AI search company Exa CEO reo3 post-launch discussion: thoughts on the eve of AGI

preamble

I recently discussed O3 (OpenAI o3) with a few friends, and their reaction can basically be summarized as, "Oh my God, is this really happening?"

Yes, this is indeed happening. The next few years are going to be crazy. This is a moment of historical significance, if not galactic level significance.

Incredibly, there is no in-depth discussion about what is happening. ai labs can't talk about it, the news media barely report on it, and the government doesn't understand it.

It's ironic that we're even discussing the future of humanity through our social media feeds, it feels like a dystopian sitcom, but that's the reality.

Here are some of my thoughts on the current situation -- my contribution to the abyss of X-Platform thought. It is important to note that these thoughts are interesting guesses that have not been fully thought through. I haven't had enough time to think deeply and examine all the ideas, and many of them may be wrong. But I hope these ideas will help those who are trying to make sense of the current situation.

O3 Impact and Development Trends

The speed of technological breakthroughs

• The appearance of O3 should not have been shocking. OpenAI showed us the test time dilation charts two months ago, and the history of computer development teaches us to trust these trend lines, no matter how unbelievable they may seem.
• What's really shocking is that this all came to fruition in two months. We've jumped from college-level AI to Ph.D.-level AI. change is exciting for humans, but change this fast is shocking.

Projections for future development

• The O3 level model performs well in optimizing any task for which a reward function can be defined.
• In the short term (1 year), we will obtain a model of capacity imbalance:
  • AGI level in math, programming and general reasoning
  • Still mediocre in areas such as novel writing
• In the long term (1-3 years), the model's blind spots will be gradually fixed as reinforcement learning training in new domains (sentiment data, perceptual data, etc.) continues to be added.

Impact on various areas

The Future of Mathematicians

• Mathematicians may be one of the groups most vulnerable to AI replacement
• Mathematical work takes place mainly in symbolic space, with less contact with the physical world
• It is predicted that within 700 days, humans may no longer be the best mathematicians in the known universe

The Fate of Software Engineers

short term impact

• Every software engineer is equivalent to getting a promotion
• By the end of 2025, coding will be more like coordinating small agents to accomplish various tasks
• Clearly regulated PR tasks will be able to be accomplished by the O4 system

Long-term outlook

• The essence of software engineering remains the translation of requirements into pure logic
• Programming languages will raise the level of abstraction from binary to English
• Non-technical people will be able to participate in programming
• The best developers are still those who are comfortable switching between different levels of abstraction

Physical labor and robotics

• AI's relatively slow impact on manual laborers
• The main bottlenecks are hardware improvements and fast and reliable perceptual action models
• Robotics breakthroughs may not come until robots are able to build robots and AI is able to conduct AI research

Computing power and competitive landscape

competition for computing resources

• All the major labs are building supercomputing clusters
• Computational power may be more important than model advantage
• Mobility of talent and exchange of information between laboratories makes it difficult to sustain technological advantage for more than one year

Special effects of the O-level model

• Incentivized the construction of large-scale computing facilities
• Could change NVIDIA's market position
• Open Source Projects May Gain Advantage by Combining Global Computing Resources

The revolution in scientific research

Theoretical sciences

• Theoretical physics will be the first to be affected
• AI may accelerate the generation and validation of scientific theories
• Physical experimentation and material acquisition can be bottlenecks to progress

Risks and challenges

Main risks

1. Human Abuse of AI
2. social unrest
3. AI is out of control
4. Exacerbation of international conflicts

Regulation and social impact

• Government regulation will be an important factor in AI development
• Risk of possible terrorism and social breakdown
• 2025 could be the last year that AI is still in a relatively free-flowing stage of development

future outlook

optimistic expectations

• Breakthrough Physics Discoveries
• Robot-built bases on Mars and the Moon
• The perfect intelligent tutoring system
• Bio-enhancing drugs with zero side effects
• Efficient and clean energy systems
• More unknown groundbreaking discoveries

ultimate goal

• Realization of Superintelligence
• All possibilities allowed by the laws of physics
• Life Extension and Interstellar Travel
• Exploring the Origins of the Universe

Recommendations for action

Recommendations for individuals

1. Develop efficient problem solving skills
2. Enhance teamwork skills
3. Adapting to a rapidly changing world
4. Moving from individual success to the pursuit of collective progress

Contribution to society

1. Developing products that contribute to social stability
2. Dissemination of quality information
3. Participation in local governance
4. Helping the world make a smooth transition to the AI era

concluding remarks

On the eve of AGI, each of us is charged with shaping the future. This is not just a task for AI labs, but a collective mission for the entire human community. The future may be full of uncertainty, but it is also full of unprecedented opportunities. Let's work together to ensure that this technological revolution brings progress, not disaster.