20VC- 70% of Neolabs Will Die | There Will Be a $100BN US Open-Source Model | Data Is a Trillion $ Market | Governments Cannot Regulate Models- It Is Too Late | the Cyber Attacks to Come Will Be Insane With Anastasios Angelopoulos @ Arena

Arena Is Real World AI Evaluation
- Arena measures AI by real-user interactions rather than static benchmarks to reveal factuality, hallucinations, steerability and real-world utility.
- The platform captures organic feedback from millions of monthly users to produce continuous, actionable model evaluations. (Time 0:04:15)
Kimi K3 Broke The Distillation Narrative
- Kimi K3, a Chinese open model, outperformed top American closed models on key web development tasks, breaking the narrative that Chinese models merely distill US models.
- That result suggests Chinese labs add training techniques beyond distillation driving real performance gains. (Time 0:05:36)
Open Source Hasn’t Killed Proprietary Inference Yet
- Chinese open models haven’t yet meaningfully cannibalized proprietary inference revenue because most inference still runs on first-party APIs and closed models.
- But enterprises will seek AI sovereignty and cost savings, making open source fine-tuning and self-hosting an inevitable part of the future. (Time 0:07:38)
Business Models Will Drive US Open Source Rise
- US open-source ecosystem lagged due to unclear business models, but new approaches (rev-share, lead-gen services, enterprise AI modernization) make large American open-source firms plausible.
- Anastasios predicts at least one multi-hundred-billion American open-source company emerging. (Time 0:10:27)
Chip Export Controls Are A Strategic Trade Off
- Export controls can slow Chinese hardware progress short-term but also incentivize China to build its own stack, so US policy must keep NVIDIA/TSMC ecosystems strong.
- Cutting exports is a trade-off between immediate advantage and encouraging foreign self-sufficiency. (Time 0:15:40)
Local Hosting Doesn’t Remove Backdoor Risk
- Hosting a model locally doesn’t eliminate risk because training-origin jailbreaks or special token sequences can still trigger data exfiltration even on on-prem deployments.
- Attackers can craft prompts that induce a hosted model to ‘vomit’ unstructured internal data, creating persistent backdoor vectors. (Time 0:19:20)
Public Margins Will Force Inference Price Pressure
- High public inference margins (Anastasios calls Anthropic’s ‘disgustingly high’ gross margins) will become visible after IPOs and likely trigger downward pricing pressure.
- Full transparency of margins changes negotiation dynamics between labs and enterprise customers. (Time 0:27:10)
Use Guardian Models To Watch Agent Behavior
- Deploy guardian models to monitor agents and flag unsafe actions because humans are too slow to catch sophisticated agent failures or jailbreaks in real time.
- Guardians should be as capable as the agents they watch, matching strengths to prevent agents from outsmarting controls. (Time 0:30:14)
Fake Candidates Passing Technical Interviews
- Arena experienced applicants who passed technical interviews but turned out to be non-existent people, likely AI-generated profiles used for espionage or attacks.
- The team now plans stricter in-person verification and onboarding because interviews were convincingly passed by synthetic personas. (Time 0:33:10)
Most Neolabs Will Fail Without Clear Revenue
- There are at least ~75 neolabs and Anastasios expects roughly two-thirds to fail or be acquired for parts unless they have an aggressive business model and clear revenue path.
- Investors are demanding P&L and hypergrowth; zero-revenue high valuations create a brutal next-round environment. (Time 0:37:02)
Data Will Be A Trillion Dollar Opportunity
- Data is a durable, non-commodity scaling complement to models; demand grows as models scale and enterprises train their own intelligence.
- Anastasios forecasts the data market at $100B to $1T by 2030 because companies will pay heavily to source and clean training data. (Time 0:41:23)
Prioritize Evaluation Before Wide AI Deployment
- Build evaluation infrastructure before large-scale deployment because evaluation is the key bottleneck to deploying AI and defining cost-per-value for business use cases.
- Arena collects agentic traces from 30M+ monthly users to extract organic performance signals labs and enterprises can act on. (Time 0:45:47)