A recent conversation between OpenAI founder and CEO Sam Altman and Conviction founder Sarah Guo at a data conference provided deep insights into the current state and future prospects of AI technology. Their practical advice on how companies should navigate the AI landscape in 2025 was particularly compelling.
Just Start: There's No Perfect Timing
Sam Altman's first piece of advice to business leaders was simple yet powerful: "Just do it." Many companies are still hesitating, waiting for the next model or watching to see how the technology will evolve. But in times of rapid technological change, the companies that win are those with the fastest iteration speed, minimal cost of mistakes, and maximum learning velocity.
This is a crucial perspective. While many companies waste time waiting for the 'perfect' solution, those that place early bets and iterate quickly are achieving much better results. Sarah Guo echoed this sentiment, emphasizing the importance of 'curiosity.' The key is to abandon preconceptions about how things used to work and start experimenting.
Technology Maturity: The Difference Between Last Year and This Year
Interestingly, Altman noted a significant shift between last year and this year. While he would have given startups the same advice last year, he would have told large enterprises, "You can experiment a bit, but for most cases, it's not fully ready for production use." Now, the situation has completely changed.
OpenAI's enterprise business has grown dramatically, with large companies actually using AI for substantial amounts of work. When companies ask "what's so different now," they're told "it works much more reliably." There's been a real inflection point in model usability over the past year.
This provides important insights for AI technology adoption, showing how crucial stability and reliability are in the transition from laboratory to real business environments.
The Dawn of the Agent Era
One of the most exciting parts of Altman's discussion was about 'Codex,' a coding agent. This system can take multiple tasks and work on them in the background, handling complex, long-term projects. Users simply sit back and say, "I like this, I don't like that, try again."
Currently, it's at the level of an intern who can work for a few hours, but eventually, it will become like a skilled software engineer who can work for days. Companies are already emerging that build agents for various fields like customer support and sales.
Many people now describe their role as "assigning work to multiple agents, checking quality, and providing feedback." This is remarkably similar to working with a team of fairly junior employees.
A New Perspective on AGI
When asked about AGI (Artificial General Intelligence), Altman offered an intriguing perspective. If we went back to 2020 and showed people today's ChatGPT, most would say, "This is definitely AGI."
Humans are excellent at adjusting expectations. What's more important than defining what AGI is, is that the annual rate of progress shown over the past five years will continue for at least another five years. Whether we declare AGI in 2024 or 2026, or superintelligence in 2028 or 2032, is far less important than one long, beautiful, surprisingly smooth exponential curve.
The Evolving Role of Memory and Search
Sarah Guo provided important insights about the role of memory and search. She explained that technologies like search have always played a key role in making generative AI grounded where it needs to be grounded. When asking factual questions and wanting trustworthy answers, you need reference points to real-time information.
Particularly interesting is the perspective of viewing search as an "attention-setting tool for models." Just as humans focus on specific things among infinite contexts, search is a tool that helps models focus on appropriate contexts.
The Amazing Capabilities of Future Models
Altman predicted that models to be released in the next 1-2 years will be "quite breathtaking." Just as we saw a huge leap from GPT-3 to GPT-4, companies will be able to do things that were completely impossible with previous generation models.
A chip design company could ask, "Design better chips that were previously impossible," and a biotech company could request, "Research ways to cure this disease." Models will understand all the context you want, connect to all tools and systems, reason really hard before bringing back answers, and be robust enough to work autonomously.
Utilizing Computing Power
When asked what he would do with 1000 times more computing power, Altman gave an interesting response. Meta-wise, he said, "I would have it work very hard on AI research to figure out how to build much better models, and then ask those better models what to do with all the compute."
More practically, he mentioned that there are real returns on test-time computing for companies currently using ChatGPT or the latest models. Making models reason more and try harder on difficult problems yields much better answers.
A Bigger Vision for Humanity
Sarah Guo presented a broader vision beyond the tech world. She mentioned the Arnold Project, describing it as similar to the DNA sequencing project from 20 years ago, but focused on figuring out RNA expression. It's been discovered that RNA controls almost everything about how proteins work in our bodies, and breakthroughs in this field could solve countless diseases and advance humanity significantly.
She suggested that the equivalent of DNA projects performed with language models would be a wonderful way to utilize massive computing power.
---
The most impressive aspect of this conversation was that AI technology has moved beyond the experimental phase to become a reliable tool in actual business environments. It's no longer "something that might be possible someday" but "technology we can leverage right now."
What companies need to do isn't wait for perfect timing, but start immediately and learn and iterate quickly. The future of AI is already here, and only prepared companies will reap its benefits. Companies that experiment with curiosity, challenge themselves without fearing mistakes, will be the winners in the AI era.