Overview of The Myth of Model Wars: Open vs Closed AI in 2026
This Practical AI episode is a wide-ranging, off-the-cuff discussion about where AI value is really heading in 2026. The hosts argue that while the open-vs-closed model debate still matters in specific cases, the bigger story is the rise of physical AI, small on-device models, and especially the agentic systems and infrastructure built around models. Their core thesis: models are becoming a commodity, and the real differentiation is shifting to workflows, orchestration, governance, and productized systems that turn AI into useful outcomes.
Physical AI and the Rise of Embedded Intelligence
The conversation opens with optimism about “physical AI” — AI embedded in the real world rather than living mostly in cloud environments.
Key examples mentioned
- Retail kiosks and in-store assistants
- Warehouse and manufacturing-floor automation
- Delivery drones
- Service robots in restaurants and stores
- Autonomous vehicles
- Consumer devices such as glasses and other wearables
Main takeaway
The hosts see this as an early-stage but rapidly expanding market. What feels futuristic is already becoming normal, and the next 1–2 years are expected to bring many more AI-enabled products into everyday life.
What “Open” vs “Closed” Models Means
A substantial part of the episode is spent clarifying terminology, since the distinction is often misunderstood.
Open / open-weight / open-source models
- The model’s weights and inference code are made available publicly
- Users or organizations can run the model themselves
- They can deploy it on their own hardware, cloud, or air-gapped systems
- It may be released under permissive or restrictive licenses
Closed / proprietary models
- The model weights and code stay on the vendor’s infrastructure
- Access is provided through:
- a hosted chat product
- an API
- a managed service from a cloud provider
- Users interact with the model, but do not possess or run it directly
Why this matters
The hosts emphasize that this distinction affects:
- privacy and sovereignty
- deployment flexibility
- regulatory suitability
- defense and government use cases
- total cost at scale
Is the Open-vs-Closed Gap Still Closing?
The hosts discuss benchmark results and the long-running perception that open models were catching up to frontier closed models.
Benchmarks discussed
- MMLU
- SWE-bench / coding benchmarks
- reasoning and math evaluations
- leaderboard-style arenas
Their assessment
- Open models still trail the best closed models by a few points in many benchmark categories
- That gap matters less in some real-world workflows than in headline comparisons
- Benchmarks can be useful, but they do not fully reflect practical value
Important nuance
The hosts suggest the model race is not the whole story anymore. For many users, especially in physical AI and enterprise automation, the key question is not “Which model wins?” but “What system are we building around it?”
Meta, Llama, and the State of Open Models
The episode highlights a major industry shift: the hosts say Meta has moved away from its open-model strategy, effectively abandoning Llama as a flagship open-source direction and turning toward a closed model family called “MuseSpark” in the conversation.
Why this is significant
- Meta had been one of the major Western champions of open models
- Llama had become a key reference point for the open ecosystem
- A shift away from open models is seen as a blow to the Western open-source AI landscape
Geopolitical angle
The hosts also note:
- China is increasingly leading in open-model development
- Western companies and governments may prefer domestically created models for security and procurement reasons
- This creates real business and compliance constraints for companies building AI products
The Real Value Is Moving to Agentic Systems
One of the strongest points in the episode is that models are becoming commoditized, and the value is increasingly in the systems around them.
The “commodity” analogy
The hosts compare models to raw materials like soy or corn:
- The ingredient matters
- But the final customer cares more about the assembled product
- In AI, the product includes the workflow, interfaces, orchestration, and business logic
Where the value is now
- RAG systems
- tool calling / function calling
- MCP integrations
- agent-to-agent communication
- automation workflows
- coding assistants and code generation systems
- governance and policy enforcement
- monitoring and observability
Core message
Even if the underlying model changes, the broader system can remain valuable. The winning companies in 2026 are likely to be the ones that solve operational problems with AI, not the ones that merely pick the best benchmark score.
Why Startups Need to Be Careful
The hosts repeatedly warn about the risk of building a company that is too dependent on a third-party model vendor.
Risks they point out
- model vendors can add your feature directly into their own product
- APIs can change or disappear
- a model capability that once differentiated you can become standard
- closed-model dependence creates strategic fragility
Practical implication
Founders should be cautious about building a business that is simply a thin wrapper around a vendor’s model. Durable AI companies need deeper workflow value, proprietary data, or system-level advantages.
Infrastructure and Observability: The Next Big Opportunity
A strong analogy is drawn between today’s agentic systems and the earlier microservices era.
The microservices comparison
- A few services become dozens, then hundreds, then thousands
- Complexity explodes
- Observability and root-cause analysis become essential
- Tools like Datadog or Splunk become sticky, mission-critical products
Applied to agentic AI
As organizations deploy:
- dozens of agents
- hundreds of workflows
- multiple models per agent
- MCP servers and APIs
- goal tracking and inter-agent communication
they will need new infrastructure to manage that complexity.
Likely high-value areas
- governance
- monitoring
- policy enforcement
- agent lifecycle management
- tool and server management
- communication between agents
- traceability and auditability
Practical Advice from the Episode
The hosts end with a strong recommendation: focus less on model hype and more on solving real business problems with AI systems.
Suggested mindset
- Use the model that fits the use case
- Treat model choice as one component, not the whole strategy
- Build around workflow value and operational needs
- Stay focused on novelty in the problem you are solving
- Use the new tools to move faster, but don’t mistake speed for differentiation
Their shared view
The tools are accelerating product development, but they have not eliminated the need for:
- novel ideas
- good product judgment
- domain understanding
- execution discipline
Closing Notes and Announcements
The episode closes with:
- a reminder to follow Practical AI on LinkedIn, X, and Blue Sky
- a mention of the upcoming Midwest AI Summit in Indianapolis on October 15
- a plug for early-bird registration
- thanks to Prediction Guard and Breakmaster Cylinder
Bottom Line
The episode’s central argument is that the AI conversation is shifting away from “open vs closed models” as the main battleground. That debate still matters for deployment, cost, and policy, but the bigger opportunity is in agentic infrastructure, AI-powered workflows, and physical-world applications. In short: the model is increasingly just one component of a much larger system.