From Pneumonia Detection to General AI: A Medical Practitioner's Perspective on a Technological Sea Change

A Medical Student's Reflection on the Times

Recently on Reddit, a tech practitioner with a medical background shared this sentiment: "Last quarter has been insane. Amazing times to be alive."

Behind those words lies a genuine story of technological evolution. The author recalls that back in 2019, while pursuing his MD, he developed a machine learning algorithm to detect pneumonia in chest X-rays. At that time, the AI capabilities we now take for granted were "an unimaginable pipe dream."

He drew a vivid analogy: if he could go back and explain to his 2019 self the capabilities of models like "GPT 5.6 Sol" today, the shock would be like describing a modern computer to someone from the early 20th century. He admitted: "I would have called that AGI without hesitation back then."

It's worth noting that names like "GPT 5.6 Sol" and "Luna" mentioned in the post may be the author's imaginings of future models, internal codenames, or forward-looking references — they are not officially released products. But that doesn't diminish the core emotion he's conveying: the pace of technological iteration has far outstripped even practitioners' expectations.

The Leap from Narrow Models to General Capabilities

This reflection resonated so widely because it precisely captures a critical inflection point in AI development: the leap from narrow models to general capabilities.

The "Handcrafted Era" of 2019

Around 2019, medical imaging AI was one of the most successful application areas for deep learning. The typical approach looked like this:

• Collect a specialized labeled dataset for one specific task (e.g., pneumonia detection)
• Train a convolutional neural network (CNN) model
• That model could only do that one thing — a different task required starting from scratch

While this approach performed well in specific scenarios, it was fundamentally "one model, one problem." Developers had to invest enormous effort in data labeling, feature engineering, and model tuning, with results that were nearly impossible to transfer to other tasks. This was the hard ceiling of medical AI in that era.

Today's "General-Purpose Era"

Modern large language models and multimodal models demonstrate an entirely different paradigm. A single model can write code, analyze medical images, perform clinical reasoning, draft research reports, and even handle tasks it was never explicitly trained on. These "emergent capabilities" are precisely what makes this practitioner feel like he's "traveled through time."

For someone who personally trained a specialized medical AI model in 2019, the qualitative shift from "narrow" to "general" carries an impact that outside observers can't fully appreciate.

Cost Collapse: Another Underappreciated Revolution

There's another detail in the author's reflection that's easy to overlook: he mentions that being able to run a model "at a reasonable price for a month" is equally astonishing.

This touches on a crucial yet often neglected dimension of the AI revolution — cost economics.

The Dual Miracle of Rising Capability and Falling Cost

When people discuss AI progress, they tend to focus on "how much smarter models have gotten" while ignoring the fact that "getting smarter also got cheaper." Looking at the trend over recent years:

• The inference cost per token has continued to drop rapidly
• Models of equivalent capability now run at prices orders of magnitude lower
• The barrier to running AI services at scale over long periods keeps falling

For practitioners in 2019, training and deploying a specialized model often meant significant compute costs. Being able to run far more powerful general-purpose models long-term at "reasonable cost" was simply inconceivable at the time. AI capabilities truly reaching the masses ultimately depends on dramatic cost reduction.

Are We Already Standing at the Doorstep of AGI?

One line from the author is worth sitting with: if you brought today's models back to 2019, "I would have called that AGI."

This raises a fascinating phenomenon — the definition of AGI keeps getting redefined.

The Ever-Moving Goalpost

Over the years, the criteria for "artificial general intelligence" have shifted continuously:

• Once, passing the Turing Test was considered the hallmark of AGI
• Once, beating humans at Go was seen as a milestone breakthrough
• Once, holding a natural conversation like a human was enough to stun the entire industry

Yet each time AI hits a milestone, people quickly treat it as "expected" and push the definition of AGI further out. Many capabilities we consider ordinary today would have been unhesitatingly labeled "AGI" just a few years ago.

This practitioner's reflection reveals exactly that cognitive inertia: our ability to adapt to technological progress is so fast that we barely feel the progress itself. What was once "science fiction" becomes "everyday reality" — often in just a few short years.

Cherish This Moment of Witnessing History

This Reddit post has no fancy data visualizations or rigorous academic arguments, yet it resonates deeply precisely because of its authentic practitioner's perspective. Someone who personally trained a medical AI model in 2019 expressed, in the plainest terms, a sense of awe at the acceleration of technology.

"Amazing times to be alive" — these words may carry a touch of excited hyperbole, but they remind us of a few important things:

• We are living through a historical moment where the technology curve is extraordinarily steep
• The transitions from narrow to general, and from expensive to accessible, are happening simultaneously
• And humanity's rate of adaptation to all of this may be even faster than the technological progress itself

Whether "GPT 5.6" is real or a vision of the future, the excitement of personally witnessing and participating in this technological upheaval is itself the most precious footnote of our times.