Why AI Suddenly Became Usable Overnight: OpenAI Post-Training Co-Lead Explains the Underlying Logic of the 2026 Perception Shift
> Source: Interview with Yann Dubois, OpenAI Post-training Frontiers Co-lead, The MAD Podcast with Matt Turck, 2026-05-21
The Gap Between Impressive and Usable
Over the past year, benchmark scores kept climbing while user experience stayed flat. Then around late 2024, something changed—not at a launch event, but as a felt shift: AI moved from "competition toy" to "productivity tool." Code actually ran, documents stopped hallucinating, complex tasks could be handed off.
Dubois offers a counterintuitive explanation: AI capability grew steadily, but the leap in user experience happened the moment reliability crossed a critical threshold.
Three Engines
1. The Reliability Threshold
AI capability curves are smooth; user trust is not. Imagine an agent that has some probability of error every two minutes—error compounds over long tasks, so users dare not delegate complex work.
OpenAI's internal assessment: in December 2024, reliability crossed the critical threshold. After that, models could take on large blocks of work rather than small, constantly-corrected fragments. The threshold came from steadily driving down the per-interval error rate—via both model-level reliability and application-layer engineering.
2. The Flywheel Spins
Once models became genuinely usable, AI began accelerating AI's own development: OpenAI uses AI-assisted coding heavily; models help researchers train other models and build internal tools. Better models → faster research → better models.
3. RL Leaves the Greenhouse
Early RL on o1/o3 mostly optimized verifiable rewards—math, competitive programming—producing models strong in closed settings but weak in open, real-world scenarios. From late 2024 into early 2025, the key shift was migrating these tools to messy real-world use cases: asking not "was this answer correct?" but "does this result actually help the user?"
GPT-5.5: Smarter Is Table Stakes; Reliable Is the Point
Codename "Spud," with company-wide involvement. Dubois's two proudest points:
- Doubled efficiency: most tasks run about twice as fast.
- Company-level coherence: vertical teams (coding, computer use, knowledge work) and horizontal teams (instruction following, function calling, thinking-time control) improve independently, then integrate.
- GPT-5.5 Thinking: the curve shifts left within normal latency ranges.
- Pro: extends the right end of the curve—longer thinking for a higher accuracy ceiling. Mathematicians love it (launch an eight-hour background task before bed); Dubois rarely uses it because he's impatient.
- SFT: imitation of ideal human answers—behavior cloning, capped by the humans providing data.
- RL: reward optimization, where models can exceed their human teachers via trial and error.
- Verifiable rewards (binary, math/code)
- Non-fully-verifiable rewards (pairwise preference judgments)
- Yann Dubois interview, The MAD Podcast with Matt Turck, 2026-05-21, https://www.youtube.com/watch?v=DhD1zZ8w8Mw
- Chinese transcript, Sina Finance, 2026-05-25, https://finance.sina.com.cn/wm/2026-05-25/doc-inhzafpt5253529.shtml
- PingWest deep dive, 2026-05-26, https://www.pingwest.com/a/313965
- Introducing GPT-5.5, OpenAI, 2026-04-23, https://openai.com/zh-Hans-CN/index/introducing-gpt-5-5/
Key concept: test-time scaling curves (thinking tokens/latency on the x-axis, accuracy on the y-axis). The core goal is pushing the curve leftward—same or better accuracy with less thinking time.
GPT-5.5's core selling point is handling messy data—ambiguous, underspecified, noisy real-world conditions. Evaluation shifted accordingly: GPQA, SWE-bench Pro, SWE-bench over Codeforces-style contests.
Paradigm Shift: Post-Training Reshapes AI
Modern training has three stages:
1. Pre-training: learning everything from the internet—mostly noise and ads; "loading the library into the brain." 2. Mid-training: targeted strengthening on high-quality data; now widely adopted in the open-source community. 3. Post-training: turning a model that "knows a lot" into one that's "genuinely useful to people"—today's frontier.
Post-training includes:
Why Benchmarks Fail
Dubois's uncomfortable truth for eval researchers: every evaluation you build is also a recipe for creating training data. Models generalize capability from eval data and soon ace the eval itself—evals decay quickly.
Model-as-a-Judge
As models surpass human judging ability, better models become superior teachers for other models—a capability flywheel: better models → better judges → better training → better models. "We're building a self-reinforcing loop."
Open Problems and Opportunity Maps
Continual Learning: The Achilles' Heel
Three years after ChatGPT, continual learning remains the industry's biggest unsolved problem. Dubois's chart: current AI has extremely high Day-Zero utility but a flattening curve—it never truly learns company-specific knowledge. Humans start low but learn steeply; over time (area under the curve), humans remain more useful. Making models grow more useful the longer they work in an environment is the goal.
The Last Mile: A Goldmine for Startups
Big labs build foundations, but the vertical "last mile"—permissions, system integration, industry workflows—holds enormous upside that large players won't or can't fully address. Startups should grind the last mile and squeeze maximal utility from large models in vertical scenarios.
Conclusion: Reliability Is the Real Capability
The 2026 shift feels sudden but is simple: AI didn't suddenly get smarter—it suddenly got more reliable. Gradual capability gains plus gradual reliability gains stacked into a step-change around December 2024, driven by a deeper paradigm shift: from leaderboard chasing to real utility, from closed problems to the open world, from behavior cloning to reward optimization.
GPT-5.5 is not the endpoint. Pre-training still expands, post-training is exploding, and continual learning awaits. But for everyone using AI, the key threshold has been crossed: AI moved from "impressive" to "usable"—and that's when it truly starts changing the world.
---
References