Technology
Wednesday, June 3, 2026
Innovation · Trends · Analysis
In this section
Technology
Researchers discover why artificial intelligence often invents facts
A new research paper from OpenAI reveals why “hallucinations” occur in LLMs. The authors argue that, in fact, they are a natural consequence of the way these systems are trained and evaluated.
Even when the training data is perfectly accurate, the objectives and measurement methods in artificial intelligence development push the model to make possible but incorrect assumptions, rather than acknowledge uncertainty.
The researchers frame hallucination as a common classification problem. Generative errors in LLMs, they explain, are very similar to classification errors in supervised learning. When the system is forced to always give an answer, it is structurally incentivized to produce confident statements even when they are incorrect.
Assessment practices further compound this problem. Most tests are based on binary accuracy metrics, where an answer is either correct or incorrect, with no reward for a “I don’t know.”
In this setup, hesitation is not rewarded, while guessing sometimes earns points. The authors point out that this system resembles the logic of multiple-choice exams: when blank answers earn zero points, rational students choose to guess. Formal analysis shows that, under these conditions, the expected outcome of a guess is always higher than that of hesitation.
To illustrate the mechanism of this stimulation, the team proposed a simplified “Arbitrary Facts” model. In this case, each question had a correct answer chosen at random, so there was no model to train.
The researchers showed that the number of “single facts” (which appeared only once in training) sets a lower limit on the rate of hallucinations. When facts are rare, the model faces high uncertainty and is practically forced to guess.
The study also highlights several other factors that contribute to hallucinations. A model with insufficient capacity can fail to understand patterns even when they exist, producing errors.
The difference in distribution between training and testing questions often leads to failures. Some questions are inherently difficult, and incorrect or noisy data directly produce incorrect results.
The authors' main recommendation is to review how artificial intelligence performance is measured.
Instead of penalizing hesitation, evaluation frameworks should reward models when they do not provide answers if they are not sure, and impose stronger penalties for incorrect answers given with high confidence. One proposal includes explicit confidence thresholds, where models should respond only when confidence exceeds a certain level, with graduated penalties for errors. Such a system would encourage models to say “I don’t know” more often when necessary. The study’s conclusions are that hallucinations are not unexplained anomalies, but predictable outcomes of the current training and testing environment.
Even when the training data is perfectly accurate, the objectives and measurement methods in artificial intelligence development push the model to make possible but incorrect assumptions, rather than acknowledge uncertainty.
The researchers frame hallucination as a common classification problem. Generative errors in LLMs, they explain, are very similar to classification errors in supervised learning. When the system is forced to always give an answer, it is structurally incentivized to produce confident statements even when they are incorrect.
Assessment practices further compound this problem. Most tests are based on binary accuracy metrics, where an answer is either correct or incorrect, with no reward for a “I don’t know.”
In this setup, hesitation is not rewarded, while guessing sometimes earns points. The authors point out that this system resembles the logic of multiple-choice exams: when blank answers earn zero points, rational students choose to guess. Formal analysis shows that, under these conditions, the expected outcome of a guess is always higher than that of hesitation.
To illustrate the mechanism of this stimulation, the team proposed a simplified “Arbitrary Facts” model. In this case, each question had a correct answer chosen at random, so there was no model to train.
The researchers showed that the number of “single facts” (which appeared only once in training) sets a lower limit on the rate of hallucinations. When facts are rare, the model faces high uncertainty and is practically forced to guess.
The study also highlights several other factors that contribute to hallucinations. A model with insufficient capacity can fail to understand patterns even when they exist, producing errors.
The difference in distribution between training and testing questions often leads to failures. Some questions are inherently difficult, and incorrect or noisy data directly produce incorrect results.
The authors' main recommendation is to review how artificial intelligence performance is measured.
Instead of penalizing hesitation, evaluation frameworks should reward models when they do not provide answers if they are not sure, and impose stronger penalties for incorrect answers given with high confidence. One proposal includes explicit confidence thresholds, where models should respond only when confidence exceeds a certain level, with graduated penalties for errors. Such a system would encourage models to say “I don’t know” more often when necessary. The study’s conclusions are that hallucinations are not unexplained anomalies, but predictable outcomes of the current training and testing environment.
Technology
GPT-5 Launched: The Model That Brings the World Closer to Artificial General Intelligence
OpenAI has officially launched GPT-5, the latest generation of its artificial intelligence model, making it available to all ChatGPT users. The new version offers improved skills in writing, programming, math, health, and finance, with expert-level performance and advanced multimodal processing across text, images, audio, and video.
The company's CEO, Sam Altman, described GPT-5 as "a conversation with a PhD-level expert." The model includes a new deep reasoning system and has significantly reduced errors known as "hallucinations" compared to previous versions.
Users can choose different personalities to interact with and, through advanced subscriptions, integrate GPT-5 with services like Gmail and Google Calendar. Free subscribers have limited access, while Pro users get unlimited use and the GPT-5 Pro version for complex tasks.
The launch of GPT-5 is seen as a major step towards general artificial intelligence, with rapid integrations into Microsoft platforms and other business tools.
The company's CEO, Sam Altman, described GPT-5 as "a conversation with a PhD-level expert." The model includes a new deep reasoning system and has significantly reduced errors known as "hallucinations" compared to previous versions.
Users can choose different personalities to interact with and, through advanced subscriptions, integrate GPT-5 with services like Gmail and Google Calendar. Free subscribers have limited access, while Pro users get unlimited use and the GPT-5 Pro version for complex tasks.
The launch of GPT-5 is seen as a major step towards general artificial intelligence, with rapid integrations into Microsoft platforms and other business tools.
Technology
OpenAI and NVIDIA break record: 1.5 million tokens per second with open-licensed AI models
OpenAI and NVIDIA have joined forces to release two powerful artificial intelligence models with the GPT‑OSS‑120B and GPT‑OSS‑20B, marking a key moment in democratizing access to advanced AI. These models are designed to support a wide range of applications in content generation, reasoning, healthcare, industrial manufacturing, and more. Distributed under the Apache 2.0 license, the models are free for commercial and research use, providing developers, enterprises, governments, and startups with the tools to build transformative AI-based solutions.
The GPT‑OSS models are trained on NVIDIA H100 GPUs and optimized to run on NVIDIA’s global CUDA platform, which powers hundreds of millions of GPUs in the cloud, personal computers, and workstations worldwide. This strategic alignment ensures that developers around the globe can integrate these models into their existing infrastructure.
At the heart of this advancement is NVIDIA’s Blackwell architecture, purpose-built for high-throughput AI inference. The GB200 NVL72 mainframe achieves an unprecedented 1.5 million tokens per second when running the GPT‑OSS‑120B model, making it one of the most powerful inference platforms in the world. Blackwell brings innovations like NVFP44-bit precision, which enables extremely efficient execution with high precision, while significantly reducing power consumption and memory requirements, a giant step towards real-time use of models with trillions of parameters.
This collaboration also reflects a long-standing partnership between OpenAI and NVIDIA, dating back to 2016, when NVIDIA founder Jensen Huang personally delivered the first DGX-1 supercomputer to OpenAI’s headquarters in San Francisco. Since then, the two companies have collaborated on some of the world’s most ambitious AI training. Today’s release builds on this legacy, bringing cutting-edge AI capabilities to millions of developers globally, supported by an ecosystem of over 6.5 million developers in more than 250 countries.
Fundamentally, the release of GPT‑OSS represents a significant step towards making advanced AI more transparent, efficient, and accessible to everyone. With scalable infrastructure, open licensing, and broad hardware support, OpenAI and NVIDIA are not only accelerating the pace of innovation, but also building the foundation for the next industrial revolution driven by open, accountable, and high-performance AI.
The GPT‑OSS models are trained on NVIDIA H100 GPUs and optimized to run on NVIDIA’s global CUDA platform, which powers hundreds of millions of GPUs in the cloud, personal computers, and workstations worldwide. This strategic alignment ensures that developers around the globe can integrate these models into their existing infrastructure.
At the heart of this advancement is NVIDIA’s Blackwell architecture, purpose-built for high-throughput AI inference. The GB200 NVL72 mainframe achieves an unprecedented 1.5 million tokens per second when running the GPT‑OSS‑120B model, making it one of the most powerful inference platforms in the world. Blackwell brings innovations like NVFP44-bit precision, which enables extremely efficient execution with high precision, while significantly reducing power consumption and memory requirements, a giant step towards real-time use of models with trillions of parameters.
This collaboration also reflects a long-standing partnership between OpenAI and NVIDIA, dating back to 2016, when NVIDIA founder Jensen Huang personally delivered the first DGX-1 supercomputer to OpenAI’s headquarters in San Francisco. Since then, the two companies have collaborated on some of the world’s most ambitious AI training. Today’s release builds on this legacy, bringing cutting-edge AI capabilities to millions of developers globally, supported by an ecosystem of over 6.5 million developers in more than 250 countries.
Fundamentally, the release of GPT‑OSS represents a significant step towards making advanced AI more transparent, efficient, and accessible to everyone. With scalable infrastructure, open licensing, and broad hardware support, OpenAI and NVIDIA are not only accelerating the pace of innovation, but also building the foundation for the next industrial revolution driven by open, accountable, and high-performance AI.
More headlines
0607
Meta is transforming the advertising industry with an ambitious new approach, using generative artificial intelligence to automate the creation and personalization of advertising content. The company's latest tools allow advertisers to produce headlines, product descriptions, images and even videos, simply by providing a few instructions such as the type of product, target audience and tone of message. The AI systems will then create different versions of an ad, optimized for different displays.
One of the most impressive plans is AI's ability to personalize ad visuals based on a user's location or context in real time. For example, a car manufacturer running a single campaign could show different versions of an ad to different users: someone in a mountainous area might see the car traveling along a rugged mountain road, while a user in the city might see the same model moving in an urban environment, reports the WSJ. This level of environmental adaptation aims to make the advertising experience more natural, increasing engagement and conversion rates.
While the potential for personalized, high-performing advertising is huge, there are concerns about AI-generated content operating without sufficient oversight. However, this move represents a major shift toward automation that could reshape the role of creativity in digital marketing.080910
Apple Pay now available in Kosovo
The Central Bank of the Republic of Kosovo (CBK) has officially announced that the advanced contactless payment service, Apple Pay, is now available in Kosovo, marking a key moment in the process of digitalizing payments and integrating the Kosovo financial market into global standards.
Through this development, citizens and businesses in Kosovo will be able to make secure, fast and modern payments through Apple devices such as iPhone and Apple Watch. Banks and financial institutions in the country will be able to link their cards to the Apple Pay platform, depending on individual agreements and relevant technical integrations.
The CBK considers this an important step towards modernizing the payment system, increasing financial inclusion and fulfilling the objectives of the Strategic Plan 2024–2028. According to the institution, after the integration of Google Pay, the addition of Apple Pay reinforces the approach towards the internationalization of financial services and approximation with European standards for digital payments.
In this regard, the CBK also commemorates the meeting of May 2025 with senior representatives of the Apple company, where the commitment to creating favorable conditions for modern financial services in Kosovo was reconfirmed.
This success is the result of close cooperation between the CBK, the banking sector and international partners, with particular emphasis on the Embassy of the United States of America.
Through this development, citizens and businesses in Kosovo will be able to make secure, fast and modern payments through Apple devices such as iPhone and Apple Watch. Banks and financial institutions in the country will be able to link their cards to the Apple Pay platform, depending on individual agreements and relevant technical integrations.
The CBK considers this an important step towards modernizing the payment system, increasing financial inclusion and fulfilling the objectives of the Strategic Plan 2024–2028. According to the institution, after the integration of Google Pay, the addition of Apple Pay reinforces the approach towards the internationalization of financial services and approximation with European standards for digital payments.
In this regard, the CBK also commemorates the meeting of May 2025 with senior representatives of the Apple company, where the commitment to creating favorable conditions for modern financial services in Kosovo was reconfirmed.
This success is the result of close cooperation between the CBK, the banking sector and international partners, with particular emphasis on the Embassy of the United States of America.
Meta plans to launch visual ad personalization based on user location
Meta is transforming the advertising industry with an ambitious new approach, using generative artificial intelligence to automate the creation and personalization of advertising content. The company's latest tools allow advertisers to produce headlines, product descriptions, images and even videos, simply by providing a few instructions such as the type of product, target audience and tone of message. The AI systems will then create different versions of an ad, optimized for different displays.
One of the most impressive plans is AI's ability to personalize ad visuals based on a user's location or context in real time. For example, a car manufacturer running a single campaign could show different versions of an ad to different users: someone in a mountainous area might see the car traveling along a rugged mountain road, while a user in the city might see the same model moving in an urban environment, reports the WSJ. This level of environmental adaptation aims to make the advertising experience more natural, increasing engagement and conversion rates.
While the potential for personalized, high-performing advertising is huge, there are concerns about AI-generated content operating without sufficient oversight. However, this move represents a major shift toward automation that could reshape the role of creativity in digital marketing.
Bitcoin climbs to new historical record
Bitcoin has crossed another historical threshold, reaching a value of over $111,738 on May 22, 2025. The growth was driven by a strong wave of institutional investment in Bitcoin ETFs. In addition, recent developments in the regulation of the crypto market in the US, legislation for stablecoins, have created a sense of legal certainty that is attracting capital to this market.
At the same time, geopolitical developments have positively affected the global investment climate. A temporary de-escalation in trade tensions between the US and China, materialized with the temporary suspension of tariffs, has reduced economic uncertainties and calmed markets. This situation, combined, has pushed investors towards alternative assets such as Bitcoin, especially in the face of fiscal uncertainties and inflation.
At the same time, geopolitical developments have positively affected the global investment climate. A temporary de-escalation in trade tensions between the US and China, materialized with the temporary suspension of tariffs, has reduced economic uncertainties and calmed markets. This situation, combined, has pushed investors towards alternative assets such as Bitcoin, especially in the face of fiscal uncertainties and inflation.
Grand Theft Auto VI delayed until May 2026
Rockstar Games has officially announced that Grand Theft Auto VI (GTA 6) will be released on May 26, 2026, pushing the release date back from its previous planned fall 2025 release date.
The company stressed that it needed more development time to ensure the game meets the high standards of quality expected by fans. In a statement, Rockstar expressed gratitude for the community's support and patience, emphasizing its commitment to delivering a product that exceeds expectations.
The delay marks a 13-year gap since the release of Grand Theft Auto V in 2013, which has sold over 210 million copies and is one of the best-selling games of all time. The next installment in the saga will be set in the fictional city of Vice City and the country of Leonidas, following the story of Lucia and her male partner in an ambitious heist.
Fans have expressed disappointment over the extended delay, but continue to support Rockstar's commitment to quality. Expectations for GTA 6 remain high, with hopes for more updates or trailers during the extended development period.
The company stressed that it needed more development time to ensure the game meets the high standards of quality expected by fans. In a statement, Rockstar expressed gratitude for the community's support and patience, emphasizing its commitment to delivering a product that exceeds expectations.
The delay marks a 13-year gap since the release of Grand Theft Auto V in 2013, which has sold over 210 million copies and is one of the best-selling games of all time. The next installment in the saga will be set in the fictional city of Vice City and the country of Leonidas, following the story of Lucia and her male partner in an ambitious heist.
Fans have expressed disappointment over the extended delay, but continue to support Rockstar's commitment to quality. Expectations for GTA 6 remain high, with hopes for more updates or trailers during the extended development period.
Google brings Gemini AI to kids under 13
Google is set to launch a customized version of its artificial intelligence chatbot, Gemini, for users under the age of 13. The service will be accessible through accounts supervised by parents through the Family Link app, giving them the ability to control access and designate Gemini as their primary assistant on Android, iOS, or the web. Gemini Kids will help kids with homework, creative writing, and general questions in a safe and controlled environment.
To promote the responsible use of artificial intelligence, Google has launched several educational programs. One of the initiatives is the “Generative AI for Educators” course, developed in collaboration with MIT RAISE, which offers teachers hands-on experience with AI tools to improve teaching. Google also introduced NotebookLM, a search assistant that helps students better organize and understand information, as well as AI features on Chromebooks to help with writing and teaching.
These efforts reflect Google's commitment to making AI a useful tool in children's education, while ensuring safety and responsible use. By integrating AI into school environments and providing resources for teachers and students, the company aims to personalize the learning experience and improve the quality of education.
To promote the responsible use of artificial intelligence, Google has launched several educational programs. One of the initiatives is the “Generative AI for Educators” course, developed in collaboration with MIT RAISE, which offers teachers hands-on experience with AI tools to improve teaching. Google also introduced NotebookLM, a search assistant that helps students better organize and understand information, as well as AI features on Chromebooks to help with writing and teaching.
These efforts reflect Google's commitment to making AI a useful tool in children's education, while ensuring safety and responsible use. By integrating AI into school environments and providing resources for teachers and students, the company aims to personalize the learning experience and improve the quality of education.