Sora 2 - OpenAI's Tool for Next Generation AI Slop?

This MIT Technology Review article highlights the negative impact of AI-based language translators on lesser used languages.

• UNESCO, the United Nations Educational, Scientific and Cultural Organization, says that a language is declared extinct every two weeks.
• Wikipedia, “the most ambitious multilingual project after the Bible”, is seen by many as a means for keeping a language alive. For instance, a Greenlandic-language version of Wikipedia now exists with over 1’500 articles. Greenlandic is spoken by around 57’000 people in Arctic villages.
• The problem is that, while the first Wikipedia articles were written by Greenlandic speakers, many articles today are AI-translated from English. These translators are very poor at Greenlandic and other lesser-spoken languages. The result is that the most public and largest collection of texts in Greenlandic are AI slop.
• Wikipedia is the primary source of content in lesser-spoken languages for AI model training. The slop penalizes the training of AI language models with content in that natural language. Thus, the next generation of AI translators will naturally be worse quality.
• Wikipedias are susceptible to “bigger-Wikipedia arrogance”. When mistakes are made in in articles written in English, the community of contributors is sufficiently large from one contributor to correct errors made by another. Wikipedias for lesser-spoken languages do not have the critical mass of contributors for this.

Meta has announced that data collected on its social media platforms from user interactions with its AI services will soon be used to build targeted advertisements to users and user groups.

• Meta says that there is no opt-out option for users.
• Target advertising will be applied globally, except in South Korea, the United Kingdom, and the European Union. This exception is due to privacy laws in those regions, e.g., the GDPR.
• The data analyzed by Meta will include text and videos. Even content processed by Meta’s smart glasses will be used for AI training.
• Meta nonetheless says that sensitive data – data about ethnicity, health, religion or sexual orientation – will not be included in training data.

AI agents has been a key topic in 2025. Microsoft is releasing the Agent Framework for public review which should become the company’s sole agent and orchestration framework. The release highlights a multitude of offerings around the theme of AI agents, as companies are still figuring out what precisely these frameworks are meant to do. In the case of Microsoft, there have been the following offerings:

• The Azure AI Foundry is a cloud platform for building, testing, deploying and orchestrating AI models and agents. Users have access to a catalog of AI models and RAG tools. One use case is the development of an LLM pipeline that connects to data sources and implements agents with RAG from these sources.
• The Semantic Kernel is a software development kit (SDK) for programmatically embedding AI features into applications. For instance, traditional applications can call “skills” modules for AI functions like text summarization or search, and can call embedding based storage.
• Autogen is a multi-agent orchestration and coordination framework where developers can define how agents interact with one another. For instance, a coder agent might review code which is then reviewed by a tester agent, and then passed on to a deployer agent.

The goal of the Agent Framework is to unify these offerings. The platform will integrate measures to verify task adherence (that agents are doing what they are supposed to do), as well as measures to detect personal information leakage and defend against prompt injection attacks.

The EU AI Act came into effect on August 1st 2024, and will be fully operational by August 2026.

• Calls have been made on the EU to delay the implementation. One reason is the current lack of technical standards for evaluating AI. Another reason cited is that many member states do not yet have the competent bodies to oversee the regulation in their country.
• The EU is preparing a reporting template for providers of high-risk AI systems to report incidents, in an effort to increase public confidence in AI. The approach aligns with that taken by the OECD’s AI Incidents Monitor and Common Reporting Framework.
• The EU is launching an initiative to simplify legislation in the digital domain, in an effort to reduce the administrative overhead on companies.

Recent AI models like GPT-5, Gemini 2.5 and Sonnet 2.4 are considered to be better than their predecessors at software coding tasks. OpenAI’s Sora 2 video generation software is also seen as a much improved version of video generation chatbots.

• Reinforcement learning with human graders is arguably the biggest driver that explain these improvements.
• On the other hand, improvements in AI models are not necessarily visible for tasks like writing quality mails, because of the absence of objective and clear pass-fail metrics.
• The definition of clear pass-fail metrics will determine the degree to which manual work processes will improve with the aid of AI chatbots.

The Guardian newspaper recorded a discussion between Yuval Noah Harari (historian and author of Sapiens and Homo Deus), Maria Ressa (former winner of the Nobel peace prize) and Rory Stewart (former UK minister and the host of Rest is Politics podcast). The three discuss the current world order, and broach the question of the impact of technology and AI.

• Technology is evolving currently in a world of increased authoritarian populism. Modern liberalism emerged in the 18th century, and its fundamental tenet was that in a democracy, people could “agree to disagree”. This consensus is eroding today and technology is playing a role through its possibility of manipulating opinions.
• AI is contributing to public anxiety, much like climate change. Nothing is being done to address this anxiety since in the West, there are still people who refuse to believe that climate change exists. One overall effect of anxiety is to erode confidence in public institutions.
• Gaza and Ukraine are examples of how a country with technology can inflict great damage on another country while taking relatively little risk.
• The increased risk of conflict in the world is causing countries to invest more in military spending, meaning that healthcare investments will be smaller – even though aging populations will require more care. AI will evolve in a world of smaller median wages and increased conflict.
• Historically, greater intelligence – artificial of natural – does not correlate with greater wisdom or compassion in humans.

A study from Shanghai Jiao Tong University and SII Generative AI Research Lab (GAIR) argues that the quantity of training data for AI models can be significantly reduced as long as its quality is high.

• Their LIMI (Less Is More for Intelligent Agency) framework used a curated dataset of just 78 examples to train LLMs that outperform models trained on thousands of examples by a considerable margin on key industry benchmarks. Specifically, LIMI achieves 73.5% on AgencyBench which is significant improvement on Kimi-K2-Instruct (24.1%), DeepSeek-V3.1 (11.9%), Qwen3-235B-A22B-Instruct (27.5%), and GLM-4.5 (45.1%). The framework shows 53.7% improvement over models trained on 10000 samples, thereby showing superior agentic intelligence with 128 times less samples.
• Another observation is that LLMs must exhibit “agency” – the ability not just to think but to act. The models were trained using demonstrations where each has a query (e.g., “build a chat application”) and a trajectory. The trajectory shows through an example how the query may be solved and verified.

OpenAI released Sora 2 – a TikTok-style app that allows users to post exclusively AI-generated videos up to 10 seconds long.

• The app immediately became the top downloaded app on Apple’s US AppStore in North America – where it is available.
• A former OpenAI researcher has referred to Sora as an “infinite AI TikTok slop machine”.
• OpenAI will have to add safeguards so that users can refuse to have themselves “avatarized” saying certain phrases or in certain situations.
• Videos posted on Sora include trademarked characters (e.g., Scooby Doo). OpenAI is believed to have contacted several copyright holders recently, asking them to opt-out of having their works appear on Sora. This is controversial since most authors prefer an opt-in approach – where their work is only used with their explicit consent.
• OpenAI has not said what the energy footprint for a Sora video is.
• OpenAI is still not a profitable company, so it is not clear how Sora can be funded in the long term. ChatGPT could soon integrated advertisements to generate revenue for OpenAI.

Pope Leo has urged news sites to avoid using click-bait – sensational, misleading or exaggerated content, often written using generative AI, designed to attract visitors to websites.

• He said that “communication must be freed from the misguided thinking that corrupts it ... and from the degrading practice of so-called click-bait".
• He also reiterated the importance of journalists in today’s world: “If today we know what is happening in Gaza, Ukraine, and every other land bloodied by bombs, we largely owe it to them”.
• However, he added, “artificial intelligence is changing the way we receive information and communicate, but who directs it and for what purposes? We must be vigilant in order to ensure that technology does not replace human beings.".

This article reports on the author’s impression of her interview with Chris Lehane, OpenAI’s VP of global policy at the Elevate conference in Toronto.

• For Lehane, Sora 2 is a “general purpose technology” like the printing press, democratizing creativity for people without talent or resources.
• He remained evasive on the problems in US towns where data centers are using up power supply and water resources. He did admit that OpenAI needs a gigawatt of power each week. China brought 450 gigawatts online last year as well as 33 nuclear facilities. For Lehane, democracies can only compete with China by modernizing energy infrastructures.
• Meanwhile, OpenAI has been accused of intimidation tactics against employees who may have been collaborating with Californian legislators in preparation of the SB 53 AI Safety Bill. Officially, OpenAI supports the bill, but insiders contradict this.
• Former and current OpenAI employees are disappointed with Sora 2 in that it is far from OpenAI’s original objective of an AI that benefits humanity. One employee said the company needed to avoid the “pitfalls of other social media apps and deepfakes”.