Nvidia reportedly acquires synthetic data startup Gretel

 Nvidia Acquires Gretel: A Strategic Move to Revolutionize Synthetic AI Training Data

In a bold move that underscores its determination to maintain a competitive edge in the burgeoning artificial intelligence (AI) market, Nvidia has reportedly acquired San Diego–based startup Gretel. This acquisition is set to redefine the landscape of AI training data by harnessing synthetic data—a frontier that is becoming increasingly crucial as tech giants race to build more sophisticated and capable models.

The Acquisition in Detail

According to reports, the acquisition of Gretel, a company known for its cutting-edge platform that generates synthetic AI training data, is valued in the nine-figure range. This figure notably exceeds Gretel’s last reported valuation of $320 million, as highlighted in a recent Wired article. While the specific terms of the deal remain undisclosed, the size of the transaction speaks volumes about the strategic importance Nvidia places on synthetic data in the next wave of AI development.

Gretel’s integration into Nvidia is anticipated to bolster Nvidia’s suite of generative AI services, providing developers with enhanced tools to overcome one of the most pressing challenges in AI development: the scarcity and limitations of real-world data. With Gretel’s technology, Nvidia is positioning itself to not only expand its current capabilities but also to pioneer new approaches in AI model training that are more robust, scalable, and efficient.

The Importance of Synthetic Data in AI

As the race for creating the next generation of AI models intensifies, the limitations of real-world data have become increasingly apparent. Real data can often be incomplete, biased, or simply insufficient to cover the diverse scenarios that AI models might encounter. Synthetic data offers a compelling alternative by allowing developers to generate vast quantities of highly customized data that simulate real-world conditions without the associated drawbacks. This can lead to improvements in model accuracy, robustness, and overall performance.

Nvidia’s acquisition of Gretel is emblematic of a broader industry trend where leading tech companies like Microsoft, Meta, OpenAI, and Anthropic are investing heavily in synthetic data solutions. These companies have already begun integrating synthetic data into their training regimens for flagship AI models, ensuring that as traditional sources of data become saturated or limited, their models can continue to learn and adapt effectively.

Gretel’s Journey and Innovations

Founded in 2019 by Alex Watson, Laszlo Bock, John Myers, and Ali Golshan—who also serves as the company’s CEO—Gretel quickly emerged as a significant player in the AI startup ecosystem. The company’s mission was to revolutionize how training data is generated and utilized. By fine-tuning existing models and layering proprietary technology on top, Gretel was able to create bespoke synthetic data solutions that could be packaged and sold to a wide range of industries.

The company’s platform was designed with versatility in mind. It provided users with the ability to generate tailored datasets that met the specific requirements of various AI applications, from natural language processing to computer vision. This capability is particularly valuable in industries where obtaining high-quality, annotated data is both expensive and time-consuming.

Gretel’s innovative approach and technology resonated with investors. Prior to its acquisition, the startup had raised more than $67 million in venture capital from a robust cohort of investors including Anthos Capital, Greylock, and Moonshots Capital. This significant capital infusion underscored the market’s confidence in Gretel’s potential to disrupt traditional data generation methods.

Nvidia’s Strategic Vision

Nvidia is no stranger to the AI arena. With its industry-leading graphics processing units (GPUs) and robust AI software ecosystem, the company has long been at the forefront of AI innovation. The integration of Gretel’s technology into Nvidia’s portfolio is a strategic maneuver aimed at reinforcing its position as a one-stop solution for developers seeking advanced AI capabilities.

The acquisition is particularly timely given the rapid evolution of AI technology. As AI models grow in complexity and require ever-larger datasets for training, the ability to generate synthetic data at scale becomes an invaluable asset. Nvidia’s existing suite of generative AI services will be significantly enhanced by integrating Gretel’s platform, allowing developers to access a more comprehensive set of tools and services designed to accelerate innovation in AI.

Moreover, Nvidia’s move reflects a broader strategic shift within the tech industry. With an increasing number of companies looking to leverage AI to transform their operations, the demand for high-quality training data is surging. Nvidia’s acquisition of Gretel not only addresses this demand but also positions the company as a critical partner in the AI development lifecycle, from data generation to model deployment.

The Broader Impact on the AI Ecosystem

The implications of Nvidia’s acquisition extend beyond the immediate benefits to the company’s AI service portfolio. In an era where the quality and diversity of training data are paramount, synthetic data is emerging as a key enabler of progress in AI research and application. By harnessing synthetic data, developers can overcome the limitations posed by traditional datasets, enabling the creation of AI models that are more generalizable, reliable, and ethical.

For instance, synthetic data can help mitigate issues related to bias and privacy. Since synthetic data is generated by algorithms, it can be crafted to reflect a more balanced and representative sample of real-world scenarios. This not only improves the fairness of AI models but also reduces the risks associated with using sensitive personal data. Nvidia’s enhanced capabilities in synthetic data generation could thus have far-reaching implications for the development of ethical AI systems.

Furthermore, the acquisition is likely to stimulate further innovation in the synthetic data space. With a major player like Nvidia integrating Gretel’s technology, we can expect to see increased investment and interest in this area. This could lead to new breakthroughs in how synthetic data is generated, validated, and applied across various industries, from healthcare and finance to automotive and retail.

Financial and Market Considerations

From a financial perspective, the acquisition represents a significant investment in the future of AI. With a nine-figure price tag, Nvidia is making a substantial bet on the value of synthetic data in driving the next wave of AI innovation. This investment is expected to yield long-term benefits by enhancing Nvidia’s competitive edge and expanding its market share in the rapidly growing AI sector.

The deal also highlights the dynamic nature of startup valuations in the tech industry. Gretel’s most recent valuation of $320 million, which has now been surpassed by the acquisition price, reflects the high stakes and rapid pace of innovation in the AI domain. Investors and market analysts will be watching closely to see how this acquisition influences future investment trends and the overall valuation landscape for AI startups.

Future Prospects and Industry Reactions

Industry experts have lauded Nvidia’s acquisition of Gretel as a forward-thinking move that positions the company at the forefront of AI innovation. By acquiring a company with proven capabilities in synthetic data generation, Nvidia is not only enhancing its current offerings but also laying the groundwork for future advancements in AI model training and deployment.

Developers and businesses alike stand to benefit from this integration. With more sophisticated and accessible tools at their disposal, companies can accelerate their AI initiatives and achieve more accurate and reliable results. This, in turn, could drive broader adoption of AI technologies across various sectors, fueling further growth and innovation.

Critics, however, caution that the success of the integration will depend on how seamlessly Gretel’s technology can be incorporated into Nvidia’s existing framework. There are concerns about potential integration challenges, including aligning corporate cultures and ensuring that the combined technologies deliver the promised enhancements in AI training capabilities. Nonetheless, Nvidia’s track record in acquisitions and integrations suggests that the company is well-equipped to manage these challenges and realize the full potential of the deal.

Strategic Implications for Nvidia

The acquisition of Gretel is not an isolated event but rather part of a larger strategic vision at Nvidia. As the company continues to expand its footprint in AI and machine learning, it is increasingly focused on developing end-to-end solutions that address every stage of the AI development process. From powerful GPUs and specialized hardware to advanced software platforms and data generation tools, Nvidia is building a comprehensive ecosystem designed to support the next generation of AI applications.

This holistic approach is particularly important in today’s competitive landscape, where companies are not only racing to develop more powerful AI models but also to create platforms that can support these models at scale. Nvidia’s investment in synthetic data through Gretel is a key component of this strategy, ensuring that its developers have access to the high-quality training data needed to build and refine sophisticated AI systems.

Looking Ahead: The Future of Synthetic Data

The integration of Gretel’s technology into Nvidia’s generative AI services is likely to have a transformative impact on the way synthetic data is utilized in AI research and application. As more companies adopt synthetic data solutions, we can expect to see significant advancements in model accuracy, efficiency, and ethical considerations.

In the coming years, the demand for synthetic data is expected to grow exponentially. With real-world data sources becoming increasingly limited or fraught with challenges such as bias and privacy concerns, synthetic data offers a viable alternative that can drive innovation while mitigating risks. Nvidia’s acquisition of Gretel is a clear signal that synthetic data is poised to become a cornerstone of AI development, with far-reaching implications for industries across the board.

Moreover, this move is likely to spur further research and development in the synthetic data space. Academic institutions, research labs, and tech startups may accelerate their efforts to explore new methodologies for generating, validating, and applying synthetic data. This could lead to breakthroughs that not only enhance AI training processes but also open up new avenues for AI applications in previously untapped domains.

Conclusion

Nvidia’s acquisition of Gretel marks a significant milestone in the evolution of AI technology. By integrating Gretel’s innovative synthetic data platform, Nvidia is reinforcing its position as a leader in the AI industry and paving the way for the next generation of AI model development. The move is emblematic of a broader trend in the tech world, where synthetic data is increasingly recognized as a critical resource for overcoming the limitations of traditional datasets.

As the integration process unfolds, all eyes will be on Nvidia to see how effectively it can harness Gretel’s capabilities to drive forward its ambitious AI agenda. With a clear strategic vision and a robust track record of innovation, Nvidia is well-positioned to lead the charge in this rapidly evolving field—transforming challenges into opportunities and setting new benchmarks for excellence in AI.

In summary, Nvidia’s acquisition of Gretel is more than just a financial transaction; it is a strategic investment in the future of AI. As synthetic data continues to play an integral role in training increasingly sophisticated AI models, Nvidia’s enhanced capabilities in this area will be crucial for maintaining a competitive edge in a market defined by rapid technological evolution and relentless innovation. The coming months and years will be critical as the integration matures, potentially reshaping how AI systems are developed, trained, and deployed on a global scale.

EU AI Act: Latest draft Code for AI model makers tiptoes towards gentler guidance for Big AI

 

Ahead of the May deadline to finalize guidance for providers of general-purpose AI (GPAI) models under the EU AI Act, a third draft of the Code of Practice was published on Tuesday. This latest draft is expected to be the final version after a year of development.

A dedicated website has also been launched to improve the Code's accessibility. Stakeholders are encouraged to submit written feedback on the latest draft by March 30, 2025.

The EU AI Act establishes a risk-based framework with specific obligations for the most powerful AI model providers, focusing on areas like transparency, copyright, and risk mitigation. The Code of Practice aims to help GPAI model developers understand and comply with these legal requirements, mitigating the risk of penalties for noncompliance. Violations of GPAI-related obligations under the AI Act could result in fines of up to 3% of global annual revenue.

Streamlined Structure The latest draft of the Code features a "more streamlined structure with refined commitments and measures" in response to feedback on the second draft released in December. Further input from working groups and workshops will contribute to the development of the final guidance, with experts seeking to enhance "clarity and coherence" in the adopted version.

The document is organized into several sections, including commitments for GPAIs, detailed guidance on transparency and copyright, and safety and security measures for the most powerful models with systemic risk (GPAISR).

The transparency section outlines a model documentation form that GPAI developers may need to complete. This form aims to ensure downstream users of GPAI technology have access to key information for their own compliance. The copyright section remains a point of contention, particularly for large AI companies.

The draft uses language such as "best efforts," "reasonable measures," and "appropriate measures" when addressing commitments like respecting rights during data collection for training or mitigating the risk of producing copyright-infringing outputs. This flexible phrasing could allow AI companies significant leeway in their compliance approaches.

Notably, language from previous drafts requiring GPAIs to provide a "single point of contact" for rightsholders to file complaints has been softened. The current text only states that "signatories will designate a point of contact" and provide accessible information about it. Furthermore, GPAI providers may refuse to act on copyright complaints deemed "manifestly unfounded or excessive," raising concerns that automated detection efforts by rightsholders could be ignored.

Regarding safety and security, the EU AI Act already limits systemic risk mitigation requirements to the most powerful models (those trained with over 10^25 FLOPs). This draft further narrows previously recommended measures in response to feedback.

US Pressure The latest draft does not mention mounting pressure from the United States to relax AI regulations. At the Paris AI Action summit, U.S. Vice President JD Vance criticized European AI regulations, arguing that overregulation could stifle innovation. This aligns with broader efforts by the Trump administration to promote an "AI opportunity" agenda over strict regulation.

In response to such pressure, the EU has already scrapped the AI Liability Directive and announced an "omnibus" package of regulatory reforms aimed at reducing bureaucracy. However, with the AI Act still being implemented, further changes to dilute its requirements may emerge.

At the Mobile World Congress in Barcelona, French GPAI developer Mistral voiced concerns about meeting some regulatory demands. Founder Arthur Mensh indicated the company is collaborating with regulators to resolve these challenges.

While the GPAI Code is being developed by independent experts, the European Commission's AI Office is simultaneously preparing additional "clarifying" guidance. This guidance will define GPAI responsibilities and could offer a means for EU lawmakers to address U.S. lobbying efforts.

Further updates from the AI Office, which the Commission says will "clarify ... the scope of the rules," are expected in the near future.

Learn what VCs want to see from founders at TechCrunch Sessions: AI

 AI has dominated the funding landscape in recent years, with investments in the sector surging 62% to $110 billion in 2024 alone, even as overall startup funding declined by 12%.

It might be tempting for startups to slap "AI" onto their names to attract investors. However, as the initial excitement around foundation models shifts toward real-world applications, AI agents, and long-term profitability, investors are prioritizing companies that can turn technical innovation into sustained business traction.

At TechCrunch Sessions: AI on June 5, 2025, in Zellerbach Hall at UC Berkeley, top venture capitalists Zeya Yang (IVP), Jill Chase (CapitalG), and Kanu Gulati (Khosla Ventures) will share insights on what they look for at every investment stage—from seed rounds to Series C. With extensive investment histories and hands-on expertise, they’re ready to break down how to stand out in the competitive AI landscape.

Meet the Speakers

Zeya Yang has backed major successes like Grammarly and Figma, and he specializes in helping founders refine product-market fit and drive growth.

Jill Chase leads CapitalG's investments in Magic, /dev/agents, Motif, and Abridge, focusing on emerging AI and ML use cases, data infrastructure, and enterprise technology.

Kanu Gulati has invested in AI leaders like PolyAI, Kognitos, and Moonhub. With over a decade of experience as a research scientist at Intel and Cadence and as a founding engineer at Heavy.ai, Spyglass, and Nascentric, she brings deep technical and operational expertise.

Don’t Miss TechCrunch Sessions: AI

Join the conversation and get a front-row seat to the future of AI. Secure your tickets now for TechCrunch Sessions: AI on June 5, 2025, in Zellerbach Hall at UC Berkeley. Take advantage of Early Bird deals to save up to $210!

Last chance! Last 24 hours to save up to $325 on TechCrunch Sessions: AI

 Time is running out! Tonight is your last chance to register for TechCrunch Sessions: AI at our Super Early Bird rates. Sign up before 11:59 p.m. PT and save up to $325 on your pass!

Join us on June 5 at Zellerbach Hall, UC Berkeley, for a full day packed with cutting-edge insights into the world of AI. You’ll experience expert-led main stage discussions, hands-on demos in the Expo Hall, interactive breakout sessions, and unparalleled networking opportunities. Whether you're an AI veteran or just curious about the technology shaping our future, this event is for you.

Secure your Super Early Bird savings by registering before 11:59 p.m. PT tonight!

Meet the AI Trailblazers

At TC Sessions: AI, industry leaders will take the stage to share their insights. Here are just a few of the speakers you won’t want to miss:

Jae Lee, CEO, Twelve Labs
Jae Lee is transforming how developers and enterprises analyze and understand massive video datasets. As co-founder and CEO of Twelve Labs, he leads the development of advanced multimodal foundation models that push the boundaries of AI-powered video intelligence. Jae will share the stage with Sara Hooker, VP of Research at Cohere, in a session titled "How Founders Can Build on Existing Foundational Models."

Oliver Cameron, Co-founder and CEO, Odyssey
At his startup Odyssey, Oliver Cameron is pioneering "world models" that generate cinematic, interactive worlds in real-time. Previously, he was co-founder and CEO of autonomous vehicle startup Voyage. Oliver will discuss how small companies can compete with industry giants in the rapidly evolving AI landscape.

Kanu Gulati, Partner, Khosla Ventures
As a partner at Khosla Ventures, Kanu Gulati invests in transformative AI, robotics, and autonomous systems. Her portfolio includes PolyAI, Regie, and Waabi. With a decade-long career as a scientist at Intel and Cadence and multiple successful startups, Kanu brings deep expertise to the table. She will join Jill Chase, Partner at CapitalG, for an insightful talk on "From Seed to Series C: What VCs Expect from Founders."

Artemis Seaford, Head of AI Safety, ElevenLabs
Artemis Seaford leads AI safety efforts at ElevenLabs. Previously, she guided OpenAI's safe deployment strategies and managed Meta's global response to geopolitical and adversarial threats. With a PhD in political science and a JD from Stanford, Artemis brings a multidisciplinary approach to responsible AI development.

Jill Chase, Partner, CapitalG
Jill Chase heads the AI investment practice at CapitalG, Alphabet's independent growth fund. She has led investments in innovative companies like Magic, /dev/agents, and Motif. Jill and Kanu Gulati will offer a deep dive into "From Seed to Series C: What VCs Expect from Founders."

Sara Hooker, VP of Research, Cohere
Sara Hooker leads cutting-edge research at Cohere for AI, focusing on efficient, safe, and interpretable large language models. Before Cohere, she worked at Google Brain, specializing in efficient model training. TIME recognized her as one of the 100 Most Influential People in AI in 2024. Sara will team up with Jae Lee to discuss how startups can build on existing foundational models.

Share Your AI Expertise

Are you an AI expert ready to shape the future of the industry? We want to hear from you! Apply by March 7 to lead thought-provoking discussions and mentor the next generation of AI innovators.

Last Chance for Super Early Bird Pricing

Don't miss your opportunity to save up to $325. Register for TechCrunch Sessions: AI before 11:59 p.m. PT tonight to lock in these unbeatable rates!

Want to stay informed about future TechCrunch events? Subscribe to our TechCrunch Events newsletter for exclusive deals and announcements.

Become a Sponsor

Interested in sponsoring or exhibiting at TechCrunch Sessions: AI? Connect with our sponsorship sales team by filling out this form.

The TechCrunch AI glossary

Artificial intelligence (AI) is a vast and intricate field. Researchers in this domain often use specialized jargon to describe their work. As a result, we frequently incorporate these technical terms when covering developments in the AI industry. To help clarify these concepts, we have created a glossary defining key terms and phrases that commonly appear in our articles.

We will update this glossary regularly to include new entries as researchers continue to push the boundaries of AI while addressing emerging safety concerns.

AI Agent

An AI agent is a tool that utilizes AI technologies to perform complex tasks on behalf of a user. These tasks go beyond basic chatbot functions and may include filing expenses, booking reservations, or even writing and maintaining code. The concept of an AI agent implies an autonomous system capable of executing multi-step tasks by integrating multiple AI systems. However, as the field evolves, interpretations of what constitutes an AI agent may vary, and the infrastructure to support their full capabilities is still under development.

Chain of Thought

Humans can answer simple questions instinctively, but more complex problems often require step-by-step reasoning. For example, solving a puzzle about the number of chickens and cows on a farm might involve writing an equation to find the answer.

In AI, chain-of-thought reasoning involves breaking down a problem into intermediate steps to improve the accuracy of the final result. This approach typically takes longer but produces more reliable answers, especially for logic-based or coding tasks. Chain-of-thought reasoning is enhanced in specialized models through reinforcement learning.

(See: Large Language Model)

Deep Learning

Deep learning is a subset of machine learning where AI algorithms are designed using artificial neural networks (ANNs) with multiple layers. This structure enables the model to identify complex patterns in data without requiring human engineers to define these features manually. Through repetition and adjustment, deep learning systems improve their outputs over time.

Deep learning models require large datasets (millions of data points or more) and significant computational resources. Although they are more capable than simpler machine learning algorithms, they are also more expensive and time-consuming to train.

(See: Neural Network)

Fine-Tuning

Fine-tuning refers to the process of further training an AI model to optimize its performance for specific tasks. This is typically achieved by providing the model with new, specialized data relevant to the desired area of focus.

Many AI companies use fine-tuning to adapt large language models (LLMs) for particular industries or applications, enhancing their utility with domain-specific knowledge.

(See: Large Language Model)

Large Language Model (LLM)

A large language model (LLM) is a type of AI model that powers AI assistants like ChatGPT, Claude, Google’s Gemini, Meta’s AI Llama, Microsoft Copilot, and Mistral’s Le Chat. These models interpret user inputs and generate responses by predicting the most likely next word in a sequence based on extensive training on text from books, articles, and other sources.

LLMs are built on deep neural networks with billions of parameters (or weights) that capture relationships between words and phrases. This allows them to generate coherent and contextually appropriate responses across a wide range of topics.

(See: Neural Network)

Neural Network

A neural network is the foundational algorithmic structure behind deep learning and generative AI technologies like large language models. Inspired by the interconnected neurons in the human brain, these networks process data through multiple layers, enabling complex pattern recognition.

Although neural networks were conceptualized in the 1940s, advancements in graphical processing units (GPUs) have recently enabled the training of more sophisticated models. This has led to breakthroughs in areas such as voice recognition, autonomous navigation, and drug discovery.

(See: Large Language Model)

Weights

Weights are the numerical parameters that determine the importance of various features in the data used to train an AI model. These values influence how the model interprets inputs and generates outputs.

During training, an AI model starts with randomly assigned weights. As it processes more data, the model adjusts these weights to improve its accuracy. For instance, in a model predicting house prices, weights might be assigned to features like the number of bedrooms, the presence of a garage, or whether the property is detached, reflecting their impact on property value.

DeepSeek: Everything you need to know about the AI chatbot app

 

DeepSeek's Meteoric Rise

Chinese AI lab DeepSeek has surged into the global spotlight, propelled by the rapid success of its chatbot app, which recently topped the charts on both the Apple App Store and Google Play. With AI models trained using compute-efficient techniques, DeepSeek has sparked concerns among Wall Street analysts and technologists about the U.S.'s ability to maintain its lead in the AI race and the future demand for AI chips.

The Origins of DeepSeek

DeepSeek traces its roots to High-Flyer Capital Management, a Chinese quantitative hedge fund leveraging AI for trading decisions. High-Flyer was co-founded by AI enthusiast Liang Wenfeng, who started experimenting with trading during his time at Zhejiang University. In 2019, he officially launched the hedge fund with a focus on developing and implementing AI algorithms.

By 2023, High-Flyer established DeepSeek as a separate AI research lab. With initial funding from High-Flyer, DeepSeek eventually spun off into an independent company while maintaining a focus on building advanced AI systems. Despite U.S. export bans on advanced hardware, DeepSeek has continued its development using Nvidia H800 chips—a less powerful alternative to the H100 available to U.S. companies.

DeepSeek's Powerful AI Models

In November 2023, DeepSeek introduced its first suite of models: DeepSeek Coder, DeepSeek LLM, and DeepSeek Chat. However, it was the release of the DeepSeek-V2 model family the following spring that cemented the company's reputation. This advanced, general-purpose system demonstrated exceptional performance across industry benchmarks while maintaining lower operational costs than comparable models. This competitive edge pressured domestic rivals like ByteDance and Alibaba to reduce model usage prices and offer some models for free.

DeepSeek's momentum continued with the launch of DeepSeek-V3 in December 2024. Internal benchmarks indicated that DeepSeek V3 outperformed both open models like Meta's Llama and proprietary models like OpenAI's GPT-4o.

The company also made waves with the release of its R1 "reasoning" model in January 2025. DeepSeek claims that R1 rivals OpenAI's o1 model on key benchmarks. R1's ability to self-check its responses enhances reliability in areas such as science, physics, and mathematics. However, reasoning models typically require more processing time to arrive at solutions.

Despite these technical achievements, DeepSeek's models are subject to censorship by Chinese regulators, ensuring their responses align with "core socialist values." For instance, DeepSeek's chatbot will not address sensitive topics like Tiananmen Square or Taiwan's independence.

A Disruptive Business Model

DeepSeek's business strategy remains ambiguous. The company offers products and services at significantly lower costs than competitors and provides some tools for free. DeepSeek attributes its price advantage to efficiency breakthroughs, although some experts question the validity of these claims.

Developers have embraced DeepSeek's models, which are available under permissive licenses for commercial use. On Hugging Face, a popular AI platform, developers have produced over 500 derivative models of R1, accumulating 2.5 million downloads.

DeepSeek's rapid ascent has unsettled the AI landscape. Its influence contributed to an 18% drop in Nvidia's stock price in January and prompted public comments from OpenAI CEO Sam Altman. Microsoft has since integrated DeepSeek models into its Azure AI Foundry platform, while Meta CEO Mark Zuckerberg emphasized ongoing AI infrastructure investments in response to DeepSeek's competitive pressure.

Despite praise from Nvidia CEO Jensen Huang for DeepSeek's innovation, concerns about national security have led to bans on DeepSeek's technology in certain regions. South Korea and New York state, for example, have prohibited its use on government devices.

The Road Ahead for DeepSeek

DeepSeek's future remains uncertain amid growing scrutiny from the U.S. government. While continued advancements in AI models are expected, geopolitical tensions and regulatory barriers could shape the company's trajectory in the coming years.

Mozilla responds to backlash over new terms, saying it’s not using people’s data for AI

 

Mozilla has responded to user backlash over Firefox's new Terms of Use, which critics claim contain overly broad language that seemingly grants the company rights to user data. Mozilla asserts that the updated terms do not reflect a change in how it handles data but aim to clarify the relationship between the company and its users.

On Wednesday, Mozilla introduced new Terms of Use and an updated Privacy Notice for Firefox. According to the company, these changes are meant to provide greater transparency regarding user rights and data practices.

“We tried to make these easy to read and understand — there shouldn’t be any surprises in how we operate or how our product works,” Mozilla wrote in a blog post.

Despite these intentions, the new language caused confusion. The backlash was strong enough that Mozilla updated its blog post to explicitly state that the company does not claim ownership of user data or any rights beyond what is outlined in its Privacy Notice.

Users were particularly concerned with a section of the new terms stating: “When you upload or input information through Firefox, you hereby grant us a nonexclusive, royalty-free, worldwide license to use that information to help you navigate, experience, and interact with online content as you indicate with your use of Firefox.”

Critics argue that this phrasing is overly vague and far-reaching. Brendan Eich, co-founder and CEO of rival browser Brave Software, reacted strongly on social media, suggesting the language could be connected to a strategy to monetize user data for AI applications.

When asked whether Firefox user data is shared with AI companies or advertisers, Mozilla clarified that its Privacy Notice still applies to AI features. Content data is not sent to Mozilla or external parties, and data shared with advertisers is de-identified, according to the company.

“These changes are not driven by a desire by Mozilla to use people’s data for AI or sell it to advertisers,” Mozilla’s VP of Communications, Brandon Borrman, told TechCrunch. “As it says in the Terms of Use, we ask for permission from the user to use their data to operate Firefox ‘as you indicate with your use of Firefox.’ This means that our ability to use data is still limited by what we disclose in the Privacy Notice.”

The Privacy Notice specifies that Firefox may collect technical and interaction data when users engage with AI chatbots. Mozilla clarified that if users opt to use third-party AI chatbots, those third parties handle the data under their own policies. Other AI features in Firefox operate locally on users' devices and do not transmit content data externally.

Regarding advertising, Mozilla acknowledged that it sells advertising within Firefox to support browser development. The company emphasized that it focuses on privacy-preserving ad products and only shares data with advertisers in a de-identified or aggregated form, as outlined in the Privacy Notice. Users can opt out of having their data processed for advertising by disabling the “technical and interaction data” setting on both desktop and mobile versions of Firefox.

Mozilla also explained the language choices in its updated terms. The term “nonexclusive” ensures that users retain the ability to share their data elsewhere. “Royalty-free” reflects that Firefox is a free product, with no financial obligations between Mozilla and users for data handling. “Worldwide” reflects the global availability of Firefox and its access to the broader internet.

Despite Mozilla's assurances that these updates do not represent a shift in data practices, the vague language has left some users skeptical. This reaction could pose a challenge for Firefox, which currently holds just 2.54% of the global browser market, trailing behind Chrome (67%), Safari (17.95%), and Edge (5.2%).

This article was updated after publication to more accurately attribute statements to Mozilla's VP of Communications, Brandon Borrman, rather than spokesperson Kenya Friend-Daniel.

Artificial intelligence (AI) | Definition, Examples, Types ai

 

Artificial Intelligence (AI): Definition, Examples, and Types

Artificial intelligence (AI) refers to the ability of a digital computer or computer-controlled robot to perform tasks that are typically associated with intelligent beings. This field of technology encompasses the creation and development of systems capable of mimicking intellectual processes similar to those of humans. These processes include reasoning, learning from past experiences, generalizing information, and discovering meaning.

Since their inception in the 1940s, digital computers have evolved to execute increasingly complex tasks, such as proving mathematical theorems and playing chess, with remarkable proficiency. Despite these advances in computational speed and memory capacity, no AI system to date possesses the full range of human cognitive flexibility. However, specialized AI programs have achieved performance levels comparable to human experts in specific fields. These applications range from medical diagnosis and search engine algorithms to voice recognition, handwriting interpretation, and interactive chatbots.

Understanding Intelligence

Intelligence is generally defined as the capacity to acquire and apply knowledge and skills. While humans exhibit a broad spectrum of intelligent behaviors, the actions of other species, such as insects, are usually categorized as instinctual rather than intelligent. For example, the behavior of the digger wasp (Sphex ichneumoneus) demonstrates an absence of adaptability. When the wasp returns to its burrow with food, it performs a repetitive routine of checking the burrow before bringing the food inside. If the food is moved during this process, the wasp restarts the entire sequence, highlighting its rigid, non-adaptive behavior. In contrast, human intelligence involves the ability to adjust to new and unfamiliar situations.

Psychologists describe human intelligence as a combination of various cognitive abilities, including learning, reasoning, problem-solving, perception, and language use. AI research focuses on replicating these components through advanced programming and machine learning algorithms.

Learning in Artificial Intelligence

Learning is a critical function in AI systems, and there are various forms of learning that computers can perform. The simplest type is learning through trial and error. For instance, an AI program designed to solve chess puzzles might attempt random moves until it discovers a successful solution. Once the solution is found, the program stores it for future reference. This method is known as rote learning.

More advanced AI systems use a process called generalization. Instead of memorizing individual solutions, these systems identify patterns and apply them to new, similar situations. For example, an AI learning the past tense of English verbs may observe the "add -ed" rule and apply it to previously unseen words. This ability to generalize allows AI systems to handle unfamiliar scenarios by leveraging past experiences.

Reasoning and Inference in AI

Reasoning involves drawing logical conclusions from available information. AI systems can perform two main types of reasoning: deductive and inductive.

• Deductive Reasoning: This form of reasoning derives specific conclusions from general premises. For example, "Fred is either in the museum or the café. He is not in the café; therefore, he is in the museum."

• Inductive Reasoning: This involves making generalizations based on specific observations. For instance, "Previous accidents of this sort were caused by instrument failure. This accident is similar; therefore, it was likely caused by instrument failure."

While AI systems can draw both deductive and inductive inferences, true reasoning requires the ability to identify and apply the most relevant inferences to solve particular problems. This remains a significant challenge in AI research.

Types of Artificial Intelligence

AI is typically categorized into the following types based on its capabilities and applications:

1. Narrow AI (Weak AI): This form of AI is designed for specific tasks and cannot generalize beyond its programmed functions. Examples include virtual assistants (e.g., Siri or Alexa), spam filters, and facial recognition systems.

2. General AI (Strong AI): This hypothetical form of AI would possess cognitive abilities comparable to human intelligence. It could perform any intellectual task that a human can do, including reasoning, problem-solving, and learning across multiple domains. General AI remains a theoretical goal and has not yet been achieved.

3. Super AI: This is a speculative concept where AI surpasses human intelligence across all fields. Such an AI could outperform humans in creativity, decision-making, and emotional intelligence. While super AI is a common theme in science fiction, it does not currently exist.

Applications of Artificial Intelligence

AI technology is integrated into numerous industries and everyday applications, including:

• Healthcare: AI assists in medical diagnostics, personalized treatment plans, and drug discovery.
• Finance: AI algorithms analyze market trends, detect fraudulent activities, and manage investments.
• Transportation: Autonomous vehicles use AI for navigation and decision-making.
• Communication: Natural language processing (NLP) enables AI to interpret and generate human language, facilitating translation services and customer support chatbots.
• Entertainment: AI enhances user experiences through personalized content recommendations and video game design.

Ethical Considerations and Societal Impact

The increasing role of AI raises important ethical and societal questions:

• Job Displacement: Automation through AI may lead to the displacement of certain jobs while creating new opportunities in emerging fields.
• Privacy Concerns: AI systems collect and analyze vast amounts of personal data, raising concerns about surveillance and data protection.
• Bias and Fairness: AI models can reflect and amplify societal biases if trained on unrepresentative datasets, leading to unfair outcomes.
• Autonomy and Control: As AI systems become more sophisticated, ensuring that they remain under human control and aligned with human values is crucial.

Governments and organizations are actively exploring regulations to ensure the responsible development and deployment of AI technologies. Ethical frameworks and interdisciplinary collaboration are essential to address these challenges and harness AI's potential for societal benefit.

The Future of Artificial Intelligence

The future of AI holds tremendous promise and potential risks. Advances in deep learning, neural networks, and quantum computing may bring us closer to achieving general AI. Meanwhile, ongoing research focuses on enhancing AI's interpretability, safety, and ethical alignment.

As AI continues to evolve, it will undoubtedly shape how we live, work, and interact with the world. The challenge lies in balancing innovation with ethical responsibility, ensuring that AI technologies benefit society while minimizing potential harms.

Large Language Models (LLMs) and Natural Language Processing (NLP)

Large Language Models (LLMs) are a transformative advancement in artificial intelligence, particularly within the field of Natural Language Processing (NLP). These models, including OpenAI's GPT (Generative Pre-trained Transformer) series, Google's BERT (Bidirectional Encoder Representations from Transformers), and Meta's LLaMA, are designed to understand, generate, and manipulate human language.

What Are Large Language Models?

LLMs are advanced AI systems trained on massive datasets of text from books, websites, scientific articles, and other sources. Using deep learning techniques, especially transformer architectures, these models learn to recognize patterns in language and generate human-like responses.

Key characteristics of LLMs:

• Scale: Trained on billions of parameters and terabytes of text data.
• Context Understanding: Can analyze and respond to text with contextual accuracy.
• Generative Ability: Produce coherent, creative, and diverse text outputs.
• Transfer Learning: Adapt to new tasks with minimal additional training.

Natural Language Processing (NLP) Overview

NLP is the branch of AI focused on the interaction between computers and human language. It enables machines to understand, interpret, and produce human communication in various forms.

Core tasks of NLP include:

• Text Classification: Organizing documents by categories (e.g., spam detection).
• Sentiment Analysis: Identifying emotions and opinions in text.
• Machine Translation: Converting text between languages (e.g., English to French).
• Named Entity Recognition (NER): Identifying entities like people, locations, and dates.
• Question Answering: Providing precise responses to user queries.
• Speech-to-Text and Text-to-Speech: Converting spoken language to written form and vice versa.

Applications of LLMs and NLP

LLMs and NLP technologies have broad applications across industries, including:

1. Virtual Assistants: Powering AI-based assistants (e.g., Siri, Alexa) to respond to voice commands.
2. Chatbots: Enhancing customer service through real-time conversational agents.
3. Content Creation: Generating articles, summaries, and creative writing.
4. Medical Analysis: Assisting with clinical documentation and diagnostics.
5. Search Engines: Improving information retrieval accuracy with semantic understanding.

Challenges and Ethical Concerns

While LLMs and NLP offer significant advancements, they present challenges, including:

• Bias and Fairness: Models can perpetuate societal biases present in training data.
• Misinformation: AI-generated content may produce factually incorrect or misleading information.
• Privacy Issues: Processing and storing vast user data raises privacy concerns.
• Interpretability: Understanding why models make specific decisions remains complex.

Analysis Forecasts More Vulnerabilities In 2025

 

Record Number of Vulnerabilities Expected in 2025, Urging a Shift to Proactive Security

A recent analysis predicts that the number of reported vulnerabilities will reach unprecedented levels in 2025, reflecting the ongoing rise in cybersecurity threats and an increase in vulnerability disclosures.

Analysis by FIRST

The Forum of Incident Response and Security Teams (FIRST), a global organization dedicated to coordinating cybersecurity responses, published the analysis. Their forecast estimates nearly 50,000 vulnerabilities will be reported in 2025—an 11% increase from 2024 and a staggering 470% rise compared to 2023. The report emphasizes the urgent need for organizations to move beyond reactive security measures and adopt a proactive, risk-based approach. This includes prioritizing vulnerabilities based on their threat level, streamlining patching efforts, and preparing for waves of disclosures instead of reacting after incidents occur.

Factors Driving the Increase in Vulnerabilities

Three key trends are contributing to the rapid growth in reported vulnerabilities:

1. AI-Driven Discovery and Open-Source Expansion Advances in artificial intelligence and automated tools are accelerating vulnerability detection. These technologies enable researchers to analyze vast amounts of code and uncover flaws that might otherwise remain hidden. As a result, the number of Common Vulnerabilities and Exposures (CVEs) continues to rise.

2. Cyber Warfare and State-Sponsored Attacks The growing prevalence of state-sponsored cyber attacks is leading to the discovery of new vulnerabilities. These advanced, persistent threats are exposing weaknesses in both public and private sector systems.

3. Shifts in the CVE ecosystem security companies like Patchstack, which focuses on WordPress vulnerabilities, are contributing to the surge in reported flaws. Patchstack offers vulnerability detection and virtual patching services, enhancing security but also increasing the number of disclosed vulnerabilities.

Eireann Leverett, the FIRST liaison and lead member of the Vulnerability Forecasting Team, emphasized the accelerating pace of vulnerability disclosures and the necessity for organizations to adopt proactive risk management practices.

Looking Ahead to 2026 and Beyond

The forecast projects over 51,000 vulnerabilities will be disclosed in 2026, reinforcing the notion that cybersecurity risks will continue to escalate. This trend highlights the need for a forward-thinking security strategy that focuses on identifying and mitigating threats before they are exploited.

For users of platforms like WordPress, adopting proactive security measures is crucial. Solutions from companies such as Patchstack, Wordfence, and Sucuri offer various approaches to strengthening defenses against emerging threats.

Key Takeaways:

• Vulnerabilities are on the rise: FIRST predicts up to 50,000 CVEs in 2025, reflecting an 11% increase from 2024 and a 470% increase from 2023.
• AI and open-source adoption are driving higher vulnerability disclosures.
• State-sponsored cyber activity is uncovering more security weaknesses.
• A proactive security strategy is essential to manage and mitigate future risks.

Read the full 2025 vulnerability forecast for a comprehensive analysis and recommendations.

Data Suggests Google Indexing Rates Are Improving

 

New Research on Google Indexing Trends (2022-2025)

A recent study analyzing over 16 million webpages reveals that while Google indexing rates have improved, many pages remain unindexed, and over 20% of indexed pages are eventually deindexed. These findings highlight significant challenges faced by websites focused on SEO and indexing.

Research by IndexCheckr Tool IndexCheckr, a Google indexing tracking tool, allows users to monitor the indexing status of their content and external pages hosting backlinks. While the research may not reflect global Google indexing patterns, it closely aligns with trends observed by website owners focused on SEO and backlink tracking.

Understanding Web Indexing Web indexing involves search engines crawling and filtering internet content, removing duplicates or low-quality pages, and storing the remaining pages in a structured database called a Search Index. Google initially utilized the Google File System (GFS) and later upgraded to Colossus, a more advanced system capable of handling vast search data across thousands of servers.

Indexing Success Rates The study indicates that a significant portion of pages remain unindexed, although indexing rates have improved between 2022 and 2025. Notable findings include:

• 61.94% of pages in the dataset were not indexed.
• Google indexes 93.2% of successfully indexed pages within six months.
• Indexing rates have steadily improved from 2022 to 2025.

Deindexing Trends The research also sheds light on Google's rapid deindexing processes. Of all indexed pages, 13.7% are deindexed within three months, with an overall deindexing rate of 21.29%. However, 78.71% of indexed pages remain consistently in Google's index. The time-based cumulative deindexing percentages are as follows:

• 1.97% deindexed within 7 days.
• 7.97% deindexed within 30 days.
• 13.70% deindexed within 90 days.
• 21.29% deindexed after 90 days.

The research underscores the importance of early monitoring and optimization to mitigate deindexing risks. Although the risk decreases after three months, periodic audits remain crucial for maintaining long-term content visibility.

Effectiveness of Indexing Services The study also evaluates the efficacy of manual submission strategies through indexing tools. It found that only 29.37% of URLs submitted via these tools were successfully indexed, leaving 70.63% of submitted pages unindexed. This suggests limitations in current manual indexing approaches.

High Percentage of Pages Not Indexed While less than 1% of tracked websites were entirely unindexed, only 37.08% of all tracked pages were fully indexed. The data, derived from IndexCheckr subscribers, may not reflect broader internet-wide trends but offers valuable insights for SEO-focused website owners.

Google Indexing Improvements Since 2022 Despite some concerning statistics, the study reveals a positive trend: Google's indexing rates have steadily improved from 2022 to 2025. This suggests enhanced efficiency in Google's ability to process and include webpages.

Summary of Findings Complete deindexing of entire websites remains rare. However, over half of the pages analyzed struggle with indexing, likely due to site quality issues. Factors that may contribute to indexing challenges include:

• Commercial product pages with bulked-up content for search engines.
• Sites designed primarily to "feed the bot" rather than to provide value to users.

Google's search results, particularly for e-commerce, are becoming increasingly precise. SEO strategies that focus solely on entity optimization, keywords, and topical maps may fail to address the user-centric ranking factors that drive long-term success.

Data Shows Perplexity Cites Sources 2.5x More Than ChatGPT

 

AI Search Engine Citation Analysis: Key Insights

A recent report by xfunnel.ai reveals new insights into how major AI search engines reference web content. The study, which analyzed 40,000 responses containing 250,000 citations, highlights key differences in citation frequency, content types, and source quality. Here are the main findings:

Citation Frequency Varies by Platform

Researchers tested AI search engines across different buyer journey stages and observed variations in how frequently each platform cites external content:

• Perplexity: 6.61 citations per response
• Google Gemini: 6.1 citations per response
• ChatGPT: 2.62 citations per response

ChatGPT's lower citation frequency is attributed to its standard mode testing, which did not utilize search-enhanced features.

Third-Party Content Dominates Citations

Citations were classified into four categories:

• Owned Content: Company domains
• Competitor Content: Rival company domains
• Earned Content: Third-party and affiliate sites
• User-Generated Content (UGC): Reviews and forum posts

Earned content constitutes the largest share of citations across all platforms, with UGC increasing in prominence. Affiliate sites and independent blogs also play a significant role in AI-generated responses.

Citation Patterns Shift Along the Customer Journey

The study found that citation patterns evolve based on the query stage:

• Problem Exploration & Education: Higher citation rates from third-party editorial content
• Comparison Stage: Increased UGC citations from review platforms and forums
• Final Research & Evaluation: More direct citations from brand and competitor websites

Source Quality Distribution

AI search engines prioritize higher-quality sources but still reference a range of content levels:

• High-quality: ~31.5%
• Upper-mid quality: ~15.3%
• Mid-quality: ~26.3%
• Lower-mid quality: ~22.1%
• Low-quality: ~4.8%

UGC Source Preferences by Platform

Different AI platforms show distinct preferences for UGC sources:

• Perplexity: Favors YouTube and PeerSpot
• Google Gemini: Frequently cites Medium, Reddit, and YouTube
• ChatGPT: Often references LinkedIn, G2, and Gartner Peer Reviews

Leveraging Third-Party Citations for SEO

The findings highlight an underutilized opportunity for SEO professionals. While optimizing owned content remains important, the dominance of earned media citations suggests a broader strategy:

• Foster relationships with industry publications
• Create compelling content others want to reference
• Contribute guest articles to reputable sites
• Engage with preferred UGC platforms for each AI engine

By focusing on creating valuable, shareable content, brands can increase their chances of being cited across AI search engines.

Why It Matters

As AI search engines continue to shape how users find information, understanding citation patterns is crucial for maintaining visibility. Diversifying content strategies across owned, earned, and UGC platforms can enhance your presence while preserving SEO best practices.

Key Takeaway

To maximize visibility in AI search engines, invest in a balanced approach:

• Maintain high-quality owned content
• Secure mentions on trusted third-party sites
• Establish a presence on relevant UGC platforms

The data suggests that earning third-party citations may offer greater value than solely optimizing your own content for AI search visibility.

TechCrunch Disrupt 2025: Lowest prices of the year end in 7 days

 Time is Running Out – Save Big on TechCrunch Disrupt 2025 Tickets!

You read that right! The best deals on TechCrunch Disrupt 2025 tickets are ending soon — just 7 days left to save. Secure your spot and enjoy savings of up to $1,130 on individual passes and up to 30% on group tickets. Don’t miss out — these exclusive offers disappear on February 28 at 11:59 p.m. PT.

Celebrate 20 Years of Innovation at TechCrunch Disrupt

Join us from October 27-29 at Moscone West in San Francisco as we mark two decades of groundbreaking tech and entrepreneurship. Connect with over 10,000 tech leaders, engage in 250+ sessions, and hear from 200+ industry experts. Plus, witness the iconic Startup Battlefield 200 and explore the latest in AI innovation.

Register now to lock in the biggest ticket savings of 2025!

What Awaits You at Disrupt 2025?

✨ AI Deep Dives: Explore cutting-edge AI advancements across industries like healthcare, transportation, SaaS, policy, defense, hardware, and more.

🔍 Expert Insights: Learn from 200+ leaders on business scaling, leadership, and emerging fields such as space tech, fintech, IPOs, and SaaS to accelerate your growth.

🔄 Interactive Sessions: Engage in live Q&As and dive deeper with expert-led roundtables and breakout discussions.

🌟 Startup Pitch Battle: Watch innovative startups compete in the renowned Startup Battlefield 200 for a $100,000 equity-free prize and the coveted Disrupt Cup. Learn from world-class VC judges and discover the next big thing—past winners include Dropbox, Fitbit, Trello, and Cloudflare.

Think your startup has what it takes? Or know one that should compete? Add them to the Startup Battlefield waitlist to be the first to know when applications open.

🤝 Build Powerful Connections: Network with investors, mentors, and tech leaders shaping the future. Whether you're seeking funding, partnerships, or mentorship, Disrupt is the place to make it happen.

Secure Your Spot Before Prices Rise!

Don’t miss your chance to attend TechCrunch Disrupt 2025 at the lowest prices of the year. Save big before these exclusive discounts end on February 28 at 11:59 p.m. PT.

20 Years of Tech Innovation

For two decades, TechCrunch Disrupt has been the premier destination for visionary founders, industry leaders, and top investors driving the future of tech. It’s where groundbreaking ideas are born and industry connections are made.

Past Disrupt speakers include:

• Alex Pall & Drew Taggart (The Chainsmokers), Co-founders, MANTIS Venture Capital
• Ashton Kutcher, Co-founder, Sound Ventures
• Assaf Rappaport, Co-founder & CEO, Wiz
• Bridgit Mendler, CEO, Northwood Space
• Colin Kaepernick, Founder & CEO, Lumi
• Denise Dresser, CEO, Slack
• Erin & Sara Foster, Co-founders, Oversubscribed Ventures
• Mary Barra, CEO, General Motors
• Matt Mullenweg, Co-founder, WordPress; CEO, Automattic
• Peter Beck, Founder & CEO, Rocket Lab
• Serena Williams, Managing Partner, Serena Ventures
• Shaquille O’Neal, Entrepreneur & Philanthropist
• Vinod Khosla, Founder, Khosla Ventures
• RJ Scaringe, CEO, Rivian

Don't Miss Out—Register Today!

Save up to $1,130 on your TechCrunch Disrupt 2025 ticket before February 28 at 11:59 p.m. PT. Lock in the best rates now and join us for this milestone event!

Interested in sponsoring or exhibiting?

OpenAI rolls out its AI agent, Operator, in several countries

 OpenAI announced on Friday that it is expanding access to Operator—its AI agent capable of performing tasks on behalf of users—to ChatGPT Pro subscribers in several countries, including Australia, Brazil, Canada, India, Japan, Singapore, South Korea, and the U.K.

Operator will be available in most regions where ChatGPT is accessible, excluding the European Union, Switzerland, Norway, Liechtenstein, and Iceland.

Initially launched in the U.S. in January, Operator is part of a growing category of AI agents that can handle tasks like booking tickets, making restaurant reservations, filing expense reports, and shopping online. Currently, the tool is exclusive to ChatGPT Pro subscribers on the $200-per-month plan and can only be accessed through a dedicated web page. OpenAI has stated that it plans to integrate Operator with all ChatGPT clients in the future. The tool operates through a separate browser window, which users can manually control if needed.

The AI agent space is becoming increasingly competitive. Companies like Google, Anthropic, and Rabbit are also developing similar tools. However, Google's AI agent remains on a waitlist, Anthropic's agentic interface is only accessible via API, and Rabbit's action model is limited to users of its proprietary device.

This Week in AI: Maybe we should ignore AI benchmarks for now

 

Elon Musk's xAI Releases Grok 3

This week, Elon Musk's AI startup, xAI, unveiled its latest flagship AI model, Grok 3, which powers the company's Grok chatbot apps. Trained on approximately 200,000 GPUs, Grok 3 outperforms several leading models, including OpenAI's, in benchmarks for mathematics, programming, and other domains.

The Benchmark Debate

While benchmarks are a common measure of AI progress, their relevance is often questioned. They tend to evaluate niche knowledge and provide aggregate scores that may not reflect real-world performance. As Wharton professor Ethan Mollick highlighted, there is an "urgent need for better tests and independent testing authorities." Given that AI companies typically self-report results, scepticism around these metrics is justified.

Despite the emergence of independent benchmarks, consensus on their value remains elusive. Some experts suggest aligning benchmarks with economic impact, while others believe real-world adoption and utility are better indicators. As the debate continues, some voices, like X user Roon, advocate paying less attention to benchmarks unless there are major technical breakthroughs.

News Highlights

• OpenAI's "Uncensored" ChatGPT: OpenAI is shifting its development approach to embrace "intellectual freedom," allowing discussions on controversial topics.

• Mira Murati's New Startup: Former OpenAI CTO Mira Murati has launched Thinking Machines Lab to create AI tools tailored to individual needs and goals.

• LlamaCon Announcement: Meta will host its first generative AI developer conference, LlamaCon, on April 29, focusing on its Llama model family.

• AI and Europe's Digital Sovereignty: The OpenEuroLLM initiative, involving 20 organisations, aims to develop AI models that preserve the linguistic and cultural diversity of EU languages.

Research Paper of the Week

OpenAI introduced SWE-Lancer, a new benchmark to assess AI coding abilities. This dataset includes over 1,400 freelance software engineering tasks, from bug fixes to technical proposals. The leading model, Anthropic's Claude 3.5 Sonnet, achieved 40.3% accuracy, indicating room for improvement. Notably, newer models like OpenAI's o3-mini were not tested.

Model of the Week

Chinese AI company Stepfun released Step-Audio, an open model capable of understanding and generating speech in Chinese, English, and Japanese. Users can customise the emotion, dialect, and even produce synthetic singing. Stepfun, founded in 2023, recently secured several hundred million dollars in funding from investors, including Chinese state-owned private equity firms.

Grab Bag

Nous Research unveiled DeepHermes-3 Preview, a model combining reasoning and intuitive language capabilities. It can toggle between fast, intuitive responses and more computationally intensive, accurate reasoning. Similar models from Anthropic and OpenAI are reportedly on the horizon.

Until Next Time

As "This Week in AI" goes on hiatus, thank you for joining us on this ever-evolving journey. Stay tuned for future updates.

people ask questions.

What are the benchmarks for AI?

AI benchmarks are similar to exams for humans. They are standardised tests designed to evaluate specific skills, knowledge, or abilities of AI systems. These benchmarks produce a score or grade, enabling systematic comparisons between different AI models that undergo the same assessment.

Google pulls Gemini from main search app on iOS

 

Google is removing its AI assistant, Gemini, from the main Google app on iOS devices. This move aims to encourage users to download the standalone Gemini app, allowing Google to better compete with other AI chatbots like ChatGPT, Claude, and Perplexity. However, the decision could limit Gemini's reach, as the Google app already has millions of users who may be reluctant to download a separate application.

The tech giant informed customers of the change through an email stating, "Gemini is no longer available in the Google app." The message directed users who wish to continue using Gemini on iOS to download the dedicated Gemini app from the App Store. While the standalone app became available to iOS users worldwide late last year, Gemini had remained accessible through the main Google app until now.

With the Gemini app, users can engage in voice conversations through Gemini Live, integrate Google services like Search, YouTube, Maps, and Gmail, ask questions, explore topics, plan trips, receive AI-generated summaries and deep dives, create images, and more. Interaction is available via text, voice, or camera input.

The email also cautioned users that Gemini may still produce errors, advising them to verify its responses. Additionally, users can upgrade to Gemini Advanced through the iOS app by subscribing to the Google One AI Premium plan, available as an in-app purchase.

iOS users attempting to access Gemini through the main Google app will now encounter a full-screen message stating, "Gemini now has its own app," with a link to download it from the App Store.

Google's decision to separate Gemini into a standalone app is a calculated risk. While it could streamline the rollout of new AI features, it may also lead to a decline in usage as some users choose not to make the transition.

Google’s ‘Career Dreamer’ uses AI to help you explore job possibilities

 

Google Unveils "Career Dreamer"—An" AI Tool to Explore Career Possibilities

Google is introducing a new experimental tool called Career Dreamer, designed to help people discover career opportunities using artificial intelligence. Announced in a blog post on Wednesday, the tool identifies patterns from your experiences, education, skills, and interests to suggest careers that may be a good fit.

With Career Dreamer, users can create a personalised career identity statement by selecting their current and past roles, skills, education, and interests. This statement can be added to a résumé or used as a guide during job interviews.

The tool also offers a visual web of career possibilities tailored to your background. If a particular career catches your interest, you can dive deeper to learn more about its responsibilities and requirements.

Additionally, Career Dreamer integrates with Gemini, Google’s AI assistant, allowing users to collaborate on crafting cover letters, refining résumés, and brainstorming new job ideas. However, unlike job platforms such as Indeed or LinkedIn, Career Dreamer does not connect users directly to job listings. Instead, it focuses on helping people quickly explore career paths without conducting multiple Google searches.

Currently, Career Dreamer is available only as an experimental feature in the United States, with no details on when or if it will expand to other regions.

“We hope Career Dreamer can be helpful to all kinds of job seekers,” Google shared in its blog post. The company collaborated with organisations supporting diverse groups, including students, recent graduates, adult learners, and the military community, to develop the tool. Whether you’re considering a career change or simply curious about new possibilities, Career Dreamer aims to make the process easier.

Google also highlights a World Economic Forum report, which notes that people typically hold 12 different jobs throughout their lifetime, with Gen Z expected to navigate 18 jobs across six careers. The company believes Career Dreamer can help users frame their diverse experiences into a cohesive narrative and better showcase how their existing skills align with new opportunities.

Thinking Machines: Ex-OpenAI CTO’s new AI startup

 

Mira Murati Launches Thinking Machines to Democratize AI

Former OpenAI CTO Mira Murati has unveiled Thinking Machines, a new AI research and product company focused on making AI more accessible, customisable, and collaborative.

A Mission to Make AI More Inclusive

Thinking Machines aims to bridge key gaps in the AI landscape by ensuring these technologies are not just powerful but also adaptable and widely understood. The company is committed to democratising AI, enabling users to tailor AI tools to their unique needs.

“We’re building a future where everyone has access to the knowledge and tools to make AI work for their unique goals,” the company states.

Addressing AI’s Key Challenges

Despite AI's rapid advancements, knowledge remains concentrated within a select few research labs, making it difficult for broader communities to innovate. AI systems also lack customisation, limiting their real-world application. Thinking Machines aims to address these issues by creating open, customisable, and capable AI systems that foster both innovation and accessibility.

The company plans to achieve this by leveraging intellectual openness, cutting-edge infrastructure, and advanced AI safety practices to empower both researchers and end users.

A Team of AI Pioneers

Murati has assembled a team of leading scientists, engineers, and technologists who have previously worked on groundbreaking AI tools, including OpenAI’s ChatGPT, Character.ai, PyTorch, and OpenAI Gym. Their expertise positions Thinking Machines to make a substantial impact in the AI space.

A Human-Centric Approach

Unlike companies focused solely on autonomous AI, Thinking Machines prioritises human-AI collaboration. It embraces an open research culture, sharing research papers, technical insights, and code with the wider AI community.

“Scientific progress is a collective effort,” the company emphasises. “By collaborating with the broader research community, we can advance AI for the benefit of all.”

The company is also investing in multimodal AI systems that integrate text, video, and imagery, allowing seamless interaction between humans and AI. This personalisation will enable AI to support diverse fields, from scientific research to creative industries.

Building Strong AI Foundations

While many AI startups rush to deploy new systems, Thinking Machines is taking a deliberate and methodical approach. Its strategy rests on two key pillars:

1. Intelligent AI Models—Developing frontier AI models capable of breakthroughs in programming, scientific discovery, and engineering.
2. High-Quality Infrastructure—Creating efficient, secure, and user-friendly platforms that enhance AI usability and scalability.

The company is also focusing on advanced multimodal capabilities, integrating language, imagery, and sensory data to build AI that better understands and interacts with the world.

Ethical AI Through Product-Driven Learning

Thinking Machines is committed to AI safety and ethical innovation by integrating research with product development. This iterative approach ensures AI remains both practical and responsible.

Key AI safety principles include:

• Maintaining a high safety bar while preserving user freedoms.
• Sharing best practices for secure AI development.
• Supporting external research on AI alignment through open access to code, datasets, and models.

The company recognises that true AI progress requires rethinking objectives, not just optimising existing metrics. By focusing on real-world impact, Thinking Machines aims to develop AI that delivers tangible societal benefits.

A New Era for AI

Thinking Machines represents the next chapter for Mira Murati, whose leadership helped shape some of OpenAI’s most successful projects. With a team of top AI innovators, a commitment to openness, and a long-term vision, the company seeks to challenge the status quo in AI development.

By prioritising collaboration, accessibility, and responsible innovation, Thinking Machines aims to empower industries and individuals to harness AI’s full potential—on their own terms.

Grok 3: The next-gen ‘truth-seeking’ AI model

---

xAI Unveils Grok 3: A Powerful AI Model with Advanced Reasoning and Image Analysis

xAI introduced its latest artificial intelligence model, Grok 3, on Monday, bringing enhanced capabilities such as image analysis and improved question-answering.

The company developed Grok 3 using a massive data center housing approximately 200,000 GPUs. According to xAI’s founder, Elon Musk, this initiative utilized ten times the computing power of its predecessor, Grok 2, incorporating a significantly expanded dataset—including legal case filings.

Musk described Grok 3 as a "maximally truth-seeking AI," emphasizing that it prioritizes accuracy over political correctness.

A Family of Grok 3 Models

The Grok 3 release includes multiple versions tailored to different needs. For instance, Grok 3 Mini offers faster response times, while the reasoning-focused models—Grok 3 Reasoning and Grok 3 Mini Reasoning—aim to simulate human-like thought processes. These models actively analyze problems and verify their answers, reducing errors. Similar to OpenAI’s o3-mini and DeepSeek’s R1, Grok 3 Reasoning models attempt to fact-check themselves, improving reliability.

Grok 3 vs. the Competition: Benchmark Results

xAI claims Grok 3 outperforms OpenAI’s GPT-4o in certain key benchmarks, including AIME and GPQA, which assess problem-solving across subjects like mathematics, physics, biology, and chemistry.

Early rankings place Grok 3 at the top of Chatbot Arena, a user-driven AI evaluation platform, making it the first model to surpass a score of 1400.

The company also asserts that Grok 3 Reasoning outperforms competing models from Google, DeepSeek, and OpenAI in a range of reasoning benchmarks.

These advanced reasoning models are now available in the Grok app, where users can activate features like "Think" or the high-powered "Big Brain" mode for complex problem-solving. xAI has positioned them as particularly useful for STEM fields, including coding, mathematics, and scientific research.

Protecting Intellectual Property from AI Distillation

One of the more controversial aspects of AI development is "distillation," where rival companies extract knowledge from proprietary models. To counter this, xAI has deliberately obscured certain internal processes within Grok 3 to safeguard its technology.

The issue of distillation gained attention recently when Chinese AI company DeepSeek was accused of leveraging OpenAI’s models to build its own system, R-1.

Introducing DeepSearch: AI-Powered Internet Analysis

A major new feature accompanying Grok 3 is DeepSearch. This tool scans the internet—including Musk’s social platform, X—to compile and synthesize information for users. By leveraging Grok models, DeepSearch aims to provide more comprehensive and accurate responses.

Access and Subscription Plans

Grok 3 is currently accessible through X’s Premium+ subscription, priced at $50 per month (~£41), offering priority access to the latest AI features.

Additionally, xAI is launching a SuperGrok subscription at $30 per month or $300 per year. This tier provides enhanced reasoning capabilities, increased DeepSearch queries, and unlimited image generation.

In the coming weeks, xAI plans to roll out a voice interaction feature similar to Gemini Live, allowing users to engage with Grok through synthesized speech. Musk also announced that Grok 3 models will soon be available via an enterprise-ready API, including DeepSearch functionality.

Open-Source Plans for Grok 2

While Grok 3 remains proprietary for now, xAI intends to open-source its predecessor, Grok 2, once the latest model reaches maturity. Musk stated that this could happen within a few months, reinforcing xAI’s commitment to transparency and open development.

A Controversial “Truth-Seeking” AI

From its inception, Grok has been positioned as an unfiltered AI model, willing to tackle politically sensitive topics. Musk has previously described it as “anti-woke,” setting it apart from more cautious competitors.

However, earlier versions of Grok sometimes exhibited political biases, prompting xAI to refine its approach. Musk now claims efforts are underway to make Grok more politically neutral. Whether Grok 3 succeeds in striking this balance remains to be seen.

The Future of Grok 3

As AI competition intensifies, xAI’s latest advancements push the boundaries of reasoning and information synthesis. However, questions remain about bias, transparency, and the broader ethical implications of deploying such technology.

With OpenAI, Google, and DeepSeek continually refining their models, Grok 3’s success will depend on its ability to deliver accurate, fast, and reliable responses while navigating the challenges of public perception and responsible AI use.

Would you like any additional refinements?

AI helps prevent fraud with intelligent document processing

 

The Rising Threat of Fraud Across Industries

Fraud is a growing concern across all industries, with increasing cases in finance, retail, and loyalty programs. From counterfeit invoices and falsified receipts to identity theft and synthetic accounts, traditional fraud detection methods struggle to keep up with evolving tactics.

Many organisations still depend on manual fraud detection processes, which are slow, error-prone, and often identify fraud only after the damage is done. As fraud tactics become more sophisticated, businesses require a more intelligent approach. AI-powered automated document fraud detection offers a proactive solution, enabling real-time document verification, anomaly detection, and fraud prevention before it occurs.

How AI-Powered Intelligent Document Processing (IDP) is Transforming Fraud Detection

AI-driven intelligent document processing (IDP) is reshaping fraud detection by leveraging machine learning (ML), optical character recognition (OCR), and real-time data verification. This combination allows businesses to analyse, authenticate, and flag fraudulent documents within seconds. Unlike traditional methods, AI-based fraud detection is faster, more accurate, and continuously evolving to detect fraud patterns before they lead to financial and reputational damage.

In this blog, we will explore IDP, how AI enhances fraud detection, and its applications across industries.

---

What is Intelligent Document Processing & How Does AI Improve Fraud Detection?

Businesses process vast volumes of documents, invoices, receipts, and identity records daily. However, manual handling and outdated fraud detection techniques cannot effectively keep pace with the increasing complexity of fraud attempts. Intelligent document processing (IDP) addresses this challenge.

What is Intelligent Document Processing?

Intelligent document processing is an AI-powered technology that automates data extraction, classification, and verification from documents. It integrates machine learning (ML), natural language processing (NLP), and OCR to analyse structured and unstructured documents far beyond the capabilities of traditional rule-based systems.

Unlike manual reviews or keyword-based matching, IDP understands context, patterns, and anomalies, making it an indispensable tool for detecting fraudulent activity.

How AI Enhances Fraud Detection with IDP

AI-driven IDP strengthens fraud detection by

• Instant Anomaly Detection: AI scans thousands of documents in real time, identifying irregularities in invoices, receipts, and identity records that humans might miss.
• Verifying Document Authenticity: AI cross-references data across multiple sources to detect manipulated text, forged signatures, and fake documents.
• Identifying Duplicate or Altered Submissions: Fraudsters often modify genuine receipts or submit duplicate claims. AI detects inconsistencies and flags them.
• Reducing False Positives: Unlike rule-based systems, which mistakenly flag legitimate transactions, AI continuously learns and improves accuracy over time.
• Scaling Fraud Detection: AI processes millions of documents, allowing businesses to detect fraud without increasing human workload.

Why Traditional Fraud Detection Falls Short

Most fraud detection systems rely on manual audits, fixed rules, and pattern-matching techniques, which are:

• Time-consuming and costly: Manual document verification demands significant resources.
• Prone to human error: Fraudsters exploit loopholes and inconsistencies that humans may overlook.
• Limited in scope: Rule-based systems struggle to detect new and evolving fraud tactics.

By leveraging AI and IDP, businesses can implement a faster, more reliable, and scalable fraud detection system that adapts to emerging threats.

---

AI-Powered Fraud Detection Across Industries

Fraud affects businesses in various ways, from loyalty program abuse to invoice fraud and identity theft. Traditional fraud detection methods fail to counter increasingly sophisticated fraud tactics. AI-powered IDP is revolutionising fraud detection across industries. Here’s how AI is tackling fraud in key sectors:

Preventing Loyalty Fraud in Rewards Programs

Loyalty programs reward genuine customers, but fraudsters exploit these systems for personal gain. Common fraud tactics include:

• Creating multiple accounts to claim sign-up bonuses repeatedly.
• Submitting fake or altered receipts to earn rewards without real purchases.
• Abusing refund and return policies to retain loyalty points after reversing a transaction.
• Hacking accounts to steal and redeem someone else’s loyalty points.

AI-powered fraud detection prevents loyalty fraud by:

• Verifying Receipts: AI detects forgeries, duplicates, and altered receipt information.
• Identifying Suspicious Patterns: Machine learning algorithms recognise unusual transaction behaviours, such as multiple claims from the same user under different identities.
• Automating Account Authentication: AI-driven identity verification ensures real customers benefit from rewards while preventing bot-driven abuse.

With real-time fraud detection, businesses can minimise losses and ensure that rewards go to legitimate customers.

Stopping Invoice & Expense Fraud in Finance & Accounting

Fraudsters target invoice and expense management systems by submitting fake, inflated, or duplicate claims. Common invoice fraud tactics include:

• Invoice Tampering: Altering invoice amounts or vendor details to redirect payments.
• Duplicate Claims: Submitting the same invoice multiple times for reimbursement.
• Fake Receipts: Generating counterfeit receipts to justify fraudulent expenses.

AI and OCR technology help detect fraudulent activities by

• Extracting & Verifying Invoice Data: AI cross-checks invoices against existing records, vendor details, and past payments to detect duplications or alterations.
• Spotting Irregular Patterns: Machine learning identifies inconsistencies like inflated amounts, mismatched dates, and suspicious vendor behaviour.
• Automating Compliance Checks: AI ensures invoices meet expense policies and tax regulations, reducing errors in financial audits.

By integrating AI-driven document processing, finance teams can accelerate invoice verification, prevent fraudulent payouts, and eliminate manual review bottlenecks.

Banking Fraud: Loan & Mortgage Fraud Prevention

Fraudsters manipulate loan and mortgage applications using falsified documents, stolen identities, or synthetic identities. Common fraud techniques include:

• Document Forgery: Altering bank statements, pay stubs, or tax records to exaggerate income.
• Identity Theft: Using stolen personal information to apply for loans under a false identity.
• Synthetic Identity Fraud: Mixing real and fake information to create a fraudulent credit history.
• Straw Borrower Schemes: Using a third party to conceal the true borrower’s financial risk.

AI-powered fraud detection in banking prevents loan and mortgage fraud through:

• Advanced Document Verification: AI analyses financial documents for inconsistencies, altered text, and signs of forgery.
• Identity Verification & Biometric Matching: AI-powered facial recognition and ID authentication ensure applicants are legitimate.
• Cross-Referencing Financial Data: AI scans multiple data sources, like credit history and banking records, to detect unusual patterns.
• Real-Time Risk Assessment: Machine learning evaluates loan applications for fraudulent indicators, reducing high-risk lending.

By integrating AI into banking fraud detection, financial institutions can enhance security, reduce loan defaults, and comply with regulations.

---

The Future of Fraud Prevention: AI is the Key

Fraud tactics continuously evolve, making traditional detection methods less effective over time. Manual reviews and rule-based systems are too rigid to counter increasingly sophisticated fraud schemes. AI, however, offers a dynamic, self-learning approach that adapts to emerging threats.

Unlike static fraud detection models, AI continuously analyses patterns, detects anomalies, and refines its accuracy. By automating document authentication, verifying identities, and flagging suspicious transactions, AI minimises human error and enhances fraud prevention across industries. Its ability to process millions of documents instantly ensures fraud is detected before financial damage occurs.

---

Conclusion: AI-Driven Fraud Detection is the Future

Businesses can no longer afford to rely on outdated fraud prevention strategies. AI-powered intelligent document processing provides a scalable, efficient, and highly accurate way to detect and prevent fraud, reducing financial losses and compliance risks. By adopting AI, companies can automate fraud detection, enhance security, and stay ahead of emerging threats.

As fraud tactics evolve, businesses must evolve with them. AI is no longer the future of fraud prevention—it is the present. The question is: Is your business ready to embrace it?