← All speakers

Bio, Work & Ideas

Lachlan Ainley

Conference affiliation: Microsoft · 2024

Lachlan Ainley is a product-marketing professional specializing in Azure AI infrastructure and its applications in enterprise computing. A Microsoft senior product marketing manager at AI Engineer World’s Fair 2024, he previously worked across technology marketing and sales in Australia, Japan, and the United States.

Infrastructure and applications

  • Infrastructure as an interconnected stack: Ainley emphasizes coordinating data centers, host processors, virtual machines, accelerators, networking, and system topology to support specialized AI workloads. His approach to workload-specific accelerator selection spans NVIDIA GPUs, AMD accelerators, and Microsoft’s Maia silicon. In a conversation with Snorkel AI researcher Humza Iqbal, he described how lessons from large-scale model training can inform infrastructure for enterprise customers; Iqbal contributed Snorkel’s distributed-training results and GPU comparisons.

His professional interests also include sovereign AI infrastructure, particularly the control organizations require over sensitive data, regulated workloads, and where their computing systems operate.

Read the topics behind these talks

1 conference talk

Key ideas

Scroll to read ↓

Enterprise model quality depends on the right domain data—and on infrastructure that can turn that data into repeated training, evaluation and synthesis cycles.

  • What gets a bank’s model ready to deploy?
    0:16 ↗
  • Scale expert knowledge, then evaluate the actual task
    3:53 ↗
  • Evaluate long context and align models to the domain
    5:17 ↗
  • Optimize the whole infrastructure path
    6:53 ↗
  • PyTorch, Horovod and a shared filesystem
    8:56 ↗
  • Find out why the GPUs are waiting
    11:05 ↗
  • Research needs capacity that can change
    13:00 ↗
  • Compare hardware at approximately equal cost
    15:10 ↗
  • Separate batch-size effects from hardware effects
    17:02 ↗
  • Match evolving accelerators to the workload
    18:18 ↗
  • Turn preference signals and expertise into better data
    19:46 ↗

References