Ideas

In an interview, Sam Altman said if he were 22, he would feel like the luckiest kid in all of history. For the empowered individual, there is simply an overwhelming abundance of projects worth pursuing and a frustrating scarcity of time to do them.

Since I haven’t yet found a way to build everything myself (still trying to), I’m open-sourcing my backlog. This is a list of (half-baked) project concepts I’d love to see brought to life. Some of these are taken from somewhere else. Feel free to steal anything.

Information understanding

Betting as a substitute for polling

As a famous saying goes, “A bet is a tax on bullshit”.
Use distributions on betting platforms to understand real mass opinion, better than polls.
Remember the 2024 US presidential election - instead of asking “what do you feel”, asking “what do you think your neighbor would feel” gave much more accurate responses and the French trader Theo made a fortune. There is headroom in mass opinion understanding.

A weekly newsletter where you ask founders or business leaders for their take on an important question and share their takes with readers.
Reach out to influential voices in the tech or business space asking them some big questions e.g. When will we reach AGI, are we in an AI bubble etc.
Send out a weekly edition where you share the responses to one of these questions from the different thought leaders polled.
Business Model: Sponsored placements in each issue + premium membership for archive / search / tools.
End Goal: Potential acquirers include Substack / Morning Brew / Axios / HubSpot, LinkedIn, or X. Target 4-6x revenue once brand and data moat are established.

User data collection

How to collect data from individuals (e.g. product reviews) when not everyone is willing to manually write reviews?
The value of human beings might just boil down to data collection and data annotation. Annotation might be more effective and noise-free if it’s done via passive participation such as simple things like games and captcha, or simply observing the decisions they make. This should be done without formally paying someone and asking them to intentionally label for your tasks.
Because if you think about it, humans are doing data annotation tasks all day on their own anyways! That’s all they ever do.

Noise as a fingerprint of origin

If a certain dataset contains a certain type of noise, what can we say / do?

Why are charts effective?

Given the same information, why is a chart more productive compared to text description of the same thing? think about how machines & humans process this info.

General visualization of outputs

Why do we need all-text output when logic can be more clearly expressed using graphs?
For example, if I ask for an explanation of the difference between big endian and little endian, the best response should include a graph or slide instead of text. The same can be said about conditional sentences, processes, conversations, and most topics.

Modeling

Distillation

We can use a model to design a rubric & to judge, & for it to be used as a reward model for RL; this is a proven recipe that gives us a better model.
But this means all that knowledge has been in the model in the first place, it’s just amplified through iterative self-reward.
Is there a way to mine this knowledge upfront?
Also, where can new learning come from? Maybe new learnings don’t need to come from a new dataset, they just need to come in another way from an existing model.

Learning from the nature

Shortest paths chosen by electric currents, whale songs, etc. These carry great amounts of information (think about the number of decisions they prune) but are under-utilized just because they can’t be very conveniently serialized as text. How to collect the data and learn from them?

Using LLMs

LLMs operate on next-token natural language prediction.
As I’m writing in 2025, a lot of startups have offered LLMs in the quality of a chatbot or a conversation predictor.
Selling LLMs in this way is like selling computers as calculators.
- A CPU is built to do arithmetic (think about all a CPU can do, they are instructions like ADD, SUB, etc.).
- In fact, arithmetic abilities are a general class ability because many problems can be framed in it. This makes a CPU a versatile machine that can solve a general class of problems. This makes computers generally valuable; not just for arithmetic but for this extension of it.
Now LLMs are built to do next-token natural language prediction. But does this mean the most powerful products which can be built on LLMs have to stay on this surface?
Most of the LLM startups of 2025 still try to package the LLM in a product that uses the natural language interface directly. Either as chatbots (to predict dialogs) or as coding agents (to predict code). This is very limiting.
Hint: LLMs are valuable because they capture inter-token relations that very cheaply encode “common sense”, above all else. Common sense used to be expensive.

LLMs are post-trained with alignment

We know that training on one front almost always results in loss in ability on other fronts.
By “alignment” (making the model easier to use), we make the model less capable.
So the best models might be intentionally not post-training aligned. The question is just how we will perform inference on them.
- For example, we can’t just ask such a model a question directly.

Programming

Balance between code and no-code

Code is too hard to learn, but no-code is too unpredictable.
Data types (chat, int32, int64, etc.) are just an artifact of how the computer likes to think about things. It is not how humans think. So it needs to be removed.
But the current (2025) no-code tools where humans just issue one command and the whole thing gets built - is also suboptimal because when users can lose track of what happens.
They actually want to enjoy the process of building, like putting together LEGO pieces. At the same time they don’t want to be totally removed. In this process they want to involve as little computer-native thinking as possible.
So, they want to code - just no data types, but still conditionals, functions, etc.
The devil is in the detail, and sometimes humans don’t want to think of the detail. It is in this case that the AI can step up.
It’s one thing to code up an algo but another thing to use database (Redis), multithreading, async, etc. to fully implement it.
Distillation or best practices in this regard has the potential to be made into a new programming language, just like how Python was distilled as best practice for C programming, in part.

Code understanding

If you want to understand and contribute to the code base, LLM should give you a good starting point.

Value analysis

Analyze GitHub code and observe signals of commercialization & purchase. Startup founders need early benchmarks on their valuation.

Contribution analysis

Want to measure output of an engineer but how? Use a model to make the value of a code segment more measurable.

Languages

Kid’s bilingual toy

Lots of immigrant parents in the US with cultural anxieties.
A lovable interactive plush that gently introduces children (ages 2–6) to a second language through playful audio prompts, engaging storytelling, and responsive language games.

Browser plugin

Selectively translate some parts of page, choosing the most pertinent parts to translate based on your level.

Online traffic

Informatics video

Embed informatics in a song! Like @learningwithlyrics on instagram

Trading

Bartering

Money was invented because people did not have enough information to find partners (or chains of partners) for bartering. Does this condition still hold in 2025?
When we use money to buy things, the price we are paying is impacted by how other people want the same thing. We don’t necessarily want this.
In bartering, a lot more trades can happen that will make everyone happier. See the Red Paperclip House.
AI should allow bartering to happen in a lot more occasions.

Compute as a currency

Compute will be the most valuable commodity / currency in the world. Because it will be the most scarce commodity in the world. Because it won’t saturate too much - more compute will always be more valuable.
By a similar logic, digital storage will also be the most valuable commodity / currency in the world.

Personal shareholding

Sell a part of yourself to shareholders at a price you agree on. They are entitled to a portion of your lifelong earnings in a designated period and will provide you with personal mentorship / resources.

Entrepreneurship

With AI there are countless good ideas. The real challenge is in the execution. But having ideas is very easy.
Nevertheless, it is still valuable to have good ideas. See the success of the Half Baked newsletter for reference.
Found a platform to share ideas & go one step further - make it easy & straightforward for people to contribute to ideas (jobs, brainstorm, capital, compute, etc)

Misc

Inter-agent communication

HTTP for agents - highly effective & condensed communication between agents. Define Inter-Agent Communication Protocol (IACP). And then commercialize it by providing ecosystem infra like HTTP

Too many browser tabs

In a browser, when clicking a link to a new page, instead of going to a new tab, use a pop-up window - use the model to show the most relevant information on a new page, but stay on the current page

Git for PPT

As title

NBU (Next Billion Users) for ChatGPT

Essentially everybody can benefit from using an AI chatbot to offload some decision-making burden. But why isn’t ChatGPT having a 8-billion DAU yet? It’s because some people still either don’t know it exists, or don’t have reliable access to it (think laptops, electricity, internet, etc.)
These people might not necessarily know how to download new apps & merge it to their daily routines.
How is AI Mode a success? The core underlying model is not too different from what’s on the Gemini endpoint. But it is closer to the user’s existent routine.
Chat model endpoint via SMS, or a phone hotline. Users prefer simplicity.

Email agent

How do CEOs manage their inbox? Assistants reply to emails for them. And by EOD or EOW they go over these decisions made on their behalf: meetings scheduled, pending items, etc. together in a standup meeting.
Do the same with email agents - this will be better than an AI smart reply.
As well, if everyone’s using AI, why not directly facilitate the exchange of information? Why still rely on text as if it’s written by someone?

Messaging platform

Turn on notification only if a condition is met. This condition can be natural language (e.g. notify only if the msg is relevant to me).
Or could this be a theme / notif center capacity in Android?

Voice mode for consumer apps

People don’t like to start into a small screen. Think Duolingo, but voice only.

Hiring

Hiring is a promising market, at least for now. Think Mercor.
LeetCode doesn’t work anymore.
Competition-based hiring, think Kaggle.

Vibe photography

Vibe programming solves the problem of programming for people who don’t know how to program.
For people who don’t know how to take good photos but have the need (think husband & wife), do we have vibe photography?