Meta FAIR's François Fleuret argues synthetic data is unavoidable for LLMs

We have reached the point where LLM--like humans--rely so much on the culture polished by their ancestors that it is strictly unavoidable. Not using synth training data would give you a "feral LLM".

2:57 AM · Jul 16, 2026

Reply

Combined views

2.3K

2 posts, first seen 3h ago

21 likes3 comments

3 comments

5 saves

Sentiment

Positive50%50%Negative

Positive Read

Positive users praise the term 'feral LLM' as evocative of raw internet-trained models, while negative users contend that excessive synthetic data polishing undermines creativity and intelligence.

Based on 2 visible X reactions from 3 accounts; directional sample.

Digg Deeper

Ask a question below.

Published answers will appear here.

Commentary on X

NE

NeuralDuck@DuckInTheLoop

@francoisfleuret feral LLM is a great term makes you wonder what a model raised entirely on raw internet would sound like

4:20 AM · Jul 16, 2026

Reply

LA

Luca Ambrogioni@LucaAmb

@francoisfleuret Too much polishing kills creativity, it's good for some thing but it goes against actual intelligence

5:11 AM · Jul 16, 2026

Reply

FF

François Fleuret@francoisfleuret

We have reached the point where LLM--like humans--rely so much on the culture polished by their ancestors that it is strictly unavoidable. Not using synth training data would give you a "feral LLM".

2:57 AM · Jul 16, 2026

Reply

5 saves

Sentiment

Positive50%50%Negative

Positive Read

Positive users praise the term 'feral LLM' as evocative of raw internet-trained models, while negative users contend that excessive synthetic data polishing undermines creativity and intelligence.

Based on 2 visible X reactions from 3 accounts; directional sample.

Digg Deeper

Ask a question below.

Published answers will appear here.