Three major announcements took center stage this week: OpenAI's reveal of Operator, the company's first attempt at an AI system that can autonomously take on tasks within a web browser; the official announcement of the Stargate Project, an enormous joint undertaking by OpenAI and SoftBank which aims to raise and invest as much as $500 billion in energy and data center infrastructure for OpenAI; and the launch of DeepSeek-R1, an openly available 'reasoning' model with a benchmark performance comparable to OpenAI's o1.

At the beginning of the week, seemingly innocuous reports emerged that DeepSeek had released Deepseek-R1 under an MIT License, making the model free to use for commercial and non-commercial applications. Initially, the announcement did not have a lot of impact. After all, DeepSeek had already launched a preview version of the R1 model (R1-Lite-Preview) in the DeepSeek Chat interface. DeepSeek was also far from the first Chinese company to claim that its reasoning model rivaled OpenAI's offerings; Alibaba's Qwen with Questions had already been launched claiming superior performance than OpenAI's o1-mini.

However, DeepSeek R1 slowly started garnering attention, mainly because it can be used free of charge even for commercial applications. This contrasts starkly with OpenAI's o1, which is available for paying users only. The price tag and the already highly-criticized decision by OpenAI to hide o1's 'thought' process started swaying user preference toward DeepSeek-R1. This was despite emphatic commentary about how, unlike OpenAI's models, DeepSeek-R1 had to comply with Chinese regulations, which mostly amounted to censorship on several sensitive topics.

Then, barely a day after the R1 official release, OpenAI and SoftBank announced the ambitious Stargate Project, a $500 billion private infrastructure venture backed by Oracle and Abu Dhabi's MGX fund to serve OpenAI's computing needs. The Stargate Project was simultaneously praised by President Trump as a "declaration of confidence in America's potential" and criticized by Elon Musk, who swiftly remarked that the project did not have the capital it claimed it did. CEO Sam Altman has established Stargate as a necessary step to provide OpenAI with the energy and computing resources it needs to develop AGI, and more importantly, to conquer that goal before China does.

While independently impressive, the Stargate Project also helped cast an even more positive light on DeepSeek's mission. Indeed, one of the most commented-on facts about R1: DeepSeek-V3, which is R1's base model, was developed on a handful of NVIDIA GPUs for approximately $6 million (a joke of a budget, according to AI researcher Andrej Karpathy). To put things into perspective, OpenAI CEO Sam Altman has claimed that GPT-4 cost over $100 million to train. If anything, the rise of DeepSeek has encouraged a critical look at the multi-billion investments driving the US-based AI industry forward.

Lately, tech giants like OpenAI, Google, and Anthropic have delved into the agentic AI narrative to justify the climbing costs of their ventures. Broadly, these organizations claim that agentic AI will unlock an essential component of the highly sought-after artificial general intelligence: the capability of AI systems to perform increasingly complex workflows unsupervised. Anthropic was among the first to take a stab at the challenge with Claude with computer use.

In light of the recent developments, but also following a highly discussed blog post by Sam Altman, in which the OpenAI CEO recorded his prediction that 2025 could well be the year in which AI agents finally "join the workforce", OpenAI released the Operator research preview, which enables US-based Pro subscribers to experiment with an agent capable of performing tasks like booking flights and accommodations, reserving restaurants, and even shopping for groceries, for them. OpenAI has shared its plans to expand Operator access to its other subscription tiers and eventually incorporate Operator's capabilities into ChatGPT.

Other noteworthy headlines this week

Synthesia recently became the UK's most valuable generative AI company: Synthesia, a London-based AI video communications platform, has raised $180M in Series D funding led by NEA, making it the UK's most valuable generative AI media company with a $2.1B valuation and over 60,000 customers.

Microsoft's rStar-Math technique enhances SLMs with complex math problem-solving skills: Microsoft's new rStar-Math technique enables small AI models to outperform OpenAI's o1-preview on complex mathematical problems, achieving 90% accuracy on the MATH benchmark. This innovative approach overcomes challenges associated with alternatives such as model distillation and reward models.

Microsoft Reverses Changes to Bing Image Creator Amid User Backlash: Microsoft responds to user feedback by reversing controversial changes to Bing Image Creator, ensuring a better user experience for creative AI tools.

The new Sonar API enables anyone to build with Perplexity's generative search capabilities: Perplexity has launched its Sonar Pro API, offering developers access to an AI-powered search solution that combines real-time internet connectivity with language model capabilities to deliver highly accurate, citation-backed responses.