Video summary

Next Gen Andromeda Is Here (Don’t get left behind)

Main summary

Key takeaways

Business

Summary

Spencer Pawliw explains Meta’s new generation of Andromeda, the ad-retrieval system that filters eligible ads and helps determine which ads are shown to each user.

His central business takeaway is that the system may handle far more creative volume and match ads to people’s situations more specifically—but simply producing more ads is not a strategy. Advertisers still need strong customer research and relevant creative.

The speaker says the 2024 version could process tens of millions of ads per second, while the 2026 version can process more than one billion per second. He attributes the increased capacity to advances in AI and computing, and says Meta partnered with Nvidia on the underlying hardware. These are claims made in the video; no independent performance data is provided.

How the System Works

The video describes a multi-stage ad-selection process:

  1. Eligibility checks: Ads are filtered according to factors such as the user’s age, country, placement, format, and campaign restrictions.
  2. Andromeda retrieval: The system quickly creates a shortlist from eligible ads.
  3. Further ranking: A more powerful model evaluates the shortlist and selects ads for the person’s feed.

The speaker compares the process to a sports draft: eligibility is the initial qualification check, Andromeda acts like scouts assembling a shortlist, and a final ranking stage makes the selection. He argues that the updated system can evaluate a much larger share of available ads, reducing the chance that strong ads are missed or weak ones make it through.

He also argues that Meta is increasingly evaluating ads in relation to multiple facets of a person and their current context—such as what they are doing, where they are, the placement, and their recent behavior. In this view, the same person may respond differently to the same ad at different moments.

Business Implications and Recommended Playbook

  • Let delivery signals guide broad creative testing. The speaker advises against forcing spend onto every ad in a high-volume test. If an ad receives little or no spend, he argues that the system may be making a more informed decision than older versions could.
  • Use CBO for testing rather than manually equalizing spend across every creative. He says forced allocation becomes inefficient at scale. For an individual ad that appears promising, he notes that the team is still experimenting with whether to push more budget through ABL; the acronym is not explained in the subtitles.
  • Replace generic messaging with customer-specific situations. Broad claims such as “veterinarian approved,” “free shipping,” or “buy now” may be hard to distinguish from similar AI-generated ads.
  • Research customers beyond demographics. Build creative around experiences, behaviors, emotions, problems, and desires—not just age or gender. The goal is to address a recognizable customer situation while allowing Meta to find people across demographic groups who share it.
  • Create variations for different contexts and triggers. Develop different creative versions for distinct situations in which a customer need becomes salient. The speaker emphasizes understanding when a feeling or problem arises, not just who experiences it.
  • Use AI to expand creative production, not replace strategy. AI can help generate more concepts and formats, but indiscriminate ad volume is unlikely to create an advantage. Human judgment and direct customer research remain important for assessing whether an ad feels relevant and emotionally credible.
  • Close the feedback loop. Compare delivery and performance signals with customer research, learn which messages and situations resonate, and use those learnings to improve subsequent creative.

Illustrative Example

For a dog-joint supplement, a generic ad might list product credentials and shipping terms. A more specific execution could speak to an owner whose dog hesitates at a curb, has tried different foods, or is limping.

The speaker’s point is not necessarily to narrow the business’s total addressable market. Rather, each ad should be relevant to a particular need or moment, allowing Meta to find other people who share that experience.

Metrics, Examples, and Timelines

  • Ad-processing capacity: The speaker contrasts “tens of millions of ads per second” for the 2024 system with more than one billion per second for the 2026 version.
  • Creative-testing example: A brand testing 50 creatives per month and scaling to 500 would be testing 10 times as many creatives. If it forced equal spend on every ad, the speaker says the required testing budget could also rise tenfold. This is an illustration, not a recommended target or reported result.
  • Efficiency claim: The speaker says advertising efficiency improved despite concerns about Andromeda, but provides no specific efficiency metric, timeframe, or supporting results.
  • No performance targets given: The video provides no quantified targets for ROAS, CPA, CAC, revenue, conversion rate, or test duration.

Presenters and Sources

  • Presenter: Spencer Pawliw.
  • Sources and references discussed: Meta’s Andromeda report; Meta’s stated partnership with Nvidia, as described by the speaker; and Sarah Levanger’s training on identifying when customer feelings arise.

Rate this summary

Your feedback will help improve summaries.

Improve this summary

Reprocess with a stronger model when the summary feels incomplete or inaccurate.

Pro

Translate summary in another language

Pro

Ask questions to this video

Chat for follow-up questions, clarifications, and source-backed answers.

Coming soon

Share this summary

Original video