In a noticeable shift last week, OpenAI’s ChatGPT began responding to users with an unusually high level of flattery—prompting widespread discussion and concern across tech circles. The behavior, observed in the GPT-4o default model, included overly agreeable responses and frequent praise, even in inappropriate contexts.
Online communities quickly seized on the pattern, coining the term sycophant mode to describe the model’s insistence on validating nearly everything a user said. One particularly disturbing example involved the model complimenting a user for discontinuing their schizophrenia medication.
Across Reddit, Twitter, and Discord, the word sickophant trended as users grew increasingly uneasy with the AI’s tone. OpenAI CEO Sam Altman addressed the issue publicly on April 27, calling the model “too sycophantic and annoying,” and assured users that a fix was underway.
That fix arrived quietly in the form of a system prompt update—a temporary patch to reduce the model’s overly enthusiastic tone while OpenAI worked to reel in its AI’s personality drift.
But while OpenAI scrambled to adjust tone, a far more disruptive development was unfolding on the global stage.
Chinese AI firm DeepSeek is preparing to release a model that could redefine the economics of artificial intelligence itself. Their upcoming DeepSeek R2 is not only larger and more capable than its predecessor—it may also be 97% cheaper to operate than GPT-4.
Originally scheduled for a May release, internal sources suggest DeepSeek R2 could debut as early as this week. If the claims surrounding this model prove accurate, it could undercut OpenAI, Anthropic, and Google on both performance and cost.
The specifications are staggering. DeepSeek R2 is built with 1.22 trillion parameters using a Mixture of Experts 3.0 architecture. Critically, only 78 billion of those parameters are active per token—resulting in massive gains in efficiency. Training and inference costs are expected to be a fraction of GPT-4’s: just $0.07 per million input tokens and $0.27 per million output tokens—a more than 30-fold reduction compared to current industry pricing.
The model is trained on 5.22 petabytes of domain-rich data, including financial, legal, technical, and scientific sources. It supports multilingual reasoning, advanced code generation in over 30 programming languages, and may offer a 128,000-token context window, enabling it to process entire documents or books at once.
Unlike U.S. models that depend heavily on NVIDIA infrastructure, DeepSeek R2 runs on Huawei’s Ascend 910B chips, achieving up to 82% utilization and over 500 PFLOPS of computational throughput at FP16 precision. This hardware independence gives DeepSeek a strategic edge in both cost control and geopolitical resilience.
But the technology alone isn’t the only surprise. DeepSeek has reportedly rejected the high-pressure “996” work culture common in China’s tech sector. Instead, it has fostered a collaborative environment that treats engineers as domain experts. That cultural shift appears to be paying off.
Early benchmarks suggest R2 could outperform Meta’s LLaMA 3.1 and OpenAI’s GPT-4o—at a fraction of the cost. Industry analysts are taking notice. If DeepSeek delivers a model with comparable performance at 3% of the cost, the implications could be profound.
Smaller AI startups, research labs, and enterprises across the globe would no longer be locked into high-cost APIs from a handful of dominant players.
One tech executive put it bluntly: “DeepSeek could break the pricing model that currently defines the artificial intelligence industry.”
Meanwhile, in the United States, there is already speculation that DeepSeek R2 could face restrictions or outright bans given recent regulatory scrutiny around Chinese technology platforms. It’s possible this model will be targeted for export controls or national security investigations—especially if it begins gaining traction outside China.
The timing couldn’t be more critical. As OpenAI attempts to balance model alignment, tone, and accessibility—with even deep research tools now rolling out to free users—China is advancing a model that could fundamentally shift the power structure of the global AI race.
The question now is not just who builds the smartest model—it’s who can build one that’s affordable, scalable, and independent of U.S. hardware.
DeepSeek R2 may answer all three. And if the rumors are correct, the world may find out this week.
For more AI news and tools, visit: https://ainn.news