Jarrod Kahn (@kahnvex) header
Jarrod Kahn (@kahnvex) profile picture

research at @GoogleDeepMind ➜ /home/kahnvex/logs cat *

Bay Area
Joined May 2022
Photo by Jarrod Kahn (@kahnvex)Photo by Jarrod Kahn (@kahnvex)Photo by Jarrod Kahn (@kahnvex)Photo by Jarrod Kahn (@kahnvex)Photo by Jarrod Kahn (@kahnvex)Photo by Jarrod Kahn (@kahnvex)Photo by Jarrod Kahn (@kahnvex)Photo by Jarrod Kahn (@kahnvex)Photo by Jarrod Kahn (@kahnvex)Photo by Jarrod Kahn (@kahnvex)
Gemini Next
What are you building with Gemini 4 Neon this weekend?
9
1
169
4,833
Jarrod Kahn retweeted
I'm sorry but Gemini 4 Neon is not on Opus 5.5 level. It's above it, and it's cheaper.
254
209
38
6,920
336,637
The $100T question: If our models learn all that there is to be learned from verifiable domains, to what extent does that knowledge transfer to non-verifiable domains? Does the generalization extend to physics, chemistry, biology? Can we skip 10 years of scaling automated physical science labs?
Why's everyone so shocked about the Riemann hypothesis? I get that it's the holy grail in math, but math is a highly verifiable task so you should *expect* this breakthrough. The big jolt would be if AI’s prowess can generalize to unverifiable tasks—so far that’s an open question
3
1
8
3,257
I’ve been thinking about how 2026 will look a century from now. Will it stand out as a seminal year for math and computer science, or just a steeper slope in the steady march of technological progress? I have this weird sense of incongruence either way. Knowing how the progress is actually being made, and how much low-hanging fruit is still left, it’s hard to see how it could possibly stop. At the same time, how can this possibly continue?
7
831
Historic buying opportunity btw.
1
20
1,163
Did not expect to see Demis, Nobel laureate, tweeting the words "cooked" and "cracked" in reference to GDM researchers. @demishassabis next we need: aura, lock in, goated, ... Would pay for these tweets.
🤣 💙 Apology accepted (so this was @_sholtodouglas then? 😅 ), but I think you missed checking the 'Underestimated the cracked Google DeepMind team' box. Amazing work from the Gemini & GDM teams, always great to see folks enjoy our cooking!👨‍🍳🧪
1
43
4,290
Jarrod Kahn retweeted
My team at Google DeepMind completely pivoted to safety about 2 months ago, and I think you should too. Interpretability is still dear to my heart, but it's always been a means to an end—and right now, I can't justify working on anything other than directly trying to improve frontier model safety. We're hiring FTEs in Seattle only to work on urgent safety problems directly on Gemini. I cannot emphasize how important this is right now in human history! Apply here: google.com/about/careers/app… 👇 (unfortunately, I won't be able to reply to any emails, but please apply!)
61
93
16
1,325
89,286
Last thought on this. When friends join GDM, one thing I try to help them understand is how deep the Google AI/ML bench is. For many it is hard to comprehend, but the pool of talent across Alphabet is almost comical. GDM is obvious, but there is also: 1. Waymo 2. Google Research 3. Ads 4. YouTube 5. Search 6. Cloud 7. Isomorphic 8. Core and Technical Infra 9. ... All of these teams have extensive R&D efforts and for many people, their inner loop is training models. So when people say GDM are cooked because they lost X talent, realize that there is a massive fluid talent pool that will fill in the gaps.
I want to thanks @SemiAnalysis_ for the motivation. In the Gemini RL scaling team, we printed & put that article on our desks to be reminded daily of what used to be true at the time. But the times they are changing :)
9
4
2
212
15,959
Apparently the gpt team never got the message.
Holy crap, someone tell the gpt 5.4 mini team that Gemini 4 seems to be tight on their heels
2
1
146
6,432
Say what you will about this being a moment of triumph for Google, the bigger story behind this screenshot is the amount of hubris in the AI community, even coming from those who are on the outside. If you’re writing what gets clicks, or what people want to hear at that moment, you’ll eventually lose credibility in the minds of those who are actually paying attention.
6
6
1
146
5,330
This is basically correct, could quibble about which model supersedes Argon, but the gist is right.
so in the gemini 4 model family hydrogen = sonnet/terra neon = opus/sol carbon = fable/astra argon is older and superseded by neon am i getting this right?
1
9
2,240
New daily driver just dropped, come get yours at antigravity.google.
Introducing Gemini 4 Neon! Pushing the frontier across domains and punching above its model class in a meaningful way. To me, Neon feels like the right balance of intelligence, speed, and usability. Available now across @GoogleAIStudio, the API, and more!
1
16
587
Jarrod Kahn retweeted
Happy WedNEsday!
417
353
157
8,527
1,217,724
Gemini
537
470
733
9,807
1,679,858
Thinking about tweeting a randomly selected element from the periodic table every night at 11pm to see this place go mad but my friends say it's a bad idea.
1
18
2,018
It's good. It's fast. Go build.
Introducing Gemini 4 Hydrogen, an extremely capable and fast model built to bring Gemini 4 to everyone! Available to all tiers of Google AI Studio, Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro + Ultra subscribers in Google Antigravity.
2
18
473
Introducing Gemini 4 Hydrogen, an extremely capable and fast model built to bring Gemini 4 to everyone! Available to all tiers of Google AI Studio, Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro + Ultra subscribers in Google Antigravity.
807
832
462
11,544
2,925,552
Gemini
910
249
474
5,804
761,696
Pumped to have Carbon out in the wild. I cannot wait to see what people build with it. On to the next one!
Gemini 4 Carbon is now available in Cursor! At high effort, it's the top model on CursorBench at 60.4%
1
10
453
I assure you none of my code changes take more than one hour to submit. At Google, there might be some things to complain about, but the ability to ship code is just not one of them.
I'm increasingly bearish on Reflection. This is what happens when you import Google's engineering bureaucracy into an early-stage startup.
1
1
29
3,804