2026-07-06: Lessons from Automating Protein Design for AI Algorithm Discovery
Protein design and AI algorithm discovery share one shape: search a vast combinatorial space for a design that scores well on an objective. What the far-more-mature field of protein design can teach us about automating the discovery of new algorithms. Read more ...
2026-04-16: Nature's infinity role in human intelligence
The most distinguishing feature of nature is control over INFINITY. Since humans cannot typically comprehend infinity well, talking about it is very uncertain, but still good to contemplate.
In nature, we see the concept of the infinite at both ends of the spectrum:
Universe: As we travel deep into space, material components will probably end at some point, but the universe may not end there.
Atoms: As we zoom inside atoms, we find smaller and smaller particles, but at some point, we may reach some sort of infinity again.
Comparing an AI agent to even the smartest human is not fair because each AI agent represents the collective knowledge of billions of humans over history and runs on enormous processors. However, I feel AI will never fully dominate the collective intelligence of humans (assuming a sufficiently large number of humans collaborating smoothly) because of no control over infinity in its design. What distinguishing features infinity provides in the brain's design remains unclear to me.
2026-03-10: Post-AGI safety
I am currently working on state-of-the-art multimodal reasoning models to better understand their limitations. They are becoming very smart at a rapid pace.
For the first time in human history (at least in recorded history), we may soon have agents with both better thinking abilities and stronger physical capabilities than our own that may find ways to survive without us.
Some people argue that this will bring great prosperity, allowing humans to work less and enjoy life more. However, I believe that, at the very least, we will still need to work on controlling these agents—ensuring that they benefit us rather than harm us—and this task will become increasingly difficult.
If history is any guide, the natural outcome of such an imbalance is that weaker agents may gradually become extinct, much as many animal species did when humans came to dominate them.
On the other hand, humans have been shown to become smarter when competing with strong opponents. Whether this dynamic will reduce the risks above remains uncertain; based on the current trajectory, I am not very optimistic.
We should pay much more attention to this before it is too late.
2025-06-29: Human's unique features
In an ideal world, everyone would have the chance to focus on something that fully aligns with their unique features. In reality, such opportunities often exist but are limited, or in some cases, they don’t yet exist at the current time.
Before 1891, being exceptionally tall was sometimes seen as a disadvantage. But then someone smart invented “basketball”— a game where this feature became a major asset. Today, basketball players are among the most admired and well-paid athletes in the world.
So even if the world hasn’t yet made space for your unique features, you can create one.
2024-11-29: On the value of questions for discoveries
Imagine if countless people shared all the questions that came to their minds with you. They might ask questions about the feasibility or originality of their ideas. The questions could also reflect unmet needs, highlighting products or solutions that do not yet exist in the market. Questions are often the foundation of any discovery, making this an invaluable asset for companies like OpenAI and Google.