Llama 4 Launch: 36 Hours of Backlash and Employee Claims

Select Language:

Meta’s latest foundational model, Llama 4, was released just 36 hours ago, but the response on social media has been far from enthusiastic. Many users have expressed their disappointment, with comments such as “very disappointing” and “it doesn’t seem to work well.” Users have noted that the model failed multiple tests and even turned the disappointing experience into memes, leading to an avalanche of negative reviews.

Critics have focused particularly on Llama 4’s coding capabilities. A notable instance from a well-known “atmosphere programming” test saw a virtual ball simply pass through walls instead of bouncing off them, sparking further frustration among users. Official testing scores from the LiveCodeBench at launch appeared favorable, but third-party benchmarks showed a stark contrast, placing Llama 4 at the bottom of the rankings. This discrepancy has raised questions about whether the high scores were a result of data overfitting or even manipulation.

Just days before Llama 4’s launch, Joelle Pineau, the head of Meta AI research, announced her sudden departure after eight years with the company, adding to the air of uncertainty surrounding the release.

Further complicating the situation, an anonymous tip from within the Meta GenAI department reported that an employee submitted their resignation and requested not to be credited on Llama 4’s technical report. This claim circulated on a platform popular among job seekers and international students, generating significant discussion both in the U.S. and abroad.

While this allegation hasn’t been confirmed, Meta GenAI head Ahmad Al-Dahle’s posts indicated that the model running in the competitive benchmarks was a special version. Furthermore, former employees pointed out ongoing data leakage issues that have plagued Llama models since Llama 1.

Concerns extend beyond just coding abilities; Llama 4 also received low scores for long-form writing in the EQBench assessments. The model exhibited a tendency to repeat content in later sections of its writing and appeared overly formulaic. Experts speculated that recent copyright litigation involving Meta may have led to the removal of substantial data from the internet and books, forcing the company to rely more on synthetic data.

In this litigation, numerous authors discovered their works used for AI training without consent, prompting protests outside Meta’s London office. The fallout from Llama 4’s release has evoked memories of earlier anonymous employee complaints, with users reflecting on how the situation has turned from casual interest into serious skepticism regarding the model’s capabilities.

In the wake of Llama 4’s troubled launch, comparisons have been made to the recent success of the DeepSeek v3 model, which raises a poignant question regarding the future of Meta’s AI development. As an era of rapid advancements continues, the shake-up within Meta may be indicative of broader challenges ahead for the tech giant in the competitive landscape of artificial intelligence.