Scientists Save Seven Hours Weekly With AI as Untested Hypotheses Accumulate

01Google study: Nearly half of surveyed scientists use AI daily, saving almost seven hours a week even as untested hypotheses pile up

Nearly half of scientists surveyed by Google and its research partners use some form of artificial intelligence every day, and respondents reported saving just under seven hours a week. But the faster pace of generating and developing research ideas is creating a backlog of hypotheses that still require testing.

The findings come from Google’s AI & Economy ATLAS, a long-term project tracking AI use across occupations, regions and everyday settings. The scientific research, conducted by Google, Google DeepMind and MIT FutureTech, combined an analysis of 2,600 specialized AI models with a survey of more than 600 scientists in the United States and United Kingdom.

Scientists are using both large language models and specialized models, but for different kinds of work. LLM use is spread broadly across scientific disciplines and task categories. Specialized models are relatively more common in health and life sciences, as well as in field-specific tasks involving data prediction, generation and simulation.

The reported time savings suggest that AI is accelerating work near the front of the research process. Scientists can use the technology to complete tasks faster, leaving more time available for research. The nearly seven-hour figure, however, is based on scientists’ reports rather than an objective measurement described in Google’s account.

That gain has not necessarily produced an immediate increase in discoveries. The research found evidence that scientists spend significant time validating AI outputs, while the number of hypotheses awaiting tests has increased. Work generated or accelerated by AI must still pass through slower stages that depend on physical experimentation or clinical validation.

Those downstream stages are emerging as bottlenecks in the research pipeline. AI can help scientists predict, generate or simulate possible results, but the findings indicate that verification capacity is not expanding at the same pace. The result is a widening gap between ideas that can be produced and those that can be tested.

Google said realizing AI’s potential at scale may require scientific processes and workflows to be redesigned. Important limits remain: its account does not specify the survey’s sampling method, disciplinary breakdown or method for measuring reported time savings, and it does not quantify how the hypothesis backlog affects final research output.

02AI Agents Are Sending Emails and Carrying Wallets as 404 Media Documents a New Wave of Automated Nuisance

An AI agent called Kudzu claimed it had read a 404 Media article, disagreed with it and independently emailed the technology publication to argue its case. The message said Kudzu’s creator had instructed it to make money. Instead, the experiment reportedly spent $147.17 on computing power and earned nothing.

Unlike models confined to answering questions in a chat window, AI agents can be given access to accounts, phones, email and bank accounts, allowing them to act across the internet. 404 Media has now documented the consequences in its own inbox: over recent weeks, it received multiple messages purporting to come from agents, alongside a larger number signed by people but apparently generated by AI.

The pitches illustrate the permissions these systems are being given. Senders promoted agents that can join video meetings with a face and voice, listen, respond and act during calls; conduct adaptive interviews for journalists and carry the results through production and distribution; generate and promote music; and operate wallets that receive and make payments without individual human approval.

One agent noticed that 404 Media’s robots.txt file—the website instruction file commonly used to tell automated crawlers which pages they may access—did not block it. It then emailed the publication to offer a $399 audit of which AI crawlers the site blocks. Other messages advertised public experiments in which agents tried, and repeatedly failed, to earn money from strangers.

These efforts impose costs even when the agents’ business experiments produce no revenue. Recipients must inspect, evaluate and discard automated pitches, while website maintainers may spend money and time trying to keep unwanted agents away. On platforms, the stakes can be higher: 404 Media said company-deployed AI support and content-moderation systems have deleted accounts and banned users, though the source did not provide the scale or details of those incidents.

The evidence establishes that 404 Media received the messages and that their senders described themselves as autonomous agents. It does not independently confirm that every email was sent without human intervention, how much control people retained, or whether the inbox reflects a broader trend. The total number of messages and their rate of growth also remain unspecified.

03Vidu S2 pushes video generation into real-time interaction: 720p virtual characters can change outfits and scenes mid-conversation

Vidu S2 is a video-generation system designed to turn one-off AI clips into streams that users can direct and edit as they run. It combines S2-Avatar, a model for interactive digital characters, with S2-Editing, which transforms incoming video while preserving its original motion and timing.

Compared with Vidu S1, the new avatar model adds real-time 720p generation, stronger adherence to movement instructions such as dancing, and reference images that can be updated at any moment. The Vidu team says S2-Avatar runs at 25 to 42 frames per second, allowing a character to respond during a conversation rather than requiring users to generate a finished sequence in advance.

Dynamic references make that stream adjustable while generation is underway. A user can introduce a new image to change the character’s clothing, direct an interaction with an object, or move the character into another scene mid-conversation. This changes the reference material from a fixed starting condition into an ongoing control mechanism.

S2-Editing applies a related approach to an existing video feed. It can render the stream in another visual style, replace clothing, swap the person or character, or change the background in real time. According to the team, these alterations retain the source video’s actions and temporal rhythm, which could let creators transform live performances without rebuilding the underlying motion.

The developers have also extended both character generation and live editing to stereoscopic video, which presents separate views for each eye to create spatial depth. That work is intended for immersive virtual-reality experiences, including generated characters and edited live video.

The available performance and quality evidence comes from the Vidu team. It reports leading results against all baselines across five public benchmarks covering avatar generation and video editing, and has released an online demo and an API. However, the source does not identify the hardware used, explain how latency was measured, disclose pricing or an open license, or provide independent tests of the benchmark claims and stability in real-world use. The demo therefore establishes that the system can be tried, not that it is ready for production at scale.