Oct 18, 2026 · 31 min · 28 segments
Gemini 3 Pro is out, and records fell like snowflakes in Svalbard. No long description, chapters or links today, huge technical difficulties, including with audio, so just want to publish asap…
artificial intelligence.
Not only because Google is now clearly ahead, but also because it will be pretty hard for other companies to match their rate of acceleration.
I have tested Gemini 3 hundreds of times, including through early access, and it is indeed a significant leap, not just a nudge forwards.
On my own private independent benchmark, Simple Bench, it crushed its rivals, or I should say beat its own record to be clearly number one in this benchmark.
I will show you a sample question in a moment, but you may think that's a fluke.
Well, that would be a pretty hard line to maintain with the 20 other benchmarks in which it reaches record performance.
I'm also going to touch on benchmarks where it didn't perform as well, as well as the fascinating new tool, Google Anti-Gravity.
Above all, I'm going to try and give you at least 11 details that you wouldn't get from just reading the headlines that are going viral about the new Gemini 3.
And the reason the author of that benchmark, whom I've spoken to, called it that was because he solicited the hardest possible questions that he could derive using any expert out there.
They paid for any question at the time, which is around a year ago, that the Frontier models couldn't get right.
And sticking with knowledge for a second, what about scientific knowledge in STEM subjects? That's tested in the Google-proof Q&A, GPQA Diamond.
Even the creator of this benchmark thought that model performance had plateaued, but no, Gemini 3 Pro sets a record, 92% almost.
Now, I know what many of you are thinking.
Ah, well, that's only a 4% improvement.
Don't go too wild.
But imagine that 5% of that benchmark is noise, as in there's no real correct answer for those 5% of questions.
That puts the ceiling at 95%.
So just that delta from 88% to 92% represents eliminating over half of the remaining genuine errors that a model makes on that benchmark.
Average PhD performance from experts in those respective domains was around 60%.
Ah, well, that's just knowledge, though.
artificial intelligence.
Not only because Google is now clearly ahead, but also because it will be pretty hard for other companies to match their rate of acceleration.
I have tested Gemini 3 hundreds of times, including through early access, and it is indeed a significant leap, not just a nudge forwards.
On my own private independent benchmark, Simple Bench, it crushed its rivals, or I should say beat its own record to be clearly number one in this benchmark.
I will show you a sample question in a moment, but you may think that's a fluke.
Well, that would be a pretty hard line to maintain with the 20 other benchmarks in which it reaches record performance.
I'm also going to touch on benchmarks where it didn't perform as well, as well as the fascinating new tool, Google Anti-Gravity.
Above all, I'm going to try and give you at least 11 details that you wouldn't get from just reading the headlines that are going viral about the new Gemini 3.
And the reason the author of that benchmark, whom I've spoken to, called it that was because he solicited the hardest possible questions that he could derive using any expert out there.
They paid for any question at the time, which is around a year ago, that the Frontier models couldn't get right.
And sticking with knowledge for a second, what about scientific knowledge in STEM subjects? That's tested in the Google-proof Q&A, GPQA Diamond.
Even the creator of this benchmark thought that model performance had plateaued, but no, Gemini 3 Pro sets a record, 92% almost.
Now, I know what many of you are thinking.
Ah, well, that's only a 4% improvement.
Don't go too wild.
But imagine that 5% of that benchmark is noise, as in there's no real correct answer for those 5% of questions.
That puts the ceiling at 95%.
So just that delta from 88% to 92% represents eliminating over half of the remaining genuine errors that a model makes on that benchmark.
Average PhD performance from experts in those respective domains was around 60%.
Ah, well, that's just knowledge, though.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.