OpenAI said that it stopped five covert influence operations over the last three months that used its AI models for deceptive activities across the internet. These operations, which originated from Russia, China, Iran and Israel, attempted to manipulate public opinion and influence political outcomes without revealing their true identities or intentions, the company said on Thursday. "As of May 2024, these campaigns do not appear to have meaningfully increased their audience engagement or reach as a result of our services," OpenAI said in a report about the operation, and added that it worked with people across the tech industry, civil society and governments to cut off these bad actors. OpenAI's report comes amidst concerns about the impact of generative AI on multiple elections around the world slated for this year including in the US. In its findings, OpenAI revealed how networks of people engaged in influence operations have used generative AI to generate text and images at much higher volumes than before, and fake engagement by using AI to generate fake comments on social media posts.

Today, OpenAI released its first threat report, detailing how actors from Russia, Iran, China, and Israel have attempted to use its technology for foreign influence operations across the globe. The report named five different networks that OpenAI identified and shut down between 2023 and 2024. In the report, OpenAI reveals that established networks like Russia's Doppleganger and China's Spamoflauge are experimenting with how to use generative AI to automate their operations. And while it's a modest relief that these actors haven't mastered generative AI to become unstoppable forces for disinformation, it's clear that they're experimenting, and that alone should be worrying. The OpenAI report reveals that influence campaigns are running up against the limits of generative AI, which doesn't reliably produce good copy or code.

OpenAI identified and removed five covert influence operations based in Russia, China, Iran and Israel that were using its artificial intelligence tools to manipulate public opinion, the company said on Thursday. In a new report, OpenAI detailed how these groups, some of which are linked to known propaganda campaigns, used the company's tools for a variety of "deceptive activities." These included generating social media comments, articles, and images in multiple languages, creating names and biographies for fake accounts, debugging code, and translating and proofreading texts. These networks focused on a range of issues, including defending the war in Gaza and Russia's invasion of Ukraine, criticizing Chinese dissidents, and commenting on politics in India, Europe, and the U.S. in their attempts to sway public opinion. While these influence operations targeted a wide range of online platforms, including X (formerly known as Twitter), Telegram, Facebook, Medium, Blogspot, and other sites, "none managed to engage a substantial audience" according to OpenAI analysts.

OpenAI is bringing ChatGPT to college campuses across the country. On Thursday, the company announced ChatGPT Edu, a version of ChatGPT built specifically for students, academics, faculty. "ChatGPT Edu is designed for schools that want to deploy AI more broadly to students and their campus communities," the company said in a blog post. ChatGPT Edu includes access to GPT-4o, OpenAI's latest large language model that the company revealed earlier this month. OpenAI claims that the model is much better than its previous versions at interpreting text, coding, and mathematics, analyzing data sets, and being able to access the web.

OpenAI's report detailed how the five groups used the company's tech in their attempted influence operations. Spamouflage, a previously known group originating in China, used OpenAI's tech to research activity on social media and write posts in Chinese, Korean, Japanese and English, the company said. An Iranian group known as the International Union of Virtual Media also used OpenAI's tech to create articles that it published on its site.

For all the bluster about generative artificial intelligence upending the world, the technology has yet to meaningfully transform white-collar work. Workers are dabbling with chatbots for tasks such as drafting emails, and companies are launching countless experiments, but office work hasn't undergone a major AI reboot. Perhaps that's only because we haven't given chatbots like Google's Gemini and OpenAI's ChatGPT the right tools for the job yet; they're generally restricted to taking in and spitting out text via a chat interface. Things might get more interesting in business settings as AI companies start deploying so-called "AI agents," which can take action by operating other software on a computer or via the internet. Anthropic, a competitor to OpenAI, announced a major new product today that attempts to prove the thesis that tool use is needed for AI's next leap in usefulness.

Last Thanksgiving, Brian Israel found himself being asked the same question again and again. The general counsel at the AI lab Anthropic had been watching dumbfounded along with the rest of the tech world as, just two miles south of Anthropic's headquarters in San Francisco, its main competitor OpenAI seemed to be imploding. OpenAI's board had fired CEO Sam Altman, saying he had lost their confidence, in a move that seemed likely to tank the startup's 80 billion-plus valuation. The firing was only possible thanks to OpenAI's strange corporate structure, in which its directors have no fiduciary duty to increase profits for shareholders--a structure Altman himself had helped design so that OpenAI could build powerful AI insulated from perverse market incentives. To many, it appeared that plan had badly backfired.

When you picture the tech industry, you probably think of things that don't exist in physical space, such as the apps and internet browser on your phone. But the infrastructure required to store all this information – the physical datacentres housed in business parks and city outskirts – consume massive amounts of energy. Despite its name, the infrastructure used by the "cloud" accounts for more global greenhouse emissions than commercial flights. In 2018, for instance, the 5bn YouTube hits for the viral song Despacito used the same amount of energy it would take to heat 40,000 US homes annually. This is a hugely environmentally destructive side to the tech industry.

Large language models (LLMs) have shown great progress in responding to user questions, allowing for a multitude of diverse applications. Yet, the quality of LLM outputs heavily depends on the prompt design, where a good prompt might enable the LLM to answer a very challenging question correctly. Therefore, recent works have developed many strategies for improving the prompt, including both manual crafting and in-domain optimization. However, their efficacy in unrestricted scenarios remains questionable, as the former depends on human design for specific questions and the latter usually generalizes poorly to unseen scenarios. To address these problems, we give LLMs the freedom to design the best prompts according to themselves. Specifically, we include a hierarchy of LLMs, first constructing a prompt with precise instructions and accurate wording in a hierarchical manner, and then using this prompt to generate the final answer to the user query. We term this pipeline Hierarchical Multi-Agent Workflow, or HMAW. In contrast with prior works, HMAW imposes no human restriction and requires no training, and is completely task-agnostic while capable of adjusting to the nuances of the underlying task. Through both quantitative and qualitative experiments across multiple benchmarks, we verify that despite its simplicity, the proposed approach can create detailed and suitable prompts, further boosting the performance of current LLMs.

Large language models (LLMs) exhibit enhanced reasoning at larger scales, driving efforts to distill these capabilities into smaller models via teacher-student learning. Previous works simply fine-tune student models on teachers' generated Chain-of-Thoughts (CoTs) data. Although these methods enhance in-domain (IND) reasoning performance, they struggle to generalize to out-of-domain (OOD) tasks. We believe that the widespread spurious correlations between questions and answers may lead the model to preset a specific answer which restricts the diversity and generalizability of its reasoning process. In this paper, we propose Cascading Decomposed CoTs Distillation (CasCoD) to address these issues by decomposing the traditional single-step learning process into two cascaded learning steps. Specifically, by restructuring the training objectives -- removing the answer from outputs and concatenating the question with the rationale as input -- CasCoD's two-step learning process ensures that students focus on learning rationales without interference from the preset answers, thus improving reasoning generalizability. Extensive experiments demonstrate the effectiveness of CasCoD on both IND and OOD benchmark reasoning datasets. Code can be found at https://github.com/C-W-D/CasCoD.