[1/2] Neural networks have taken over social networks: how a Kazakhstani startup blew up all AI trends and became a unicorn n (AI column)
I saw it. curious Erzata DulataCTO and co-founder Higgsfield AI (First unicorn of Kazakhstan, estimate $1.3B)What he gave to Elizaveta Osetinskaya, a foreign agent. Among the interesting facts, after which the interview is even more interesting to watch that Erzat does not have a higher education, he was once called in OpenAI and he did not go. (When he published algorithms and papers on GitHub in the second half of the 2010s)He also founded one of the fastest growing AI startups in history. (Maybe even faster than Lovable, which I'm talking about. told).
Below are some details about the startup itself. 🚀 Timeline Higgsfield AI
- 2023: foundation of the company, 1.5 research
- 31 March 2025: product release
- January. 2026: $1.3B valuation, $200M ARR Growth metrics:
- $0 → $100M ARR per <9 months ahead of Cursor (reached $100M ARR for 12 months) Lovable in speed)
- $100M → $200M ARR for ~2 month
- Growth. 50-150percent
- 15M+ users, 45M video/day
And then the key insights that I found interesting in the approach of the guys.
1Speed is the only real “moat” in AI and here’s why Models are updated every few months OpenAI kills hundreds of startups with each release (Example: When they announced canvas/documents, it took out a bunch of startups.) Classical defensibility (network effects, data moats) Too slow for AI, you can see it. Interesting episode from Y Combinator about
Higgsfield Strategy: Collect Low Fruit Example: Google has released Veo3 - powerful video generation, but the model does not give accurate control of the camera. Higgsfield took Veo3 over the API and added:
- Precise control of camera movement
- Click-to-video interface (templates like in PowerPoint) Camera techniques for professional operators
Quote about why the providers of frontier models do not do it
Hegemon is blind. Google/Meta is so big that they don’t think about details for professionals. We collect these low-hanging fruits.
As a result, a similar algorithm of actions for the creators of GenAI applications is being assembled. Don't build foundation models - orchestrate them
- Use the API. (OpenAI, Anthropic, Google) Add “packaging” for a specific niche MVP release in weeks, not months → collect feedback → iterate
2Synthesis of creatives + ML-engineers = magic From the interview, it can be seen that the breakthrough occurred when professional operators began to explain to ML engineers what kind of product is needed. How it works:
- Big Creative Department Invents AI Hooks for TikTok/Instagram (camera hits the ground, morphing clothes, turning into a flock of ravens)
- ML team trains neurons to these hooks
- Roll out → viral in social networks → millions of users Result: All AI trends in social networks in recent years 9 months created by Higgsfield
- Madonna, Will Smith, Snoop Dogg, Timbaland, Zlatan Ibrahimovic use TikTok and Meta tried to repeat, but “none of the trends created”
As a result, a similar algorithm of actions for the creators of GenAI applications is being assembled. Not just a tech problem, but a product/culture problem.
- Hire domain experts (cameramen, directors, designers)Not just ML engineers. Creativity + Engineering = Differentiation
In continuation I'm going to end the interview with a story about how the guys are driving such rapid growth in their app audience, and how they're working with models to generate video.
#AI #Engineering #Software #Management #Leadership #Startup #LLM #ML #Architecture