Tibo Sottiaux, who leads ChatGPT and Codex at OpenAI, issued a 28-day public commitment to his team on X: from now on, every day they must either ship an improvement that clearly benefits most Codex and Work users, or perform a complete quota reset. The core principle behind this commitment is that product quality is no longer measured by how many features are released, but by whether users genuinely benefit in their day-to-day work.

In a recent interview with host Lenny, Sottiaux explained the decision-making logic behind quota resets. He said he has full autonomy to hit the quota reset button without going through layers of approval, which allows him to stay closer to the community. When asked how many more quota resets users can expect, his answer carried a self-deprecating tone: “Depends on how many times we break things.” He added that resets happen when something goes wrong, and also when there’s something worth celebrating.

The product lead also admitted a little-known detail: on his third day at OpenAI, he took down the production environment. This candid admission echoes the product philosophy he repeatedly emphasized throughout the interview—turning stronger capabilities into products that are easier to use.

During the interview, Sottiaux said he’s had enough of increasingly complex choices. Which model should you pick? How high should reasoning intensity be set? Multi-agent mode or Ultra mode? What’s the difference between these options anyway? He expressed his exhaustion with a slightly exaggerated remark: “You practically need a PhD in model selection.” He wants to shed these burdens as quickly as possible and let the app itself gradually fade into the background.

This pursuit of simplicity also stems from changes in his own workflow. Today, analyzing business trends, researching the next feature, and reviewing the performance of previously shipped products—the bulk of the code required for all of this is written by Codex. Hand-writing code has become a weekend relaxation activity for him, occasionally solving a few LeetCode problems.

More Agents Isn’t Necessarily Better

Sottiaux revealed in the interview that he hasn’t been chasing the goal of running more agents simultaneously. He used to run more agents at once, but after the team made breakthroughs in ultra-high-speed capabilities, he now feels that a single faster agent is more helpful for his work. As agent speed improved, he rediscovered a state of flow: ideas no longer get interrupted by long waits, building can be driven through voice, and attention can be focused more on creation itself. He said: “This flow is different from before, but I no longer miss the old way.”

Regarding concepts like loops and execution graphs that have been hotly debated in the industry, Sottiaux believes that building loops, tweaking them repeatedly, and figuring out how they should work may have excited some people in the past, but the future won’t operate that way. He’s more bullish on the positioning of a personal agent platform like Dots: a very smart agent that works 24/7, understands the user’s goals, knows their preferences, and learns from feedback.

In his view, product design going forward needs to account for three shifts: most operations on the internet will be performed by agents; models will rapidly become cheaper and faster; and interaction across different modalities will become smoother. This is also why he questions some products on the market—if you seriously imagine these capabilities being ten times stronger a year from now, developers would build things differently.

Dots Will Become the Primary Interface for Human-AI Interaction

On the relationship between Codex, ChatGPT Work, consumer-facing ChatGPT, and Dots, Sottiaux gave a clear direction. He believes that whether or not it ultimately ends up being Dots, the fundamental goal is to free people from the constraints of technology itself. Users will have a persistent, always-active intelligent capability that knows what it needs to do and can interact through any client, any screen. You can call it, it will appear in a meeting room to take notes when you walk in, and you can continue communicating with it via email afterward.

He said: “You don’t need to be glued to your laptop, and you don’t need to be staring at your phone constantly. When you need it, it’s there; when you don’t, it stays out of your way.” He described how people now carry laptops everywhere like carrying a brick—people are tethered to technology, rather than technology working for people.

On the integration of Codex and ChatGPT, Sottiaux said the team has heard user feedback loud and clear: people like Work’s capabilities, but sometimes want to use chat mode because it’s faster and more comfortable. So the team is integrating these experiences to reduce complexity. Ultimately, all of Dots’ capabilities will be brought directly into chat mode, raising the baseline capabilities available to 1.2 billion users. He specifically mentioned one thing about Dots that excites him: it has no model selector and requires no configuration—users just talk to it, and the only thing they need to configure is which channels they want to communicate with it through.

Work Patterns Will Still Change Dramatically

When asked whether work patterns will undergo even more dramatic changes over the next two years, Sottiaux gave an affirmative answer. He believes that even today, technologies like voice and multimodal input and output finally seem to be coming together, but they still feel somewhat clunky to use. This feeling will persist until one day it suddenly stops being clunky. At that point, users will be able to communicate with AI as naturally as face-to-face conversation, AI will accurately remember everything, users can brainstorm and ideate through shared whiteboards, and collaboration can happen between people and people, and between agents and agents.

He said this will transcend conversation and transcend many of today’s client-side interfaces, because these things aren’t truly working yet—so there’s still a major new wave of transformation ahead.

At the end of the interview, Sottiaux mentioned science fiction works that have deeply influenced him, including Neuromancer and the original Star Trek. He specifically pointed out one setting in Star Trek that left a strong impression: the technology is highly practical—you can talk directly to the computer and it does things for you; you can talk to the ship and it takes action. He said: “These things are becoming reality. We are entering that era.”

Why ChatGPT Keeps Giving Generic Advice

Echoing Sottiaux’s pursuit of simplicity, ChatGPT in practice is often criticized for being too “averaged out.” OpenAI disclosed in February 2026 that ChatGPT’s weekly users have surpassed 900 million, with over 9 million paying enterprise users. Such massive scale means the model is almost inevitably tuned to satisfy the average user’s needs, rather than the specific circumstances of a particular company or individual.

The reason ChatGPT’s advice often feels like generic content borrowed from elsewhere isn’t because the model lacks the ability to think about individual business situations—it’s because during training, through Reinforcement Learning from Human Feedback (RLHF), it was trained across millions of conversations to prioritize answers that please the average evaluator. When a user asks in a new conversation “how to grow a startup,” the model doesn’t know whether the questioner runs a three-person SaaS company or a one-person consulting firm, so it can only choose the answer least likely to be wrong.

The way large language models work further compounds this problem. The model predicts the next most likely token based on everything it has previously read. When millions of business-related questions are stacked together, the statistically safest answer becomes a blend of all answers. Advice like “know your customers better” and “publish more valuable content” appearing so frequently is a direct result of this probabilistic averaging.

To break through this generic response pattern, the key is to provide the model with specific business information. Instead of asking “how should I price my product,” try asking: “I run a 14-month-old B2B analytics tool with $8,000 in monthly recurring revenue, 40 customers on a $99/month plan, 3 large prospects requesting annual contracts, and we’re only 2 people. Based on this, propose two pricing options and explain the pros and cons of each.” By providing actual data like customer retention rates, burn rate, and current customer base, and breaking the question down into step-by-step analysis, you can pull the answer away from the statistical midpoint toward genuine analysis.

OpenAI provides three tools for this problem: custom instructions, ChatGPT memory, and projects. Custom instructions allow users to set a standing profile that applies to all new conversations, including company growth stage, industry, target customers, and current priorities. Memory automatically accumulates details during conversations but requires periodic review and correction. Projects are used to group related conversations, files, and instructions together, letting ChatGPT work with the same premises across ongoing work.

However, these tools cannot replace professional judgment. A 2026 survey of 1,000 entrepreneurs by US online legal services company LegalZoom found that 73% use ChatGPT and 42% use AI tools daily, but when it comes to actual legal and financial risk, entrepreneurs still rely on human experts. Another survey of 255 business owners by small business financing company Clarify Capital found that only 8% of solo entrepreneurs consider themselves in the “active power-user” tier who use AI comprehensively from strategy to execution, while 55% use AI only for specific tasks.

The key distinction is this: custom instructions and projects can make ChatGPT better at drafting, summarizing, and organizing thoughts, but they cannot replace a lawyer reviewing a cap table or an accountant double-checking burn rate. If a user writes incorrect premises into custom instructions, the model not only fails to detect them—it will confidently build arguments on top of that flawed foundation. The quality difference in ChatGPT’s answers ultimately depends on what premises the user provides, not on how smart the model itself is.