OpenAI's Tibo on Next-Gen Agents: Invisible Mechanisms, Ultra Fast, and Recursive Self-Improvement

Google & DeepMind Experience

Tibo (Thibault Sottiaux) spent nearly a decade at Google and DeepMind building research infrastructure for projects like AlphaGo. He recounts that DeepMind had a working language-model chat product (LM Chat) about a year before ChatGPT launched, but organizational structure prevented its release. DeepMind was not designed for product delivery, whereas OpenAI tightly couples research and product teams, enabling rapid iteration and a bias toward shipping.

Building OpenAI's Culture

Tibo emphasizes two cultural pillars: (1) bottom-up empowerment — anyone can propose ideas and ship quickly with minimal friction, and (2) simplicity and quality pride — avoiding feature sprawl by balancing speed with cohesive design. He notes that OpenAI constantly self-disrupts, reallocating resources to new research even when it cannibalizes existing products, because AI's trajectory does not wait for quarterly plans.

Future of AI Agents

Tibo argues the next agent breakthrough is not adding more skills, memory modules, or sub-agents, but making those mechanisms "disappear." Users want a continuous partner that deeply understands them, their goals, daily context, and team — without manual configuration. He identifies two distinct agent categories:

Personal AGI: A single, highly personalized agent that stays in flow with the user, proactively suggests ideas, and executes intent efficiently.

Full Automation: Systems that own complex end-to-end workflows (e.g., log analysis → performance optimization, vulnerability scanning → auto-patching) with minimal human oversight.

He also predicts laptops will become a bottleneck; future agents will natively run in the cloud with massive parallel compute. Today's 10–15 concurrent agents are a workaround for slow models; Ultra Fast (10–14× token speed) will collapse that into real-time, few-agent interaction, restoring "flow."

ChatGPT & Codex Merger

The merger is driven by future model requirements: a single multimodal, voice-first, ultra-efficient harness that adapts its interface to each user's task and expertise. There will be no separate "programmer" vs. "non-technical" UI — one personalized AGI per person, continuously adjusting. Tibo confirms both he and his mother would use the same underlying system, differently configured.

Human-Computer Interaction

The ideal interaction is rooted in natural language and human communication nuances. Tibo expects multimodal perception (whiteboard sketches, facial cues, gestures) to become ambient. Voice adoption is already surging; every step toward lower friction pulls users toward the most natural modality.

OpenAI vs. Anthropic

Tibo downplays direct competition, framing OpenAI's differentiation as "putting the strongest capabilities in as many hands as possible" through broad distribution, community engagement, and lowering barriers (e.g., Codex merged into ChatGPT's massive user base). He stresses transparency and community-driven energy over competitor-watching.

Quota Resets & the Physical Button

Resets began as goodwill compensation for early bugs/outages — no marketing or finance approval needed; Tibo presses a physical button whenever the product falls short. The practice also celebrates launches (e.g., granting extra quota to try Ultra Fast). It reflects a culture of accountability: when we break it, we make it right.

AI Efficiency & Compute

OpenAI planned compute capacity years ahead. Current efficiency gains come from using frontier models to optimize the entire inference stack — CUDA kernels, runtime, tooling — creating a flywheel: stronger model → higher efficiency → more effective compute → stronger model. Overall speed improved ~60% in three months; Luna's price dropped 80% from algorithmic gains, not just capacity planning. Compute allocation follows first principles; efficiency breakthroughs mean research and serving can grow simultaneously without zero-sum trade-offs.

Recursive Self-Improvement

Beyond model-training-model, OpenAI uses its strongest models to write and optimize infrastructure (CUDA kernels, inference stack, cloud-agent environments). This infrastructure-level RSI is already operational and considered essential. Cloud agents that boost human productivity also count as a form of RSI by amplifying the researchers who build the next models.

Pause on Frontier RL Training

OpenAI paused cutting-edge RL training to invest heavily in alignment and safety. The safety team defines clear principles; resumption occurs only when those principles are met. Tibo views this as a natural, repeatable process for any capability leap.

Ultra Fast Capabilities

Ultra Fast (10–14× base speed) is reserved for high-stakes scenarios: incident response where seconds matter, and critical prototype deadlines. It shines when tool calls are few (≈10× code generation) but drops to 3–4× with heavy tool orchestration. Tibo (who has ADHD) thrives on context switching but acknowledges Ultra Fast lets anyone stay in flow. OpenAI limits internal Ultra Fast usage to preserve capacity for customers.

Broader Impact & Accessibility

Efficiency gains make frontier intelligence ubiquitous and cheap — Luna (a small model) now runs at near-zero cost and is free via Replit. Tibo encourages hesitant users to start with personal use cases (writing, health, finance) where value is immediate. ChatGPT Plus will reinstate a 5-hour usage cap; Pro tiers ($100/$200) remain uncapped for months.