2025 Snowflake Summit Interview-Sam Altman on AGI Timetable and Next Generation AI Function

Frontier Trends

At the Snowflake Summit in 2025, OpenAI CEO Sam Altman and Snowflake CEO Sridhar Ramaswamy conducted an extensive burnside conversation under the auspices of the founder of Conviction Sarah Guo. Together, they explored the rapidly developing patterns of artificial intelligence and its impact on the present and future of enterprises. They discuss the status and future of artificial intelligence (AI), focusing in particular on the concept of enterprise-level applications and common artificial intelligence (AGI, Artificial General Intelligence).

I. PROPOSALS FOR ENTERPRISE: What should be done in the context of AI change?**

# Core point of view:

  • ** “Just do it” - Action as soon as possible is essential**

  • ** Do not wait for the next generation model**, otherwise the current dividend will be missed.

  • Companies with fast succession, low cost of error and learning skills will eventually win.

# Deep meaning:

  • ** The pace of technological evolution is much faster than the pace of adaptation of the enterprise ‘ s organizational structure** and a slow step will lose competitiveness.

  • One of the key competitivenesss of enterprises will be the “AI landing speed” rather than the “watch strategy”.

  • Unlike the traditional IT strategy, AI has to go ahead in the chaos when the product matures.

# practical advice:

  • Establishment of AI pilot unit, beginning with light tasks such as customer service, process automation, search, summary etc.

  • Open-testing error, fast-tracking small-scale testing, building organizational experience.

II, AI technical maturity: 2024 vs. 2025**

The status quo:

  • Before 2023, big companies were reluctant to use AI for productive tasks.

  • Now: the models of ChatGPT and OpenAI are more stable and are used by enterprises on a large scale.

  • Moving from the “trigger phase” to the “core business collaboration phase”.

# meaning:

  • Business perception has shifted from “AI is a good toy” to “AI is a productivity tool”.

  • AI became a controlled partner within the organization.

# III. The importance of memory and retrieval: Why is AI getting more and more “know you?”**

The key point:

  • AI no longer relies on “knowledge in model training”, but has ** search** and ** memory** capabilities: **Retrieval: ** For real-time access to facts, news.

  • Memory: for personalized interactive and long-term understanding of user needs.

# # # interactively upgraded:

  • Like human assistants: remember user preferences, habits, history decisions.

  • Closer to the “Agent-like” experience, where models can answer your questions more precisely.

IV, the reality and future of the AI proxy system

  • At present it’s like a “smart intern” and in the future it’s like a “senior engineer”.

  • It is expected that next year the proxy system will help enterprises to solve complex problems and even create new knowledge.

# Current ability:

  • Existing agents such as Codex can handle tasks automatically: GitHub operations, code writing, context judgement.

  • Capable of handling ** low-level repetitive cognitive work, e.g. automatic customer support, sale of mail, etc.

The next phase (for the next year) predicts:

  • Agents not only carry out tasks but also identify new solutions and generate knowledge**.

  • Accompanies the solution of the enterprise’s “most critical destructural problems”.

#

  • The role of personnel within the enterprise has shifted towards “allocation of tasks, evaluation of outputs, provision of feedback” and managing AI agents like managing the team of interns.

# V, AGI definition of distance from reality: how far are we from it?**

# Sam Altman’s point of view:

  • “AGI is not a point in time, but a continuous evolution”.

  • If you see ChatGPT today in 2020, people will say, “It’s already AGI.”

  • There is no need for a definition, but more importantly, attention** to a continuous, fast-tracked progress curve**.

Possible criteria for AGI:

  • It automatically discovers new scientific principles.

  • A significant increase in the rate of human knowledge creation (e.g., four times the rate of scientific progress).

# # meaning transformation:

  • The discussion of “AGI” is actually about “AI’s consciousness”, but it is a question of ** philosophical **.

  • The real concern is ** what “AI can solve”** and not whether it meets a definition.

Sixth and next generation model (Next-gen Modes) breakthrough**

#

  • Future models will achieve ** superb understanding, reasoning** and higher contextual capacity (e.g., millions token).

  • The model provides access to tools, databases, systems within and outside the enterprise and integrates knowledge to perform tasks.

  • To perform complex tasks of high quality and robustness and to become a true “manifest of minds”.

  • Businesses can hand over the most complex issues to AI, such as chip design, drug research and development.

  • Models can connect tools, read a great deal of context and “think for a long time”.

# Example scene:

  • Chip Company: let AI design a new chip.

  • Biopharmaceuticals: AI Mechanism for the Analysis of New Medicines.

  • Enterprises: allow AI to analyse multi-year financial statements and export strategy reports.

Seven, how does the future super-calculation resource (1,000x Compute) work?**

# # Sam Altman’s answer:

  • ** “Let AI study for itself how to build a stronger AI”**.

  • High-cost resources are used mainly for: Increase the depth of reasoning.

  • Execute multiple rounds and try to find the best solution.

  • Enabling complex scientific research.

Sridhar Ramaswamy (Snowflake CEO) adds:

  • If there is an unlimited amount of money, it should be invested in the study of such “significant human dilemmas” as the RNA expression.

  • Similar to the “Language Model Human Genome Project”.

# Full Chinese word for word: OpenAI CEO Sam Altman Dialogue

I’ll introduce you

** Moderator** (00:00–1:12): Looking to the future, we know that AI has the potential to change the world – and it can change in a better direction. The generator AI unlocks the capabilities from writing codes to philosophical reasoning. It starts with a “base model” that allows billions of users to use data that they could not otherwise access. They change our rules of the game, and the most far-reaching thing that drives change is OpenAI. The outbreak rise of ChatGPT, which is now used by more than a billion people every day, has completely reshaped the way we interact with data, intelligence, and the way we imagine our future work. At the heart of this transition is one of the most influential voices in science and technology. He leads OpenAI from the frontier to the products that affect billions of people, driving the AI globally. Please welcome OpenAI’s founder and CEO-Sam Altman! (applause and music ringing, 1:13–1:30)

# Sam Altman started a conversation with Sarah Guo

Sarah Guo (1:32): I’m Sarah Guo, founder and management partner of Conviction, and I’m happy to host today’s conversation. Welcome, Sam. ** Sam Altman** (1:42): It’s nice to be here again. I was just backstage saying – it’s like a rock concert for data people! Sarah (1:46): Ha-ha, yeah. You were here two years ago. Sam (1:48): I’ve been here before, but it wasn’t that much fun!

Enterprise AI Strategic Recommendation (2025)

Sarah (1:54): So let’s start here – Sam, what would you suggest to the corporate leader who wants to deploy AI in 2025? Sam(2:04): My suggestion is: ** Directly dry, do not wait. Many people are still hesitating, and models are being updated too quickly, waiting for the next generation of products to come out. But the faster change in technology, the easier it is to win ** firms with the fastest succession, the lowest test cost, and the most efficient learning. **Sarah(2:46): One thing I would like to add is: ** Keep curious. Many of the things we used to think of as “this is what” have changed completely. Thanks to tools such as OpenAI and Snowflake, the cost of testing is now very low, and you can do a lot of small experiments, gain value, and improve on that basis. Again, let’s echo Sam: **Who goes fast, who gets the most.

What’s the change from last year?

Sarah(3:36): So, what’s the difference between your last year’s proposal and this year? Sam (3:40): But it’s different now – our business is growing very rapidly. The big companies are actually using our products on a large scale. They tell us, “This thing is working now; it’s doing what I didn’t believe.” Sarah (4:56): So what do you think we’re gonna say next year? Sam(5:14): Next year, we might say: not only can businesses use AI automated processes or build new products, but we can also say: “** This is my company’s most important problem, and I’m going to invest a lot of money in it, and ask AI to solve it.”** These models will solve the problems that human teams are not capable of doing.** The companies that now use models and accumulate experience will have a huge lead.**

The role of memory and retrieval

Sarah (6:02): With regard to memory and retrieval, what role do you think they play in the future of AI? Sam (6:16): Like Retrieval technology, it has always been the key to making the generation AI more “downland”. For example, you ask a factual question, and models can be wrong if they are not supported by context. So we did the web search system back in the GPT-3 period to supplement the background needed for answers such as current events. And “Rememory” allows the system to know what you’ve talked to it before, so that it can do better in the future. These capabilities will become more important in the future, especially when it comes to more complex tasks.

Agent abilities and prospects

Sarah(7:16): Can you provide a “Agent Competency Framework” for corporate leaders? What can we do now, and what will happen next year? Sam(7:25): Yes, the programming agent that we have just introduced, Codex, is one example. You can give it a bunch of tasks, and it’ll be handled in the backstage, very smart. It can connect to your GitHub, and it may be watching meetings, reading Slack or internal files in the future. It’s like an intern who can work a few hours a day, but ** it will become a senior engineer who can work a few days a day.

What’s AGI? How far are we from it? Sarah(9:23): You said Codex made you feel AGI’s proximity, so how do you define AGI now? How far are we? Sam (9:44): If you go back to 2020, when the GPT-3 is not published, and show the present ChatGPT to those who are there, ** they’re sure to say, “This is AGI.” ** We humans are good at keeping up with “high standards,” which is great. But I don’t think “AGI” is that important, and the definition of different people is different. ** The models are going up every year for the past five years, and will continue for at least five years. ** Whether the word AGI is “proclaimed success” in 2024, 2026 or 2028, it’s not that important. The most convincing definition of AGI is: ** AGI is that a system can discover new science on its own or help human science break several times **.

What will the next generation model achieve? Sarah (15:50): How does it affect product construction if your ability to master the next generation model changes? Sam (16.00): The next one to two years will be amazing. A leap like the GPT-3 to GPT-4 will reappear again. Businesses can say, “Let’s give you a calculus, help me design a better chip / cure for new diseases.” Models will connect tools, understand the business context, carry out in-depth reasoning, and do their job independently. I didn’t think it would happen so soon, but ** is really close.**

# model capability range frame

Sarah (17:05): How do you judge whether a model solves a problem? Do you have a judgment framework? Sam (17:25): My “ideal model” assumes this: ** is super small + extraordinaire + can handle trillions of levels of context + connects to all tools. ** The model is wrong to use as a database – it is slow, expensive and unreliable. ** The true strength is “the ability to reason.” ** You can plug in every context of a person or an enterprise and then let the model use a tool to think and solve problems.

What if you have 1,000 times the power?

Sarah (18:24): What would you do if you had a thousand times the size of a calculus? Sam (18:33): Yuan said: I’ll use it to study better models, and let that better model tell me how to use all the calculations. Practical answer: Now you can invest more in hard questions to get better results. Although you don’t have 1,000 times more, ** it makes sense to understand it and try it.** Snowflake CEO: I’m going to invest it in RNA expression research. It’s like a genome program that can make a huge breakthrough in the treatment of diseases.** It’s an advance in the human class, and it’s very valuable to use a big model.**

# Final remarks (20:44)

Sarah: Thank you very much, Sridhar. Thank you very much, Sam. Thank you.

← Previous OpenAI announces the introduction of two core functions that integrate the business or your in-house knowledge with ChatGPT in depth.
Next → Google Programming Assistant Gemini Code Assist Update Integrating the latest Gemini 2.5 model and supporting personal customization

Related Posts

No related posts found