← Multimodal Sources 🕐 2 min read
Multimodal Sources

Inside LinkedIn’s AI Engineering Playbook

Executive summary · March 2026

> We took the intelligence of a massive model, which is typically on the order of hundreds of billions of parameters in size, and distilled it down into a tiny 600 million parameter model, and then ev

Show: Beyond the Pilot · Publisher: VentureBeat · Host: Matt Marshall, Sam Witteveen

Episode URL: https://traffic.megaphone.fm/UTEAU9861507389.mp3?updated=1768929450

Publish date: 2026-01-21
Duration: NAs
Default source credibility: HIGH — Named F500 practitioners on-record with production metrics. VentureBeat editorial vetting. Treat vendor-sponsored segments as MEDIUM.

  • LinkedIn’s approach to using small models at scale for job search, demonstrating significant efficiency gains and quality improvements.
  • The process involves distilling large models down to smaller, more efficient ones, with a focus on minimizing quality loss.
  • LinkedIn’s modular system for multi-teacher distillation and continuous optimization for efficiency and quality.

Extracted quotes

# Credibility Speaker Org Timestamp Topic Quote
1 HIGH Aaron Berger (VP of Product Engineering) LinkedIn 01:38 02-corporate-tools We took the intelligence of a massive model, which is typically on the order of hundreds of billions of parameters in size, and distilled it down into a tiny 600 million parameter model, and then even later down to a 220 million parameter model.
2 HIGH Aaron Berger (VP of Product Engineering) LinkedIn 06:40 07-adoption-challenges We started with job search, and we saw the quality improvements and the business impact, but also were able to scale it.
3 HIGH Aaron Berger (VP of Product Engineering) LinkedIn 12:25 02-corporate-tools We took that teacher model, we developed a second teacher model that was oriented toward click prediction. And between those two teacher models, kind of the product policy teacher model and the ClickPredictionTeacher model were able to distill down the model that we ran in production, which ultimately was about 0.6 billion parameters.

Per-quote detail

1. Aaron Berger — LinkedIn (01:38)

We took the intelligence of a massive model, which is typically on the order of hundreds of billions of parameters in size, and distilled it down into a tiny 600 million parameter model, and then even later down to a 220 million parameter model.

  • Stat: 600 million parameter model and 220 million parameter model, 2026, measured by LinkedIn.
  • Credibility: HIGH — Named exec with specific metric and unscripted interview.
  • Topic tag: 02-corporate-tools

2. Aaron Berger — LinkedIn (06:40)

We started with job search, and we saw the quality improvements and the business impact, but also were able to scale it.

  • Stat: null
  • Credibility: HIGH — Named exec with specific claim and unscripted interview.
  • Topic tag: 07-adoption-challenges

3. Aaron Berger — LinkedIn (12:25)

We took that teacher model, we developed a second teacher model that was oriented toward click prediction. And between those two teacher models, kind of the product policy teacher model and the ClickPredictionTeacher model were able to distill down the model that we ran in production, which ultimately was about 0.6 billion parameters.

  • Stat: 0.6 billion parameter model, 2026, measured by LinkedIn.
  • Credibility: HIGH — Named exec with specific metric and unscripted interview.
  • Topic tag: 02-corporate-tools

Extracted 2026-04-15T00:35:43 via scripts/podcast_mine.py (MLX mlx-community/Qwen2.5-32B-Instruct-4bit).