Kihagyás

Practical AI — The Myth of Model Wars

Overview

Daniel Whitenack (Prediction Guard CEO) and Chris (Principal AI & Autonomy Research Engineer) debate the open vs closed source AI model landscape in May 2026. Meta abandoned Llama for the closed MuseSpark, Yann LeCun left, China now leads open source models. Central thesis: models are now a commodity — real value lies in agentic infrastructure, MCP server management, and workflow orchestration.

Physical AI — The 2026 Trend

  • AI embedded in everyday devices: retail kiosks, factory floors, smart glasses, drones, robots
  • Walmart drone deliveries; robots in restaurants/stores — futuristic 2 years ago, normal now
  • Exploding across all industry segments: defense, retail, marketing, healthcare
  • Parallel microelectronics revolution: GPU/CPU boundaries blurring, smaller low-power chips
  • Anyone can experiment with a few hundred dollars → "Wild West of physical AI"

Open vs Closed: Current State

How Models Work

  • Training produces weights (parameters) + architecture code
  • Closed: SaaS product or API — weights never leave vendor servers
  • Open (open weights): weights + code on Hugging Face, anyone can run on their own infra
  • Managed open: third party hosts open model (AWS Bedrock, together.ai)

Meta: Llama → MuseSpark Pivot

  • Meta was Western champion of open source AI for years (Llama family)
  • Yann LeCun (one of "three godfathers of AI") spent decade at Meta on condition models stayed open
  • LeCun left → Meta abandoned Llama development, pivoted to MuseSpark (closed model family)
  • Llama increasingly fell behind frontier models
  • Consequence: no more major Western corporate open source AI champion

China Takes the Lead

  • Leading open source models now all Chinese: DeepSeek, Qwen 3.5, Kimi K2
  • US national security dilemma: lack of Western open models vs dependence on Chinese models
  • US government won't buy solutions built on Chinese open source models

Benchmarks vs Reality

  • Gap measured by benchmarks (MMLU, SWE, MATH, scoring arenas)
  • Closed models (Opus, GPT 5.5) still a few percentage points ahead on benchmarks
  • Daniel: "Who cares about benchmarks? It has nothing to do with the real world."

Models as Commodity

Central thesis: models are now a commodity — like soybeans or corn.

  • In a restaurant, you don't care about the corn brand, you care how the dish is prepared
  • Similarly: it's not the model, it's the agentic harness, workflow, and orchestration that matters
  • "The model is now a complete commodity... it's such a small piece of the puzzle"

Mythos / Anthropic Example

  • Cybersecurity world panicking about Mythos
  • Daniel: "Whether Mythos is released or not, there's more than enough AI and agentic capability to already disrupt cybersecurity"
  • Transformation driven by agentic systems, not individual models

Risk of Building on Closed APIs

  • Claude Design (Anthropic) released 2 weeks ago → anyone building similar on their API got "eaten"
  • Same with OpenAI image generation capabilities
  • "I would be very cautious before building a business exclusively on somebody else's business"

Agentic Complexity as Next Big Value

Microservices Analogy

  • Microservices boom → first 5, then 50, 500, thousands → Datadog, Splunk (monitoring, RCA)
  • These products: you'd never build yourself + extremely sticky
  • Same coming for agents: first 1 agent, then 10, 100, 1000 → complexity explosion

What Needs Management

  • Multiple models (LLM, embeddings, rerank)
  • MCP servers
  • API calls
  • Workflow code
  • Agent-to-agent communication
  • Goal tracking
  • Governance & policy enforcement
  • Monitoring

"The overall project is much more important than that individual dependency, and that dependency could be swapped out for any number of things."

Where Value Lies in 2026

  1. Don't focus on model hype — models are swappable
  2. Build agentic infrastructure that solves real business problems
  3. Novel ideas — tooling accelerated (months → days/weeks), but fundamental value creation hasn't changed
  4. Open models clearly win: air-gapped environments, privacy-sensitive use cases, high-volume processing (economics)
  5. Closed models clearly win: managed service SLAs, high reliability, low-trust environments

Key Quotes

"The model is now a complete commodity."

"The question of what model you're using is kind of irrelevant."

"We should be talking not so much about the open versus closed model gap, but the development of the agentic workforce and where the value lies."

"This is the moment. We're definitely in the Wild West of physical AI."

Vissza a tetejére