The Problem of Seeing the Future
The project we are currently building is a frontier information radar for small businesses and individuals.
By the so-called “ChatGPT moment,” I mean the day ChatGPT-3.5 was released: Sam Altman simply said, “welcome to chat with it,” even OpenAI itself had not fully anticipated that ChatGPT would pass one million users in just five days and reach around one hundred million monthly active users in about two months.
In hindsight, we can explain that moment from many angles: the Transformer architecture, scaling laws, large-scale compute, researcher mobility, the partnership between OpenAI and Microsoft…
But the problem is that before things actually happen, ordinary individuals and small businesses have a hard time seeing the connections among these signals.
Information Gathering Was Once Reserved for Institutions
A person’s attention is limited. We cannot read all papers, news, and social media discussions consistently, and it is even less possible for us to turn them into a coherent view. Many times, we do not even know that a certain field is changing. In the past, this kind of information analysis capability usually belonged to large companies, VCs, research institutions, and consulting firms. They had dedicated analysts, industry consultants, technology scouting networks, and internal databases. Now, the emergence of LLMs has changed this cost structure for the first time.
A New Kind of Information System
If you do not understand the specific principles of LLMs, you only need to remember one thing: one of the things they are best at is reading, summarizing, classifying, and reorganizing information. Like humans, they have a memory limit, which is the context window; but as long as we arrange the workflow well, let it read in batches, extract information structurally, save checkpoints, and update continuously, it can gradually accumulate a fairly substantial digital asset.
O-DataMap is an example worth referring to. It tries to reorganize information scattered across papers and experimental data into a visualized data map. Users see a scientific map they can zoom into, click through, and explore, but what is truly important behind it is an agentic AI pipeline: continuously collecting new data, parsing papers and experimental results, placing information into a unified coordinate system, and continuously updating the entire map. I want to apply the same idea to “frontier technology judgment in a particular field.”
Tracking the Conditions for Breakthroughs (and Failures)
My professional background is neuroscience, so naturally I would start thinking from neuroscience: in the next five years, which directions might have their own “ChatGPT moment”? Is it BCI? Is it neuromorphic computing? Is it brain foundation models? Is it neuro-diagnostics? Or is it connectomics?
There are no clear answers yet. But each of them has several signals worth tracking: paper growth, hardware progress, capital investment, clinical trials, company formation, talent migration, regulatory actions, product prototypes, and public narratives. The question is not “which one will definitely succeed,” but whether we can establish a continuously updated set of indicators to help us understand these changes earlier and more systematically.
Looking back at the landing of ChatGPT, it was not simply because capital marketing succeeded. Behind it, at least several types of conditions were satisfied: first, the theoretical and technical path had accumulated to a certain point. Transformer, pretraining, instruction tuning, RLHF, and scaling laws collectively advanced large-model capabilities. Second, hardware and infrastructure had matured. GPUs, cloud computing, data centers, and large-scale training frameworks made training and deployment at scale possible. Third, the product interface was simple enough. Users did not need to understand how the model worked; they only needed to open a webpage and chat in natural language. Fourth, the capability was general enough. It was not only able to complete one vertical task, but changed writing, programming, translation, Q&A, learning, office work, and many other scenarios at the same time. Fifth, user feedback formed a flywheel. The use, discussion, secondary development, and business attempts of a large number of users made ChatGPT no longer just a model, but an entrance to a platform.
Looking at it in reverse, why did the metaverse not succeed? It was not because it had no capital (over $80 billion invested), not because it lacked endorsement from giants (Facebook even changed its name to Meta), and not because it lacked media volume (needless to say). Its problem was that many basic conditions had not truly matured: the hardware experience was still cumbersome, the content ecosystem was insufficient, high-frequency user scenarios were unclear, implementation costs were too high, and even presenting leg movements stably and naturally in a VR world was very difficult. In other words, the metaverse had a strong narrative, but lacked the foundations for widespread adoption.
How Should We Judge?
This is exactly the kind of thing I hope this system can help judge. It is not meant to make investment decisions for users, but to continuously read frontier materials, organize signals with a clear set of indicators, and tell users why a certain direction is worth watching, which evidence supports it, which evidence refutes it, which parts are still hype, and which parts have already begun forming a viable industry.
Preliminarily, each field can be broken down into several core indicators: 1) technological maturity: is there a repeatable and scalable core breakthrough? 2) hardware and infrastructure maturity: does it have the compute, equipment, supply chain, or experimental conditions needed for scaled implementation? 3) industrial maturity: have companies, standards, development tools, supply chains, and real customers appeared? 4) investment momentum: is funding chasing a short-term narrative, or is it continuously entering the infrastructure and product layers? 5) talent migration: are top researchers, engineers, and entrepreneurs beginning to migrate toward this direction? 6) market adoption: are there ordinary users or real customers who can directly feel value? 7) regulatory and ethical resistance: are there strong constraints in medicine, safety, privacy, or policy? 8) cost curve: are key costs declining? 9) platform potential: is it a single product, or could it become an entrance to a new ecosystem?
In this way, what we establish is a continuously updatable, traceable, and correctable frontier information radar. It reads papers, news, patents, company dynamics, financing, hiring, product releases, and regulatory documents, then converts information into structured events and maps them into a unified evaluation framework.
Each update should answer several questions: What happened? What is the evidence source? Which core indicator or indicators did it affect? What uncertainties or open questions remain?
And also: if we put it into a virtual portfolio, how would it shift our focus toward which directions, companies, or industry chains?
LLMs as Human Extension
The focus of this system is not to “let AI make decisions for you.” Quite the opposite, it should avoid making judgments detached from reality on behalf of users. It should be based on real materials, updating views in the most cautious and smallest-step way. All conclusions can be traced, reviewed, rolled back, and restarted from checkpoints. More importantly, after users gradually form their own understanding of a certain field, they should be able to give feedback to the system in return.
For example, the system may raise its attention weight for a company because that company has raised a lot of funding. But if the user has industry insight and raises an opposing view, then the system should absorb this feedback and adjust its subsequent judgment rules. The user only needs to tell the agent in a chat interface like ChatGPT, and the system can incorporate this feedback into the subsequent workflow.
In other words, this AI adviser is not a substitute for the user’s judgment, but an extension of the user’s cognition.
It helps individuals and small businesses accomplish things that were very difficult in the past: continuously tracking complex fields, turning scattered signals into structured understanding, and establishing their own judgment framework before opportunities become widely recongnised. In the past, doing this required expensive data services, industry consultants, and research teams. Now, with LLMs, agent harnesses, public data sources, cloud APIs, or local inference devices, it only takes a few thousand dollars, or even a lower startup cost, to build an initial version.
Case Study: Neuroscience
It does not need to cover all fields at the beginning. The first stage can choose only one specific track, such as frontier technologies in neuroscience. It can first track BCI, neuromorphic computing, brain foundation models, neuro-diagnostics, and connectomics. If this method works in neuroscience, it can naturally expand to other fields, including some more niche industries in which small and medium-sized businesses themselves operate.
In short, the question I want to explore is: if ChatGPT allowed ordinary people, for the first time, to have language and information-processing capabilities close to those of experts, then in the next step, can it also allow ordinary people and small businesses to have the technology-intelligence capability that used to be possessed only by large institutions? It cannot eliminate uncertainty. Its goal is to help you understand why something matters, how to evaluate it, and how to build your own perspective before the next “ChatGPT moment” arrives.
The End
If you are interested in this project, the questions I most want to discuss further are:
From the perspective of investors or industry observers, is this division of core indicators scientific enough? Can it truly help us judge whether a direction is approaching an outbreak?
If you could have a low-cost, 24-hour online AI adviser that continuously reads the latest materials and organizes judgments, what would it ideally look like? What should it proactively remind you about, how should it explain complex fields to you, and where should it remain restrained?
Which field would you most want to apply such a system to? AI, robotics, neuroscience, energy, materials, biomanufacturing, education technology, or a more specific niche direction in your own industry?
And I would very much like to hear any insight, risk, application scenario, or possibility that you think I have overlooked.