Anudhyan Boral
@bloopsie
Followers
122
Following
538
Media
5
Statuses
82
@GoogleDeepMind: Gemini Pretraining
Redwood City CA
Joined June 2010
Super proud to have co-led the development of new training algorithms that were critical for pre-training of the awesome Gemini 2.0 models. Looking forward to users and devs unleashing their creativity and driving innovations using these models!🚀♊️⚡️.
Gemini 2.0 is now available to everyone. ✨. ⚡ Start using an updated 2.0 Flash in @Google AI Studio, @GoogleCloud’s #VertexAI and in @GeminiApp. We’re also introducing:.🔵 2.0 Pro Experimental, which excels at coding. 🔵 2.0 Flash-Lite, our most cost-efficient model yet. 🔵
0
0
3
RT @WenhuChen: Gemini-2.0 makes a huge leap on our MEGA-Bench leaderboard to beat all the competitors! .With the other benchmarks being eit….
0
30
0
RT @ashish_vt: I am in awe of Gemini's ability to solve JEE level maths questions with stepwise reasoning as well as with the option elimin….
0
9
0
RT @DynamicsSIAM: Short Stories: "Rational Approximation" (by Lloyd N. Trefethen, in Notices of the @amermathsoc):
0
3
0
Just planned a Seattle getaway with - it's an absolute game changer! Offering everything from handpicked destinations to perfectly timed itineraries and seamless booking. It's one of the most unique and well-put-together AI products I've seen recently.
We are announcing the launch of Airial Travel’s open-to-all beta version for desktop today. Airial is your personal travel agent with AI superpowers which makes planning and booking trips as easy as dreaming them up. Me and Sanjeev co-founded Airial
0
1
2
Thrilled to have helped train this awesome model. Algorithmic efficiency FTW!⚡️⚡️⚡️.
Flash8B General Availability:. We originally trained Flash 8B giving it all our algorithmic efficiency improvements to pack as much as possible in a small form factor which then was scaled up to Flash. On benchmarks, it is closely matching Flash announced during May at I/O
0
0
33
RT @jainprateek_: Excited to share that the Machine Learning and Optimization team at @GoogleDeepMind India is hiring Research Scientists a….
0
59
0
RT @zacharynado: Gemini Flash is now tied with gpt-4o for #2 on the lmsys *vision* leaderboard!. combine that with a 1M context length and….
0
20
0
RT @JonasAAdler: A modest 59% win-rate over the original GPT4 too. I remember when that model was considered halfway to AGI. Now the price….
0
5
0
RT @tianle_cai: Just finished reading the Gemini 1.5 report and I'm blown away by the depth of information shared in such a competitive env….
0
42
0
For more, come visit our poster #603 at 5pm at #NeurIPS2023 or check out our paper (. Joint work with my awesome colleagues @GoogleAI Zhong Yi Wan, Leonardo Zepeda-Núñez, James Lottes, Qing Wang, Yi-fan Chen, John Anderson, @feishaAI . (9/N; N=9).
0
0
0
. and let SGD and empirical risk minimization discover the latent low-dimensional dynamics!. The latent dim is low enough that we can solve the SDE much faster than a numerical solver. We get stable rollouts across 1000s of steps while maintaining high fidelity throughout. (8/N)
1
0
0
However, can we bake in the inductive bias of ideal LES in a probabilistic ML architecture? Yes! We use a *latent neural SDE* to probabilistically model the evolution of the small scales in a low dimensional latent space. (7/N)
1
0
0
. by marginalizing over all possible small scales *consistent* with the large scales. While optimal closures can be computed in closed form for 'simple' settings such as homogeneous turbulence in a box, it is analytically intractable to compute in more complex scenarios. (6/N).
1
0
0
One of the promising approaches for closure modeling is ideal large eddy simulation or Ideal LES, proposed by Langford and Moser in 1999. Given an underlying distribution of full scale trajectories, it computes the deterministic large-scale state at the next time step. (5/N)
1
0
0
It has also been postulated that the truly important dynamics of the large scales are governed by a slowly varying dynamics along a low dimensional latent space --- but it's hard to discover this elusive slow manifold for a general turbulent system! (4/N).
1
0
0