Armin W. Thomas
@ai_with_brains
Followers
792
Following
2K
Statuses
623
Member of Technical Staff @LiquidAI_ | Prev: Data Science Fellow @StanfordData working with @russpoldrack and @HazyResearch | He/him
San Francisco, CA
Joined August 2016
Today, we report advances in automated neural network architecture design and customization that we have been working on @LiquidAI_, offering an end-to-end process to tailor architectures to the demands of any task, metric, and hardware.
2
6
18
RT @JosephJacks_: We are very excited to announce that @LiquidAI_ has raised $250 million led by AMD. Consider this a $2.5 billion + raise…
0
20
0
RT @LiquidAI_: We raised a $250M Series A led by @AMD Ventures to scale Liquid Foundation Models and accelerate their deployment on-device…
0
50
0
RT @maximelabonne: ⭐ STAR: Synthesis of Tailored Architectures Really cool work from my colleagues @LiquidAI_ on model architecture. This…
0
16
0
RT @MichaelPoli6: We tackle the complex problem of architecture design for a world where AI is deployed everywhere. How do we balance trade…
0
13
0
Our new work on methods for automated neural network architecture design for any objective/task/hardware is out:
New Liquid research: STAR -- Evolutionary Synthesis of Tailored Architectures. At Liquid we design foundation models with two macro-objectives: maximize quality and efficiency. Balancing the two is challenging. To make progress towards this goal, we built a new algorithm — STAR. Read more about it here:
0
0
10
RT @xanamini: Excited to unveil STAR: Synthesis of Tailored Architectures! 🌟 We develop an evolutionary algorithm to automate neural archi…
0
8
0
This work would’ve not been possible without an amazing team at @LiquidAI_ , including @romu_nishi , @xanamini , @Massastrello , and @MichaelPoli6 .
0
0
1
We evaluated STAR in the design of improved language modeling architectures, optimizing for various mixtures of objectives, such as quality (perplexity after training), quality and parameter-efficiency, and quality and cache-efficiency.
1
0
1
By leveraging fundamental principles of evolutionary optimization–evaluation, recombination, and mutation–STAR refines genomes, iteratively shaping the encoded architectures towards any set of pre-defined static and/or dynamic objectives.
1
0
1
STAR is representing architectures as hierarchical numeric sequences called "STAR genomes", enabling it to define architectures at different levels of granularity, from the ordering and interconnection of computational units to their specific characteristics.
1
0
1
Importantly, STAR's evolutionary process can be guided by both static objectives (e.g., parameter count, cache size; defined by the architecture) and dynamic objectives (e.g., training perplexity, hardware latency; requiring evaluation).
1
0
1
Our methods for the synthesis of tailored architectures (STAR) are based on evolutionary algorithms applied to a numerical representation for model architectures derived from a new design theory. With STAR, we are able to push the frontier of model performance and efficiency.
1
0
1
If you’re looking for a CS PhD position, Dan is amazing! 👇
Excited to share that I will be joining UCSD CSE as an assistant professor in January 2026! I'll be recruiting PhD students from the 2024 application pool - if you're interested in anything ML Sys/efficiency/etc please reach out & put my name on your application! Until then I'll be finishing up some requirements at Stanford (long story...) and hanging out at @togethercompute. Stay tuned for more!
0
0
1
John is great! 👇
I’m hiring PhD students in computer science at Columbia! Our lab will tackle core challenges in understanding and controlling neural models that interact with language. for example, - methods for LLM control - discoveries of LLM properties - pretraining for understanding
0
0
0
Our work on Evo is now out in @ScienceMagazine and made it onto the cover! Grateful to have worked with an amazing team led by @exnx @MichaelPoli6 @BrianHie @pdhsu
A new Science study presents “Evo”—a machine learning model capable of decoding and designing DNA, RNA, and protein sequences, from molecular to genome scale, with unparalleled accuracy. Evo’s ability to predict, generate, and engineer entire genomic sequences could change the way synthetic biology is done. Learn more in this week's issue:
1
3
27
Our first models are out, offering state-of-the-art performances and small memory footprints; expect more to come soon! We are re-thinking all parts of the AI pipeline, from architecture design to post-training, and are just getting started!
Today we introduce Liquid Foundation Models (LFMs) to the world with the first series of our Language LFMs: A 1B, 3B, and a 40B model. (/n)
1
2
19