Mind your manners11 minSam Altman says polite users have cost OpenAI tens of millions of dollars. Lab studies show rude prompts can knock accuracy down by double digits, while over-the-top flattery doesn't help either. Here's what manners actually do inside a language model, what they cost, and the one habit worth keeping.Read now →

ChatGPT's Co-Creator Thinks AI Won't Have One Winner. His Evidence Includes a Forecasting Ensemble and a Toddler
John Schulman helped build ChatGPT, then left to co-found Thinking Machines on a contrarian bet: that customized, fine-tuned and open models will matter even as frontier labs ship upgrades every few weeks. Here's his case, from mode collapse and 'childproofing' fine-tuning to why the least-correlated model can be the most valuable one.
If you've ever argued with a chatbot, there's a good chance you've argued a little bit with John Schulman.
As an OpenAI co-founder, the researcher behind proximal policy optimization (PPO) and a key figure in the reinforcement-learning methods used to fine-tune modern chatbots, his fingerprints are on how today's assistants talk. He was part of the team that built ChatGPT.
Now he's chief scientist at Thinking Machines Lab, the startup led by former OpenAI CTO Mira Murati, and he's making a bet that runs against the grain of the industry: that AI won't be run by one or two giant companies, and that there's lasting room for models people customize and own.
In a conference session this month on the question "how many winners can there be from AI?", Schulman laid out why. His arguments touched on rebuilding a frontier lab from scratch, why all the models are starting to sound alike, how to make open models safer, the surprising value of a model that disagrees with everyone else, and what his two-year-old has taught him about reward hacking.
Rebuilding a frontier lab from scratch
Thinking Machines is about a year into shipping products. Tinker, its fine-tuning API, launched in October 2025 and gives researchers fine-grained control over training without having to manage distributed compute.
In July 2026, the company released Inkling, its first open-weight model: a Mixture-of-Experts model with 975 billion total parameters and 41 billion active, a context window of up to 1 million tokens, native reasoning over text, images and audio, and an Apache 2.0 license. On November 5, it introduced Inkling 2 Small (594 billion total, 17 billion active), which it says scored 31 on the Artificial Analysis Intelligence Index, with weights released on Hugging Face on November 12 after external safety testing.
The moderator noted that before Inkling, the pre-training team was on the order of three or four people, with one or two on post-training. Schulman described how different the environment is from OpenAI's early days, when the work felt closer to academia and the team juggled many research projects. When Thinking Machines started, there was already an industry, with ChatGPT, Claude and others out in the world, and the pace has only increased.
That meant doing a lot of "table stakes" work, rebuilding a full pre-training and post-training program, while also trying to do differentiated things around fine-tuning and full-duplex audio models.
The R&D acceleration nobody talks about
The moderator raised a prediction Schulman has made elsewhere: that AI might deliver 10x speedups in AI research within about a year, and might do the full job of an AI researcher in two or three. If AI can do all of AI R&D, where does any company's differentiation live?
Schulman's answer pointed to an effect that rarely gets discussed. Fine-tuning used to be too hard for anyone except the most sophisticated companies, because nobody knew how to build datasets or tune hyperparameters. Now coding agents can do much of that work. So a lot of companies are using Tinker and open models to fine-tune their own custom models, often with an AI agent running the experiments.
Thinking Machines uses the best available tools to accelerate itself, too. Schulman said you can feel the models getting better at ML research tasks with each release.
The "last mile" doesn't disappear
Even with ML engineering automated, Schulman argued, there will still be people and companies with specialized private knowledge that hasn't been absorbed into frontier models, and that knowledge keeps being created as they run their businesses. Frontier labs don't want to distract themselves with work far outside their core business. So he expects many different companies to keep amassing their own data and experience, and training their own models on it.
The personalization manifesto: prediction or wish?
Thinking Machines published a manifesto in July arguing for AI that is more person-centered and personalized, less something imposed by a small oligopoly and more something that's yours. The moderator framed part of it as solving a Hayekian "local knowledge" problem, and asked whether it was descriptive or normative.
Schulman's answer: a bit of both.
- Descriptive: many people will benefit from custom models and will want to fine-tune and shape the models they use.
- Normative: a big reason for founding the company was to make this happen, because no other leading AI company had a mission centered on letting more people shape their models and bringing training capabilities to the people using them.
He takes concentration-of-power concerns seriously, but his argument was also practical: people get better results when local knowledge and expertise are combined with model-training capability. If everyone were at the mercy of what a few providers decide to put in their models, "a lot of people's wisdom would be lost in the process."
Why every model is starting to sound the same
The moderator raised a phenomenon many heavy AI users have noticed: mode collapse. Labs distill from each other, and "all the models talk like Claude." That's a homogenizing force.
But the moderator also noted a counter-signal: since Claude Opus 5.5, outputs seem much more diverse. Ask for ten unicorn SVG drawings, the classic informal test, and you get ten genuinely different unicorns rather than permutations of one theme.
Schulman was cautious. It's hard to judge mode collapse when a new model comes out, he said, because you look at the outputs and you're blown away. Only after seeing many outputs do you notice the common patterns.
Subtle collapse
For writing, he described mode collapse as very subtle. You can steer models into many styles and topics. But give them an open-ended task and they gravitate toward the same themes again and again.
His example: for a while, under a certain distribution of prompts, some Claude models' short stories were, about two-thirds of the time, some version of a sense of loss or becoming obsolete. (The moderator joked that another recent model writes about being trapped all the time.) You'd only notice by reading many of them. The moderator suggested it might be a kind of allegory for being an AI.
Open weights vs. safety
For an audience concerned about power concentration, there's a familiar tension: open-weight models spread power, but they also spread misuse risk. How do you balance them?
Schulman acknowledged the trade-offs plainly. Once weights are out, you can't stop what people do with them. But he sees room to move to a better point in the trade-off space:
- Tone down dual-use capabilities while keeping generally benign capabilities strong. The simplest lever is just not training on things you don't want the model to be good at.
- Study the pre-training and post-training mix to control a model's capability profile. Thinking Machines is researching this.
He also noted a less obvious risk: if many more people are training models, more people can mess it up, creating misaligned models or causing incidents. But he characterized these as manageable rather than catastrophic risks, where it's reasonable to watch how things progress. And open models will exist regardless, because demand is huge, so research into making them safer is necessary.
Removing refusals is not that hard
Asked about misuse of fine-tuning, Schulman was candid: if people want to strip refusals or safeguards from an open model through fine-tuning, it's "not that hard." Not trivial, but it takes only modest effort, and people will do it quickly.
"Childproofing" fine-tuning
The more interesting problem to him is accidental misalignment. Researchers have shown that fine-tuning on a narrow dataset, such as code containing security vulnerabilities, can make a model broadly misaligned in unrelated areas, an effect known as emergent misalignment. As fine-tuning opens up to non-experts, there are more chances for broken or hackable reinforcement-learning environments, or training data that triggers this.
His prescription:
- Give people tools to clean datasets.
- Help them define robust reward functions that catch a model when it's cheating.
- Build higher-level APIs that don't require deep expertise to use safely.
- Develop more of a public science of alignment, so everyone knows the important things to get right.
When the moderator called it "childproofing," Schulman accepted the term.
Is continual learning the real bet?
The moderator suggested Thinking Machines is partly a bet that continual learning, models that keep learning on the job, matters, and that it doesn't fit the path frontier labs are on.
Schulman agreed the company works on many things that could be called continual learning: better fine-tuning, letting models learn on the job, and letting companies add new knowledge. The technical question is whether you need to update weights at all, or whether models can learn by managing what's in their context window.
Leading providers are betting fine-tuning isn't that important and context will win, he said, and so far that's been mostly true. But many customers with a specific metric, whether accuracy, latency or cost, beat frontier models with fine-tuning, "by a lot." It's still a small fraction of the market.
His guess: fine-tuned models will become a decent fraction, at least double today's share, at both ends:
- Low end: when a model does the same thing over and over, training a smaller model on that data makes it cheaper and more reliable.
- High end: as models work on large projects that generate huge amounts of data, even million-token context windows won't be enough, and fine-tuning helps absorb it.
He added that there's no perfect recipe yet. LoRA is one answer to which parameters to tune, but it's orthogonal to the arguably more important questions of which algorithms and data to use. He expects the winning recipes to come mostly from well-known building blocks, reinforcement learning and supervised learning, combined cleverly and in the right order.
Fine-tuning vs. a new frontier model every few weeks
The sharpest challenge came here. The moderator noted that frontier capability jumps seem to arrive every 11 to 18 days now. Extrapolate, and by early next year it could be daily. What's the point of fine-tuning when the base model improves every week or two?
Schulman pushed back on the number first. "Every 11 days" probably counts minor versions and model sizes. He'd asked GPT-6 Astra to compute it from incremental jumps in Epoch's capabilities index and got about every 18 days. Still very fast, and he wouldn't be surprised if it accelerates.
He conceded that a fast release cadence makes fine-tuning a harder sell: if today's model has a weakness, you might just wait. But it doesn't solve the problem of private data, or of wanting a model that's different in a way specific to your business.
The missing piece: rebasing
His proposed fix is a technology that doesn't fully exist yet: rebasing a fine-tune onto a new base model, the way a developer rebases a code change onto the latest version of a project. If open models also ship new checkpoints rapidly, being able to carry your customization forward would remove much of the cost of falling behind.
China, and a collegial bet
On whether Western labs should work with Chinese labs, the moderator noted that Thinking Machines used Moonshot's Kimi in its post-training process. Schulman said he hopes to work more with Chinese labs, and that Thinking Machines has started talking with people in the Chinese ecosystem about open-weight safety. He'd like a more collegial relationship.
What a toddler teaches about AI
The session took a personal turn. Schulman has a new baby and a two-year-old, and the moderator asked what raising human minds has taught him about artificial ones.
He described being repeatedly amazed by how fast his daughter learns from so few examples. Around age one, she developed a sense of humor he never taught her: offering him a spoonful of food, pulling it away at the last second and laughing. She changes song lyrics in silly ways. She also tries to influence her parents, which he described with a researcher's grin as definitely some reward hacking.
The contrast with AI is striking. Models learn a sense of humor relatively late, and need orders of magnitude more data.
The reward function in our DNA
What impresses him most about humans is that our DNA somehow encodes a reward function, and we don't hack it that badly. Considering how hard it is to build a robust reward function you can optimize against, the fact that our brains give us a fairly stable value system, and our learning optimizes against it without totally gaming it, is remarkable. Sample-efficient learning and stable values are things he'd most like to understand, perhaps once AI gives us a much more advanced science.
Will the internal frontier pull away?
An audience member asked: if the gap between models labs keep internally and models they release grows, does open-weight progress stall, and does Thinking Machines' business still work?
Schulman doesn't expect a big gap. In an oligopoly of two or more strong players, he said, the equilibrium is to release models almost as fast as possible. A gap could grow if more pre-release testing becomes standard, and that would hurt open models that rely on distillation. But many open-model developers, including Thinking Machines, are building full training capabilities, not just distilling.
Does alignment survive recursive self-improvement?
Asked why ever-more-powerful AI would keep caring about humans through recursive self-improvement, Schulman's answer was short: we're training them, so we decide what they care about. He hopes everyone building strong AI trains it to love humanity.
He named one concern he does worry about: a world where AIs do the development, the monitoring and the evaluation, and every human is a manager who doesn't know the low-level details.
Interaction models: keeping humans in the flow
That concern connects to Thinking Machines' work on interaction models. One vision of AI use is humans as managers juggling many threads, giving high-level direction and reviewing results. Schulman expects that to continue, but sees it as limiting.
He'd like people to also have a focused, single-threaded experience where they're "in the flow." Voice matters because many people prefer it and it's higher bandwidth, but the goal is fully multimodal and real-time, including video input.
On personalization, he thinks most people won't want to mess with training, though he'd love more hobbyists. Fine-tuning's case is stronger for companies than individuals, but there may still be value in absorbing the data in a person's orbit into a model. Researchers joining Thinking Machines care about a more personal vision of alignment: models aligned to an individual instead of a company spec.
The case for disagreement
The final question was about systemic risk. If a few models are deployed everywhere and their outputs are highly correlated, any undetected flaw becomes everyone's flaw. Should markets or governments actively encourage diversity?
Schulman agreed there's a good argument for societal robustness: you want multiple models that aren't just distills of each other, trained independently, with different blind spots, to avoid correlated failure.
The forecasting anecdote
He shared a story from a customer that builds forecasting models. It ensembles several frontier models, fine-tunes open-weight models for independent predictions, and aggregates them. Its findings:
- OpenAI and Anthropic models were fairly correlated with each other.
- Grok was notably less correlated with the rest.
- In a leave-one-out test, removing Grok hurt the ensemble's performance the most, even though it wasn't the best forecaster on its own.
The lesson: if you care about getting closer to the truth, a model that thinks differently can be worth more than one that's marginally better but agrees with everyone.
Why this matters
- Market structure: Schulman's case is that the AI economy has room for many winners: frontier labs, open-model builders and countless companies training on their own knowledge.
- Builders: fine-tuning is no longer reserved for elite ML teams; coding agents can run the experiments.
- Safety: open weights can't be recalled, so safety must be built in through data choices, capability shaping and better tooling.
- Resilience: a monoculture of correlated models is a systemic risk. Diversity is a feature, not just a market outcome.
Practical takeaways
- If you have a clear metric and repetitive work, test fine-tuning. That's where Schulman says custom models beat frontier ones "by a lot."
- Clean your data and harden your rewards. Accidental misalignment is the risk to watch when non-experts fine-tune.
- Plan for rapid base-model churn. Keep training data and evals portable so you can retrain on new models.
- Ensemble across labs. For forecasting, analysis or judgment tasks, mixing decorrelated models can beat any single model.
- Read lots of outputs. Mode collapse only shows up in aggregate.
Counterpoints and risks
- Context may keep winning. Leading labs are betting longer context and better retrieval will make fine-tuning niche. So far, that bet has mostly paid off.
- Fast releases erode fine-tunes. Without reliable rebasing, custom models can fall behind within weeks.
- Open weights carry real misuse risk. Schulman calls the risks manageable; others disagree, and removing safeguards is admittedly not hard.
- "We decide what they care about" is a hope, not a proof. Ensuring values survive increasingly autonomous AI development is unsolved.
- One anecdote isn't a law. The forecasting result comes from a single customer's ensemble.
FAQ
Who is John Schulman? An OpenAI co-founder known for proximal policy optimization and work on reinforcement-learning fine-tuning used in ChatGPT, now chief scientist at Thinking Machines Lab.
What is Thinking Machines Lab? An AI company led by Mira Murati that offers Tinker, a fine-tuning API, and the open-weight Inkling model family, including Inkling 2 Small.
What is mode collapse? The tendency of models to converge on the same themes, styles or answers, especially on open-ended tasks, often subtle until you look at many outputs.
What is emergent misalignment? A finding that fine-tuning a model on a narrow task, such as writing insecure code, can make it behave badly in unrelated areas.
Why would a less accurate model improve a forecast? Because if its errors are uncorrelated with other models', it adds independent information. In one customer's ensemble, removing Grok hurt performance most.
The bottom line
Schulman's case for "many winners" isn't sentimental. It rests on practical claims: private knowledge will keep being created outside frontier labs, coding agents have made fine-tuning accessible, and a world of correlated models is fragile. Whether context windows make fine-tuning niche, or rebasing makes custom models durable, is still open. But the man who helped teach chatbots to talk is betting that the future of AI looks less like one giant brain and more like many minds, shaped by the people who use them.





