weekend ai reads for 2024-02-16

But in case you want to be more precise, let’s examine how enshittification works. It’s a three-stage process: first, platforms are good to their users. Then they abuse their users to make things better for their business customers. Finally, they abuse those business customers to claw back all the value for themselves. Then, there is a fourth stage: they die.

I get how this helps users. How does it help creators? Without them there is no web…

“I hate cars. They’re making my planet bad,” he says. “But I’m not riding a horse anymore, right? I’m driving a car.”

Vyalshakaeva learned about the project a few months into the relationship. She wasn’t angry when she learned her fianceé had been using ChatGPT to talk to her, just shocked.

Tools shape genres, though the tool marks may only be visible to those who are trained in its use.

The oppression of indigenous languages is why we have concerns about non-indigenous groups building tools like Whisper.

A few of our data scientists tried Whisper on te reo Māori videos from YouTube. Their initial reaction was, "Wow it works!" A more critical assessment of Whisper by our Māori data experts saw that it sort of worked but it was terrible. Still, this is concerning, that a non-Māori organisation thought it was okay to create a Māori speech recognition model and open it to the public.

In this blog, we present a successful attempt to intercept and “hijack” a live conversation, and use LLM to understand the conversation in order to manipulate the audio output unbeknownst to the speakers for a malicious purpose.

TRAYS uses historical data to forecast the number of passengers who will not show up for their flights. This helps KLM match the number of meals stocked on the flight with the exact passenger count, reducing food waste. The program was trialed for three months and helped to reduce food waste by 63%.

“You cannot allow that to be done by other people,” Huang said at the World Government Summit in Dubai.

Huang, whose firm has catapulted to a $1.73 trillion stock market value due to its dominance of the market for high-end AI chips, said his company is ‘democratizing’ access to AI due to swift efficiency gains in AI computing.

Every year, the company puts out a report aggregating insights from the billions — in 2023, the number was 6.5 billion — of messages sent across large companies, tabulating perceived risk factors and workplace sentiment scores. Schumann refers to the trillions of messages sent across workplace communication platforms every year as “the fastest-growing unstructured data set in the world.”

The goals of AI-related strategic planning are primarily related to supporting students. The three highest-ranking goals of AI-related strategic planning are preparing students for the future workforce, exploring new methods of teaching and learning, and improving higher education for the greater good (selected by 64%, 63%, and 41% of respondents, respectively).

Second, an urgent push to incorporate genAI literacy in classrooms might lead to a low quality of tools, content and teaching as companies prioritize quickly getting their products to market over ensuring the rigor and educational integrity of their offerings.

Our attack relies on MoE routing that uses finite sized buffer queues for each individual expert. When these queues are filled, tokens could be dropped 1 or passed to another expert whose queue is not full. This opens up a new attack vector; if adversarial data is mixed with benign user data in a batch, the adversarial data can influence the expert choice of the benign data by filling the buffer of certain experts.

Phidata is a toolkit for building AI Assistants using function calling.

Function calling enables LLMs to achieve tasks by calling functions and intelligently choosing their next step based on the response, just like how humans solve problems.

Common Crawl and AI builders have a shared responsibility for making generative AI more trustworthy. While Common Crawl was never primarily about providing AI training data, it now positions itself as an important building block for LLM development. However, it continues to provide a source that AI builders need to filter before model training.

Our current work focuses on developing a joint image and video encoder and aligning this joint encoder to existing foundation models. This has several notable benefits: firstly, it allows for the use of both action, image, and video with language datasets for pre-training. Secondly, it increases the capabilities of the model across a variety of downstream tasks (e.g., video understanding, temporal reasoning, action prediction, interaction with human feedback, etc.). Finally, by using a joint encoder, we can reduce the overall model size (instead of using two separate encoders), which can be useful for edge deployments or in limited computing scenarios such as robotics, gaming, and interactive healthcare tasks.

With today’s release, Smaug-72B becomes the first open-source model to achieve an average score of 80 on the Hugging Face Open LLM leaderboard, which is considered a remarkable feat in the field of natural language processing and open-source AI.

Introducing the first recipe book for artificial intelligence. And humans too.