AI Engineer Melbourne 2026 Keynote Livestream | Day 1

Key Concepts

• AI Engineering: The shift from merely using models to building "harnesses," services, and agentic workflows.
• Agentic Workflows: Systems where AI models perform multi-turn tasks, self-correct, and interact with tools (e.g., coding agents, email triage).
• Tokenomics: The economic strategy of managing inference costs, balancing model capability against token consumption.
• Open Weights vs. Proprietary Models: The ongoing debate regarding the flexibility and cost-effectiveness of open-weight models versus the frontier intelligence of proprietary ones.
• Jevons Paradox in AI: As AI becomes more efficient and cheaper, the total consumption of AI (tokens/compute) increases rather than decreases.
• Memory Architecture: The challenge of providing agents with long-term, episodic, and semantic memory to prevent "forgetting" during long-running tasks.
• Optionality: The strategic necessity of not locking into a single AI vendor to maintain leverage and cost control.

---

1. The State of AI and Market Dynamics

The conference speakers emphasized that AI progress is accelerating, not slowing down. Key observations include:

• Intelligence Benchmarking: Artificial Analysis (George Cameron) tracks an "Intelligence Index" across 10 benchmarks. Claude Opus 4.8 and GPT-5.5 are currently leading, but the "Pareto curve" of cost-to-intelligence shows that users can often achieve 10x–100x cost savings by selecting models appropriate for specific tasks rather than defaulting to the most expensive frontier model.
• The "Fortune 5 Million": Sarah Saxs (Notion) argued that while Fortune 50 companies have the leverage to negotiate custom deals, the rest of the market must build for "optionality." Relying on a single vendor creates vendor lock-in, which is dangerous as model capabilities and pricing shift rapidly.
• Commoditization of Software: Jeff Huntley noted that software development has been commoditized. With AI, the barrier to entry for writing code has collapsed, meaning "everyone is now a software developer." The value has shifted from execution (which is now cheap) to ideation and architectural design.

2. Frameworks for AI Engineering

• The Agentic Loop: An agent is defined as a simple "while-true" loop: it takes a prompt, adds it to an array, sends it for inference, checks if a tool needs to be executed, and repeats.
• Memory Systems: Igor Costa (AutoHand) highlighted that current agents are largely stateless. He proposed a hierarchical memory architecture involving:
  • Semantic Memory: Rules of engagement and framework definitions.
  • Episodic Memory: Storing past experiences to prevent drift in long-running tasks.
  • Reflective Memory: Allowing agents to evaluate their own performance and adapt.
• The "Guitar" Analogy: Jeff Huntley compared AI tools to musical instruments. Just as a musician must practice to master a guitar, engineers must engage in "deliberate intentional practice" to master AI agents rather than just consuming them.

3. Actionable Strategies for AI Implementation

• Build for Multi-Model: Architect systems to switch models easily. If a model is deprecated or a cheaper, equally capable model is released, the system should be able to pivot without a total rewrite.
• Evaluate on Value, Not Tokens: Don't just look at the cost per API call. Evaluate the entire task, including latency, accuracy, and the number of retries/errors.
• Use Determinism Where Possible: For repetitive tasks (e.g., converting CSV to PDF, or specific conversational flows), use traditional programming (CPU-based, regex, state machines) rather than expensive LLM reasoning tokens. This can reduce costs by up to 80%.
• The "Curiosity Test": When hiring or evaluating engineers, look for those who understand how AI works "under the hood" (e.g., can explain an agent's sequence diagram) rather than those who only use AI as a black-box tool.

4. Notable Quotes

• Jeff Huntley: "Software development now costs less than minimum wage... Ideas are more important than execution because execution is now commoditized."
• Sarah Saxs: "Optionality is leverage. You should be ready to walk at all times for a durable business."
• Igor Costa: "I believe intelligence should be distributed, not concentrated in four players... We believe that you should own your AI."

5. Synthesis and Conclusion

The overarching theme of the conference is that the "AI gold rush" is transitioning into a phase of operational maturity. The primary takeaway for engineers and leaders is to stop treating AI as a magical, monolithic black box. Instead, they should:

1. Adopt a multi-model strategy to maintain financial and technical independence.
2. Prioritize "outcome-maxing" over "token-maxing" by offloading deterministic tasks to traditional code.
3. Invest in internal evaluation frameworks to objectively measure which models provide the best value for specific, real-world use cases.
4. Focus on architectural design—specifically memory and agentic orchestration—to build durable, scalable AI systems that can survive the rapid churn of the current model landscape.