Why Open Models Will Beat GPT and Claude
This episode explores the strategic shift of enterprises towards open AI models driven by cost and data sovereignty, highlighting the evolving tooling and hybrid deployment strategies. It also delves into the "attenuation funnel" phenomenon, where increased code generation by AI tools doesn't proportionally translate to shipped software or end-user value, and introduces "neo-consulting"—a new business model challenging traditional consulting with outcomes-based pricing and AI-focused expertise.
This discussion, featuring Evangelos Simoudis of Synapse Partners, delves into three key trends shaping the AI and Machine Learning landscape for enterprises: the rise of open models, the productivity paradox in coding with generative AI, and the emergence of ’neo-consulting'.
The Enterprise Case for Open Models: Cost, Control, and Data Sovereignty
Simoudis argues that while enterprises initially adopt frontier models, they quickly encounter high costs and a need for corporate data sovereignty. This drives a shift towards open models, which allow for:
- Modification with proprietary data: Enhancing models with internal datasets.
- Cost control: Managing expenses more effectively than with proprietary APIs.
- Corporate sovereignty: Maintaining full control over their IP and data.
He envisions a progression from full cloud-based proprietary models to a hybrid deployment, where some models run in-house, eventually leading to enterprises running their own ‘AI factories’ in their data centers. This approach is essential for building internal expertise and control.
Ben reinforces this by noting the rapid advancement of open models, with the performance gap between open-weight models and frontier models narrowing from 12 months to 3-6 months. Anecdotal evidence suggests open-weight models can realistically handle 80% of routine enterprise tasks, reserving frontier models for the remaining 20% that require deep reasoning (from labs like Anthropic, OpenAI, DeepMind).
While many open models originate from China, questions about provenance and associated risks are noted. However, enterprises are increasingly understanding how to match a task’s requirements with a model’s capabilities, leading to a more holistic view of workflow integration and cost-benefit analysis of open models.
Simoudis emphasizes cost and corporate sovereignty as paramount reasons. Ben adds that owning your own AI is critical, enabled by evolving tools and infrastructure for both specialized model pre-training (e.g., DataDog, healthcare startups using models in the hundreds of millions to a few billion parameters) and post-training (e.g., supervised and reinforcement fine-tuning tools). This ownership enables a compounding loop of continuous model improvement based on production usage and user feedback.
Building these AI factories requires engagement beyond just IT, involving:
- Financial organization: For managing token costs, task completion costs, and attribution.
- HR: For talent acquisition and educating end-users.
- Process owners: To ensure ROI and guide AI application in workflows.
The Attenuation Funnel: More Code, Barely More Shipped Software
The discussion then shifts to the impact of generative AI and agents on programming. While developers are undoubtedly writing 2-3x more code (based on surveys and telemetry), this increased activity doesn’t translate proportionally to shipped software or end-user adoption. This is described as an **