The last month of the year 2023 was a busy one despite the holidays as the end neared. Coming into 2024, Google has released two new models, and the EU has finally reached an agreement with their proposed AI act.
There is also a raging debate around the use of media-produced content in training AI models. There are media companies that actively cooperate with AI-forging organizations, as well as ones that wage court battles for allegedly stolen content.
12.05.2023 Intel: 10% of companies introduce generative AI this year
With the ChatGPT-induced race to generative AI-powered innovation, it is not a surprise that companies around the world are looking for ways to implement the new tech and infuse their processes with LLMs. The survey shows that companies are eager to implement the new tech, citing better customer experience (58%), improved efficiency (53%), enhanced product capabilities (52%), and cost savings (47%) as key drivers of the adoption of the new technology.
More information can be found on the Intel website.
12.06.2023 Google released Gemini
A new model that is said to outperform human experts in LLM-evaluating tasks (MMLU – Massive Multitask Language Understanding). The model appears to outperform not only human evaluators but also all other models available on the market, including GPT-4.
Gemini is, in fact, a multimodal model that can not only process language but also interpret the images it sees or generate others using text-to-image capabilities.
More about the model can be found on the company’s official website.
12.09.2023 EU AI Act passed
After two years of heavy negotiations, the European Union has reached an agreement regarding the AI Act, arguably the first and most comprehensive legal framework for AI-based solutions. The proposed law forbids profiling users with information about their ethnicity or religion, among other restrictions and imposes multiple fines for companies unwilling to comply.
More about the law can be found on the Euro Parliament website.
12.12.2023 Microsoft launches Phi-2
Phi and Phi-2 are small models (2,7 billion parameters) that can be used for research works, for example, to improve interpretability or reduce the risk of hallucinations without the need to work on full-scale networks.
Phi-2, recently released by Microsoft, appeared to outperform up to 20-times larger networks when performing language reasoning and logic tasks.
More information about the model can be found in the company’s official statement.
12.13.2023 OpenAI strikes a deal with Axel Springer
The company behind ChatGPT has signed a deal with German publishing giant Axel Springer to allow the AI to be trained on and share the information found in articles and stories published by the media company. This content will be used to prepare answers to questions about the modern world and politics. The AI system will also include links to sources that would otherwise be behind the paywall.
This new deal is not exclusive – OpenAI will be able to form deals with new publishers, and Axel Springer can sign agreements with other chatbot providers.
More about the story can be found in the Guardian.
12.13.2023 Midjourney releases Alpha web version
One of the most popular AI image generation tools has been released in a web version. Before this, users could access the tool only by using the Discord-based interface. As of the end of December, the tool has been made available for the public, who have created over 10 thousand images using Midjourney. The developers ensure that access to the tool will be broadened with time.
More can be found on VentureBeat.
12.14.2023 Deepmind uses LLMs for mathematical and computer science discoveries
According to a recent research paper, published by the Deepmind team in Nature, it is possible to use LLM’s innate tendency to hallucinate to actually discover new information. The team has introduced FunSearch, a method that pairs pre-trained LLMs that are tasked with searching for innovative solutions for math-related problems in the form of computer code. The process is supervised by an “evaluator” system that prevents the models from hallucinating.
The process is based on iterating back-and-forth between the networks and searching for solutions unseen before. The whole process is based on searching for new functions – that’s why it is called Fun-search.
More details can be found in this Deepmind blog post and the Nature paper.
12.14.2023 H&R Block launches AI tax-filing tool
The tax prep company has released a conversational AI bot that supports users in filing their taxes. The tool is designed to answer tax-related questions, including rules, possible exemptions, and quotas. The bot also encourages users to reach human tax advisors. The company states that it is working on AI-powered personal advisors as well.
More information can be found in The Verge.
12.21.2023 Google releases VideoPoet – a text-to-video and image-to-video model
VideoPoet is one of at least a handful of text-to-video models recently released. What makes it unique is that the model is based on a Large Language Model (LLM) instead of the more common Diffusion-based approach. LLMs are often seen in tools like ChatGPT, Claude, or Llama, which generate texts or computer code but are rarely seen in other tasks. This time, the Google Research team has managed to train the tool to actually generate videos.
To do so, it was necessary to pre-train the model using one billion “text and video” pairs from “the public internet and other sources.”
More about the model can be found in this research paper.
12.27.2023 New York Times sues OpenAI for using its content to train chatbots
The New York Times, one of the most renowned newspapers in the world, has filed a federal lawsuit against OpenAI and Microsoft with the goal of ending the practice of using stories to train their chatbot. According to the media company, the AI developers have effectively stolen the texts used in the chatbot’s training. The NYT claims that it could amount to “billions of dollars worth of journalism work.” With users increasingly more willing to ask the chatbot for answers instead of searching for them on the publisher’s website, this practice can lead to a vast reduction in the newspaper’s and news service’s income.
More can be found in the Associated Press report.
