Pinned Tweet
Excited to share my very first article on AI policy & scientific discovery! 🧬🔬 At @GoogleDeepMind, we are building powerful "conjecture machines"—agentic AI that generates ideas at scale. But as AI accelerates, we face a critical new challenge: the Validation Bottleneck. 🧵👇
Idet, but testing them in the physical world remains slow and costly. For policymakers, funders, and researchers to truly unlock superhuman scientific discovery, we must understand and address this growing gap between generation and verification.
These thoughts build upon our works in AI for Math over the past years—from neuro-symbolic systems such as AlphaGeometry and AlphaProof, to thinking models such as Gemini Deep Think, and most recently Aletheia. My belief is advancements in AI technology for addressing the verification problems in maths will give great insights for other fields! 📐💡🔍
Idet, but testing them in the physical world remains slow and costly. For policymakers, funders, and researchers to truly unlock superhuman scientific discovery, we must understand and address this growing gap between generation and verification.
These thoughts build upon our works in AI for Math over the past years—from neuro-symbolic systems such as AlphaGeometry and AlphaProof, to thinking models such as Gemini Deep Think, and most recently Aletheia. My belief is advancements in AI technology for addressing the verification problems in maths will give great insights for other fields! 📐💡🔍
From proposing hypotheses to designing experiments, AI agents are starting to reshape scientific discovery. But the hardest part is testing these ideas in the real world.
Our essay explores the growing validation bottleneck and outlines four priorities for policymakers and o/biz1QlHNtm
11
38
197
44,704
Excited to have this release of Gemini 4 Carbon, #1 on @ArtificialAnlys, just 2 weeks after Argon! Our superhuman reasoning team and the broader discovery strike have contributed significantly to the model reasoning capabilities and the RSI recipe ;) Stay tuned!
Just two weeks after rolling out Gemini 4 Argon, excited to share a preview of Gemini 4 Carbon - our next frontier model. It has state-of-the-art long horizon coding capabilities driving SOTA performance across the cybersecurity defense lifecycle including detection, validation, o ranks #1 overall on the recently updated v5.0 of Artificial Analysis Intelligence Index at 59.
We're starting by rolling out to a subset of cyber defenders through our Fairwind Program this week and soon expanding access further. Read more at: blog.google/innovation-and…
21
38
619
24,718
@ArtificialAnlys Full details in the blog post with frontiner performances in cybersecurity, safety, and security! blog.google/innovation-and…
2
0
17
1,971
Google’s Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5, and second only to Claude Opus 5.5 on the Artificial Analysis Coding Agent Index
Gemini 4 Carbon scores 59 on Intelligence Index v5, driven by leading performance in complex reasoning, ic tasks. It also has large gains in coding and knowledge work, particularly when compared to other Gemini models, but it is not leading in all these areas.
As with launch of Gemini 4 Argon, Gemini 4 Carbon is being rolled out to selected users and is not yet publicly available. Google is offering a 50% launch discount, with no confirmed end date. All our charts and figures use standard pricing for both Gemini 4 Carbon and Argon.
Key results for Gemini 4 Carbon with high reasoning:
➤ Gemini 4 Carbon leads the Intelligence Index v5, with the #1 score in 4 of the 9 Index evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, AutomationBench-AA. Carbon makes substantial gains elsewhere but is not leading in coding and knowledge work
➤ Behind only Claude Opus 5.5 on coding agents: With Antigravity CLI, Gemini 4 Carbon (high) scores 59 only 3 points behind Claude Code with Claude Opus 5.5 (max, 62) on the Artificial Analysis Coding Agent Index. Carbon ranks second on both Terminal-Bench 4.0 (64%) and FrontierCode v1.1 (51%), behind Claude Opus 5.5 (max) at 56%. Google has historically underperformed in this area and so this improvement stands out
➤ Improved agentic knowledge work, but not yet leading: On AA-Briefcase, Gemini 4 Carbon (high) scores 1609 Elo, a 126-point improvement over Gemini 4 Argon (high, 1483), but still behind models including Grok 4.7 (xhigh, 1638), Muse Wonder (max, 1659) and Kimi K3.1 Pro (1678). The largest gain comes from Presentation, up 293 Elo to 1598, alongside a 120-point improvement in Analytical Quality to 1671
➤ Lower Cost per Task than Claude Opus 5.5 for comparable overall intelligence: Gemini 4 Carbon costs $7.76 per Intelligence Index task, ~23% less than Claude Opus 5.5 (max, $10.04) but ~1.2x GPT-6 Astra (max, $6.25). It uses 123k output tokens per task, 29% fewer than Claude Opus 5.5 (max, 173k). At the same $4/$20 standard pricing, its Cost per Task is in line with Gemini 4 Argon ($7.68), while scoring 9 points higher on the Intelligence Index
➤ Leading AA-Omniscience v1.1 score, with low hallucination: Gemini 4 Carbon scores 54 on our recently updated AA-Omniscience v1.1, 6 points ahead of Claude Opus 5.5 (max, 48) and 8 points ahead of Gemini 4 Argon (46). Accuracy (% correct) rises 10 points over Argon to 58%, while its hallucination rate rises from 4% to 11%. This is still the second lowest of any model, behind only Argon, compared with 50% for both Claude Opus 5.5 (max) and GPT-6 Astra (max)
Key model details:
➤ Context Window: 1M tokens
➤ Multimodality: Text, image, video, and speech input, with text output
➤ Pricing: $4/$20 per 1M input/output tokens, the same as Gemini 4 Argon’s standard pricing, with a 50% discount. Cached input tokens receive a 95% discount ($0.20 per 1M)
Congratulations to the @GoogleDeepMind team!
0
1
12
1,425
Thang Luong retweeted
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
135
795
6,561
826,911
Thang Luong retweeted
I have to admit, watched this a few times, especially to the end:) the creativity in the prompting and the model is awesome to see
Argon has this amazing natural understanding of the world. It can do so many things from first principles.
Here's a pelican riding a bike, using fourier epicycles.
Video made by Argon.
Time-synced Lyria 3.5 music prompted by Argon.
318
529
9,176
951,504
Gemini-4 Argon ranks #5 on Agent @arena. Getting there!
Excited to launch Gemini 4 Argon, designed to help with hardest coding and knowledge work tasks with its frontier capabilities..
For developers and enterprises, it’s available in the paid tier of the Gemini API and the Gemini platform, for Ultra subscribers in Google Antigravity, and Ultra and Pro subscribers in the Gemini app and Search.
Read more: blog.google/innovation-and…
6
5
134
10,011
Gemini 4 era - Glad to be pushing frontier capabilities together with amazing people! Still more to to improve and we keep pushing!
Lots of discussion out there about our next model(!), so I wanted to give an early look as soon as possible. Introducing Gemini 4 Argon!
It shows frontier performance in complex workflows, cyber defense and software engineering. Teams are using it extensively at Google, from ting, great feedback.
Here’s a look at the benchmarks:
12
11
220
12,133
State of the art in real-world long-horizon software engineering tasks such as DeepSWE and many complex tasks.
More details in blog
blog.google/innovation-and…
More details in blog
blog.google/innovation-and…
3
0
19
1,643
#1 on Text Arena x.com/arena/status/2…
Strong agentic performances across AutomationBench-AA, Terminal Bench 4, AA-Briefcase, and more. Details in Artificial Analysis x.com/ArtificialAnly…
Strong agentic performances across AutomationBench-AA, Terminal Bench 4, AA-Briefcase, and more. Details in Artificial Analysis x.com/ArtificialAnly…
Big news: Gemini 4 Argon (High) by @GoogleDeepMind just landed #1 in Text Arena with 1525 pts, and #8 in Code Arena: WebDev with 1679 pts!
This release has reshaped the Text Arena Pareto frontier with a blended $8/MToken! Gemini 4 Argon (High) is now the most cost efficient t on Pareto frontier below.
In the Text Arena, Gemini 4 Argon (High) ranks #1 in Coding, Hard Prompts, Instruction Following, Longer Query, and Creative Writing. It also leads every occupational domain evaluated, with additional #1 spots in English, Non-English, Chinese, and Russian.
This model is +20 points above the #2 ranked Claude Opus 4.6 (High), and a huge leap from Google’s previous release, Gemini 3.8 Flash (High) at #11!
In Code Arena: WebDev, Gemini 4 Argon (High) gained +96 points from Gemini 3.8 Flash (High), and went from #29 to #8.
Congrats to the @GoogleDeepMind team on this impressive frontier release!
0
0
16
1,417
Sharing math results done right in the AI era :) Congrats to @Carlein90 from our team & Mark!
Instead of dumping AI outputs, they derived insights, expanded the scope & gave the community reusable work. This is how science should be done—setting the tone towards Gemini 🚀
Instead of dumping AI outputs, they derived insights, expanded the scope & gave the community reusable work. This is how science should be done—setting the tone towards Gemini 🚀
Mark Shusterman and I have posted our resolution of the probabilistic version of Shafarevich's conjecture, posed by Liu--Wood and Sawin--Wood: arxiv.org/abs/2609.30200
The main idea of the proof was discovered independently by GPT5.5 pro and Aletheia, an internal agent at Google DeepMind. While digesting the proof we discovered several new applications. Among these, we settled a conjecture of Nikolov on the existence of universal perfect groups. This latter result came from an interesting experiment in letting the agent pattern-match problems to methods in reverse.
I wanted to see the random group question settled since I studied Melanie and Yuan's paper back in 2021. Very happy this is finally done!
2
4
26
4,860
@Carlein90 chasing for historic conjectures, Carlo & Mark "doubly" solved recent ones.
In Carlo’s words: “We built a very useful bridge between two theories and communities which was precedently unknown to the big people that posed that conjecture. And with this we proved an important foundational fact for one of these two theories.”
In Carlo’s words: “We built a very useful bridge between two theories and communities which was precedently unknown to the big people that posed that conjecture. And with this we proved an important foundational fact for one of these two theories.”
1
1
1
1,131
Thang Luong retweeted
Congratulations to Carlo and Mark on this paper! It contributes to one of my favorite themes in mathematics: the web of analogies between number rings, algebraic curves, and 3-manifolds. Recently these analogies have been used by by Wood, Sawin, Liu, Zureick-Brown, etc. to develop probabilistic models for number-theoretic invariants. Carlo and Mark establish some important consistency properties of these models, in particular answering questions of Liu--Wood and Sawin--Wood.
The Aletheia project has been relatively quiet in the past few months. We were excited to see these problems solved by a model we (including @jj_at_brown, @jon_lee0, @DawsenHwang, @gjb_ai and led by @lmthang ) post-trained at the beginning of July, but we were also happy to wait for Carlo and Mark to write up the paper properly, adding their own followup contributions and also documenting the new insights clearly. In this case, the key was a connection to deep results in group theory, specifically the theory of crowns and monoliths (full disclosure: I didn't even know those were mathematical terms before this paper!), which provide the "right" moments to bound the probability of needing many generators.
Keep watching @Carlein90 who is cooking up more interesting results 🔥🔥🔥
The Aletheia project has been relatively quiet in the past few months. We were excited to see these problems solved by a model we (including @jj_at_brown, @jon_lee0, @DawsenHwang, @gjb_ai and led by @lmthang ) post-trained at the beginning of July, but we were also happy to wait for Carlo and Mark to write up the paper properly, adding their own followup contributions and also documenting the new insights clearly. In this case, the key was a connection to deep results in group theory, specifically the theory of crowns and monoliths (full disclosure: I didn't even know those were mathematical terms before this paper!), which provide the "right" moments to bound the probability of needing many generators.
Keep watching @Carlein90 who is cooking up more interesting results 🔥🔥🔥
Mark Shusterman and I have posted our resolution of the probabilistic version of Shafarevich's conjecture, posed by Liu--Wood and Sawin--Wood: arxiv.org/abs/2609.30200
The main idea of the proof was discovered independently by GPT5.5 pro and Aletheia, an internal agent at Google DeepMind. While digesting the proof we discovered several new applications. Among these, we settled a conjecture of Nikolov on the existence of universal perfect groups. This latter result came from an interesting experiment in letting the agent pattern-match problems to methods in reverse.
I wanted to see the random group question settled since I studied Melanie and Yuan's paper back in 2021. Very happy this is finally done!
1
4
63
12,786
Thang Luong retweeted
Glad to see another live training run! Here's how Marin 535B-A23B is doing today: wandb.ai/marin-communit…
Anyone else want to share?
Anyone else want to share?
Nearly half a year of silence. We spent it studying one problem: how far RL can scale.
MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task agentic RL, mixed across multiple harnesses in one run), and grader compute (agentic in-group credit assignment, with test-case and rubric-based rewards). We'll open-source the details piece by piece over the coming weeks.
Streaming the run: mimo.xiaomi.com/rl/
32
70
842
79,494
Thang Luong retweeted
For 20+ years @ShaneLegg and I've discussed AGI’s potential impact on the economy, science & society. With the DeepMind Institute, we're expanding interdisciplinary research on key questions for the AI era. We hope it spurs the discussions needed to get the next steps right: deepmind.google/institute
341
630
3,985
626,340
Thang Luong retweeted
We built high-throughput materials labs in Menlo Park to create a loop between experiments and models. The labs generate fresh data, the models learn from it, and then help us decide what to try next.
Using only 1,300 H200s, plus months of our experimental data, we mid-trained model to surpass GPT-6 Astra on our analysis benchmark. We call it Neon.
This is real footage from our lab. We’re focusing first on hard problems in materials science, including superconductors, magnets, and semiconductor materials.
Read our blog posts below.
Using only 1,300 H200s, plus months of our experimental data, we mid-trained model to surpass GPT-6 Astra on our analysis benchmark. We call it Neon.
This is real footage from our lab. We’re focusing first on hard problems in materials science, including superconductors, magnets, and semiconductor materials.
Read our blog posts below.
274
522
5,097
1,665,724
Thang Luong retweeted
Today we have set out how we’re building AI to accelerate science and improve people’s lives. Just some examples in the last week or so:
- Mapped all 9B possible single letter genetic changes across the human genome with AlphaGenome Atlas and made it openly available to of decisions depend on weather predictions so we introduced WeatherNext 3, our most accurate and capable global weather AI model to date.
- We published AI & Economy ATLAS, a comprehensive open-access look at how people are using AI globally.
- AI has enabled extraordinary advances in language translation. Today our services are available in nearly 300 languages, spoken by 7B people
We’re focusing our efforts on four key areas: health, natural disaster and weather resilience, learning, and economic opportunity. blog.google/innovation-and…
- Mapped all 9B possible single letter genetic changes across the human genome with AlphaGenome Atlas and made it openly available to of decisions depend on weather predictions so we introduced WeatherNext 3, our most accurate and capable global weather AI model to date.
- We published AI & Economy ATLAS, a comprehensive open-access look at how people are using AI globally.
- AI has enabled extraordinary advances in language translation. Today our services are available in nearly 300 languages, spoken by 7B people
We’re focusing our efforts on four key areas: health, natural disaster and weather resilience, learning, and economic opportunity. blog.google/innovation-and…
438
678
6,279
821,297
Thang Luong retweeted
The resolution of Navier-Stokes is an amazing accomplishment. Some comments …
First, a brief mathematical postmortem. Millennium problems are important, but when you confront an unsolved problem, it is always unclear how far the solution is out of reach. The initial opinion of at least some experts (not myself – I have no expertise on this problem) seems to be that, in retrospect, the conceptual program underlying this solution happened to already be in place*, and the contribution of AI was a very intricate technical execution. Technical execution is one of the axes on which we've already seen AI become superhuman, though never before with such a dramatic demonstration.
This is not to diminish the solution or OpenAI’s achievement. But there are many problems which are out of reach of current models, no matter how much inference-time compute one invests. I’m still convinced that human mathematicians are laying indispensable groundwork for future progress.
*Terry Tao describes one possible program at mathstodon.xyz/@tao/117207851…. As I understand, the one Tao describes is NOT the Cordoba--Martınez-Zoroa program that OpenAI (and Alpoge/Buckmaster) followed. However, the successful approach was not as esoteric as Buckmaster’s statement might make it seem – there were certainly other people (not me) also prompting that approach; it just isn’t the kind of thing that one goes around saying publicly.
As far as I know, Navier–Stokes is the only one among all Millenium problems with a well-defined program towards a solution. Concretely, for the Birch and Swinnerton-Dyer Conjecture (over number fields) I am not aware of even a vague proposal of what a proof might look like, so I doubt that current AI models will get anywhere close to a resolution. (My opinion is that this conjecture cannot be resolved in the negative; perhaps the most refined formulation is false, but then the real problem would be to correct the formulation and then prove it.) I'd be curious to see what progress was made on it!
First, a brief mathematical postmortem. Millennium problems are important, but when you confront an unsolved problem, it is always unclear how far the solution is out of reach. The initial opinion of at least some experts (not myself – I have no expertise on this problem) seems to be that, in retrospect, the conceptual program underlying this solution happened to already be in place*, and the contribution of AI was a very intricate technical execution. Technical execution is one of the axes on which we've already seen AI become superhuman, though never before with such a dramatic demonstration.
This is not to diminish the solution or OpenAI’s achievement. But there are many problems which are out of reach of current models, no matter how much inference-time compute one invests. I’m still convinced that human mathematicians are laying indispensable groundwork for future progress.
*Terry Tao describes one possible program at mathstodon.xyz/@tao/117207851…. As I understand, the one Tao describes is NOT the Cordoba--Martınez-Zoroa program that OpenAI (and Alpoge/Buckmaster) followed. However, the successful approach was not as esoteric as Buckmaster’s statement might make it seem – there were certainly other people (not me) also prompting that approach; it just isn’t the kind of thing that one goes around saying publicly.
As far as I know, Navier–Stokes is the only one among all Millenium problems with a well-defined program towards a solution. Concretely, for the Birch and Swinnerton-Dyer Conjecture (over number fields) I am not aware of even a vague proposal of what a proof might look like, so I doubt that current AI models will get anywhere close to a resolution. (My opinion is that this conjecture cannot be resolved in the negative; perhaps the most refined formulation is false, but then the real problem would be to correct the formulation and then prove it.) I'd be curious to see what progress was made on it!
15
40
272
25,316