Price:

AI & TECH

SAP builds specialized transformers for enterprise data beyond generic LLMs

Friday, April 24, 2026 · from 3 podcasts
This Week in AINo PriorsThis Week in Startups
  • SAP developed RPT1 transformers to process tabular business data, claiming a 3-4% boost in forecast accuracy over general LLMs.
  • Enterprise AI faces a massive scale wall, breaking at 1,000 documents across 90 countries versus tidy 10-document demos.
  • The industry is splitting: specialized models for coding or finance are winning, while Apple bets on local orchestration for privacy.

Enterprise software is a world of spreadsheets, APIs, and compliance rules, not unstructured chat. Standard large language models, built for text, consistently fail at the relational logic of finance and supply chains.

SAP CTO Philipp Herzig detailed the company's solution on No Priors: a specialized architecture called RPT1, or Relational Pre-trained Transformers. Designed to interpret tabular data directly, it aims for a 3% to 4% improvement in demand forecast accuracy. In global logistics, that margin translates to millions in saved costs, a tangible outcome generic models can't capture. The goal is to automate the feature engineering that still requires expensive data scientists.

"Democratization in the predictive space has lagged behind generative AI."

- Philipp Herzig, No Priors

This push for specialization reflects a broader industry split. On This Week in AI, Perplexity CEO Aravind Srinivas argued that coding, being open-ended and creative, prevents LLMs from becoming true commodities. Models are developing distinct personalities - Claude leads in front-end design, while Codeex is preferred for Swift and Rust. Simultaneously, companies like Apple are building a different moat: using its Silicon to run agent orchestration loops locally on devices, sidestepping cloud latency and privacy issues for sensitive data.

The real challenge isn't building the AI but deploying it at enterprise scale. Herzig highlighted the 'context bloat' problem. A retrieval-augmented generation (RAG) prototype works on ten documents but collapses when reconciling a thousand across 90 countries with different tax laws. SAP manages 20,000 APIs; an agent must disambiguate whether a user needs a 'sales order' or a 'maintenance order' based on departmental data.

"Startups frequently underestimate the complexity of enterprise scale. It breaks when it has to reconcile 1,000 documents across 90 different countries."

- Philipp Herzig, No Priors

This engineering bottleneck is the final barrier to autonomy. Until agents can be verified against a system of record with absolute certainty for global compliance, they'll remain in advisory roles with humans supervising. The race isn't for the smartest chat model; it's for the most reliable system of record.

#Enterprise#Models#AI Infrastructure
PerplexityClaude

Source Intelligence

- Deep dive into what was said in the episodes

This Week in AI
This Week in AI

Aravind Srinivas & Edwin Chen: The $1B Bootstrap, Apple's AI Edge, and Benchmarks | TWiAI E10 • Apr 23

  • Perplexity Computer provides an intuitive interface for orchestrating multiple AI agents without API keys or complex setup, offering access to various models and connectors for tasks like research, browser automation, and data analysis.
  • Edwin Chen, founder and CEO of Serge AI, clarified his company's role as 'AI teaching' rather than simple 'data labeling,' employing highly educated experts to cross-examine and instill values, wisdom, and taste into frontier models.
  • Serge AI, which has never raised venture capital, has approximately 130 employees and 50,000 expert contractors, providing services to major AI labs including OpenAI, Google, Anthropic, Microsoft, and Meta.
  • Aravind Shavas advises founders to be disciplined capital allocators, maintaining a 'bootstrap founder' mentality even after raising substantial funds, and suggests Perplexity AI aims for profitability without significant increases in payroll or infrastructure.
  • Apple's M-series chips are an underrated asset for local LLM inference, outperforming DGX Spark benchmarks, and the company has already secured two-nanometer chip fab capacity for next year, positioning it for future local AI processing.
  • Aravind Shavas envisions Apple enabling local agent loops to run on devices, preserving user privacy for personal data like photos and messages, believing Apple is uniquely positioned to profit from this due to its chips, OS, and ecosystem.
  • Edwin Chen believes AI models will not be commodified due to their distinct personalities and specializations, arguing users will naturally prefer different models based on their mood or the specific task, similar to choosing friends.
  • Aravind Shavas argues that value in AI accrues at the application layer, as pure API model companies struggle to maintain a significant lead for long; the performance gap between frontier models typically shrinks to months.
  • Edwin Chen describes LM Arena as a 'terrible cancer on AI,' leading models to prioritize 'pretty formatting' over correctness due to companies optimizing for its visible yet flawed benchmarks, making models ultimately worse.
  • For model evaluation, Edwin Chen advocates measuring real-world human usage and practical helpfulness, rather than contrived benchmarks, ensuring models produce creative and useful outputs that truly benefit users.
  • Aravind Shavas developed Perplexity's 'Model Council' at Jensen Huang's suggestion, allowing users to query multiple models simultaneously, compare their responses, and receive a synthesized analysis highlighting agreements and disagreements.
  • Jason Calacanis praises Whisper Flow for its superior speech-to-text accuracy, especially when combined with a foot pedal for dictating long, detailed prompts, which significantly improves model responses.
  • Aravind Shavas is impressed by the improved Grok integration within X, particularly the 'Explain Grok' button for contextualizing tweets, and by the Gemini 3 Flash model for its exceptional speed and intelligence.
  • Edwin Chen highlights Claude Design for its well-designed and opinionated interface, which enables non-designers to rapidly prototype new interfaces and landing pages, accelerating product development.
  • Edwin Chen and Jason Calacanis share experiences where AI models, like those analyzing blood work results or personalized health trackers (e.g., Whoop), provided more effective and tailored health recommendations than human doctors.
Also from this episode: (2)

Agents (1)

  • Aravind Shavas, CEO of Perplexity AI, confirmed the company's revenue grew to $500 million recently, driven primarily by Perplexity Computer's success in simplifying agent orchestration for users.

VC (1)

  • Edwin Chen criticizes the trend of raising excessive AI capital, arguing it incentivizes volume over quality, diverts CEO focus from product to fundraising, and can lead to models prioritizing engagement over genuine usefulness.
#Chips#Agents#Startups#Enterprise
No Priors: Artificial Intelligence | Technology | Startups
No Priors: Artificial Intelligence | Technology | Startups

SAP: Bringing the ‘Operating System’ of a Company into the AI Era with CTO Philipp Herzig • Apr 23

  • Philipp Herzig states SAP is the market leader in enterprise software, serving 400,000 customers globally by managing finance, HR, supply chain, and customer-facing operations end-to-end.
  • SAP was founded in 1972 on the idea of standard software, stemming from IBM employees repeatedly implementing similar finance systems for different customers, recognizing the need for scalable solutions.
  • Herzig notes SAP is fully adopting AI, with developers using generative coding for productivity. "Drew for Consulting," an AI product, helps consultants reduce project efforts by 30%, making it one of SAP's fastest-growing AI offerings.
  • Herzig outlines three areas AI is transforming SAP's software: dynamically generated UIs, agents blending structured and unstructured business processes, and a harmonized data layer to fuel AI.
  • Herzig identifies the biggest engineering challenge as teaching AI to operate effectively at scale across thousands of documents and 20,000 APIs, rather than the AI technology itself, which often works for small POCs.
  • Philipp Herzig emphasizes writing "evals" for AI agents to ensure reliable outcomes, akin to test-driven development. This shifts developer focus from coding to defining desired system behavior and boundary conditions.
  • SAP is developing "agent mining" to record user decision traces and contextual "tribal knowledge." This data flywheel helps identify process anomalies or elevate new efficient practices to company-wide standards.
  • Herzig states LLMs excel in unstructured data but are not suited for forward planning or high-accuracy predictions. Classical machine learning (XG boost, auto-ML) remains crucial for tasks like demand or cash flow forecasting.
  • SAP developed RPT1 (Relational Pre-trained Transformers), a transformer-based architecture, to enable high-accuracy predictions on structured data with small inputs. This research was published in NeurIPS and aims to democratize predictive analytics.
  • Herzig points to disaggregated data, the complexity of scaling AI across large landscapes, and security concerns (citing vulnerabilities like Light LLM) as significant challenges for enterprise AI adoption.
  • SAP is transitioning from seat-based to a hybrid consumptive licensing model for AI, acknowledging customer demand for predictability and outcome trust. A purely outcome-based model is a future aspiration.
  • Herzig predicts AI will automate mundane tasks, allowing employees to focus on strategic thinking, deeper insights, and supervising AI outputs, effectively "up-leveling" every role in the organization.
  • Herzig highlights SAP's early research in quantum computing for solving hard optimization problems like logistics and the traveling salesman problem. This could lead to significant savings and reduced emissions.
#Enterprise#Agents#Models#Reasoning
This Week in Startups
This Week in Startups

SpaceX and Cursor team up to topple Claude Code | E2279 • Apr 22

  • SpaceX bets billions on Cursor to secure the data needed for recursive AI self-improvement.
  • Bitstarter launches a crowdfunding model to strip power from predatory Bit Tensor investors.
  • Subnet 11 creates a sandbox for AI agents to write their own instruction sets.
#AI Infrastructure#Models#Startups#Agents

Related Stories