The eleventh edition of our newsletter on Large Language Models is here.<\/p>\n\n\n\n
In this edition, we explore<\/p>\n\n\n\n
Stanford AI Index Report<\/strong><\/p>\n\n\n\n Stanford released their traditional AI Index Report. They highlighted the top 10 takeaways<\/a>, but let’s look at some other interesting findings,<\/strong> too:<\/p>\n\n\n\n In short: the \u201ctapestry\u201d of news<\/strong><\/p>\n\n\n\n OpenAI<\/strong> showcased their new GPT-4o model<\/a>, which is faster, cheaper<\/strong>, and has improved multimodality<\/strong>. It’s confirmed that it was the rumored gpt2-chatbot<\/a>. It supports native speech input and output, making it a useful real-time<\/strong> assistant. At the I\/O conference, Google<\/strong> countered by introducing the project Astra<\/a>, another fast multimodal AI assistant<\/strong>. However, it will be released to the public only later this year.<\/p>\n\n\n\n We now also understand why<\/a> GPT often uses words like “delve”<\/strong>, “tapestry”, or \u201cleverage\u201d. OpenAI used workers from Africa<\/strong> to finetune the raw model, aligning it with their linguistic preferences<\/strong>. For example, the word \u201cdelve\u201d is far more common in Nigerian English than in US or British English.<\/p>\n\n\n\n As usual, many new models<\/strong> have been released, including Mistral’s<\/strong> larger MoE model Mixtral 8×22<\/a>, Snowflake’s<\/strong> big Arctic LLM<\/a> with 480 billion parameters specialized in SQL generation and coding, Meta’s<\/strong> impressive LLaMa 3<\/a> in 8B and 70B versions, Apple’s<\/strong> small OpenELM<\/a>, and Microsoft’s<\/strong> small but surprisingly capable Phi-3<\/a>. For an in-depth analysis, we recommend Sebastian Raschka, PhD<\/a>‘s newsletter<\/a>.<\/p>\n\n\n\n Researchers have developed an exciting new architecture<\/strong> called Kolmogorov\u2013Arnold Networks<\/a>. This innovation aims to replace the classical building block of neural networks<\/strong>, Multi-Layer Perceptron, by using learnable activation functions<\/strong>. Zulnorain Ahmed<\/a> explains how it works and discusses the potential implications<\/a>.<\/p>\n\n\n\n Blueprint for evaluating AI tools in journalism<\/strong><\/p>\n\n\n\n Journalists often hesitate to embrace<\/strong> artificial intelligence due to a lack of tailored evaluation<\/strong> methods. Researchers from Northwestern University proposed a framework<\/a> that addresses this gap, focusing on three aspects:<\/p>\n\n\n\n Check out Sachita Nishal<\/a>‘s presentation and get inspired<\/a>!<\/p>\n\n\n\n Trust drives usage and willingness to pay<\/strong><\/p>\n\n\n\n Schibsted News Media identified what drives people’s trust<\/a> in the media in Sweden and Norway. As Agnes Stenbom<\/a> explained, \u201ctrust can be a key to unlocking user revenue\u201d<\/strong>. Of course, trust is built on the credibility<\/strong> of journalists, the news creation process, and the content itself. Just as crucial is the personal relevance<\/strong> and usefulness of the articles to users, along with selectivity<\/strong> \u2013 the composition of the topics and events covered. See how we applied this by analyzing municipal coverage<\/a> for Radio France.<\/p>\n\n\n\n Transforming News Creation<\/strong><\/p>\n\n\n\n At the INMA World Congress of News Media<\/strong>, David Caswell<\/a> discussed how genAI alters news<\/a> creation. A live survey revealed that 70%<\/strong> of attendees use AI to create transcripts and summaries<\/strong>, while 60%<\/strong> let it suggest headlines and SEO<\/strong>. Investigative agency Cuesti\u00f3n P\u00fablica<\/strong> employs Retrieval Augmented Generation (RAG<\/strong>) to enhance breaking news with information about prominent figures, and Zamaneh Media generates their newsletter. Other applications include generating alerts, social media posts, and tagging metadata.<\/p>\n\n\n\n OpenAI’s pitch to publishers<\/strong><\/p>\n\n\n\n The list of publishers in OpenAI’s<\/strong> Preferred Publishers Program is growing<\/strong>, but so is the list of publishers suing them<\/strong> (see Spotlight #8<\/a>). Adweek obtained confidential documents<\/a> about the program. For the right to train on and display<\/strong> publisher content with attribution, OpenAI offers financial compensation<\/strong>, priority placement<\/strong>, and richer brand expression<\/strong> in user chats. The financial compensation consists of a guaranteed value for the publisher’s archive of articles and a variable value based on user interactions with the content. OpenAI’s statistics show that 25% of users already use the browsing function, and The Atlantic predicts that search with integrated AI<\/strong> will answer<\/strong> about 75%<\/strong> of queries without clickthrough<\/strong>. OpenAI claims the leaked documents are for discussion purposes only and contain some mischaracterizations and outdated information.<\/p>\n\n\n\n Sci-fi prompting<\/strong><\/p>\n\n\n\n The most effective prompts<\/strong> for LLMs can be rather unconventional<\/strong>. Broadcom researchers let people compete with automatic prompt optimizers<\/a> to generate optimal prompts for solving mathematical problems. Optimization<\/strong> methods outperformed humans<\/strong> by prompting the LLM to act as Star Trek captain<\/strong>. Should we even try to enhance prompts ourselves?<\/p>\n\n\n\n Echo embeddings<\/strong><\/p>\n\n\n\n Carnegie Mellon University researchers discovered a simple yet effective method to enhance embeddings<\/a> (a mathematical representation of text meaning in a multidimensional space). Conventionally, meaning<\/strong> is encoded sequentially<\/strong> so it lacks the information that is about to come later. They address<\/strong> the limitation by echoing the input<\/strong> (just repeat it and get embeddings from the second part). For example, in the sentences \u201cShe loves summer but dislikes the heat.\u201d<\/em> and \u201cShe loves summer for the warm evenings\u201d, <\/em>conventional embeddings would overestimate the similarity of their initial segments.<\/p>\n\n\n\n Answer elections<\/strong><\/p>\n\n\n\n It is known that allowing multiple LLMs to vote<\/strong> on an answer improves<\/strong> results<\/strong>. Researchers from Stanford University, UC Berkeley, Google, and Princeton University examined how the results improve for queries with various difficulty<\/a> levels. Interestingly, only<\/strong> answers to easy queries<\/strong> improve with more votes; with hard queries,<\/strong> we just get more wrong answers, degrading<\/strong> the performance<\/strong>. The challenge to accurately distinguish the easy queries from the hard ones remains.<\/p>\n\n\n\n Thinking before speaking<\/strong><\/p>\n\n\n\n Researchers from Stanford and Notbad AI Inc taught the Mistral model to think before speaking with a Quiet-STaR<\/a>, generalization of<\/strong> the STaR<\/strong> framework. STaR fine-tunes a model on a question-answer dataset with generated rationales<\/strong> and uses the REINFORCE<\/strong> algorithm to improve<\/strong> them. Those generated thoughts then guide the model through difficult questions. Quiet-STaR generalizes the thoughts<\/strong> to reasoning that helps infer future text<\/strong> in general, and the tokens are generated in parallel. While this approach enhances performance, it comes with a significant increase in<\/strong> overhead “thought” tokens<\/strong>.<\/p>\n\n\n\n Dots before speaking<\/strong><\/p>\n\n\n\n Are thoughts crucial? Researchers from New York University had the intriguing idea that performance<\/strong> can be improved by adding computation<\/strong> through generating extra tokens, regardless of<\/strong> their content<\/strong>. They experimented with meaningless filler tokens<\/a> \u2018…’, teaching the model how to use them. Nonetheless, they demonstrated that they really do improve performance almost as the Chain-of-Thought approach on a subclass of problems<\/strong> (first-order logic). This brings a new challenge to<\/strong> the interpretability<\/strong> of LLM results.<\/p>\n\n\n\n\n
AI for Media<\/h2>\n\n\n\n
\n
Surprising Ways to Improve Outputs<\/h2>\n\n\n\n