OpenAI's Tibo Predicts ChatGPT as an Open Platform, Ultrafast Models, and Personal Agents Within a Year

Key Takeaways

1. OpenAI product lead Tibo predicts ChatGPT will become an open platform with external login, plugins, and multi-agent collaboration within 12 months.
2. Tibo also forecasts speed as the new frontier with Ultrafast, personal agents like Dots to reclaim attention, and medical/scientific breakthroughs within a year.
3. A Coatue AI lead describes hedge fund investing as a decision function where each step should use the best executor—deterministic code, classic ML, LLMs, or humans—not automation for its own sake.
4. Robert F. Smith says every software company is now Agentic, Rule of 70, or has no right to exist, citing an AI agent that cut a call center's cost from $4.8M to $2.4M while raising resolution from 60% to 95%.
5. Tony Dinh replaced a $124/month i18n SaaS with an in-app translation tool costing about $0.75 for 2,705 keys across 16 locales.
6. A German developer's MonAI went from $300/month to $35K/month after a Colombian creator partnership, with one WWDC video bringing 1.7M views and $5K MRR.
7. Patrick McKenzie found multi-agent development workflows clearly the future, consuming significant inference compute and feeling less like traditional software engineering.

1. AI Models and Infrastructure

1. OpenAI product lead Tibo gave four predictions for the next 12 months to Alex Lieberman. ChatGPT will become an open platform where external products log in with ChatGPT, carrying usage and tokens, with 16 initial partners; internally, products can be built as plugins and extensions. A year from now, he envisions multiple agents—possibly not all OpenAI's—collaborating with humans inside ChatGPT, with OpenAI providing intelligence and partners providing products. 1
2. Tibo's second prediction is that speed becomes the new frontier: OpenAI released Ultrafast, which was previously limited to few users due to capacity but can now run at scale and is "almost affordable." Faster models will enable real-time applications rather than back-and-forth conversational interaction. His third prediction is personal agents like OpenAI's internal Dots, a virtual colleague with a defined role reporting to a person, applicable to procurement and other repetitive but intelligent back-office work. Calling a Dot currently takes about 5 seconds to pick up and is being improved; future versions will understand physical space, such as knowing you just got home. The challenge is balancing proactive help with interruption, and he asks what happens when 300,000 employees each have a chief of staff. 1
3. Tibo's fourth prediction is medical and scientific breakthroughs within a year, with surprising progress in medicine, science, and math making "a large number of problems unsolved for decades suddenly within reach," potentially changing the fear narrative around AI. 1
4. Patrick McKenzie, after first exposure to multi-agent development workflows, said it is clearly the future, will consume large amounts of inference compute, and already feels less like traditional software engineering, though he still felt intellectually tired after a day. 1

2. Enterprise AI and the Future of Software

1. Robert F. Smith of Vista Equity said every software company now falls into one of three categories: Agentic, Rule of 70, or no right to exist. He cited a client paying $4.8 million annually for a 120-person call center resolving 60% of issues; Vista's AI agent took over, raised resolution to 95%, and changed pricing from $250,000 per year to half the savings, or $2.4 million. Smith said AI lowers software build and support costs, shifting his company's target from "growth plus margin equals 40" to above 70, and one company's support team went from 300 people to about 3 with customer satisfaction rising. 1
2. Smith argued that software merely repackaging public information has "no right to exist" because AI is the best search engine; survivors will be software that owns customer workflows and data, with specialization beating generalization. He advocated bringing models to data rather than data to models, meaning AI should run inside enterprise systems to avoid handing data to others' models where competitors could learn from it. He said his company lost less than 1% of customers to AI startups. On valuation multiples, he noted on-premise software trades at 4-6x, cloud at 8-12x, and he expects AI to re-rate again, adding that every new technology brings "software is dead" headlines, as with the internet, cloud, and now AI. 1
3. The AI lead at Coatue, a $93 billion hedge fund, described hedge fund investing as a function that takes external and internal context and outputs investment targets, positions, and timing—essentially compressing all world information into a single decision. The principle is that everything visible on an investor's screen should be visible to agents via API: external data includes news, prices, and documents; internal data includes current positions, liquidity, risk requirements, past cases, and history with an industry or management team. Coatue's advantage is decades of rigorously recording internal context, but he admitted it is impossible to stuff all data into a context window and get three magic numbers. 1
4. The decision function was simplified into three steps—collect data, analyze, decide—each with inputs and outputs where the next step's input comes from the previous step's output, and each step contains dozens of sub-steps. For each sub-step, the best executor must be chosen: deterministic code is fast, reliable, and cheap for rules like "if P/E is below 15"; classic ML models more complex relationships; LLMs suit complex tasks like summarizing earnings calls; humans remain irreplaceable, such as meeting company management where "you probably don't want to send an LLM or AI agent, nor an if statement." He stressed there is no glory in automating for automation's sake; the goal is maximizing absolute quality at each step, not handing as much as possible to LLMs. 1
5. The real function is not linear but has many branches, and each step may have multiple optimal executors in practice, with many tunable knobs like model size, thinking effort, hooks, and skills. The key question is "what exactly can I carve out so we can understand and evaluate it"; without a north star to climb, you are just flailing. He estimated at least 10 knobs per node and at least 20 effective values per knob, making the search space "basically infinite." The real difficulty is presenting this in a technically understandable and organizationally absorbable way, which is the biggest challenge of applying AI to investing. He described the essence as "creating a state machine that can receive world information and understand it well enough to make coherent decisions," a pattern applicable to policy, nonprofits, drug discovery, and corporate strategy. 1
6. Alex Lieberman himself believes agents are steadily raising society's collective ambition; models and harnesses are imperfect but good enough, and the unrealized potential of agent use is limited by imagination of what is possible, not by technology itself. 1

3. Developer Tools and Practical AI Adoption

1. Tony Dinh said he replaced his previous i18n translation SaaS with an in-app translation tool, saving $124 per month. His app has 2,705 translation keys across 16 locales; using GPT-Luna 6 costs about $0.75, and he said translations are more accurate and contextually appropriate. 1
2. Ben Tossell shared a case of using FactoryAI's coding agent to scale design workflows: batch-rearranging design assets for global billboard campaigns in New York, Tokyo, and London, handling grouping, cropping, and unexpectedly placed panels. Previously, formatting all panels to spec took hours to days of non-creative work. 1
3. In an Already Here podcast conversation, Ben Tossell described himself as "non-technical" and afraid of the terminal, yet he has shipped code using FactoryAI. He keeps agent context in folders, does not use Claude Code, and said AI makes him want to spend more time on the computer. The episode also covered assigning different tasks to Pi, Factory, and Codex, preferring to guide agents while working, using a shared vocabulary to reduce communication gaps with agents, making personal context portable across agents and models, reducing agents' memory of him, and his view that calling AI adoption a "skills problem" lets AI labs off the hook. 1
4. Arvid Kahl praised Linus Torvalds' pragmatism, saying Torvalds views curiosity and willingness to learn new tools as useful qualities. The quoted tweet said Torvalds called rules banning AI use "extremely stupid" and "idiotic," and that the Linux kernel will not make stupid rules saying "you can't use a certain tool to do something." 1
5. Arvid Kahl said some still do not see agentic engineering as a quantum leap; the quoted tweet said Anthropic "outgrew" git's capabilities because it could not handle 25 million CI tasks per day. 1

4. Startup Growth and Monetization

1. Independent developer Flo, a German developer, built MonAI three years ago—a voice-input, Apple Pay-automated, minimalist bookkeeping app. For the first 18 months he ran it alone with monthly revenue of only about $300. A Colombian content creator found his social account in the app and messaged him; they agreed on a "fixed monthly fee plus profit share" partnership, after a pure revenue-share model failed to scale. Their first video was a simple walkthrough, and MRR grew 10x within a week and approached $8,000 within a month. They post only 3 videos per month, emphasizing quality and storytelling; one WWDC invitation video got 1.7 million views and brought $5,000 MRR alone. The app now earns $35,000 per month, a 10,000% increase from the start, and Flo quit his job earlier this year. His tips for finding creator partners include matching lifestyle and audience, warming the relationship with genuine interaction first, citing specific videos in outreach, showing fit rather than transaction, stating willingness to pay early with profit share preferred over fixed fees, and keeping the whole tech stack under $1,000 per month. 1
2. Arvid Kahl said libgen and AA remain effective top-of-funnel distribution channels for his books: many people download and never pay, but they would not have bought anyway, and some read and then purchase, with thank-you emails as evidence. He calls this "Freemium." 1
3. Arvid Kahl said he is on his first flight in the agentic AI era and finds it incredible to have free in-flight WiFi and check PRs being handled by a Claude code instance via the Claude app on his phone. 1
4. Arvid Kahl said he could watch a quoted clip for hours; the quoted tweet said its dystopian level is beyond what even Black Mirror predicted. 1
5. Arvid Kahl used a prompt telling Claude to implement a checklist item by item, satirizing the idea of vibe-coding a beehiiv alternative over a weekend. The quoted tweet listed features including 99.3% email deliverability, over 3 million emails per minute, three-tier suppression lists, DKIM/SPF/DMARC, REST and sending APIs, MCP server, automation, payments, community, podcasts, and ad network, and noted that self-building still requires assembling many third-party tools like SendGrid/SES, Stripe, Clerk, Cloudflare, and Litmus. 1
6. Arvid Kahl argued the key is not whether a usable alternative can actually be built—he called it "too hard"—but that there is currently strong willingness to try, which is enough for capable entrepreneurs. A quoted reply said he is using a self-built solution to cut many SaaS tools; beehiiv was originally on the list, but even without 90% of the listed features, he could not reliably send 65,000 emails at a lower price. 1