Recent type Among the artificial intelligence agent, trained in understanding how the software is built by browsing the company’s data and learning how this leads to the final product, can be both a more talented software assistant and a diminutive step in the direction of much smarter AI.
The recent agent, called Asimov, was developed by Reflection, a diminutive but ambitious startup mixed by the best AI researchers from Google. Asimov reads the code, as well as E -Maile, Slack messages, project updates and other documentation to learn how all this leads to the creation of ready software.
The ultimate goal of Reflection is to build superintelligent artificial intelligence – something that according to other leading AI laboratories work. Meta recently created a recent Superinteligencja Laboratory, promising huge amounts for scientists interested in joining a recent effort.
I visited the Reflection headquarters in the Brooklyn district in Williamsburg, Recent York, on the other side of the street from the elegant Pickleskall Club club to see how reflection plans to achieve superintelligence before the competition.
The company’s general director, Misha Laskin, claims that the ideal way to build AI Supermart agents is to have their real coding, because it is the simplest, most natural way to interact with the world. While other companies build agents who apply human users interfaces and are looking through the network, Laskin, who previously worked on Gemini and agents in Google Deepmind, claims that this almost does not naturally come to a huge language model. Laskin adds that teaching artificial intelligence to understand the development of software will also bring much more useful coding assistants.
Laskin says that Asimov aims to spend more time reading the code instead of writing. “Everyone really focuses on generating code,” he told me. “But how to make agents useful in team conditions are not really resolved. We are in this semi -automatic part in which agents are just starting to act.”
Asimov actually consists of several smaller factors in the coat. All agents cooperate to understand the code and respond to user queries. Smaller agents download information, and one larger reasoning agent synthesizes this information in a coherent answer to the inquiry.
Reflection claims that ASIMOV is already seen to surpass some leading AI tools according to some measures. In a survey conducted by reflection, the company stated that developers working on huge open source projects who asked questions, preferred answers from Asimov in 82 percent compared to 63 percent of Claude Anthropik’s Claude code with the Sonnet 4 model.
Daniel Jackson, an IT specialist at the Massachusetts Institute of Technology, claims that the Reflection approach seems promising, taking into account the wider scope of its information collection. Jackson adds, however, that the benefits of this approach remain to see, and the company’s survey is not enough to convince him of broad benefits. He notes that the approach can also escalate calculations and potentially create recent security problems. “It would read all these private messages,” he says.
Reflection says that a multi -aggregate approach soothes calculations and that it uses a secure environment that provides greater security than some conventional SaaS tools.