In San Jose, a crowd gathered for TPC 25’s hackathone/tutorial plane session. (Source: TPC 25)
On Monday, a day of TPC 25 marked, the trillion parameters of the consortium and San Jose, an exhibition in Calif, a group of more than 130 participants gathered for a full session in the morning to focus the conference on the hectic/tutorial of the agent for science. Before the key note begins, Pacific North West National Laboratory Chief Data Scientist Neeraj Kumar introduces the scene for a conference on July 28-31, reminding the Standing Room only the crowd that the Kharban parameter consortium was only two years ago with a sole purpose for speeding with a sole purpose.
Kumar said that TPC now has about 1,1500 members in universities, national laboratories and industry, which is proof, Kumar said, the theory of open access to consortium is now under arrest in everyday research. He pointed to the recent milestone, from the agent’s systems that are drafting assumptions, including production models that are accelerating the discovery of content. “These are not extra improvements, but how we run science, an example is a change,” he said. “These achievements are just the beginning. The real change happens when we put these tools into the hands of scientists, engineers, all of you who understand deep questions in your field that can use AI to find answers.”
Today’s keynote was a research lead in Google Deep Mind Vivek Natarajan, which introduced the lab’s “AI co-scientist”, a multi-agent AI system made on the Gemini 2.0 model and a tool with mutual cooperation. Early work with medical language models revealed an obstacle: Thousands of random samples are needed to create the same useful assumption. AI co – – scientist runs a network of Gemini -based agents to deal with the incompetence that share four basic tasks: create, review, classify and improve. This is a mirror of the way the human group can improve mind -making, criticism and ideas, but works in seconds instead of day.
Natarajan explained how the architecture of this system pushes the papers to read, test ideas, and classification options into AI agents, while scientists continue to make every decision to review. An agent reads recent papers and suggests extension, another stage text -based debate, and a reflection agent examines claims against basic data or alpha folds. A classifying agent scores every idea on novelty, testing, potential impact and safety, then the results detect the tournament – the style reaches so strong ideas. An evolutionary agent revishes promising ideas, to speed up new literature or targeted data. He accelerates them. The full track of each step is saved, a choice that Natarajan chose that the reproductive capacity can be improved by giving the reviewers a transparent chain of reasoning and the use of the device.
Design by Deep Mind’s co -scientist. Vivek Natarajan’s courtesy slides.
A member of the audience sought details from Natarajan how the system scored ideas for “novelty”. He said that every assumption has been broken in basic ideas and assumptions, then the current literature has been examined to confirm the elements. The creative combination of known results receives some credit, but nothing less than a published example. Since fresh ideas are often unexpected assumptions, it balances novelty against testing and accuracy, which shows scores in all axes so that researchers can weigh different aspects before choosing which paths are to chase.
Case Studies founded the architecture of AI’s co -scientists. In severe myelide leukemia, AI co -scientists suggested three re -drugs: one showed strong tumor prevention in cell tests, two did not show the distance between drug promise and reality. In partnership with Stanford, this system has been used to identify vuranostat -based therapy for liver fibrosis, which damaged chromatin, which has a significant reduction in human organizations by 91 %. In another amazing case study, the system closely re -reproduced an unpublished anti -microbial -resistance progress from Imperial College London in two days, indicating the lead researcher to ask Google whether he had accessed his private files.
Natarajan noted that the agent of each AI co -scientists currently operates on the same Gemini model, yet future versions can be found for chemistry, genomics or specialized models for physics. Access through a reliable tester program is limited because every run compute is extremely, but it encouraged participants to propose pilot projects.
LTR: Google Deep Mind’s Vivek Natarajan, Charlie Catel of Argon, and PNL’s Neeraj Kumar. (Source: TPC 25)
Fully closed with practical guidance for the hacket/tutorial session. Kumar reminds newcomers that TPC’s working sessions have been created for about three goals: Increase an open community, launch projects, no lab can be dealt with alone, and AI’s researchers train the next generation. Administrators promise complete transparency on codes, data, and diagnostic methods, with the expectation that each partner meets with the group.
For Hikathon, teams will meet in six sessions from the basics to the final demo. Argon, Berkeley Lab, and the patronage agent of the industry sponsor will rotate to help access, debugging, and polydia Nvidia and Serbras Compete Credit. Parallel, the tutorial offers a structure in the agent for science, which includes the topics such as fine in scientific literature – toning, retrieval – generation of solidarity, and grounding model responses. Organizers stressed that participants could move between the tracks, but the purpose is the same: San Jose has a prototype, new colleagues, and a clear sense of how the agent system can advance everyday research.
The four -day program continues on Thursday with key notes, power conversations and breakout sessions. We will bring you daily highlights and deep divers to the discoveries that will appear in TPC 25.
Relevant
