Researchers develop new ways for AI models to work together
The development, by researchers at the Weizmann Institute of Science (WIS) makes it possible to combine the strengths of different AI systems, speeding up performance and reducing costs, Xinhua news agency reported.
The new method significantly improves the speed of large language models, or LLMs, which power tools like ChatGPT and Gemini.
On average, it increases performance by 1.5 times, and in some cases by as much as 2.8 times, the team said, adding that it could make AI more suitable for smartphones, drones, and autonomous vehicles.
In those settings, faster response times can be critical to safety and accuracy. For example, in a self-driving car, a faster AI model can mean the difference between a safe decision and a dangerous error. Until now, AI models developed by different companies could not easily communicate or collaborate because each uses a different internal 'language,' made up of unique tokens.
The researchers compared this to people from different countries trying to talk without a shared vocabulary.
To overcome this, the team developed two algorithms. One allows a model to translate its output into a shared format that other models can understand. The other encourages collaboration using tokens that have the same meaning across different systems, like common words in human languages.
Though initially concerned that meaning might be lost in translation, the researchers found that their system worked efficiently.
The new tools are already available through open-source platforms and are helping developers worldwide create faster and more collaborative AI applications.
The finding was presented at the International Conference on Machine Learning being held in Vancouver, Canada.
Hashtags
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
Hindustan Times
30 minutes ago
- Hindustan Times
AI Search Is Growing More Quickly Than Expected
Chatbots are becoming the go-to source for online answers for many consumers, chipping away at the dominance of traditional web search and adding another avenue of outreach that brands must cultivate to connect with customers. An estimated 5.6% of U.S. search traffic on desktop browsers last month went to an AI-powered large language model like ChatGPT or Perplexity, according to Datos, a market intelligence firm that tracks web users' behavior. That pales beside the 94.4% that still went to traditional search engines like Alphabet's Google or Microsoft's Bing, which have tried to fight off the new competition by adding artificial intelligence summaries to the top of their search results. But the percentage of traffic that went to browser-based AI search has more than doubled since June 2024, when it was 2.48%, according to Datos, which is part of marketing software company Semrush. It has more than quadrupled since January 2024, when the figure was just under 1.3%. Datos says it draws its data from more than 10 million panelists worldwide who agree to have their behavior on desktop browsers observed anonymously in exchange for rewards such as free access to software products from the company's partners. The numbers exclude activity on mobile browsers and apps including those from OpenAI's ChatGPT and Google. The rapid growth in AI searches could mark a sea change in online behavior comparable to the emergence of Google's web browser and the first social-media platforms, according to Eli Goodman, chief executive and co-founder of Datos. The numbers are more striking among so-called early adopters, or consumers who had already started using LLMs in desktop browsers when Datos began tracking their behavior in April 2024. Forty percent of desktop browser visits among these early adopters went to LLMs, up from just over 24% in June 2024, Datos research found. Traditional search engines' share of the desktop browser traffic from these early adopters fell significantly in the same period, to 61% in June from around 76% a year earlier. The amount of time that consumers worldwide spent on traditional search apps and websites declined 3% from April 2024 to April 2025, according to a report released last month by market research firm Sensor Tower. The drop was twice as high among early adopters, defined as people who first downloaded ChatGPT in 2023, the report said. Use of Google's traditional search product is still growing, and the AI overviews that now appear on Google searches are leading to more queries that connect consumers to businesses, a Google spokeswoman said. The growth in AI searches, plus the appearance of AI overviews in traditional search browsers, is likely to further depress search traffic to the websites where brands have long directed energy, resources and search advertising, according to Neil Vogel, CEO of Dotdash Meredith, the publisher of magazine and digital brands including People and Better Homes & Gardens. Dotdash Meredith Chief Executive Neil Vogel. Organic search traffic to large news sites such as Business Insider, the Washington Post and HuffPost has fallen by large margins over the past three years, a drop publishers blame on AI responses that fulfill searchers' curiosity without requiring a click. Marketers in response are racing to make sure their names show up in AI searches. For brands, the most important difference between LLMs and browsers is that LLMs reveal one answer rather than a list of links, giving businesses fewer openings to appear before consumers, said Vogel. 'All the brands are terrified they're not going to be included in these AI answers,' he said. The trend has inspired a new wave of AI-optimization startups that promise to help marketers adapt. Businesses should move cautiously despite the apparent speed of change, according to Goodman, the Datos CEO. Search engines still handle an overwhelming majority of overall search traffic and remain firmly embedded in the smartphones that occupy so much of our time, he said. Brands must also remember that AI searches often serve different needs than a traditional google search, he said. 'Over 90% of all of the AI searches are what we call informational or productivity-based: Help me solve this problem, help me answer this question,' said Goodman. Traditional search result pages, on the other hand, were designed to lead consumers out to other destinations online, he said. AI chatbots for now are less a replacement for traditional search engines than another emerging responsibility for marketers, according to Andrew Lipsman, founder of consulting firm Media, Ads + Commerce. Lipsman compared the emergence of AI to the rise of smartphones, which led to predictions that consumers would abandon computers altogether. Desktop traffic numbers over the past decade have barely budged, however, even as mobile use grew, according to research from Comscore. The spread of advertising in AI search replies may become the next turning point because it will demand new spending from brands that want to stay competitive online, Lipsman said. OpenAI says it has no plans to develop ad products. Perplexity has begun experimenting with sponsored searches, however, and last year hired a head of advertising and shopping. Perplexity this year also released its own web browser, where searches are handled by its AI instead of a traditional engine like Google or Bing. The company says its search gives users 'the choice to navigate the web.' OpenAI is close to releasing a browser as well, Reuters reported this month. The company declined to comment. News Corp, owner of The Wall Street Journal, has a content-licensing partnership with OpenAI and a commercial agreement to supply content on Google platforms. Write to Patrick Coffee at
Time of India
an hour ago
- Time of India
Meta, Microsoft hire top AI talent from DeepMind
Sonal Gupta, a former engineering lead at Google DeepMind, who has joined Suleyman's team as a technical staff member. Adam Sadovsky, previously a software engineer and senior director at DeepMind, with nearly 18 years at Google, who now serves as a corporate VP at Microsoft AI. Jonas Rothfuss, who spent a year at Google DeepMind as a research scientist. He joined Microsoft AI's technical staff in May. Meta has again stepped up efforts to hire top AI experts, reportedly bringing in three more researchers from Google to The Information, Tianhe Yu, Cosmo Du, and Weiyue Wang are the latest additions to Meta's growing AI trio previously worked on a variant of Google's Gemini model, which demonstrated gold-medal-level problem-solving capabilities at this year's International Math new hires bring the total number of former Google employees now working at Meta to at least six, highlighting the company's aggressive recruitment strategy in the AI Microsoft has also made significant talent acquisitions from Google DeepMind. According to CNBC, around two dozen former DeepMind employees have joined Microsoft in recent months, expanding its AI division, which is led by DeepMind cofounder Mustafa Suleyman, who now reports directly to CEO Satya latest hires include Amar Subramanya, a Google veteran of 16 years and former vice president (VP) of engineering, who worked on the Gemini assistant. On Tuesday, he announced on LinkedIn that he has joined Microsoft AI as a corporate who recently joined Microsoft's AI group include:The aggressive hiring among tech giants underscores the intense competition to secure top AI experts. OpenAI CEO Sam Altman recently claimed that Meta had offered signing bonuses of up to $100 million to lure away talent from his company. However, Meta CEO Mark Zuckerberg stated that the reported figures are 'inaccurate' Meta's hiring spree has also included high-profile names such as Scale AI CEO Alexandr Wang and former GitHub CEO Nat Friedman, alongside numerous OpenAI the other hand, Google hired the CEO and team of AI coding startup Windsurf in a deal reportedly worth $2.4 billion.
Time of India
3 hours ago
- Time of India
OpenAI, Oracle deepen AI data center push with 4.5 gigawatt Stargate expansion
OpenAI and Oracle will develop another 4.5 gigawatts of data center capacity, expanding a tie-up that has promised hundreds of billions of dollars in infrastructure investment to keep the U.S. ahead in the global artificial intelligence race. The ChatGPT maker did not disclose the locations or funding details for the new facilities in Tuesday's announcement. The move builds on the Stargate initiative , an up to $500 billion and 10 gigawatt project that also includes Japanese technology investor SoftBank Group and is setting up its first AI data center in Abilene, Texas. OpenAI, as well as its backer Microsoft , are among the technology companies pouring billions of dollars on data centers to power generative AI services such as ChatGPT and Copilot that require huge amounts of computing power. The growing use of AI in sensitive sectors such as defense, as well as China's push to catch up, has made the nascent technology a top priority for U.S. President Donald Trump, who unveiled Stargate at the White House in January. The new data centers will bring Stargate's total capacity under development to more than 5 gigawatts, which will run on over 2 million chips, OpenAI said in a blog post, adding that the tie-up now expects to exceed its initial commitment. Oracle did not immediately respond to a Reuters request for comment, while the White House declined to comment. Analysts have raised doubts about the venture's ability to secure the funding, including $100 billion for immediate deployment. In January, xAI owner Elon Musk, dismissed the group, saying "they don't actually have the money." OpenAI and SoftBank will each commit $19 billion to fund Stargate, reports said in January. The Wall Street Journal reported on Monday the two companies have been at odds with each other and Stargate is now setting a more modest goal of building a small data center at end-2025, likely in Ohio.
