VentureBeat

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More A new report from AI data provider Appen reveals that companies are struggling to source and manage the high-quality data needed to power AI systems as artificial intelligence expands into enterprise operations. Appen’s 2024 State of AI report, which surveyed over 500 U.S. IT decision-makers, reveals that generative AI adoption surged 17% in the past year; however, organizations now confront significant hurdles in data preparation and quality assurance. The report shows a 10% year-over-year increase in bottlenecks related to sourcing, cleaning, and labeling data, underscoring the complexities of building and maintaining effective AI models. Si Chen, Head of Strategy at Appen, explained in an interview with VentureBeat: “As AI models tackle more complex and specialised problems, the data requirements also change,” she said. “Companies are finding that just having lots of data is no longer enough. To fine-tune a model, data needs to be extremely high-quality, meaning that it is accurate, diverse, properly labelled, and tailored to the specific AI use case.” While the potential of AI continues to grow, the report identifies several key areas where companies are encountering obstacles. Below are the top five takeaways from Appen’s 2024 State of AI report: 1. Generative AI adoption is soaring — but so are data challenges The adoption of generative AI (GenAI) has grown by an impressive 17% in 2024, driven by advancements in large language models (LLMs) that allow businesses to automate tasks across a wide range of use cases. From IT operations to R&D, companies are leveraging GenAI to streamline internal processes and increase productivity. However, the rapid uptick in GenAI usage has also introduced new hurdles, particularly around data management. “Generative AI outputs are more diverse, unpredictable, and subjective, making it harder to define and measure success,” Chen told VentureBeat. “To achieve enterprise-ready AI, models must be customized with high-quality data tailored to specific use cases.” Custom data collection has emerged as the primary method for sourcing training data for GenAI models, reflecting a broader shift away from generic web-scraped data in favor of tailored, reliable datasets. The use of generative AI in business processes continues to expand, with notable increases in IT operations, manufacturing, and research and development. However, adoption in areas like marketing and communications has slightly declined. (Source: Appen State of AI Report 2024) 2. Enterprise AI deployments and ROI are declining Despite the excitement surrounding AI, the report found a worrying trend: fewer AI projects are reaching deployment, and those that do are showing less ROI. Since 2021, the mean percentage of AI projects making it to deployment has dropped by 8.1%, while the mean percentage of deployed AI projects showing meaningful ROI has decreased by 9.4%. This decline is largely due to the increasing complexity of AI models. Simple use cases like image recognition and speech automation are now considered mature technologies, but companies are shifting toward more ambitious AI initiatives, such as generative AI, which require customized, high-quality data and are far more difficult to implement successfully. Chen explained, “Generative AI has more advanced capabilities in understanding, reasoning, and content generation, but these technologies are inherently more challenging to implement.” The percentage of AI projects making it to deployment has steadily declined since 2021, with a sharp drop to 47.4% in 2024. Similarly, the mean percentage of deployed projects showing meaningful ROI has fallen to 47.3%, reflecting the growing challenges businesses face in achieving successful AI implementations. (Source: Appen State of AI Report 2024) 3. Data quality is essential — but it’s declining The report highlights a critical issue for AI development: data accuracy has dropped nearly 9% since 2021. As AI models become more sophisticated, the data they require has also become more complex, often requiring specialized, high-quality annotations. A staggering 86% of companies now retrain or update their models at least once every quarter, underscoring the need for fresh, relevant data. Yet, as the frequency of updates increases, ensuring that this data is accurate and diverse becomes more difficult. Companies are turning to external data providers to help meet these demands, with nearly 90% of businesses relying on outside sources to train and evaluate their models. “While we can’t predict the future, our research shows that managing data quality will continue to be a major challenge for companies,” said Chen. “With more complex generative AI models, sourcing, cleaning, and labeling data have already become key bottlenecks.” Data management emerged as the leading challenge for AI projects in 2024, with 48% of respondents citing it as a significant bottleneck. Other obstacles include a lack of technical resources, tools, and data, highlighting the increasing complexity of AI implementation. (Source: Appen State of AI Report 2024) 4. Data bottlenecks are worsening Appen’s report reveals a 10% year-over-year increase in bottlenecks related to sourcing, cleaning, and labeling data. These bottlenecks are directly impacting the ability of companies to successfully deploy AI projects. As AI use cases become more specialized, the challenge of preparing the right data becomes more acute. “Data preparation issues have intensified,” said Chen. “The specialized nature of these models demands new, tailored datasets.” To address these problems, companies are focusing on long-term strategies that emphasize data accuracy, consistency, and diversity. Many are also seeking strategic partnerships with data providers to help navigate the complexities of the AI data lifecycle. Data accuracy in the U.S. has steadily declined, dropping from 63.5% in 2021 to just 54.6% in 2024. The decrease highlights the growing challenge of maintaining high-quality data as AI models become more complex. (Source: Appen State of AI Report 2024) 5. Human-in-the-Loop is More Vital Than Ever While AI technology continues to evolve, human involvement remains indispensable. The report found that 80% of respondents emphasized the importance of human-in-the-loop machine learning, a process where human expertise is used to guide and improve AI models. “Human involvement remains essential for developing high-performing, ethical, and contextually relevant AI systems,” said Chen. Human

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More A pair of researchers at OpenAI has published a paper describing a new type of model — specifically, a new type of continuous-time consistency model (sCM) — that increases the speed at which multimedia including images, video, and audio can be generated by AI by 50 times compared to traditional diffusion models, generating images in nearly a 10th of a second compared to more than 5 seconds for regular diffusion. With the introduction of sCM, OpenAI has managed to achieve comparable sample quality with only two sampling steps, offering a solution that accelerates the generative process without compromising on quality. Described in the pre-peer reviewed paper published on arXiv.org and blog post released today, authored by Cheng Lu and Yang Song, the innovation enables these models to generate high-quality samples in just two steps—significantly faster than previous diffusion-based models that require hundreds of steps. Song was also a leading author on a 2023 paper from OpenAI researchers including former chief scientist Ilya Sutskever that coined the idea of “consistency models,” as having “points on the same trajectory map to the same initial point.” While diffusion models have delivered outstanding results in producing realistic images, 3D models, audio, and video, their inefficiency in sampling—often requiring dozens to hundreds of sequential steps—has made them less suitable for real-time applications. Theoretically, the technology could provide the basis for a near-realtime AI image generation model from OpenAI. As fellow VentureBeat reporter Sean Michael Kerner mused in our internal Slack channels, “can DALL-E 4 be far behind?” Faster sampling while retaining high quality In traditional diffusion models, a large number of denoising steps are needed to create a sample, which contributes to their slow speed. In contrast, sCM converts noise into high-quality samples directly within one or two steps, cutting down on the computational cost and time. OpenAI’s largest sCM model, which boasts 1.5 billion parameters, can generate a sample in just 0.11 seconds on a single A100 GPU. This results in a 50x speed-up in wall-clock time compared to diffusion models, making real-time generative AI applications much more feasible. Reaching diffusion-model quality with far less computational resources The team behind sCM trained a continuous-time consistency model on ImageNet 512×512, scaling up to 1.5 billion parameters. Even at this scale, the model maintains a sample quality that rivals the best diffusion models, achieving a Fréchet Inception Distance (FID) score of 1.88 on ImageNet 512×512. This brings the sample quality within 10% of diffusion models, which require significantly more computational effort to achieve similar results. Benchmarks reveal strong performance OpenAI’s new approach has undergone extensive benchmarking against other state-of-the-art generative models. By measuring both the sample quality using FID scores and the effective sampling compute, the research demonstrates that sCM provides top-tier results with significantly less computational overhead. While previous fast-sampling methods have struggled with reduced sample quality or complex training setups, sCM manages to overcome these challenges, offering both speed and high fidelity. The success of sCM is also attributed to its ability to scale proportionally with the teacher diffusion model from which it distills knowledge. As both the sCM and the teacher diffusion model grow in size, the gap in sample quality narrows further, and increasing the number of sampling steps in sCM reduces the quality difference even more. Applications and future uses The fast sampling and scalability of sCM models open new possibilities for real-time generative AI across multiple domains. From image generation to audio and video synthesis, sCM provides a practical solution for applications that demand rapid, high-quality output. Additionally, OpenAI’s research hints at the potential for further system optimization that could accelerate performance even more, tailoring these models to the specific needs of various industries. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More ChatGPT’s Canvas feature allows users to edit the chatbot’s responses on the app rather than copying and pasting them to a separate document.  However, when Canvas launched in early October for its paid tiers, it didn’t let people see what changes GPT-4o made to its responses. OpenAI’s latest update to the feature corrects that.  The show changes button will show the most recent changes to either the generated text or code on Canvas. It will highlight added information in green and deleted sections in red.  Tracking changes has always been a good feature of any editing platform; Google Docs and Word documents offer a toggle for users to check what’s been changed. But OpenAI had been planning to roll out updates to Canvas slowly as ChatGPT subscribers get used to it.  Canvas already offers familiar features like comments, where users can add suggestions or give more instructions for the AI model to follow when editing responses.  Canvas is still only available on the web version of ChatGPT for ChatGPT Plus, Teams, Enterprise and Edu users. Mac app users and anyone downloading the recently released Windows version of ChatGPT will have to wait until Canvas is rolled out to these standalone apps.  Currently, people can access Canvas on the regular ChatGPT window rather than in any custom GPTs.  A much requested feature OpenAI’s developer X account acknowledged that developer customers have requested a track or show change feature since Canvas launched.  But while many developers said this was a step in the right direction, Canvas still doesn’t immediately connect to code repositories like GitHub or let users visually see how the edited code works.  This is one area where ChatGPT competitor Claude from Anthropic and its Artifacts feature excels. Artifacts function much like Canvas; users can begin a prompt on the Claude chat interface.  When users launch Artifacts, a dedicated window opens where they can manipulate the model’s responses. Artifacts let users replicate websites using the code Claude just generated and edited, so developers can see not only which lines of code have changed but also whether it worked. Artifacts are now available to all Claude users, including those on mobile devices.  Canvas and Artifacts represent what could be the next phase in the evolution of AI chat platforms and assistants. The Interface War could see other platforms begin to explore how to keep users in the platform instead of opening other dedicated windows for different tasks.  source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Google will soon offer a paid version of its AI research tool NotebookLM, specifically targeting businesses.  NotebookLM Business will have “enhanced features for businesses, universities, and organizations.” For now, access to NotebookLM Business is through a pilot program for early access to its features, training and email support.  Google told VentureBeat in an email that participants in the NotebookLM Business pilot “will gain a significant advantage with enhanced capabilities designed to boost productivity and collaboration.” These capabilities include higher usage limits and new features such as customization and sharing notebooks with team members. The company said these features could unlock new use cases for businesses using the tool. “We’ve seen this early feature streamline onboarding, shared understanding of complex projects, and building a centralized repository of your team’s collective intelligence all within a collaborative notebook environment,” a spokesperson said.  Another feature that will be part of NotebookLM Business is Audio Overview, which lets users create a narrated study guide. Google said the paid version will continue to have robust data privacy and security.  NotebookLM, built with Gemini 1.5, lets people upload source material to “notebooks” to gather information and ask the Gemini chatbot questions about the research. First announced in July last year, NotebookLM became generally available in December.  Google will also remove the “experimental” tag on the tool.  NotebookLM product manager Raiza Martin previously told VentureBeat that the team saw many different uses for the platform, including some for enterprises. While NotebookLM was never intended for a specific audience, Martin said many researchers and students embraced the product. Many businesses have also begun using NotebookLM as a repository of information for teams.  Google will announce general availability and pricing for NotebookLM Business later this year.  Additional control over audio  Along with announcing NotebookLM Business, Google updated the Audio Overview feature of NotebookLM. Audio Overview lets people generate podcasts about their research. Google characterized Audio Overview as a spoken research or study guide rather than a podcast. However, its first version featured two voices (one male, one female) conversing about the information in the notebook, reminding many of podcasts.  Audio Overview proved popular among some users, with many posting their generated audio on social media. Martin had previously promised additional controls over Audio Overview and said the company’s research showed conversations helped people retain more information. Users can now guide more of the conversation of Audio Overview, including prompting the model to focus on specific topics or levels of expertise. Audio Overviews will also continue to play while users query their sources or ask questions with its chat feature.  I got to explore the updated capabilities of Audio Overviews early. In a notebook with sources around AI Orchestration, I told it to focus on the definition of orchestration and how different frameworks like LangChain work. The final product did talk about AI orchestration based on the different blog posts and YouTube videos I had uploaded. The two “hosts,” however, spoke about frameworks as if LangChain was the only orchestration framework out there. This might be a misunderstanding of my prompt where I specifically named LangChain because the source documents definitely talk about available tools.  Google does point out that Audio Overviews “are generated discussions and are not a comprehensive or objective view of a topic.” It only takes into account information found in the uploaded source materials.  Open NotebookLM, an open-source competitor to NotebookLM, launched last month and included an audio recap function. While Open NotebookLM does not have the same fact-checking capabilities as NotebookLM, it represents a shift in the ease of deploying complex AI-driven platforms.  source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More AI agents hold much promise, with some saying they will revolutionize the workplace itself.  But they can be a bit concept-y, and enterprises don’t always know where to begin.  One-year-old startup CrewAI has quickly become one of the most popular AI agent frameworks — it’s used by the likes of AI pioneer Andrew Ng, among many other leading companies — as it simplifies the building and deployment of multi-agent systems.  Today, the company is launching its first — and highly-anticipated — product to market, CrewAI Enterprise. The platform, which has been in beta for some months, enables users to build, deploy and iterate multi-agent “crews.” The company is also announcing an $18 million funding round.  CrewAI founder and CEO João Moura called the opportunity for AI agents “immense.”  “The AI agent is basically right now an LLM that doesn’t need to be part of conversations,” he told VentureBeat. “Instead of a conversation, you give it a task, and it has the agency to autonomously decide what to do and when to do it.” ‘The simpler the better,’ open-source critical According to Markets and Markets, the AI agent industry will grow dramatically in the next five years, from $5 billion this year to nearly $50 billion by 2030. Capgemini reports that 10% of large enterprises are already using AI agents, more than half plan to use them in the next year and 82% will adopt them within the next three years.  “Agents are the big thing everyone’s talking about right now,”said Moura. “The genie is not going back into the bottle. People want this to happen.” CrewAI, which was just founded in 2023, has already established itself as one of the most popular agent frameworks, competing with the likes of Langraph and Autogen. Moura noted that enterprises are more and more quickly moving from AI agent conception to use cases.  “What we’re noticing is that companies are graduating way faster than we expected,” he said.  The company’s new platform is built on top of its popular open-source framework and enables organizations to build crews of AI agents using any large language model (LLM) or cloud platform. Users can plan and build multi-agent systems; securely deploy those agents into a production environment with custom levels of access and control; and iterate and track ROI with testing and training tools.  When getting started with AI agents, “the simpler the better,” said Moura. That’s what sets CrewAI apart, he said; they also have “doubled down” on educating people.  “It’s a brand new category and market,” he said. “People are trying to understand how they should go about this, they want to be educated.” He noted of other competing projects, “it’s almost like they’re trying to make it complex on purpose.” CrewAI also has significant open-source traction. The company has a “very opinionated” view on how agents work. Open-source is an instrumental part of how the world builds software, Moura noted and is an “amazing distribution channel.” “The world runs on open-source, every software out there uses open-source libraries,” he said. “We don’t want to be in a world where all models are closed source, you don’t know what’s going on, you’re locked in with all these vendors.” Use cases from internal processes to marketing Moura pointed to a “big array of use cases” for agentic AI overall and called CrewAI’s offering “such a cross vertical product.” The most common use cases are around internal automations, marketing and coding, he noted. Agents can perform research, summarization and reporting, and can also help with legal analysis.  For instance, one Fortune 500 customer with consumer-facing products was looking to update legacy projects and apps (including Java and SAP). They were able to build agents that can update and test code themselves before passing them off for final review by a human engineer. They are saving hundreds of thousands of dollars as a result (by their own estimate), said Moura.  “Marketing’s another interesting one,” he noted. Agents can develop leads by interacting with  large, instant sources of information. Or, in the case of real estate companies, agents can monitor markets, produce leads and advise agents on buy-or-rent scenarios.  Moura pointed to a big beverage company that used CrewAI to build agents that handle internal requests from a portal accessible by thousands of employees. A series of very specific rules need to be reviewed before internal requests can be approved, Moura explained; agents understand and review those rules, reply to requests (whether they were approved or not or if more info is needed). Robotic process automation (RPA) systems then take over to port stored information into the company’s database.  Getting even more complex, Moura said CrewAI’s platform has been used by a big media company that fine-tuned models to act like movie directors: They can cut frames and add subtitles and music, then automatically push out to social media.  “People are always pushing the cutting edge,” said Moura. Millions of agents, significant traction among Fortune 500 CrewAI’s open-source platform executes 10 million-plus agents a month, and the company claims it is already being used by nearly half of the Fortune 500. It signed its first 150 beta enterprise customers in less than six months.  “I gotta say it has been insane. I think we’re one of the fastest-growing projects out there,” said Moura. “It’s very intense and very humbling.” Crew AI’s inception round was led by boldstart ventures, and its series A was led by Insight Partners. Additional funding comes from Blitzscaling Ventures, Craft Ventures, Earl Grey Capital and several top angels including Ng and Dharmesh Shah, co-founder and CTO of HubSpot.  Ng noted in a statement: “CrewAI makes it easy and fast to develop both simple and complex multi-agent AI workflows. Its powerful orchestration features for enterprises — including memory and self-healing — help businesses go well beyond traditional automation.” source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Canadian AI image startup Ideogram, founded last year by former AI researchers from Google Brain, has made a new for itself among AI creators with its text-to-image models that produce a wide range of styles from realistic to fantastical, and most impressively of all, highly accurate text baked into the image itself (something other leading image generators, including Midjourney, took a while to implement and still struggle to generate reliably). Now, it’s getting in on the trend of expanding its web-based user workspace to include a new interactive, infinite Canvas where users can spread newly generated images out, compare them to older generations, resize and reorder them at will, and even combine multiple AI generated images into one new composite. It also allows users to upload their own visuals. With this addition, Ideogram Canvas aims to streamline workflows and offer flexible tools to refine creative projects step-by-step. But of course, Ideogram is far from the only AI company to go beyond the simple chatbot-style text entry interface. Earlier this month, OpenAI launched an experimental new “Canvas” view for ChatGPT. Unlike Ideogram’s version, it doesn’t help with imagery. Rather, OpenAI’s version offers the ability to see text-based documents and code alongside the chat interface, and watch as their chat conversation changes the resulting output in the “Canvas” view to the right. Moreover, Ideogram’s “Canvas” view is closely reminiscent of an approach that was pioneered last year by a startup called Visual Electric, which uses open source Stable Diffusion AI image generation models, and which recently launched a mobile app. However, Ideogram trains and offers its own proprietary, ground-up image generation models such as the recently launched Ideogram 2.0, which sets it apart. Magic fill and Extend Alongside Canvas, Ideogram is also debuting two additional new features: Magic Fill and Extend. Magic Fill allows users to edit specific regions of an image by replacing objects, adding text, changing backgrounds, or fixing imperfections. The tool enables users to focus on particular areas of an image and generate high-resolution details with a simple text prompt. Extend helps users expand images beyond their original borders, keeping a consistent style. This tool is useful for resizing images, adjusting composition, or adapting content to different screen formats without losing the original structure. These tools, designed to complement each other, give users the ability to make extensive edits or modifications to images while maintaining the overall quality and coherence of the content. Subscription plans and features Ideogram Canvas is available with all of Ideogram’s various usage tiers, though naturally paid plans get you more perks and features and fewer limitations. In fact, the company posted a thread on its social account on the network X (formerly Twitter) that noted all paid plans receive unlimited Canvases. The pricing for the various options is as follows: Free Plan: Allows up to 40 images per day with 10 slow credits, access to 2 canvases, and basic features like text-to-image generation and compressed image downloads. Basic Plan ($7 USD/month, billed annually): Offers 400 priority credits per month, 100 slow credits per day, unlimited canvases, and access to Magic Fill and Extend, along with features such as PNG downloads and customizable aspect ratios. Plus Plan ($16 USD/month, billed annually): Adds 1,000 priority credits, unlimited slow credits, image uploads, private generation, and additional customization options. Pro Plan ($48 USD/month, billed annually): Includes 3,000 priority credits per month, support for up to 12,000 images, and an upcoming bulk generation feature with CSV integration. Furthermore, Ideogram offers its own API that developers can use to build third-party apps atop, yet this offers only the new Magic Fill and Extend features rather than the Canvas (which makes sense, since it is highly integrated into and dependent upon Ideogram’s website design). Pricing for accessing the models through the API ranges from $0.01 per input to simply describe images to $0.08 per input for image generations with Ideogram 2. Ideogram credits part of the development of Ideogram Canvas to its community of beta testers and members of the Ideogram Creators Club, who provided feedback during the platform’s testing phase. The company acknowledges their contributions in refining the platform’s functionality and design. Expanding its teams As part of its broader growth strategy, Ideogram also noted it is expanding its teams and open to hires in Toronto and New York City. The company is actively recruiting for various roles across AI research, engineering, marketing, and finance to continue developing its suite of AI tools. Interested candidates can apply via the company’s jobs page. With the launch of Ideogram Canvas, the company seeks to offer a platform that blends user-generated content with AI-assisted tools like Magic Fill and Extend. By making it easier to create and modify images, Ideogram aims to support creators in a wide variety of industries. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Content Warning: This article covers suicidal ideation and suicide. If you are struggling with these topics, reach out to the National Suicide Prevention Lifeline by phone: 1-800-273-TALK (8255). Character AI, the artificial intelligence startup whose co-creators recently left to join Google following a major licensing deal with the search giant, has imposed new safety and auto moderation policies today on its platform for making custom interactive chatbot “characters” following a teen user’s suicide detailed in a tragic investigative article in The New York Times. The family of the victim is suing Character AI for his death. Character’s AI statement after tragedy of 14-year-old Sewell Setzer “We are heartbroken by the tragic loss of one of our users and want to express our deepest condolences to the family,” reads part of a message posted today, October 23, 2024, by the official Character AI company account on the social network X (formerly Twitter), linking to a blog post that outlines new safety measures for users under age 18, without mentioning the suicide victim, 14-year-old Sewell Setzer III. As reported by The New York Times, the Florida teenager, diagnosed with anxiety and mood disorders, died by suicide on February 28, 2024, following months of intense daily interactions with a custom Character AI chatbot modeled after Game of Thrones character Daenerys Targaryen, to whom he turned to for companionship, referred to as his sister and engaged in sexual conversations. In response, Setzer’s mother, lawyer Megan L. Garcia, filed a lawsuit against Character AI and Google parent company Alphabet yesterday in U.S. District Court of the Middle District of Florida for wrongful death. Photos of Setzer and his mother over the years. Credit: Megan Garcia/Bryson Gillette A copy of Garcia’s complaint demanding a jury trial provided to VentureBeat by public relations consulting firm Bryson Gillette is embedded below: The incident has sparked concerns about the safety of AI-driven companionship, particularly for vulnerable young users. Character AI has more than 20 million users and 18 million custom chatbots created, according to Online Marketing Rockstars (OMR). The vast majority (53%+) are between 18-24 years old, according to Demand Sage, though there are no categories broken out for under 18. The company states that its policy is only to accept users age 13 or older and 16 or older in the EU, though it is unclear how it moderates and enforces this restriction. Character AI’s current safety measures In its blog post today, Character AI states: “Over the past six months, we have continued investing significantly in our trust & safety processes and internal team. As a relatively new company, we hired a Head of Trust and Safety and a Head of Content Policy and brought on more engineering safety support team members. This will be an area where we continue to grow and evolve.  We’ve also recently put in place a pop-up resource that is triggered when the user inputs certain phrases related to self-harm or suicide and directs the user to the National Suicide Prevention Lifeline.” New safety measures announced In addition, Character AI has pledged to make the following changes to further restrict and contain the risks on its platform, writing: “Moving forward, we will be rolling out a number of new safety and product features that strengthen the security of our platform without compromising the entertaining and engaging experience users have come to expect from Character.AI. These include:  Changes to our models for minors (under the age of 18) that are designed to reduce the likelihood of encountering sensitive or suggestive content. Improved detection, response, and intervention related to user inputs that violate our Terms or Community Guidelines.  A revised disclaimer on every chat to remind users that the AI is not a real person. Notification when a user has spent an hour-long session on the platform with additional user flexibility in progress.“ As a result of these changes, Character AI appears to be deleting certain user-made custom chatbot characters abruptly. Indeed, the company also states in its post: “Users may notice that we’ve recently removed a group of Characters that have been flagged as violative, and these will be added to our custom blocklists moving forward. This means users also won’t have access to their chat history with the Characters in question.” Users balk at changes they see as restriction AI chatbot emotional output Though Character AI’s custom chatbots are designed to simulate a wide range of human emotions based on the user-creator’s stated preferences, the company’s changes to further align the range of outputs away from risky content is not going over well with some self-described users. As captured in screenshots posted to X by AI news influencer Ashutosh Shrivastava, the Character AI subreddit is filled with complaints. As one Redditor (Reddit user) under the name “Dqixy,” posted in part: “Every theme that isn’t considered “child-friendly” has been banned, which severely limits our creativity and the stories we can tell, even though it’s clear this site was never really meant for kids in the first place. The characters feel so soulless now, stripped of all the depth and personality that once made them relatable and interesting. The stories feel hollow, bland, and incredibly restrictive. It’s frustrating to see what we loved turned into something so basic and uninspired.“ Another Redditor, “visions_of_gideon_” was even more harsh, writing in part: “Every single chat that I had in a Targaryen theme is GONE. If c.ai is deleting all of them FOR NO FCKING REASON, then goodbye! I am a fcking paying for c.ai+, and you delete bots, even MY OWN bots??? Hell no! I am PISSED!!! I had enough! We all had enough! I am going insane! I had bots that I have been chatting with for MONTHS. MONTHS! Nothing inappropriate! This is my last straw. I am not only deleting my subscription, I am ready to delet c.ai!“ Similarly, the Character AI Discord server‘s feedback channel is

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Noam Brown, a leading research scientist at OpenAI, took the stage at the TED AI conference in San Francisco on Tuesday to deliver a powerful speech on the future of artificial intelligence, with a particular focus on OpenAI’s new o1 model and its potential to transform industries through strategic reasoning, advanced coding, and scientific research. Brown, who has previously driven breakthroughs in AI systems like Libratus, the poker-playing AI, and CICERO, which mastered the game of Diplomacy, now envisions a future where AI isn’t just a tool, but a core engine of innovation and decision-making across sectors. “The incredible progress in AI over the past five years can be summarized in one word: scale,” Brown began, addressing a captivated audience of developers, investors, and industry leaders. “Yes, there have been uplink advances, but the frontier models of today are still based on the same transformer architecture that was introduced in 2017. The main difference is the scale of the data and the compute that goes into it.” Brown, a central figure in OpenAI’s research endeavors, was quick to emphasize that while scaling models has been a critical factor in AI’s progress, it’s time for a paradigm shift. He pointed to the need for AI to move beyond sheer data processing and into what he referred to as “system two thinking”—a slower, more deliberate form of reasoning that mirrors how humans approach complex problems. The psychology behind AI’s next big leap: Understanding system two thinking To underscore this point, Brown shared a story from his PhD days when he was working on Libratus, the poker-playing AI that famously defeated top human players in 2017. “It turned out that having a bot think for just 20 seconds in a hand of poker got the same boosting performance as scaling up the model by 100,000x and training it for 100,000 times longer,” Brown said. “When I got this result, I literally thought it was a bug. For the first three years of my PhD, I had managed to scale up these models by 100x. I was proud of that work. I had written multiple papers on how to do that scaling, but I knew pretty quickly that all that would be a footnote compared to this scaling up system two thinking.” Brown’s presentation introduced system two thinking as the solution to the limitations of traditional scaling. Popularized by psychologist Daniel Kahneman in the book Thinking, Fast and Slow, system two thinking refers to a slower, more deliberate mode of thought that humans use for solving complex problems. Brown believes incorporating this approach into AI models could lead to major performance gains without requiring exponentially more data or computing power. He recounted that allowing Libratus to think for 20 seconds before making decisions had a profound effect, equating it to scaling the model by 100,000x. “The results blew me away,” Brown said, illustrating how businesses could achieve better outcomes with fewer resources by focusing on system two thinking. Inside OpenAI’s o1: The revolutionary model that takes time to think Brown’s talk comes shortly after the release of OpenAI’s o1 series models, which introduce system two thinking into AI. Launched in September 2024, these models are designed to process information more carefully than their predecessors, making them ideal for complex tasks in fields like scientific research, coding, and strategic decision-making. “We’re no longer constrained to just scaling up the system one training. Now we can scale up the system two thinking as well, and the beautiful thing about scaling up in this direction is that it’s largely untapped,” Brown explained. “This isn’t a revolution that’s 10 years away or even two years away. It’s a revolution that’s happening now.” The o1 models have already demonstrated strong performance in various benchmarks. For instance, in a qualifying exam for the International Mathematics Olympiad, the o1 model achieved an 83% accuracy rate—a significant leap from the 13% scored by OpenAI’s GPT-4o. Brown noted that the ability to reason through complex mathematical formulas and scientific data makes the o1 model especially valuable for industries that rely on data-driven decision-making. The business case for slower AI: Why patience pays off in enterprise solutions For businesses, OpenAI’s o1 model offers benefits beyond academic performance. Brown emphasized that scaling system two thinking could improve decision-making processes in industries like healthcare, energy, and finance. He used cancer treatment as an example, asking the audience, “Raise your hand if you would be willing to pay more than $1 for a new cancer treatment… How about $1,000? How about a million dollars?” Brown suggested that the o1 model could help researchers speed up data collection and analysis, allowing them to focus on interpreting results and generating new hypotheses. In energy, he noted that the model could accelerate the development of more efficient solar panels, potentially leading to breakthroughs in renewable energy. He acknowledged the skepticism about slower AI models. “When I mention this to people, a frequent response that I get is that people might not be willing to wait around for a few minutes to get a response, or pay a few dollars to get an answer to the question,” he said. But for the most important problems, he argued, that cost is well worth it. Silicon Valley’s new AI race: Why processing power isn’t everything OpenAI’s shift toward system two thinking could reshape the competitive landscape for AI, especially in enterprise applications. While most current models are optimized for speed, the deliberate reasoning process behind o1 could offer businesses more accurate insights, particularly in industries like finance and healthcare. In the tech sector, where companies like Google and Meta are heavily investing in AI, OpenAI’s focus on deep reasoning sets it apart. Google’s Gemini AI, for instance, is optimized for multimodal tasks, but it remains to be seen how it will compare to OpenAI’s models in terms of problem-solving capabilities. That said, the

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More The number of platforms being released to help enterprises integrate AI agents into their technology stack is not slowing down as the year winds down.  Work management platform Asana already has an AI agent service, but with its new AI Studio feature, the company wants its customers to think about AI agents as part of their larger workflow. AI Studio lets users build workflows on Asana and then deploy multiple custom AI agents directly on the workflow. Customers can create the agents without code and allow them to “take on the busywor,” including handling project coordination.  “The difference here is that we’ve opened up the toolkit to create agents to the folks who build workflows at companies, the folks that orchestrate a large body of work,” said Alex Hood, Asana’s chief product officer, in an interview with VentureBeat. “We can now bring agents to all the places where teams show up to hand off work, inserting an AI agent to take work off people’s plates.”  Hood cited a recent Asana 2024 State of Work Innovation Report that showed 53% of an employee’s time is spent on “busy work,” with unproductive meetings doubling since 2019. By bringing in agents, Hood said, teams are freed from doing tasks crucial to a workflow, such as intake for some marketing teams, to focus on other important work.  AI Studio, which will be a tool integrated into Asana but with an additional fee for access, is built on Asana’s Work Graph data mode, which tracks cross-functional work in an organization.  “There are other platforms who are building AI agents and are doing it on top of places where they have specialties,” Hood said. “For us, our specialty is the Work Graph, the place where work happens and the workflows powering that work leverages the right data.” Customers saw an improvement One of Asana’s first customers to use AI Studio is the financial data company Morningstar.  Hood said Morningstar used AI Studio to centralize IT project requests to streamline the workflow to evaluate new projects. Belinda Hardman, director of Program Management at Morningstar, said in a press release that the new workflow helped the company “eliminate time spent on manual back-and-forth because Asana AI identifies and captures the information we need right off the bat.” Islands of AI agents Agents have become the hot topic in AI this year, with several companies announcing either a platform to customize agents or to access a library of ready-made agents.  To name a few: Microsoft announced it will release a suite of AI agents for its Dynamics 365 service this week. Salesforce released Agentforce last month, and ServiceNow launched its agent library on its Now Assist platform. Asana’s earlier released agentic system joins Agentforce, and other ready-made agents from other service providers will be integrated into Slack.  “The things we might have dreamed of and were talking about a couple of years ago are playing out now because the models are getting that much better,” Hood said. “But the models can’t create great agents on their own. They need to be hooked into software and we, as builders, have gotten good at figuring out how to best integrate deeply AI capabilities.” However, many of these agents — even those embedded in third-party applications like Slack — still function as individual islands of agents talking to other agents built on the same platform. The next frontier for agents coming from workflow systems or other enterprise-focused software will be the ability to communicate with other agents elsewhere.  We’re not there yet, but as more enterprises become comfortable with AI agents and begin deploying these into their organizations, that future may come soon enough.  source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More SambaNova Systems and Gradio have unveiled a new integration that allows developers to access one of the fastest AI inference platforms with just a few lines of code. This partnership aims to make high-performance AI models more accessible and speed up the adoption of artificial intelligence among developers and businesses. “This integration makes it easy for developers to copy code from the SambaNova playground and get a Gradio web app running in minutes with just a few lines of code,” Ahsen Khaliq, ML Growth Lead at Gradio, said in an interview with VentureBeat. “Powered by SambaNova Cloud for super-fast inference, this means a great user experience for developers and end-users alike.” The SambaNova-Gradio integration enables users to create web applications powered by SambaNova’s high-speed AI models using Gradio’s gr.load() function. Developers can now quickly generate a chat interface connected to SambaNova’s models, making it easier to work with advanced AI systems. A snippet of Python code demonstrates the simplicity of integrating SambaNova’s AI models with Gradio’s user interface. Just a few lines are needed to launch a powerful language model, underscoring the partnership’s goal of making advanced AI more accessible to developers. (Credit: SambaNova Systems) Beyond GPUs: The rise of dataflow architecture in AI processing SambaNova, a Silicon Valley startup backed by SoftBank and BlackRock, has been making waves in the AI hardware space with its dataflow architecture chips. These chips are designed to outperform traditional GPUs for AI workloads, with the company claiming to offer the “world’s fastest AI inference service.” SambaNova’s platform can run Meta’s Llama 3.1 405B model at 132 tokens per second at full precision, a speed that is particularly crucial for enterprises looking to deploy AI at scale. This development comes as the AI infrastructure market heats up, with startups like SambaNova, Groq, and Cerebras challenging Nvidia’s dominance in AI chips. These new entrants are focusing on inference — the production stage of AI where models generate outputs based on their training — which is expected to become a larger market than model training. SambaNova’s AI chips show 3-5 times better energy efficiency than Nvidia’s H100 GPU when running large language models, according to the company’s data. (Credit: SambaNova Systems) From code to cloud: The simplification of AI application development For developers, the SambaNova-Gradio integration offers a frictionless entry point to experiment with high-performance AI. Users can access SambaNova’s free tier to wrap any supported model into a web app and host it themselves within minutes. This ease of use mirrors recent industry trends aimed at simplifying AI application development. The integration currently supports Meta’s Llama 3.1 family of models, including the massive 405B parameter version. SambaNova claims to be the only provider running this model at full 16-bit precision at high speeds, a level of fidelity that could be particularly attractive for applications requiring high accuracy, such as in healthcare or financial services. The hidden costs of AI: Navigating speed, scale, and sustainability While the integration makes high-performance AI more accessible, questions remain about the long-term effects of the ongoing AI chip competition. As companies race to offer faster processing speeds, concerns about energy use, scalability, and environmental impact grow. The focus on raw performance metrics like tokens per second, while important, may overshadow other crucial factors in AI deployment. As enterprises integrate AI into their operations, they will need to balance speed with sustainability, considering the total cost of ownership, including energy consumption and cooling requirements. Additionally, the software ecosystem supporting these new AI chips will significantly influence their adoption. Although SambaNova and others offer powerful hardware, Nvidia’s CUDA ecosystem maintains an edge with its wide range of optimized libraries and tools that many AI developers already know well. As the AI infrastructure market continues to evolve, collaborations like the SambaNova-Gradio integration may become increasingly common. These partnerships have the potential to foster innovation and competition in a field that promises to transform industries across the board. However, the true test will be in how these technologies translate into real-world applications and whether they can deliver on the promise of more accessible, efficient, and powerful AI for all. source