Next Recession as Tipping Point for AI Job Losses

The startup EarthAI has announced that it has discovered potentially significant deposits of copper, cobalt, and gold in the Northern Territory and silver, molybdenum, and tin at another site in New South Wales. EarthAI was part of the Y Combinator’s 2019 beneficiaries and received 20 million USD in Series B financing in January of this year. The company uses AI to make predictions about potential mineral deposits. This is made easier in Australia since the government owns the rights to mineral deposits, and exploration companies have been obliged to submit findings on mineral deposits since the 1970s. This data was exploited by EarthAI. An interesting aspect of EarthAI’s story is that they encountered resistance from mining companies which is considered to be a “very conservative industry” where any method “outside of the approved dogma is considered heresy”. The success of EarthAI suggests that AI has an important role to play in future mineral discovery, notably in identifying regions that traditional mining companies may have overlooked.

Google has unveiled Gemini 2.5 – a new multimodal family of reasoning models. Like other reasoning models from Anthropic, DeepSeek, Google, and xAI, the Google models use extra inference computing to fact-check and reason in the construction of answers to questions. Google claims that Gemini 2.5 Pro is the company’s most intelligent family to date, notably for agentic tasks and for software development. It is said to outperform competitor models on the Aider Polyglot code editing benchmark with a score of 68.6%, though it performs less than Anthropic’s Claude 3.7 Sonnet on the software development abilities of the SWE-bench Verified benchmark. Gemini 2.5 Pro scored a worthy 18.8% on Humanity’s Last Exam – a benchmark of multimodal crowdsourced questions on topics from natural sciences, mathematics and the humanities. The model has a context window of 1 million tokens, meaning that a prompt can contain around 750’000 words – the equivalent of the “Lord of The Rings” anthology.

This article looks at the over supply of computational power in China due to the precipitated development of data centers across the country in the last three years. 500 new data centers were announced since 2023 and 150 newly built centers were in operation by the end of 2024. However, Chinese companies Jiazi Guangnian and 36Kr report that 80% of these new data centers remain unused. One reason for this waste is an overestimation of future needs and an excess of enthusiasm by companies and local governments following the decision by the Central government to make AI a national priority. There has also been a fall in the number of large models being developed: 144 companies announced to the Cyberspace Administration of China in 2024 that they were developing models, whereas only 10% were still investing in model training at the end of the year. Also, fewer companies are pre-training their own models, and are looking to DeepSeek’s R1 to leverage AI value. The downturn has even led to a huge drop in the price of smuggled Nvidia H100 chips from 28’000 USD per chip at the height of the boom. Another issue for data centers is the move towards reasoning models with less training upfront but more real-time inference computations. This requires architectures that are close to the users for reduced latency, which handicaps many of the new data centers that are far from large economic centers. Inference also requires different chips to those used up to now. Nividia’s H20 chip is a lighter chip that is optimized for AI inference and has now surpassed the H100 in popularity in China, despite US export sanctions.

This study from MIT Media Lab and OpenAI seeks to measure the emotional impact of prolonged interaction with chatbots in terms of loneliness, social isolation, dependence and problematic usage (which is defined as excessive and compulsive use, leading to negative impacts). The past months have seen a large increase in interactions with chatbot services. CharacterAI for instance now accounts for 20% of Google’s traffic, handling 20’000 queries per second, which is four times the traffic of ChatGPT. The context is the increased level of social isolation in industrialized societies, which the U.S. Surgeon General referred to as the “loneliness epidemic”. There is a 30% increased mortality risk associated with chronic loneliness. The article cites some positive benefits of chatbot interaction, such as its help in preventing suicide, but like for social media, prolonged interaction only increases social isolation which the concept was designed to prevent. This is an example of “problematic use of technology”, with loneliness being both the cause and effect of Internet usage. This study evaluated 981 users over four weeks with over 300’000 messages. Chatbot modalities were text and voice (with both emotive and neutral voice). Topics discussed were personal issues and impersonal topics (e.g., history, science). The study showed that chatbots can reduce loneliness, but extended usage can induce emotional dependence. This comes from the fact that chatbots mirror the emotional sentiment of user messages.

OpenAI has launched a new image generator in ChatGPT that has a more permissive content creation policy. The release made news with the tool being able to generate images in the style of Japanese Studio Ghibli. OpenAI now allows image creation in the style of studios or artists, though refuses to copy specific works. Another policy change is that the tool allows images of public figures to be created, though a public figure may explicitly opt out of having his images created and it refuses to create images that mock the person. OpenAI is now permitting the creation of “hateful symbols” like swastikas as long as they do not “clearly praise or endorse extremist agendas” and it has watered down its idea of “offensive” (e.g., a user can generate a picture of a person with “Asian eyes”). OpenAI nevertheless maintains strict guidelines on the creation of images with children. OpenAI fundamentally is aligning itself on the content creation policies of other platforms like Meta and X, and Meta management denies that the changes are politically motivated.

Elon Musk’s xAI has acquired Musk’s X for 33 billion USD. The deal combines the companies along with Tesla and SpaceX, and also pays back the 13 billion USD bank loans that Musk used to buy Twitter. The X platform can now facilitate the distribution of xAI products as well as providing data back to xAI for training. There had been a significant fall in advertising revenue on X (formerly Twitter) when Musk bought the platform because of his extreme right-wing rhetoric, though his closeness to Donald Trump and his role in the new US administration has led to a return of some advertisers. Also recently, the billionaire is pushing for the construction of a supercomputer cluster, called Colossus, in Memphis, Tennessee. He nevertheless failed in his court bid to prevent OpenAI transforming from a non-profit organization to a for-profit business.

This research paper from Carnegie Mellon, Stanford, Harvard and Princeton Universities highlights that contrary to popular belief, extending the pre-training of a model does not lead to better model performance. In fact, the opposite is true as extended pre-training can lead to models that are more difficult to fine-tune. The authors term this phenomenon catastrophic overtraining. There had been a rule of thumb that says the optimal ratio of tokens to model parameters is 20 to 1. However recent models have distorted this ratio with Llama-2-7B (7 billion parameters) being trained with 1.8 trillion tokens – giving a ratio of 260 to 1. The authors use several benchmarks (ARC-Easy, ARC-Challenge, PIQA, HellaSwag) and show that the instruction-tuned OLMo-1B model, pre-trained on 3 trillion tokens has over 2% worse performance than its 2.3 trillion token counterpart. Catastrophic pre-training from a progressive increase in model pre-training yields a greater sensitivity to parameter transformations, which in turn leads to increased forgetting after fine-tuning.

TechCrunch has relayed evidence that the Chinese government is using large models to increase its clamp down on dissidents. It is based on a dataset recovered from an unsecured Elasticsearch database hosted on a Baidu server. The dataset contains 133’000 examples of Internet content that needs to be flagged for censorship. These include references to military movements, Taiwan, political satire of current political leaders or complaints about party officials, or any negative comments about the social and economic situation in China. The examples are defined for “public opinion work” which is a term used by the Cyberspace Administration of China (CAC), the Internet regulator. Last February, OpenAI had reported use of its LLMs to track anti-government protests and to attack Chinese dissidents. The Chinese embassy in Washington rejected the allegations, calling them “groundless attacks and slanders against China”.