The ninth edition of our newsletter on Large Language Models is here. <\/p>\n\n\n\n
In this edition, we look at<\/p>\n\n\n\n
The European AI Act (<\/strong>see Spotlight #3<\/a>) <\/strong>was approved<\/strong> unanimously by the Council of the EU<\/a> (= the ministers from all member states) in February and by the parliament yesterday<\/a> (85% of MEPs voted for it). The newly created European AI Office<\/a> started to hire experts and should soon publish tools, methodologies, and benchmarks for the evaluation of AI systems. <\/p>\n\n\n\n We cannot trust the LLMs<\/strong><\/p>\n\n\n\n Both Google and Anthropic acknowledged to WSJ<\/a> that hallucinations<\/strong> are a serious problem in the adoption<\/strong> of LLMs. According to Eli Collins, a VP at Google DeepMind: “We\u2019re not in a situation where you can just trust the model output.” But as Jared Kaplan, a co-founder of Anthropic, says, we cannot simply make the models more cautious because they would always answer: “I don\u2019t know the context.” Both Collins and Kaplan talked about the importance of<\/strong> users validating<\/strong> any LLM response<\/strong>, and the providers should make it easy by identifying<\/strong> the sources<\/strong> of any answer. <\/p>\n\n\n\n Chatbot responsibility<\/strong><\/p>\n\n\n\n However, a court in Canada disagrees with Collins and Kaplan<\/a>, at least in the case of customer-service chatbots. Air Canada’s chatbot provided<\/strong> a passenger with incorrect<\/strong> instructions for a discount<\/strong>. It did not matter that the bot provided a link to the correct policy. The airline<\/strong> was ordered to pay<\/strong> anyway. According to the court, the airline “is responsible for<\/strong> all the information on its website<\/strong>. It makes no difference whether the information comes from a static page or a chatbot.”<\/p>\n\n\n\n Microsoft AI Bot might not be worth the money<\/strong><\/p>\n\n\n\n Microsoft<\/strong> is pushing AI to<\/strong> all of its products<\/strong>, even Notepad. It has been testing AI Copilot in MS Office (now called Microsoft 365): it summarizes emails, creates presentations, writes memos, and so on. Selected companies were able to test the tool, and according to the Wall Street Journal<\/a>, they are not really persuaded. The employees<\/strong> were eager to test it<\/strong>, but their enthusiasm waned quickly as the tool made frequent mistakes<\/strong>. The expected price of $30\/month\/user does not<\/strong> seem worth it<\/strong>.<\/p>\n\n\n\n Don’t let AI do your taxes yet<\/strong><\/p>\n\n\n\n TurboTax and H&R Block, two major tax service<\/strong> providers in the U.S., offer their users a chatbot<\/strong>-based tax expert<\/strong>. Or at least that’s what they claim. Washington Post warns<\/strong> its readers to avoid the chatbots<\/strong> unless they want to get into trouble. In their tests, the AI “experts” were wrong<\/a> in more than 50% of answers for TurboTax and 30% for H&R Block.<\/p>\n\n\n\n ASCII art<\/strong><\/p>\n\n\n\n Even though AI still cannot help you with taxes, it can help you build a bomb.<\/p>\n\n\n\n LLMs are trained on all kinds of data, so if left unchecked, they can advise on all kinds of questionable activities. Various techniques<\/strong>, such as data filtering and supervised fine-tuning, have been designed to prevent<\/strong> the models from doing it. However, all these methods have <\/strong>their limitations<\/strong>. <\/p>\n\n\n\n Researchers from the University of Washington managed to use ASCII art<\/a> to get past the defenses<\/strong> of GPT-3.5, GPT-4, Gemini, Claude, and Llama2. <\/p>\n\n\n\n NYT\u2019s Zach Seward on AI-powered journalism <\/strong><\/p>\n\n\n\n Zach Seward, the new editorial director of AI initiatives at The New York Times, recently gave a talk about AI<\/a> at the SXSW Conference in Austin, Texas. He provides concrete examples of how AI is used in journalism, both bad (such as Sports Illustrated publishing awful articles automatically written by made-up journalists; see Spotlight #7<\/a>) and good (e.g., searching large troves of documents with embeddings, rephrasing prison policies and audits for public consumption). <\/p>\n\n\n\n Columbia Journalism Review \u2013 Report on AI in the news<\/strong><\/p>\n\n\n\n Columbia Journalism Review, a magazine for journalists by Columbia University, published a detailed report by Felix Simon<\/a> on the use of AI in the news. From July 2021 to September 2023, he interviewed<\/strong> more than 130 news workers, including journalists, data scientists, and product managers in the US, UK, and German media, including The Guardian, Bayerischer Rundfunk, the Washington Post, The Sun, and the Financial Times. He also interviewed 36 independent American and European experts.<\/p>\n\n\n\n In general, newsrooms have become more open to AI<\/strong>. The adoption is driven mostly by economic pressure<\/strong>, technology readiness<\/strong>, and hype<\/strong>.<\/p>\n\n\n\n For now, AI has brought no fundamental change<\/strong> and instead only helped make some old approaches more effective. Many of the most beneficial AI applications are quite mundane (transcription, search, content categorization, etc.).<\/p>\n\n\n\n There are some important challenges:<\/p>\n\n\n\n You might also want to look at NiemanLab\u2019s summary<\/a> or the full report<\/a>.<\/p>\n\n\n\n BBC Generative AI pilots<\/strong><\/p>\n\n\n\n BBC announced they are starting 12 generative AI pilots<\/a>. Unlike Air Canada or HR Block, BBC is more cautious<\/strong>: most of the pilots<\/strong> are internal<\/strong> only, and the resulting content won\u2019t be made public. The pilots can be organized into three groups<\/strong>:<\/p>\n\n\n\n 1) Maximizing<\/strong> the value<\/strong> of existing<\/strong> content, for example, translating or reformatting existing content, e.g., writing an article based on a live sports radio commentary.<\/p>\n\n\n\n 2) New audience experiences<\/strong>, such as a chatbot providing personalized learning to students.<\/p>\n\n\n\n 3) SImplifying and speeding up <\/strong>processes, for example, suggesting headlines and summaries.<\/p>\n\n\n\n AI at Ippen Digital<\/strong><\/p>\n\n\n\n Nikita Roy<\/a> from Newsroom Robots interviewed<\/a> Alessandro Alviani<\/a>, the product lead for AI at Ippen Digital, a part of the German Ippen Media Group. In the second part of the interview, they discuss finetuning<\/strong> language models on<\/strong> the corpus of Ippen\u2019s local<\/strong> German news<\/strong> content so that it can assist journalists<\/strong> with writing headlines, lead paragraphs, summaries, etc. Another interesting topic they discuss is the role of AI in the context of modular journalism<\/strong>. It can break a story into modules<\/strong> that can be selected and mixed to create personalized content<\/strong>.<\/p>\n\n\n\n Training Data & Intellectual Property<\/strong><\/p>\n\n\n\n News publishers take different positions regarding their articles being used by AI companies to train their models. While The New York Times<\/a> has sued<\/strong> OpenAI and Microsoft (see Spotlight #8<\/a>), News Corp<\/a> (the parent of Dow Jones, The Times, and The Sun) is negotiating<\/strong>, <\/strong>and Axel Springer<\/a> has already closed a partnership<\/strong> with OpenAI. <\/p>\n\n\n\n In the meantime, more and more newspapers are blocking AI crawlers<\/strong>. According to the Reuters Institute<\/a>, 48% of 150 top news sites across 10 countries block OpenAI. Compare that with 33% of the top 1000 websites, as reported by <\/a>Originality.ai<\/a> (see also Spotlight #5<\/a>). Fewer websites block Common Crawl (18% of the top websites), Google AI crawler (24% of the top news sites and 10% of the top websites), and Anthropic (4% of the top websites).<\/p>\n\n\n\n In short<\/strong><\/p>\n\n\n\n The competition<\/strong> among AI leaders intensifies<\/strong>. Google unveils Gemini 1.5<\/a> with a huge 1 million token context window and Gemini Advanced. Anthropic debuts Claude 3<\/a>, Mistral introduces Mistral Large<\/a>, and Inflection AI launches Inflection-2.5<\/a>. All are said to be GPT-4 class models<\/strong>. OpenAI counters this surge of models with their text-to-video<\/strong> Sora model<\/a>, overshadowing Google’s text-to-video Lumiere model<\/a> and the other LLM releases. On the other hand, Gemini Advanced gathers some unwanted attention for generating too inclusive images<\/a>.<\/p>\n\n\n\n We recommend reading Ethan Mollick<\/a>‘s post about Gemini Advanced<\/a>. He dismisses the value of standard benchmarks and offers a high-level comparison with GPT-4. He also feels the model exhibits so-called “sparks” or “ghosts” of general intelligence, such as when playing a Dungeons and Dragons type of game with it.<\/p>\n\n\n\n Search and assistants<\/strong><\/p>\n\n\n\n A popular use for LLMs is AI-powered multimodal search<\/strong>, as seen with platforms like You.com<\/a>, Perplexity.ai<\/a>, and the AI browser Arc<\/a>. Google responds to this by integrating <\/a>an LLM writing assistant into the Chrome <\/strong>browser. Meanwhile, OpenAI is developing an AI agent<\/a> to automate tasks<\/strong> on user devices, while HuggingFace is challenging their <\/a>GPTs with Assistants, an open-source alternative for creating customized chatbots<\/strong>. Microsoft aims higher, constructing an Agent Foundation Model<\/a> encompassing language, image\/video encoders, and an action encoder<\/strong> trained on robotics and games.<\/p>\n\n\n\n Following OpenAI’s example, Mistral partners with Microsoft<\/a> to launch Le Chat<\/strong> chatbot. Nvidia releases<\/a> a locally run chatbot<\/strong> Chat with RTX, free from privacy concerns. <\/p>\n\n\n\n Amazon debuts its shopping assistant<\/a> Rufus<\/strong>. It will be interesting to see how it handles the wrongly AI-generated product descriptions<\/a>. <\/p>\n\n\n\n Chips<\/strong><\/p>\n\n\n\n Meta purchases 350,000 H100 graphic cards<\/a> while developing their own \u201cArtemis\u201d chip<\/strong>, and OpenAI has ambitions to open its own chip<\/a> factories<\/strong>. Meanwhile, Groq <\/strong>unveils Language Processing Units<\/a> <\/strong>that enable Mixtral to operate at a speed<\/a> of 500 tokens per second<\/strong>.<\/p>\n\n\n\n Open source<\/strong><\/p>\n\n\n\n Significant developments also occurred in the open-source domain, as the Allen Institute for Artificial Intelligence introduced OLMo<\/a>, a genuine open-source model<\/strong>. The release includes it all: weights, inference code, training data, and evaluation code. Google contributed with Gemma<\/a> family models, while the RWKV Project released the Eagle 7B<\/a>, <\/strong>an attention-free model<\/strong>. Abacus.ai<\/a> launched Smaug<\/a>, holding the top spot<\/strong> on HugginFace’s Open LLM Leaderboard.<\/p>\n\n\n\n Testing LLM can be fun<\/strong><\/p>\n\n\n\n Every AI developer would tell you how important evaluation<\/strong> is, but typically it is quite boring. Sometimes, however, the results are unexpected<\/strong>\u2026<\/p>\n\n\n\n A common approach to testing LLM is the so-called needle-in-the-haystack<\/strong> evaluation: You put a very specific statement (the needle) into a random context (the haystack) and ask a question that can only be answered using the information in the needle. <\/p>\n\n\n\n When Anthropic researchers tested Claude 3 Opus<\/a> using a needle about pizza toppings, it answered correctly, but it also added: “This sentence<\/strong> seems very out of place<\/strong> and unrelated to the rest [..]. I suspect this pizza topping “fact” may have been inserted as a joke or<\/strong> to test if I was paying attention<\/strong>.” (see Alex Albert’s tweet<\/a>). <\/p>\n\n\n\n When a model detects it is being tested, we have a problem. But it’s not the only challenge we face. Additional issues include inadequately annotated benchmarks (see Spotlight #7<\/a>), and contamination during LLM training<\/a>. For a deeper analysis, check out Rohit Krishnan<\/a>‘s comprehensive post<\/a>. We do need to design better evaluations.\u00a0<\/p>\n\n\n\n Evaluation and leaderboards<\/strong><\/p>\n\n\n\n HuggingFace introduced several new leaderboards<\/strong>, including the Enterprise Scenarios Leaderboard<\/a> for real-world applications<\/strong>, the LLM Safety Leaderboard<\/a>, and the Hallucinations Leaderboard<\/a>. To battle contamination,<\/strong> they released a dynamic benchmark, the NPHardEval Leaderboard<\/a>.<\/p>\n\n\n\n The University of Berkeley unveiled a Function-Calling Leaderboard<\/a>, together with their Gorilla OpenFunctions-v2 model. It is particularly useful given the growing integration of function calls<\/strong> into LLMs. Finally, there is an AgentBoard benchmark<\/a> for LLM agents<\/strong> that isn’t only about the final success rate but evaluates also the intermediate steps<\/strong>.<\/p>\n\n\n\nChallenges<\/h2>\n\n\n\n
AI and news media<\/h2>\n\n\n\n
\n
\n
New models and assistants<\/h2>\n\n\n\n
Model evaluation<\/h2>\n\n\n\n
New training methods<\/h2>\n\n\n\n
\n