It's remarkable, its not out of the bounds of the pattern of success that AI has had with math recently to the point that people should sound alarm bells.
A lot of the weight this holds is the fact that it's an old problem and that its difficulty hinges on the lack of investigation the disproof side of hypothesis. The model basically took a contrarian path and found tools and methods that support that a disproof is viable. So the (unquantified amount of) mathematicians out there were all dedicating their resources on the notion that this can be proved. Some with hindsight would say that if they a had team of experts who are driven to the goal of disproof that this would have been achievable by humans, and one of the mathematicians of the paper state as much,this still has value in terms of reliability measurement, and possibly human-aided endeavors when the methods scrounged by the model can be used in other solutions.
I think the intention of this paper is to build some type of culture of "math generalists" that don't quite exist in today's academia. The thing is, is that a good half of the people in that paper were actually very pragmatic on the implications of such a success and present questions in terms of the measurability of the difficulty of the problem and the generalizability of the solution provided for other questions. Gowers in particular offers no resistance and in fact resorts to the theatrics of "being the bearer of bad news" on Twitter for some reason.
As with Tao, he's always been a measured optimist even before the tools were consistently usable for his work. And even still nowadays, he adds stipulations to his statements on the successes of AI. Yes, he's part of Math Inc. now and is in close contact with Google Deepmind for some projects but his interest lies in using the tools today. Gowers has been hypothesizing on the future of math in the tone he has taken now ever since o3/GPT5. There's no comparison between the two who should attract more scrutiny.
This part of the announcement holds no value besides maybe taking a shot at the Deepmind Co-Mathematician paper. Nearly every mathematical success they've achieved around the GPT 5.2 generation has been done with general (and even public) models. Their last bountied problem solved was done with 5.4 Pro, also a general model.
There may be years of investigation as to how far you can generalize these methods. As to how central it is, it's a longstanding problem that Erdos loves to cite for that branch of math.
The thing is is that it seems a lot of the effort through the years (which is unquantifiable in scale as to how much time was spent and how many people focused their entire worklives on it if any) has gone for trying to look for the proof, and the search for the disproof seems minimal.
The accomplishment is cool. But all Erdos problems and other complicated mathematical problems they solved were accomplished with general-purpose models too. In fact for some of those problems, including bountied ones, they were public models. So I don't get saying this
It's a very complicated matter honestly. This is a new height that AI has reached, even though it follows the usual methods of success that it has had.
What strikes me as unusual though is that they do make a point of saying things like "this is a general purpose model that wasn't trained on the problem" among a few other things as if that's new. The last bountied problem they accomplished used a public model that ALSO didn't rely on specialized training. And that didn't make their blog.
I'm not sure your characterization of Tao is accurate lol. In that companion paper, only Gowers seems to extensively show no pragmatism in the implications of this accomplishment. Even the younger math experts in that paper were a lot more cautious with their statements. Tao seems to follow that same tune most of the time even though he uses AI for first-pass inspections of solutions brought to his attention.
It still has some artifacts more often than not, they are a lot subtler in nature but they still come out, whether it's texture, proportion, lighting, or perspective. Now some things are easier to fix on second pass edits, some are not. I guess it's why they consider image editing to be the next challenge.
The Gemini models and Eleven V3, and whatever internal audio model Sora 2 uses are about neck and neck in converging performance. They have some unexplainable flavor to them though. Especially Sora.
It's the skin textures. It's the slightly better lipsyncing. Maybe it will be different when us normal users get it but so far the demos with Sam don't make him look waxy.
So far the true progress it has made is getting textures right close up. It still fudges how skin looks like the more it pans away from the characters.
This whole autonomous driving levels kinda muddies the waters. Some would argue this isn't full L4 even. But it is a self driving car in the places it offers its services.
A lot of the weight this holds is the fact that it's an old problem and that its difficulty hinges on the lack of investigation the disproof side of hypothesis. The model basically took a contrarian path and found tools and methods that support that a disproof is viable. So the (unquantified amount of) mathematicians out there were all dedicating their resources on the notion that this can be proved. Some with hindsight would say that if they a had team of experts who are driven to the goal of disproof that this would have been achievable by humans, and one of the mathematicians of the paper state as much,this still has value in terms of reliability measurement, and possibly human-aided endeavors when the methods scrounged by the model can be used in other solutions.