Attention’s Reign Challenged: New ‘Power Retention’ Model Slashes AI Training Costs by 98% | SAP’s Business AI Arrives, Market Research Grapples with Trust

Attention’s Reign Challenged: New ‘Power Retention’ Model Slashes AI Training Costs by 98% | SAP’s Business AI Arrives, Market Research Grapples with Trust

A futuristic AI neural network with a prominent graphic showing 98% cost reduction, subtly integrated with SAP branding, illustrating the new 'Power Retention' model's impact.

Key Takeaways

  • Manifest AI’s Brumby-14B-Base introduces a “Power Retention” architecture, replacing attention layers for significant cost reduction and efficiency in LLMs, achieving performance parity with state-of-the-art transformers.
  • SAP launches RPT-1, a specialized relational foundation model pre-trained on business data, enabling out-of-the-box predictive analytics for enterprises without extensive fine-tuning.
  • A new survey reveals 98% of market researchers use AI daily, but 39% report errors and 37% cite data quality risks, highlighting a critical trust gap that necessitates human oversight.

Main Developments

Eight years after Google’s “Attention Is All You Need” paper laid the groundwork for modern AI, a radical new architecture is emerging to challenge the transformer’s dominance. Manifest AI has unveiled Brumby-14B-Base, a variant of Qwen3-14B-Base, that abandons attention entirely in favor of a novel mechanism called Power Retention. This recurrent, hardware-efficient architecture stores and updates information over arbitrarily long contexts without the quadratic memory growth that plagues attention-based models. In a stunning demonstration of efficiency, Brumby-14B-Base was retrained at a stated cost of just $4,000—less than 2% of what a conventional model of its scale would cost from scratch. The 14-billion-parameter model performs on par with established transformers like Qwen3-14B and GLM-4.5-Air, particularly excelling in mathematical and long-context reasoning where attention typically falters. This “Power Retention” system promises constant-time per-token computation, regardless of context length, and boasts superior hardware utilization over even optimized attention mechanisms like FlashAttention2. Jacob Buckman, founder of Manifest AI, emphasized that while building from scratch would be more expensive, the ability to retrain existing transformer weights with this new paradigm significantly accelerates its adoption, potentially democratizing large-scale AI experimentation.

This architectural shift arrives as enterprises continue to seek AI solutions tailored to their specific needs. SAP has responded with the launch of RPT-1, its first relational foundation model. Unlike general large language models trained on text and code, RPT-1 is pre-trained on business transactions and tabular data, making it semantically aware of enterprise contexts. SAP claims RPT-1 can perform predictive analytics and build business models out-of-the-box, without the extensive fine-tuning typically required for general LLMs. This specialized approach, building on SAP’s proprietary ConTextTab architecture, aims to provide more structured and precise answers for financial and business use cases, directly challenging the need for enterprises to retrain broader AI systems for specific tasks.

Meanwhile, companies like Zendesk are already pushing the boundaries of applied AI by integrating advanced models like OpenAI’s GPT-5. Zendesk’s AI agents now handle nearly 80% of customer requests autonomously, with GPT-5 significantly enhancing their ability to reason, take action, and maintain 95%+ execution reliability. This has led to a 30% reduction in workflow failures and a 20% decrease in escalations. Zendesk has further bolstered its real-time intelligence by acquiring HyperArc, an AI-native analytics platform, to merge structured and unstructured support data, providing proactive insights and recommendations for the entire company.

Despite these advancements, the real-world adoption of AI is not without its challenges. A new QuestDIY survey of market researchers reveals a striking paradox: 98% of professionals use AI daily, with 72% using it multiple times a day, yet a significant trust problem persists. Nearly four in ten researchers report “increased reliance on technology that sometimes produces errors,” and 37% cite “new risks around data quality or accuracy.” This leads to 31% spending more time re-checking and validating AI outputs, effectively creating new work while saving time. Data privacy and security concerns are also a major barrier. The industry’s current operating model treats AI as a “junior analyst,” requiring constant human oversight and judgment to ensure credibility, underscoring that raw speed must be balanced with verifiable accuracy and ethical considerations.

Analyst’s View

The simultaneous emergence of Manifest AI’s Power Retention and SAP’s RPT-1 signals a pivotal moment for AI. Power Retention’s efficiency promises to break the computational and memory bottlenecks of attention, potentially democratizing access to large-scale model development and accelerating architectural diversification beyond the transformer monoculture. If scalable to larger models, this could dramatically reshape the economics of AI. Complementing this, SAP’s RPT-1 exemplifies the growing importance of specialized, domain-aware AI that offers out-of-the-box reliability for specific business tasks, moving beyond the “one-size-fits-all” approach of general LLMs. However, the market research survey serves as a crucial reality check: widespread adoption won’t guarantee trust. The persistent issues with accuracy, transparency, and data privacy highlight that innovation must go hand-in-hand with robust validation, ethical governance, and a clear understanding of AI’s limitations. The future of AI will likely be a mosaic of architectural breakthroughs, highly specialized applications, and an evolving “human-led” partnership where critical judgment remains paramount.

Source Material

阅读中文版 (Read Chinese Version)

Comments are closed.