Tech Disruptors: Invisible Technologies on RLHF and LLM Training

Matt Fitzpatrick, CEO of Invisible Technologies, discussed reinforcement learning, RLHF, agentic AI, large language model evolution, coding agents and contact centers on Bloomberg Intelligence’s Tech Disruptors podcast. The conversation highlights how frontier model providers are using reinforcement learning for training and how Invisible Technologies is positioning its enterprise business. The article is informational and does not include financial results, guidance, or a material corporate event.

Analysis

The key second-order takeaway is that reinforcement learning is becoming a capital-allocation problem, not just a model-quality problem. As frontier labs push more training compute into post-training and self-improvement loops, the marginal advantage shifts toward firms that can generate high-signal interaction data, cheaply route human oversight, and instrument workflows end-to-end. That favors infra vendors with strong enterprise distribution and workflow ownership, while commoditizing pure model access faster than the market currently discounts.

The biggest competitive risk is not model performance, but workflow capture. If coding agents and contact-center agents become good enough to sit in the daily loop, the vendor that controls the task surface can disintermediate standalone SaaS, BPO, and even some services revenue over 12-24 months. That creates a winner-take-most dynamic in vertical AI, where data flywheels compound: more usage produces better reward signals, which improves automation, which then increases usage.

The near-term catalyst path is uneven: enterprise adoption should show up first in pilot expansions and seat-based productivity gains, while budget replacement cycles will lag by quarters. The tail risk is that “agentic AI” stalls at high error rates in edge cases, causing enterprises to cap deployment to low-value workflows and leaving model providers with high training costs but weak monetization. A second contrarian risk is that RLHF itself becomes a feature, not a moat, as open-source tooling and synthetic data narrow the gap for smaller labs.

Consensus seems too focused on model scale and not enough on operational throughput. The better framing is that the next equity winners will be the companies that convert AI into measurable labor substitution with short payback periods, not the ones with the best demos. If that thesis is right, the market is underpricing a multi-year re-rating in software and IT services names exposed to repetitive knowledge work, while overestimating the durability of standalone model-provider differentiation.

Key Decisions for Investors

• Long MSFT / short a basket of high-multiple standalone SaaS names most exposed to workflow substitution; 6-12 month horizon. Thesis: distribution + copilot attachment should compound faster than pure-play AI hype, while point-solution SaaS faces seat compression.
• Long CRM on weakness versus a basket of IT services/BPO names with high contact-center exposure; 3-9 month horizon. Risk/reward improves if AI-assisted service modules drive upsell faster than cannibalization of legacy seat revenue.
• Short WDAY or ZM against long MSFT/GOOGL as a pair trade over 6-12 months. These names are more vulnerable to agentic workflow replacement than the platform owners that capture the underlying interaction data.
• Buy medium-dated call spreads on a leading semiconductor compute supplier only on pullbacks, not breakout strength; 6-18 month horizon. The trade benefits if post-training and inference demand keep compounding, but upside should be capped because pricing power can leak to hyperscalers.
• Avoid aggressive shorts in frontier-model names until enterprise spend visibly slows; use defined-risk puts if desired. The better entry is after evidence of stalled pilot-to-production conversion, not on headline AI enthusiasm.