. 200-year-old mathematics reveals AI's mysterious black box:

P. Hassanzadeh/Rice University

The neural networks that power much of today's AI are famously black boxes. Researchers give them training data and look at the results, but they don't have much insight into what happened during that time. Researchers studying fluid mechanics are using Fourier, a mathematical technique used to identify patterns that has been around for about 200 years, to study neural nets trained to predict turbulence. I decided to use analysis.

2. How Duolingo's AI learns what it needs to learn:

eddie guy

This article is one of them spectrumA signature deep dive, featuring articles written by the experts building the technology. In this case, it's the AI ​​team behind language learning app Duolingo. They describe how they developed Birdbrain, an AI system that leverages both educational psychology and machine learning to deliver lessons with just the right level of difficulty to keep users engaged.

1. Calm down about GPT-4:

Rodney Brooks: Christopher Michel/Wikipedia; Background: Ruby Chen/OpenAI

spectrum As our readers tend to be contrarians, we really enjoyed this Q&A with Rodney Brooks, a self-proclaimed AI skeptic who has been working in this field for decades. Rather than praise GPT-4 as a step toward artificial general intelligence, Brooks drew attention to the difficulty LLMs have in generalizing from one task to another. “What large language models are good at is telling you what the answer should be sounds likethis is different from what should be answered Become familiar,” He said.