(As far as I’m concerned)
On January 16, 2026, AI solved an impossible bug. A bug in a library downloaded over a billion times every year, with a wall of comments asking for help, several hacked-together workarounds, and an open call to the entire internet to please try to solve it. A bug that had been open for 15 years, languishing as the oldest open issue on the project’s GitHub, predating its migration to GitHub even, because it had stymied every person who had tried to tackle it. I myself had already spent a full day with little progress to show. It would have taken me at least another week of full-time effort to crack it. Claude Code took 50 minutes.
Ahab chased his whale. Aguirre his gold. I had Issue #209 3D scatter plots don’t work in logscale.
This was, on its face, not a simple thing to solve. You combine logarithmic scale transforms with 3D graphics, and the audience has already tuned you out and started doodling in the corners of their pamphlets. I had a head start on it after spending years poking around in the 3D portion of the codebase, to the point that the project lead asked me to join the official maintainers team. “Surely,” I thought to myself, “open-source is a rewarding way to contribute to the grand project of science that in no way could ever put me in the middle of an international controversy.” And yet even then I was missing the other half of the puzzle. Scale transforms were their own maze of complexity built up over decades, and solving the issue meant being an expert in both.
My full day earnestly tackling the problem told me I didn’t know enough to solve it. I knew that we needed to insert scale transforms in-between the data and its projection to the 3rd dimension, but there were dozens of places over twenty thousand lines of code where this might need to happen. Progress would be largely all-or-nothing and hard to incrementally track, and what was currently happening in the code was poorly documented. It would take me days if not a week to figure out. So as alluring a prize as it was, it got pushed to the get-to-it-someday back burner.
Where it languished for years.
Until Opus 4.5 came out.
Claude Opus 4.5 was released the week of Thanksgiving, and the busy holiday season meant I was a few weeks late to trying it out. According to the industry hype on twitter, this made me a dinosaur on the brink of extinction. (A ridiculous statement – everyone knows we have 4 years until that point). But the hype this time was justified. Using Claude Code had previously required one-at-a-time prompting to generate useful code, and its “YOLO mode” where you took the blinders off was as likely to wreck your project as accomplish anything useful. Opus 4.5 was the first model to be reliable and capable enough to let it loose and chew on problems on its own. YOLO!
I threw it at the 3D logscale problem on a whim. I told it what I knew and what needed to be done, had it make a plan, and let it think. 20 minutes passed and it came back saying it had fixed the issue. But its eyes were not yet as good as its brain. “The rendering is messed up,” I responded, “There are still missing transformations somewhere.” Another 10 minutes of thinking through the code. The rendering still looked messed up. But progress was happening.
Four or five rounds of this, and then suddenly the problem was solved.
There were a dozen places that needed to be tweaked, but it turns out that expertise in the other half of the puzzle was key. Way down in a corner of the code several jumps away from the part dealing with 3D, setting the scale to logarithmic was automatically applying a 2D transformation to the data. This had to be overwritten so that we weren’t stretching the data twice, in the wrong direction. This would have taken me ages to discover.
Its solution wasn’t perfect – I spent the next few hours doing additional debugging, cleanup, and refactoring, both manually and with targeted prompting, before it was good enough to submit a pull request. Even today the models (yes, Fable 5 too) are still dramatically lacking. They need to be steered in the right direction, hallucinate often enough that manual fact checking is still crucial, and somehow write worse prose than they did a year ago.
But on January 16, AI solved a problem I couldn’t easily solve, in a domain where I was credibly the world’s expert, in less than an hour. That’s intelligence by any metric that matters. I find myself able to delegate it more and more work by the day, and it consistently hits my quality bar. Developers all over the world are having similar experiences. The models are pushing beyond their first conquest in coding to other fields. I share this story not as something unique, but as a snapshot in time.
Are the AI models “generally” intelligent now? Others might hold back the label until they snag a literary award, load a dishwasher, win an art contest, or solve a landmark 80-year-old math problem (but wait – those goalposts have been met!). For me, it’s already past the threshold. We’ve reached AGI and are continuing to climb.
Which for now is great. We all still have jobs, the looming risks from bad actors using these tools as a force multiplier have not yet materialized at scale, and they offer incredible promise to speed up drudge work and enable research and discovery. The small slice of progress made here means the world gets 3D logscale plots with the release of matplotlib 3.11 a few days ago. I am happy to leave Ahab and Aguirre wallowing in their yearning while I mark that issue closed and move on. This, in a nutshell, is the promise of AI progress.
I’m considerably less optimistic about our trajectory as we zoom towards something beyond our capabilities. What makes these things astonishing is also what makes them dangerous. For now I’ll mark the milestone.