Yes, ChatGPT can draft a tricky email. But can a general-purpose AI assistant for the masses contribute to a big scientific breakthrough, without hallucinating or screwing up the research? The Allen Institute for AI (Ai2) isn’t holding its breath.
The Seattle-based institute’s new project, dubbed Asta, lets research scientists create and deploy specialized AI agents, working autonomously to help them tackle and solve some of the world’s toughest problems.
Ai2 bills Asta as a “bold initiative to accelerate science by building trustworthy and capable agentic assistants for researchers.” It’s also an open-source project, following the nonprofit institute’s ethos of making the digital plumbing of its projects available for others to analyze, adapt, and build upon.
There are three major components to Asta:
- Asta agents: AI assistants that can find papers, analyze data, summarize literature, generate hypotheses, and cite their sources.
- AstaBench: a benchmark suite for testing and ranking scientific agents on real-world problems.
- Asta Resources: a developer toolkit for building scientific agents using Ai2’s open-source tools and data.
With entry-level jobs in some industries already under pressure, Ai2 is careful to call Asta a collaborator for scientists, not a replacement. But most grad students would envy the kind of praise the institute is heaping on the project.
Ai2’s chief scientist, Dan Weld, said the inspiration came from the institute’s own need for AI tools that can execute complex plans, explaining how they came up with their ideas, and stay grounded in scientific evidence.
“That’s what Asta delivers,” Weld said in the announcement. “It’s not just another assistant but a collaborator designed to think like a scientist.”
The institute sees this open approach — where developers use Asta’s resources to build agents and Asta’s benchmark to test them — as a way to create “a flywheel of scientific improvement” for the broader community.
The news caps what will no doubt go down as one of the busiest months in Ai2 history, including the announcement of a new AI robotics initiative and a $152 million grant from Nvidia and the National Science Foundation to lead the creation of the future AI backbone for U.S. scientific research.
In their spare time, the team has also been settling into new offices in a mass-timber building on the north end of Seattle’s Lake Union.
Founded in 2014 by Paul Allen, and funded largely by the late Microsoft co-founder’s estate, Ai2 is a non-profit research institute that builds open-source AI systems for the good of science and humanity.
Ai2’s team of 225 people includes some of the top AI researchers in the world, many of them also affiliated with the University of Washington’s Allen School of Computer Science & Engineering. Ai2 is known for OLMo (open language models), Molmo (multimodal models), Tülu (instruction-tuned language models), and other open-source AI systems.
OpenAi, Anthropic, and other Silicon Valley startups have grabbed headlines for turning their proprietary AI models into widely used consumer and commercial products, while Ai2 has worked more quietly behind-the-scenes on foundational research and fully open-source tools for scientists and developers.
Asta is an outgrowth of Semantic Scholar, Ai2’s long-running project to create a massive research database that indexes more than 200 million scientific papers. Asta’s agents use this database through a dedicated “Scientific Corpus Tool,” an extension of the Semantic Scholar API.
Ai2 has been burnishing its open-source credentials under CEO Ali Farhadi, an entrepreneur and University of Washington professor who returned to Ai2 in 2023 after leading machine learning at Apple — which acquired his Ai2 spinout Xnor.ai for an estimated $200 million in 2020.
Farhadi says the goal is to hold AI to standards grounded in the scientific method: He said in a statement that Asta represents Ai2’s “bet on a future where AI doesn’t just keep up with science, it helps drive it forward.”
Read the full article here
