Important Problems (2025)

---

Note:

As pointed out by readers, this list is very broad, perhaps as syntom of me being a student and not (yet) working on any of these problems in depth.

Another valid critique I received is that I focused on the means /Level 5 autonomy/ rather than the end /reliable transportation/. A good solution to reliable transportation (think of cities that have good transportation systems for people and/or self-driving tubes, etc.) may not involve any type of vision-based capabilities or the construction of a general physics-based world model. I still think my point holds, as I wasn't referring to reliable transport to be solved as a valuable goal per se (while it may be), but I was talking about how specifically solving Level 5 self-driving could unlock a general physics simulator that could inadvertently also turn useful for robotics, biology, and beyond. In general, as per Vapnik's Principle, working on the hardest but practical and specific problems may be more effective than working on an abstract generalization as an intermediate step. (Thanks Edo).

Also, I wrote about "Better RL" and "learning from explanations" as abstract wishes. Recent research into reasoning techniques using RL with verifiable outcomes suggests it's working.

I remain a student of these problems, but I am now less interested in the broad philosophy and more obsessed with the engineering of making useful things happen.

What problems deserve the best researchers and engineers alive?

> What are the important problems of your field? What important problems are you working on? If what you are doing is not important, and if you don’t think it is going to lead to something important, why are you working on it? — Richard Hamming

I think all of the following problems are worth solving and we need someone to work on them. This doesn’t want to draw a line between things we should do and things we shouldn’t do. This is an attempt to figure out: what problems deserve the best researchers and engineers alive?

Problems that I believe to be important and pivotal for humanity include:

Better Reinforcement Learning: learning from explanations, not millions of examples
LLMs + traditional/better RL (not RLHF): what if you fine-tune an LLM to be a chess master in an RL fashion and as a side effect it becomes a superhuman genius in something else? Repeat WLOG for every problem where traditional/better RL can be applied, and strive to figure out ways to reformulate problems so traditional RL can be applied
Level 5 Self-driving: it must not be fun to drive trucks around all day, especially if you can’t come back home to your family at night — see also the following bullet point
Generative Vision / Robotics: Solving level 5 self-driving and/or generative vision potentially also gets us a good general physics simulator as a side effect, which can then be used to do reinforcement learning on more robotics tasks. The impact of each of these problems would compound far further and contribute to solving the other pivotal problems as you can see.

Maybe these are fundamentally all the same problem. Maybe all of these connect and converge towards general intelligence (”AGI”). I don’t know. Yet.

Bioengineering: solving diseases. This is fundamentally just information science, where the rest of techniques apply. Nothing new. It’s still worth remarking and being included here because this application is timeless and important enough regardless of techniques

Laziness / Attention span: the human attention span decrease is the saddest loss of this decade. This can both lead towards slowing down productivity and innovation as well as cutting individual’s ability to enjoy life and human consciousness as a whole. It also makes us optimize for the wrong lazy objectives, addictive infinite scrolling feeds may be just the beginning
Climate change: yes, we don’t have enough data and are not too sure about this, exactly because of this uncertainty, we should move as if the worst forecasts will become true.

I also think almost all of the original Y Combinator Requests for Startup from Summer 2024 are great ideas pivotal for humanity.

On the other hand, problems that may be cool but I don’t believe to be pivotal or worth of pursuing as much as the former include:

AI application/system-level efficiency (retrieval tricks, orchestration, use case-specific wrappers): the iPhone interface was so good it stayed the same for years. The chat interface, similarly, is so good that will stay until the next innovation of similar impact. Nothing pivotal here. Yes, it is necessary to integrate AI to improve the efficiency of as many fields as possible, and it is the right thing to do. Is it worth the best researchers/engineers of the decade? No.
AR/VR: cool, impressive, but what difference will it make? And will it be net positive or net negative? Is it worth the best researchers/engineers of the decade?
Brain-Computer Interfaces: the computer keyboard is completely fine. This is just good for accessibility.
Interplanetary quests: this is of course interesting but all the former problems, if solved, would make this easy.

I also don’t believe any of the latest (as of today) YC Requests for Startups are pivotal for humanity. Sure, they would probably make for good businesses, but do they deserve the best researchers/engineers alive? Probably not. I’m just a young student. Correct me if I’m wrong at hello at danielfalbo dot com.