The Best AI Articles of 2024

IEEE SICTRUMThe most popular artificial intelligence stories in the past year show a clear topic. In 2024, the world fought to reconcile with the capabilities and disadvantages of obstetric intelligence – both of which are important.

Two artificial intelligence articles in two years dealt with codes in Chatbots, while another looked at the best way to demand chat generators and photo generators (and found that humans can be dispensed with).

In the “faults” column, one of the in -depth investigations found that the Midjournyy photo generator had a bad habit in spitting on the images that almost coincide with characters and commercial scenes that have been placed in a brand of copyright films, while another investigation into the extent of bad actors who can use a stable version of photos 1.5 to make sexual assault on children.

Two of my favorite of my best collection are distinctive articles that tell great stories. In one of them, an artificial intelligence researcher tells how workers have collected and organized data to review the employer. In another, the sociologist who merged himself into a tumultuous emerging company for 19 months describes how the engineers cut the corners to meet the expectations of the owners of investment capital. Each of these important stories brings readers inside the noise bubble to obtain a real vision of how companies that make artificial intelligence benefit from human workers. In 2025, IEEE SICTRUM Promise to give you the basic truth.

David Blonker

Although the Tructured AI boom brought fears that chatting and photo generators would give up on jobs, some have expressed hope that they will create completely new jobs – such as unique engineering, which is the exact construction of claims to obtain the AI ​​Reed tool to create the exactly required output.

Well, this article is inhibited by this hope. range The editor Dina Genkina reported new research showing that artificial intelligence models do better in building claims than human engineers.

Gary Marcus and Red Souton via Midjorn

New York Times Another newspaper has filed a lawsuit against artificial intelligence companies because of the impersonation of the text, on the pretext that Chatbots lifted its stories protected by copyrights literally. In this important investigation, Gary Marcus and Red Southin showed clear examples of visual plagiarism, using Midjourney to produce images that are completely similar to footage of the major films, as well as brands such as Darith Fader, Homer Simpson, and the Sonik The Sanad. It is worth seeing the full article just to see the pictures.

The authors write: “These results provide strong evidence that Midjourney has been trained in copyright -protected materials, and they prove that at least some artificial intelligence systems may produce suicide outputs, even when they are not asked directly, which is likely to expose users to claim copyright violations.”

Gety pictures

When Chatgpt came out of Openai for the first time in late 2022, people were surprised by his ability to write code. But some researchers who wanted an objective measure of his ability evaluated the code in terms of function, complexity and security.

They tested GPT-3.5 (a copy of the Great Language Model that operates ChatGPT) on 728 coding problems from the Leetcode test platform in five programming languages. They found that it was good in the coding problems that were on Leetcode before 2021, and it was assumed that they had seen these problems in their training data. With more modern problems, its performance decreased significantly: its degrees fell on the functional code for ease of problems in coding from 89 percent to 52 percent, and for difficult problems, it decreased from 40 percent to 0.66 percent.

It should be noted that Openai GPT-4 and GPT-4O models are superior to the old GPT-3.5. While the artificial intelligence platforms for general purposes continue to improve coding, 2024 also witnessed the multiplication of the increasing artificial intelligence tools designed for coding.

Pioneer

This third story in our list is perfectly placed fourth, which meets a good look at how professors change their approaches to teaching coding, given the aforementioned spread of coding assistants.

The preliminary computer science courses focus less on building the coding total and more on the test and correction, so students are better equipped to pick up the mistakes made by their assistants from artificial intelligence. One of the professors says: “This is a skill that must be known early because you need to divide a big problem into smaller pieces that LLM can solve.” In general, the trainers say that their students’ use of artificial intelligence tools works to free them to teach high -level thinking that was reserved in advanced classes.

Mike McCadad

This feature story was composed by Amnesty International researcher, Dana Kalkashi, who gathered with the party workers in SHIPT, TARGET’s shopping and delivery platform.

The workers knew that SHIPT changed their payment algorithm in a mysterious way, and many saw a decrease in wages, but they were unable to obtain answers from the company – so they started collecting data themselves. When they joined Calacci, worked with them to create Textbot so workers can easily send screenshots for their paid receipts. The tool also analyzed the data, and told each worker if they get their salaries more or less under the new algorithm. It found that 40 percent of the workers had obtained an undeclared wage reduction, and workers used the results to attract the media’s attention while organizing strikes, provinces and protests.

Calacci writes: “The companies whose business models depend on the party workers have an interest in maintaining their algorithms that are not transparent. This“ lack of consistency of information ”helps companies to better control their working bases – they determine terms without revealing the details, and the only choice for workers is whether there is whether there is any technical reason for the reason for the need of these algorithms to be black boxes; and the real reason is To clarify the structure of power. ”

IEEE SICTRUM

Like two Russian nesting dolls, here we have a menu on menu. Every year, Stanford puts the huge artificial intelligence index, which contains hundreds of plans to track trends within artificial intelligence; The chapters include artistic performance, responsible artificial intelligence, economics, education, and more. Over the past four years, range I have read everything and pulled these more visible plans indicating the current state of Amnesty International.

In 2024, we highlighted the investment in the artificial intelligence, the cost and the environmental footprint of the training institution models, the corporate reports on artificial intelligence help in the minimum, and the general burn of artificial intelligence.

Istock

The nerve networks have been the dominant structure of artificial intelligence since 2012, when a system called Alexnet merged GPU with a multi -layer nerve network for unprecedented performance on the image recognition mission. But they have their negative aspects, including their lack of transparency: they can provide an answer that is often correct, but they cannot show their work.

This article describes a new way mainly for making more interpretative nerve networks than traditional systems and it appears to be more accurate. When designers tested their new model on physics and differential equations, they were able to determine how the model got his answers (often).

Ed Gent

The following story leads us to the Technology Center in Bangaluru, India, which has grown faster in the population more than the infrastructure – which led to some of the most crowded streets in the world. Now, a former chip engineer has been granted the hard task of taming traffic. He turned to artificial intelligence to help, using a congestion tool, expects traffic jams, determines events that attract large crowds, and managed police officers to record accidents.

For the following steps, CZAR Place for Traffic Mixing data from safety cameras throughout the city, allowing automatic vehicle account and classification, as well as data from food delivery companies and ride sharing.

Mike Kemp/Getty Pictures

In another important investigation, exclusive to rangeAnd artificial intelligence policy researchers, David Ivan Harris and Dave Wilnener, explained how some of Amnesty International photo generators can make sexual assault materials (CSAM), although they are against the declared conditions of use. They focused in particular on the stable, stable, open source of spread model, on the platforms that embrace the face and CIVITAI that host the model and provide it for free download (in the case of Hugging Face, was downloaded millions of times a month). They were building on previous research that showed that many photo generators were trained in a data collection that included hundreds of parts of CSAM.

Harris and Willner called companies to request responses to these allegations, and perhaps in response to their inquiries, stable proliferation 1.5 immediately disappeared from the face embrace. The authors argue that the time has come for artificial intelligence companies and hosting platforms to take care of their potential responsibility.

Pottery

What happens when the sociologist includes himself in a startup in San Francisco, which has just received a preliminary investment of the investment capital of $ 4.5 million and was quickly shot in rows to become one of the “unit” Silicon Valley to evaluate more than a billion dollars? Answer: You get an attractive book with a depth called Behind the startup: How forms of investment capital work, innovation and inequalityWho rangeChapter excerpts.

The author of the sociologist, Benjamin Shestakovsky, describes how the company called Aldone (not its real name) has given the priority of growth at any cost to meet the expectations of investors, which leads to a focus on employing both employees and users instead of doing a lot of actual engineering.

Although the company’s full value is that it will automatically match people who need local services with local service providers, he ended up using external sources of the matching process for Filipina working forces that made matches manually. “Philippine contractors work effectively like industrial “Artificial Intelligence”, shestakofsky writes, “simulating the output of programs that have not yet been completed.”