Why AI Suddenly Became Usable: OpenAI Post-Training Lead Explains the 2026 Leap
*Source: Interview with Yann Dubois, Co-lead of OpenAI Post-training Frontiers, 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 the end of 2024, something changed — no launch event, no paper, but a felt shift: AI went from "competition toy" to "productivity tool." Code actually ran, documents stopped hallucinating, and people started trusting AI with complex tasks.
Dubois's counterintuitive explanation: AI capability grew steadily, but the jump in user experience happened the moment reliability crossed a critical threshold.
Three Engines Behind the Shift
1. The Reliability Threshold (December 2024). Imagine an agent that has some probability of failing every two minutes. Cumulative errors make users afraid to hand over long tasks. OpenAI internally judged that in December 2024, reliability crossed the critical threshold — models could now take on large blocks of work rather than only small, human-correctable fragments. This came from continuously driving down error probability through both model improvements and application-layer engineering.
2. The Flywheel. Once models became genuinely usable, AI began accelerating AI's own development. OpenAI heavily uses AI-assisted coding internally; models help researchers train other models and build internal tooling. Better models → faster R&D → better models.
3. RL Leaves the Greenhouse. Early reinforcement learning (o1, o3) mostly optimized for verifiable rewards — math and competitive programming with clear right/wrong answers. Strong in closed environments, weak in the open world. From late 2024 into early 2025, the key shift was migrating these RL tools to messy real-world use cases: no longer "was this answer correct?" but "did this result actually help the user?"
GPT-5.5: Not Just Smarter, More Dependable
GPT-5.5 (internal codename "Spud") delivered roughly 2x speed on most tasks and company-wide alignment between vertical teams (coding, computer use, knowledge work) and horizontal teams (instruction following, function calling, thinking-time control).
The core concept is the test-time scaling curve: performance vs. thinking tokens/latency. The goal is to push this curve left — same or better accuracy with less thinking.
- GPT-5.5 Thinking: pushes the curve left in the normal latency range — more efficient.
- GPT-5.5 Pro: extends the curve's right end, allowing much longer reasoning for a higher accuracy ceiling. Dubois says he rarely uses Pro ("I'm very impatient"), but mathematicians love it — hand off a task overnight and get results eight hours later.
- Pre-training: learning from the internet — mostly content that isn't truly useful (ads, noise). "Loading the library into the brain."
- Mid-training: continued training on high-quality data to prepare for target behavior; now widely adopted in open-source communities.
- Post-training: turning a model that "knows things" into a model that's "actually useful." This includes: 1. SFT — supervised fine-tuning, essentially behavior cloning; the model can never exceed its human data providers. 2. RL — reward optimization; through trial and error the model may surpass its human teachers. Splits into verifiable rewards (binary right/wrong) and non-fully-verifiable rewards (pairwise comparisons for open-ended tasks).
- Yann Dubois interview, The MAD Podcast with Matt Turck, 2026-05-21, https://www.youtube.com/watch?v=DhD1zZ8w8Mw
- Sina Finance transcript, 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/
The headline capability is handling messy data — ambiguous conditions, underspecified tasks, real-world noise. Evaluation shifted accordingly: GPQA, SWE-bench Pro, and SWE-bench track real engineering better than Codeforces.
Post-Training Reshapes AI
Modern training has three layers:
Why benchmarks fail: "Every time you build an evaluation, you're also building a way to create training data." Models generalize to eval data and quickly saturate it — the line between evaluation and training is blurred.
Model-as-a-Judge: as models outstrip human judgment, stronger models become excellent teachers for other models — a capability flywheel: better models → better judging → better training → better models.
Open Problems and the Opportunity Map
Continual learning remains the industry's biggest unsolved problem. Current AI has enormous Day-Zero utility — dropped into a company, it's more useful than most new hires — but its utility curve flattens because it doesn't truly learn company-specific knowledge. Humans start lower but have steep learning curves; long-term, the area under the human curve is larger. "The more time a model spends in an environment, the more useful it becomes — that's what we need to crack."
The last mile is where startups should play: permissions, system integration, industry-specific workflows. Frontier labs build the foundations, but vertical "last mile" work is a huge open opportunity.
Conclusion
The plain answer to 2026's sudden AI leap: AI didn't suddenly get smarter — it suddenly got dependable. Capability accumulated gradually; reliability improved quietly; at a threshold (December 2024), the two combined into a step change. GPT-5.5 isn't the endpoint — pre-training still scales, post-training is booming, and continual learning remains unsolved. But for everyone using AI, the key gate has been crossed: from "impressive" to "usable."
References