In the previous post, I wrote about the dark days of developers and software startups.
Nevertheless, despite this situation, I would like to explain why I am still running Omnis Labs and deepblock.net, and why I continue this work.In fact, I once paid to use Copilot myself.
The problem is that our product is a client–server-based web application, so bugs in the frontend are often related to bugs in the backend server. Interpreting and debugging such issues through Copilot or LLMs is only easy when dealing with a single-language, single-project environment.
When debugging messy legacy code like ours, where modularization and interactions are not cleanly designed—as in a web service—it is not useful at all. In the end, I gave up using Copilot.
Even now, I still use ChatGPT, but only for implementing new code modules or components, or for writing separate code. I don’t really use it to debug the DeepBlock product itself.
LLMs still show many limitations when debugging messy projects with complicated interconnections (what looks like spaghetti code in a poor-quality web service).
To supplement this problem, one could simply input all the client and server code into GPT. But why would I provide OpenAI with my company’s core business logic?
In fact, I have already heard that many companies around me have been inputting all of their code into Copilot and GPT. The business logic of many software companies has already gone into Claude and ChatGPT, and they are doing their best to increase productivity even at the cost of providing their own code.
However, even so, I am still confident that GPT alone cannot reproduce our product yet. If LLMs advance further in about two years, perhaps they could build a clone of DEEP BLOCK, but based on my experiments so far, there are still many limitations.
LLMs are weak at debugging projects where various apps—client, server, etc.—interact.
To solve this, one must design cleanly from the beginning, with modularization and consideration for low coupling and high cohesion. But for a small startup that has been developing for years, it is unrealistic to expect such design, especially with LLM usage in mind.
LLMs are more specialized in building a single module, while understanding and reproducing something complex like our product, made up of multiple software components, is still beyond them.
Of course, it is true that LLMs have significantly increased developer productivity. But we have already endured countless trial-and-error while building and operating DEEP BLOCK.
Will it really be enough to build our frontend with LLMs? Could anyone, in a short time, build a cloud-hosted web-based software and overcome all the trial-and-error that Omnis Labs has endured over more than seven years?
I still believe that to build this product even in a similar way, one would need at least 10MM USD.
Perhaps in countries like India or Pakistan, where labor cost is cheaper, one could make it at lower cost. But talented individuals in those countries are already leaving for the U.S. and other nations.
In other words, at least in countries where software engineer wages are high, buying and using this product from us is still cheaper than reproducing it themselves.
That is exactly our reason for existence. Even if LLMs are good at coding, unless they have gone through our accumulated trial-and-error, our countless bugs, and our experiences, they cannot fully replicate this product. At least for now, our product is still cheaper to buy than to make from scratch.
Our company has executed more projects than people might think. Some of them were with global corporations and national institutions. But the markets we target—microscopy and remote sensing—are extremely difficult for small software startups. We have spent quite a lot of money over more than seven years, yet we still have not secured a stable cashflow.
Still, we have confirmed that there is real demand for what we have built in these niche markets. The fact that we have not yet turned that into steady revenue is due to many complex problems.
For example, in government institutions, budget execution has many procedures and obstacles. Lawmakers(politicians) and central government hold the budget allocation authority.
Local governments in Korea, for example, do not have strong authority over budget allocation. Even if a demand department(like GIS team)within a local government wants to use Omnis Labs’ product, the budget department is separate, and the demand department is usually the weakest division.
Not only the government: the microscope market is an oligopoly dominated by a few global corporations.
Likewise, the industries that need this are also oligopolistic, dominated by large firms that are conservative and dislike risks caused by change.
And making an opportunity in these industries really difficult.
Employees at Samsung Electronics, defense companies, and many Korean public officials share the same mindset.
They prefer to sit still. They are extremely afraid of trying something new that might cause problems and ruin their personal safety and career in their company.
They value stability above all else, dislike problems, and dislike creating problems. For this reason, selling a new-concept product in such markets is extremely difficult.
The problem is our target markets are these. 😐
At first, I focused mainly on the remote sensing field, but I realized it was a very tough market. So I expanded into the microscopy field, trying to find new applications and sales opportunities for our technology.
But the microscopy market also turned out not to be easy. I feel that I have tried almost every possible field where I could apply our competitive edge.
The past seven years of my business have been a series of hardships. To build competitiveness in a business I started naively, I built data pipelines specialized for remote sensing and integrated large size image processing into our no-code AI suite.
Then, because that market was tough, I expanded into microscopy—which processes essentially the same type of data—to open new channels.
In the end, I realized both markets are not easy. But at least, there are places interested in our product. And I am convinced that, with luck, we may still find opportunities to grow a little further and stabilize.
However, I believe this will also require luck. And in the distant future, I believe LLMs may eventually be able to build software similar to ours.
But at least for now, LLMs cannot generate our trial-and-error and experience. Whether a large corporation or a small startup, any company trying to build similar technology will inevitably require at least 10MM USD in development costs.
One Korean aerospace startup—where my alumni and a former employee of my company now work—received tens of millions of dollars in investment and spent it all.
I thought they would have built something similar to our technology. But when I checked their product, they had only built a simple software plugin for DJI Terra, despite burning through all that investment.
Most Korean startup incubation centers are run with government subsidies. Startups that move in survive on VC funds (also backed by government money) and grants, then mostly fail.
Most startups that began seven years ago around the same time as me have already gone bankrupt. When I visited coworking spaces, shared offices, or free offices at these incubation centers, I often saw expensive cars with rental or corporate plates in the parking lots.
These startups often wasted their investment or subsidies leasing luxury cars, buying alcohol, and partying before shutting down.
Considering that the majority of Korean startups fail within three years without making operating profit, it is obvious how strange it is to buy or lease luxury cars under a corporate name. Yet in most startup hubs, luxury cars are abundant.
Other companies’ moral hazard does not create customers for my business. But at least, seeing 80% of Korean scam startups blow their investment like this, I know they do not have the ability or willingness to build what Omnis Labs painstakingly created. That still gives us an edge.
In the AI field, they say SOTA changes every six months, and trends every year. Semi-supervised learning and few-shot learning were popular barely two years ago, but now researchers are talking about foundation models. Soon, that trend will change again.
But in this regard, DEEP BLOCK is at least free. Our strength lies not in the AI model but in the user interface and software. If a model changes, we just replace it. The frontend can stay the same.
Chinese researchers in computer vision and remote sensing publish terrifyingly powerful new models every six months.
Reading their incomprehensible papers, I often feel the limits of my own ability. But on the bright side, we can simply use the new AI model they invent in our backend, while focusing on user experience and fixing bugs.
The visitors of DEEPBLOCK.NET are now more from the U.S. than from Korea, and traffic is also coming from India and other countries, including leads generated through LLM services.
I don’t know whether this company can survive the LLM doomsday.
But to maintain the gap against companies chasing us with LLM,
to keep up with the brilliant Chinese AI researchers and their work,
and to prevent myself from despairing at the intelligence gap I feel when reading their papers,
I intend to keep making efforts with the time I still have.
Back in university, one professor once told students:
“People think in similar ways. If you find something hard, others will find it just as hard.”
If I struggle dealing with large-scale images in building and maintaining DEEP BLOCK, then of course others will struggle too.
If anyone thinks they can make something similar, they are welcome to try this pain. 😊