Rerunning a failed experiment from two years ago gives you a totally different result today. It's the cheapest way to feel the speed of technical growth. We unpack this principle through a test — write an essay from a single photograph — run once with ChatGPT 3.5 and again with Claude 3.5 Sonnet.
If you skip AI news for just a few months, catching up feels heavy. "GPT can do what now?" "What changed this time?" The fatigue piles up. So most people end up just consuming news and nothing more. But reading news alone doesn't put the speed of the change in your body. You don't develop a feel for it.
Here's today's principle. Rerun the same experiment in a new era. This essay explains why this habit is the cheapest and most accurate way to learn in the AI age. You can follow along even if you've never used AI. Examples age; principles don't. We'll go slowly.
Let's start with the principle. How do you know how much you've grown? Mark your height on a doorframe, come back months later, stand in the same spot. It has to be the same frame. Different frame, no comparison.
Skill works the same way. Pull out a piece of sheet music from a year ago and play it again. The measure that tripped you up is easy now. The feeling only exists because the sheet music is the same.
The speed of technological change must be measured the same way. Rerun the same experiment from a year ago. That's what drawing a line on the doorframe is. One rerun beats reading 100 news articles, because news is someone else's measurement, while a rerun is your own.
Oddly, this common sense collapses with AI. When a new model drops, people try new questions. "What can it do this time?" It's not bad, but the feel doesn't land. You didn't ask the previous model that question, so there's no baseline.
To measure accurately, do the opposite. Re-ask the question that failed two years ago. That gap is the real speed of technology. "3× better" is an abstract news number; "it couldn't do this, and now it can" is something you witnessed yourself. The second one is much more accurate.
Let's go to a specific example. In early 2023, when AI was first going viral, I tried one experiment. "Give the AI a single photo and have it write an essay about that photo."
ChatGPT was on version 3.5 at the time. Text only. It couldn't accept images. When I asked, it replied:
"Unfortunately, as a text-based language model, I cannot receive images. Please describe the photo or its elements, and I will do my best."
So I built a workaround. I used Imagga, an image tagging API. Upload a photo and it returns keywords with probabilities. I uploaded a photo of a skateboarder and got:
I copied this tag list, pasted it into ChatGPT 3.5, and said "write an essay from these keywords." I got: "It was a hot summer day and the sun was shining brightly in the sky. A young man full of excitement and joy came down on his skateboard…" A pedestrian narrative. It felt magical at the time. The video reached 29,000 views.
And a year and a half later, July 2024. Same photo — but this time I pasted it straight into Claude 3.5 Sonnet. No tagging service. No keywords. Just one photo and one line. "Analyze this photo and write an essay."
Claude's first paragraph:
"Between the blue sky and the endless sea, at the boundary where the city's rough concrete meets the soft sand of nature, we witness a remarkable expression of the human spirit. The skateboard leaping into the air seems to defy the laws of gravity."
Same photo, different era, completely different result. That's the line on the doorframe.
Think about a restaurant. A year ago the food was average. You took a note. "Miso stew too salty, three side dishes." One year later you go back. The miso stew is completely changed. Seasoning is balanced. Seven side dishes now.
"They changed chefs" or "they really leveled up" — the feeling lands immediately. That feeling doesn't exist on a first visit. There's nothing to compare against.
AI is the same. Same question, same data, different moment. Those three conditions have to align for the speed of progress to enter your body.
Let me compare the same "photo → essay" experiment, before and after.
| Item | Early 2023 (ChatGPT 3.5) | July 2024 (Claude 3.5 Sonnet) |
|---|---|---|
| Image input | Impossible — API tags as workaround | Direct paste |
| Intermediate tools | 3 (tagging API + copy/paste + AI) | 1 (AI only) |
| Prompt length | ~350 chars (tags + instruction) | ~20 chars |
| Depth of output | Plain narrative | Space, symbol, urban philosophy |
The same experiment went from 3 tools to 1. The gap between the two runs: about 18 months. That's the speed. What news calls "a multimodal revolution" in abstract terms shows up as something this concrete in one rerun.
Here's the aha moment.
Rerun the same experiment, and the speed of technology becomes a number in your body.
Keep one question close.
"What AI experiment from 1-2 years ago failed or left me unsatisfied?"
The answers split three ways.
"Korean translation was weird" → rerun
Older ChatGPT was slow and awkward in Korean. Everyone worked in English. Try the same sentences in Korean today. You'll be surprised. That feeling gives you the judgment that "from now on, Korean is fine."
"It couldn't read long PDFs" → rerun
A year ago, AI lost chunks of 100-page PDFs. Try the same PDF now. Most of the time it handles the whole thing. One rerun shifts your habit: "from now on, just drop the whole document in."
"Image analysis was garbage" → rerun
This is the exact experiment I did. It used to need a workaround. Today a single paste is enough. If your workflow still routes around something, it's probably deletable now.
Mnemonic: The failed one. The same one. Again.
Make one folder. Any name. I call mine AI_retry/.
AI_retry/
├── 2024-07_photo-essay_claude35.md
├── 2025-04_coex-map_o3.md
└── 2026-04_same-experiment_newmodel.md
Store only the same experiment re-run across different eras. The filename has the date and the model name. Years later, just opening the folder shows the curve of technological evolution.
Simpler experiments are better. Start with three.
1. One photo → essay (multimodal feel)
2. Long PDF → summary (context feel)
3. Vague instruction → result (reasoning feel)
Rerun these every 6 months and in 3 years you'll have six doorframe lines. Not someone else's news. Your record.
Let's wrap up.
A feel for technology comes from comparison. Not from seeing new things, but from remeasuring the same thing. It's drawing a line on a doorframe. The true speed of technology is read clearly from your own notebook, not the news.
The two photo-to-essay experiments are just examples. In three years, Claude 3.5 Sonnet will be an old name. But the habit of rerunning the same experiment across eras works under any technology. In fact, the faster technology moves, the cheaper and more powerful this habit becomes.
When you see news about a new model, don't read the news — open your folder. Pull out the experiment that failed a year ago. Ask the same question again. One comparison beats 100 articles.
Technology changes. Principles don't.
Failed. Same. Again.