Unique Research · Interview · November 14, 2024
Source: HubSpot 2024 GenAI Summit roundtable. Unique Research editorial compilation. Moderator: Charles Elfin (Fellows Fund partner; Ficc.AI founder).
Guests
- Young Zhao — OpusClip co-founder & CEO. One-click long→short viral video. Sep: ~$30M raised (A + seed); A led by Millennium New Horizons + AI Grant, Samsung Next, GTM Fund, DCM, Alumni, Fellows Fund, Alpine, Jason Lemkin. 6M+ users; eight-figure ARR; clients incl. Billboard.com, Univision, Telefónica, Jenny Hoyos, Scott Galloway.
- Jia Li — LiveX AI co-founder, Chief AI Officer & President. Human-like VIP AI agents across the customer journey (sales, trial conversion, renewals/upsells, fewer returns).
- Dmytro Dzhulgakov — Fireworks AI co-founder & CTO. Inference + fine-tune platform for open GenAI models. Traffic ×100 in six months; Jul Series B led by Sequoia → $552M valuation. (Ex-PyTorch core.)
- Hassan Raza — Tavus co-founder & CEO. GenAI video research; Mar Series A $18M led by Scale VP (Rory O’Driscoll joins board); Sequoia, YC, HubSpot. Personal digital twins, lip-sync, dubbing, text→video; web + API; authorized clones only.
Q: How do you use AI — and what’s your founding story?
Young: OpusClip turns long footage into social-ready shorts with video understanding + generation so anyone can tell stories on social.
Jia: Multimodal AI agents for daily life / enterprise customer service — GenAI will change how we interact with products, services, hardware.
Dmytro: Developer platform for open GenAI (Llama, Mistral, Stable Diffusion, Whisper…) — customize, deploy, scale, optimize for quality and low latency; tops performance leaderboards.
Hassan: Models for personal digital avatars — text-only video without filming; healthcare partners send doctor messages in patients’ languages; even Zoom stand-ins. Consent-first, high-fidelity cloning at scale.
Q: Selling vision to investors, employees, customers — which is hardest?
Hassan: Customers hardest — investors buy a dream; customers need the dream rooted in today’s product and their business. Win by partnering so their roadmap feels like yours.
Jia: Customers often non-technical vs. AI-fluent investors — listen to pain; alignment follows solving real problems.
Young: Customers easy (KOL creators often articulate the vision better than he does). Employees hardest — no precedent product, constant mission reinforcement. Investors also hard — many aren’t founders and don’t live social video.
Dmytro: Customer focus first; for developer products, craft in details matters; also reframe the customer’s business problem, not only the tech.
Q: Niche differentiation vs. huge market?
Dmytro: GenAI market dwarfs prior deep learning; focus on customer problems over competitors — best product wins.
Jia: CX / agents are noisy; demos are ~5% of the work, production ~95% — educate, earn trust.
Hassan: Shifted from sales/marketing product models to developer-facing models — broader distribution beats locking models inside one product.
Young: Started niche, became early category leader; expanding ICP carefully — want 3–10× market but pace expansion with limited team.
Q: Real-world ethics?
Hassan: Disclosure & consent for face/voice clones; refuse unauthorized celebrity virality even when marketing wants it.
Dmytro: Huge gap between training a model and shipping a product — guardrails, safety filters, privacy (log only when customer needs fine-tuning).
Young: Millions of copyright issues early (users clipping others’ content looking like OpusClip); warn/ban; make abuse harder while smoothing UX for legitimate creators.
Jia: Job-replacement fear — experience shows AI lifts agents off repetitive work toward harder cases; transformation not pure replacement.
Q: Biggest production challenge — reliability, hallucination, prompts, guardrails…?
Dmytro: All of the above. Early: iterate fast (prompting > fine-tune speed). Later: cost, latency, TCO — specialized deploys can × performance without changing the model (e.g. phone assistants need ultra-low latency on first tokens).
Jia: Latency for real-time multimodal agents (vs. batch report agents). NVIDIA HW/SW work → >6× avg token speed — still not enough for edge dreams.
Hassan: Research→prod needs heavy testing for weird real inputs; then scale/latency/cost dominate.
Q: How important is interpretability?
Young: Visual video outputs are inherently inspectable.
Hassan: Critical for B2B embeds — predictability + knowing why a training video was rejected.
Dmytro: Importance varies by product criticality.
Jia: Customers care less about parameters than journey insights — hot topics, frequent questions that drive product upgrades.
Q: Future of foundation models?
Consensus: multimodal. Hassan: modalities preserve context lost in text-only translation. Dmytro: spatial reasoning / robotics / sensors next. Young: models still don’t match human world models — big improvement area. Jia: multimodal agent future matches LiveX’s bet.
Q: Are transformers the ceiling of intelligence?
Dmytro: Transformers dominate partly because of HW/SW investment — hard to displace even if a new idea is slightly better. Hassan: agrees (non-expert). Jia: data under-discussed vs. architecture/scale — data encodes the problem you’re solving. Young: agrees models still miss true human world perception.
— END —