VentureBeat

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More 2024 has been a banner year for Perplexity. The AI search startup, founded by former DeepMind and OpenAI researcher Aravind Srinivas, raised hundreds of millions of dollars — its latest funding round reportedly valuing the company at $9 billion — and introduced several notable features, including Pages, Spaces, and innovative shopping experiences. These developments have solidified Perplexity’s reputation as an “AI-first” knowledge discovery engine, standing apart from traditional search giants like Google and Bing, which are bolting AI capabilities onto their existing engines. However, the journey is far from over.  Facing intensifying competition, Perplexity is broadening its scope with a new addition to its portfolio: Carbon. The company has just acquired this startup, for an undisclosed sum, to address the “data gap” enterprises encounter with AI search and streamline the knowledge discovery process in their workflows. Carbon has developed a comprehensive retrieval framework that streamlines the process of connecting external data sources to LLMs. Users can tap the Carbon universal API or SDKs to sync their data sources and retrieve the data to use with LLMs. It offers native integrations with over 20 data connectors and supports more than 20 file formats, including text, audio and video files. The expanding scope of AI search From individuals to business users, almost everyone today uses AI search as part of their workflows. The idea of the technology is pretty simple — you don’t have to go through a swathe of links and content to find relevant insights and information. Instead, the information will come to you as the direct answer to your query. Perplexity has thrived on this approach, using a range of large language models to retrieve information from the web and simplifying how users work. It even allows teams to extract information from their personal or business files such as PDFs and Word documents.  But, here’s the thing. The web is home to public information, and uploading internal files — PDFs, conversations, images — individually is not feasible for business users dealing with large volumes of proprietary data. This affects the quality of answers, keeping them generic and devoid of important organization-relevant contexts. Highlighting this “data gap,” Sanjeev Mohan, the former Gartner Research VP for data and analytics, told VentureBeat that one of the biggest AI trends for 2025 will be ETL for unstructured data. It will allow teams to extract and transform data from dispersed internal sources, ultimately powering their LLMs to generate highly relevant and accurate responses. Now, this is exactly what Perplexity plans to do with the acquisition of Carbon’s comprehensive, streamlined retrieval framework. Perplexity will integrate Carbon’s retrieval engine and connectors into its tech stack, giving users of the search platform a direct way to plug in their diverse sources of data, from Google Docs and Notion to Hubspot and Slack.  This, the company says, will expand the knowledge pool powering the AI search engine, making its responses more comprehensive, relevant and personalized to users.  What can users expect from Carbon-powered Perplexity? While Perplexity has just acquired Carbon and the integration is yet to be executed, it’s pretty easy to imagine how the additional data connectors will improve the workflows of enterprise teams using the AI search engine. For instance, if one has to move the date for a launch and needs to figure out the latest deadline and guidelines set by their team, Perplexity would be able to parse through all the data in Google Docs, Notion, and Slack — and make necessary correlations — to find the information that answers the question.  In essence, there would be no more worrying about stitching together context from the web, individual apps, and messages. The platform does everything on its own to provide the answer. “The notable benefit of this setup is that our technology can find the answer without making you pinpoint the document/database where that information is stored,” Sara Platnick, who leads communications at Perplexity, told VentureBeat.  Another example, she said, could be extracting customer meeting insights. Perplexity would be able to fetch the details and focus of the conversation from connected CRMs in no time.  Notably, by leveraging Carbon’s retrieval-augmented generation (RAG) workflows, Perplexity is making enterprise search more accessible, saving companies the hassle of building their own RAG pipelines from scratch. “By finding and interpreting proprietary data with Perplexity and Carbon, companies can address a range of multi-faceted gen AI use cases. We find the leading adopters are most focused on customer service, document processing, image processing and recommendation engines, Kevin Petrie,” VP of research at BARC US, told VentureBeat. Execution will be key Acquiring Carbon is just the beginning. The real key will be execution, or how seamlessly and safely the startup’s tech is integrated. After all, we are talking about proprietary data from some of the most critical knowledge repositories that enterprises maintain. “Companies are rightly wary of exposing their intellectual property to the public. So Perplexity and Carbon will need to provide governance controls that ensure companies can keep their data inside their own firewalls. They have no interest in sharing secrets or training a public model to mimic their intellectual property,” Petrie added. On Perplexity’s part, Platnick noted that “all information from internal and private sources on the engine is encrypted, as is all data transmitted and stored in Carbon’s data connectors.” She also pointed out that the company has additional protections to ensure that private documents stay private and aren’t accessible to non-authorized users. As of now, there’s no specific timeline for the integration of Carbon with Perplexity. However, the startup will cease operations of its managed API on March 31, 2025. Existing customers using the API have already been notified for offboarding, with the Carbon team assisting them in the transition. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Salt AI, an AI workflow orchestration firm for enterprises, raised $3 million in seed funding and appointed Aber Whitcomb as CEO. Salt AI offers a unified AI collaboration environment, dubbed Salt, where organizations can securely connect their firewalled data to build AI automations, agentic workflows and bespoke AI solutions. With avisual drag and drop interface, and full-code capabilities, every member of an organization can collaborate in real time to build powerful AI on the Salt platform. Teams can deploy in one click to Salt’s cloud infrastructure that autoscales to meet the real-time needs of any use case. “We’re at an inflection point where AI can transform how companies operate, but only if we make it truly accessible and actionable,” said Whitcomb in a statement. “Salt’s platform enables teams to create powerful AI agents and workflows that automate complex tasks and drive real business impact. I’m excited to lead Salt as we help organizations build and scale their AI capabilities.” Aber Whitcomb is CEO of Salt. The investment will accelerate development of Salt’s proprietary AI orchestration platform and expand its market presence. “We’re pleased to back the Salt AI team. Aber Whitcomb’s impressive track record of success in launching and scaling businesses, paired with the immense market opportunity makes this an exciting investment for us.” said Kristian Blaszczynski, partner at Morpheus, in a statement. “Very soon, AI will power almost every industry and Salt will be the engine on which enterprises execute.” Salt integrates with all major closed-source and open-source LLMs and supports diffusion models for generative art. Users can connect to 30+ enterprise data sources for both reading and writing, with new connections being released weekly. Command menu for Salt. Whitcomb and Jim Benedetto started Salt in Los Angeles in 2023. The company now has 16 people. Whitcomb pivoted into this business from PlaiDay, the generative AI social mobile app that first targeted consumers. Chris DeWolfe, who was the previous CEO, stayed with the Web3 gaming part of the company and renamed it Rough House Games. Benedetto, Whitcomb and Charlie Basil went with Salt. “We built the Salt platform as the backend to enable rapid feature development for PlaiDay, and ultimately realized we had solved all the major pain points for developing and deploying AI, and that taking our backend to market as a B2B SaaS solution was a better and bigger opportunity than the consumer app,” said Whitcomb in an email to VentureBeat. Asked about the competition, he said the space is very noisy, with a lot of tools using similar language to describe their feature set. “It does feel crowded at first blush,” Whitcomb said. “However, there are only a small number of significant competitors. Salt differentiates itself by enabling team collaboration on AI workflows. It does this by providing a robust and powerful visual-first toolkit that enables non-technical stakeholders (executives, product managers, designers, marketers, etc) to build AI; alongside a full-code toolkit that enables engineers to get down to the bare metal and have complete control over their solutions. Salt is the only platform that has fully featured solutions for both user types.” source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Arm Holdings sued Qualcomm over the alleged breach of a licensing agreement. But a jury today found it could not reach a conclusion on one of the allegations and found in favor of Qualcomm on another. Arm sued Qualcomm after Qualcomm acquired Arm licensee Nuvia for $1.4 billion. The jury could not come to an agreement on whether Nuvia breached its license agreement, but it also said that Qualcomm did not breach Nuvia’s license with Arm. “We are pleased with today’s decision,” Qualcomm said in a statement. “The jury has vindicated Qualcomm’s right to innovate and affirmed that all the Qualcomm products at issue in the case are protected by Qualcomm’s contract with ARM. We will continue to develop performance-leading, world class products that benefit consumers worldwide, with our incredible Oryon ARM-compliant custom CPUs.” The jury said that Qualcomm had properly licensed its central processor chips. Arm shares were down in after-hours trading, while Qualcomm shares were up slightly. The case held in U.S. federal court in Delaware could be tried again. The jury also found that Qualcomm’s chips, which were created using Nuvia’s own technology, are properly licensed under Nuvia’s agreement with Arm. That means Qualcomm can continue selling them. Those chips are helping Qualcomm move into the personal computer market. In a statement, Arm said, “We are disappointed that the jury was unable to reach consensus across the claims. We intend to seek a retrial due to the jury’s deadlock. From the outset, our top priority has been to protect Arm’s IP and the unparalleled ecosystem we have built with our valued partners over more than 30 years. As always, we are committed to fostering innovation in our rapidly evolving market and serving our partners while advancing the future of computing.”  source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More AI technology is making semiconductor leaders more optimistic about 2025, but headwinds could come from geopolitical and talent retention concerns. Those are some of the predictions in KPMG‘s 20th annual Global Semiconductor Outlook report from the U.S. audit, tax and advisory firm as well as the Global Semiconductor Alliance (GSA). About 92% of the semiconductor executives interviewed for the survey forecasted overall industry growth in 2025. With promises of ongoing demand for chips thanks to AI, cloud, data centers, wireless communication and automotive applications, new data from KPMG and GSA reveals significant optimism for 2025 among semiconductor executives. The KPMG Semiconductor Industry Confidence Index rose to 59, up from 54 in 2023, indicating increased optimism (a value above 50 indicates a more positive outlook than negative), and showcasing bolstered confidence across the following factors: company revenue growth, profitability growth, workforce growth, research and development (R&D) spending, and capital expenditures.  “AI underpins the industry’s near-term growth and revenue expectations,” said KPMG technology media and telecommunications leader Mark Gibson, in a statement. “The upward trajectory for the industry in the short term is clear, but the companies that can manage their supply chains and attract and retain talent will be the ones well-positioned to sustain and benefit from the AI boom.”  Despite widespread optimism, executives still anticipate significant challenges in 2025, including geopolitical territorialism — such as tariffs and trade restrictions — and ongoing talent issues within the industry. (President-elect Donald Trump is vowing to impose tariffs on his first day in office on January 20.) Strengthening supply-chain resilience and flexibility, along with enhancing talent development and retention, will be crucial as demand for chips continues to grow. Navigating this complex landscape in 2025 will require adaptive strategies.  About the survey Key facts about KPMG’s chip market survey for 2025 In the fourth quarter of 2024, KPMG and the GSA conducted the milestone 20th annual global semiconductor industry survey, capturing insights from 156 semiconductor executives about their outlook for the industry in 2025 and beyond. More than half of the respondents were from companies with $1 billion or more in annual revenue.  Semiconductor executives have positive outlooks for 2025 across all factors, with a five-point Confidence Index increase year-over-year (from 54 to 59). Interestingly, the smaller companies, defined as the organizations with less than $100 million in annual revenue, have the most positive outlook.  Across the board, all semiconductor companies have positive Confidence Index scores, with smaller companies displaying the most optimism for 2025, potentially seeing opportunities for rapid revenue increases due to their earlier stages of development.  Among those participating in the survey, there were 58 large companies ($1 billion or more in annual revenue); 54 mid-size companies ($100 million to $999 million in annual revenue); and 68 small companies (less than $100 million in annual revenue). Semiconductor executives are very optimistic about their company and overall industry revenue growth, with more than one-third predicting revenue growth by at least 10%.   The overwhelming majority (86%) anticipate their company’s revenue will grow in 2025, with almost half (46%) expecting growth to exceed 10%. And 92% forecast overall industry revenue growth, while one-third (36%) predict industry revenue growth of more than 10%.  For the first time in the history of the outlook, AI is the most important semiconductor revenue driver, displacing automotive, which held the top spot for the past two years.  As a result, microprocessors, including graphics processing units (GPUs) used for AI, are seen as the leading product opportunity for industry growth, ahead of memory and sensors/MEMs.  AI enablers, such as high-bandwidth memory, are the production technology that is projected to have the greatest impact on the industry over the next three years. Other key revenue drivers expected in 2025 include cloud/data centers (rose to second place), wireless communications (remained in third place), and automotive (dropped to fourth place, previously the top revenue driver).  Geopolitical concerns, particularly territorial tensions and trade restrictions like tariffs, are the most important issues shaping the industry’s supply chains. Talent risk remains a persistent concern as chip demand surges.  While territorialism (including tariffs and trade restrictions) is tied with talent risk as the biggest issues facing the industry over the next three years, territorialism is the clear-cut biggest issue among large companies with $1 billion or more in annual revenue.  Semiconductor executives surveyed view armed conflicts and tariffs as the two most concerning geopolitical matters that could affect the semiconductor ecosystem over the next two years. Government subsidies and the nationalization of semiconductor technology also rank near the top.  In response, semiconductor leaders are increasing geographic diversity to improve supply chain resiliency. Making the supply chain more flexible and adaptable to geopolitical changes (tied with talent development and retention) is the top strategic priority, after being named second in last year’s survey.   Executives are also on high alert for disruption as non-traditional semiconductor companies (tech giants, platform companies, and automotive companies) carve out their own place in the industry.   While most executives (39%) still view competition for talent as the primary impact to the industry over the next three years, the emergence of new competitors has become an almost equally significant concern among execs (35%), signaling a shift in the industry’s outlook.   To compare, last year only 19% of semiconductor execs cited the emergence of new competitors as a concern.   “Tech giants and established semiconductor players are starting to battle for market share, with ongoing technical developments and optimization of chips for AI aiming to enhance and provide alternatives for AI training and inferencing capabilities,” said KPMG global semiconductor leader Lincoln Clark, in a statement. “As the industry becomes more competitive, significant investments and cutting-edge strategies will be essential for companies to not only survive but thrive in this rapidly evolving landscape.”  The full Global Semiconductor Industry Outlook report will be released in early 2025. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Two years on from the public release of ChatGPT, conversations about AI are inescapable as companies across every industry look to harness large language models (LLMs) to transform their business processes. Yet, as powerful and promising as LLMs are, many business and IT leaders have come to over-rely on them and to overlook their limitations. This is why I anticipate a future where specialized language models, or SLMs, will play a bigger, complementary role in enterprise IT. SLMs are more typically referred to as “small language models” because they require less data and training time and are “more streamlined versions of LLMs.” But I prefer the word “specialized” because it better conveys the ability of these purpose-built solutions to perform highly specialized work with greater accuracy, consistency and transparency than LLMs. By supplementing LLMs with SLMs, organizations can create solutions that take advantage of each model’s strengths. Trust and the LLM ‘black box’ problem LLMs are incredibly powerful, yet they are also known for sometimes “losing the plot,” or offering outputs that veer off course due to their generalist training and massive data sets. That tendency is made more problematic by the fact that OpenAI’s ChatGPT and other LLMs are essentially “black boxes” that don’t reveal how they arrive at an answer.  This black box problem is going to become a bigger issue going forward, particularly for companies and business-critical applications where accuracy, consistency and compliance are paramount. Think healthcare, financial services and legal as prime examples of professions where inaccurate answers can have huge financial consequences and even life-or-death repercussions. Regulatory bodies are already taking notice and will likely begin to demand explainable AI solutions, especially in industries that rely on data privacy and accuracy. While businesses often deploy a “human-in-the-loop” approach to mitigate these issues, an over-reliance on LLMs can lead to a false sense of security. Over time, complacency can set in and mistakes can slip through undetected. SLMs = greater explainability Fortunately, SLMs are better suited to address many of the limitations of LLMs. Rather than being designed for general-purpose tasks, SLMs are developed with a narrower focus and trained on domain-specific data. This specificity allows them to handle nuanced language requirements in areas where precision is paramount. Rather than relying on vast, heterogeneous datasets, SLMs are trained on targeted information, giving them the contextual intelligence to deliver more consistent, predictable and relevant responses. This offers several advantages. First, they are more explainable, making it easier to understand the source and rationale behind their outputs. This is critical in regulated industries where decisions need to be traced back to a source.  Second, their smaller size means they can often perform faster than LLMs, which can be a crucial factor for real-time applications. Third, SLMs offer businesses more control over data privacy and security, especially if they’re deployed internally or built specifically for the enterprise. Moreover, while SLMs may initially require specialized training, they reduce the risks associated with using third-party LLMs controlled by external providers. This control is invaluable in applications that demand stringent data handling and compliance. Focus on developing expertise (and be wary of vendors who overpromise) I want to be clear that LLMs and SLMs are not mutually exclusive. In practice, SLMs can augment LLMs, creating hybrid solutions where LLMs provide broader context and SLMs ensure precise execution. It’s also still early days even where LLMs are concerned, so I always advise technology leaders to continue exploring the many possibilities and benefits of LLMs.  In addition, while LLMs can scale well for a variety of problems, SLMs may not transfer well to certain use cases. It is therefore important to have a clear understanding upfront as to what use cases to tackle.  It’s also important that business and IT leaders devote more time and attention to building the distinct skills required for training, fine-tuning and testing SLMs. Fortunately, there is a great deal of free information and training available via common sources such Coursera, YouTube and Huggingface.co. Leaders should make sure their developers have adequate time for learning and experimenting with SLMs as the battle for AI expertise intensifies.  I also advise leaders to vet partners carefully. I recently spoke with a company that asked for my opinion on a certain technology provider’s claims. My take was that they were either overstating their claims or were simply out of their depth in terms of understanding the technology’s capabilities.  The company wisely took a step back and implemented a controlled proof-of-concept to test the vendor’s claims. As I suspected, the solution simply wasn’t ready for prime time, and the company was able to walk away with relatively little time and money invested.  Whether a company starts with a proof-of-concept or a live deployment, I advise them to start small, test often and build on early successes. I’ve personally experienced working with a small set of instructions and information, only to find the results veering off course when I then feed the model more information. That’s why slow-and-steady is a prudent approach. In summary, while LLMs will continue to provide ever-more-valuable capabilities, their limitations are becoming increasingly apparent as businesses scale their reliance on AI. Supplementing with SLMs offers a path forward, especially in high-stakes fields that demand accuracy and explainability. By investing in SLMs, companies can future-proof their AI strategies, ensuring that their tools not only drive innovation but also meet the demands of trust, reliability and control.  AJ Sunder is co-founder, CIO and CPO at Responsive. DataDecisionMakers Welcome to the VentureBeat community! DataDecisionMakers is where experts, including the technical people doing data work, can share data-related insights and innovation. If you want to read about cutting-edge ideas and up-to-date information, best practices, and the future of data and data tech, join us at DataDecisionMakers. You might even consider contributing an article of your own! Read More From DataDecisionMakers source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Elvis Presley once said, “Ambition is a dream with a V8 engine.” Brendan Greene, the creator of PlayerUnknown’s Battlegrounds (PUBG), has a lot of ambition. His battle royale game, inspired by the Japanese film Battle Royale (2000), has sold more than 80 million copies. And one of Greene’s ambitions is doing something important like that again in video games. And so he just announced that his PlayerUnknown Productions is resurfacing after years of development with a three-game plan to bring on the next generation of survival games. And it’s ambitious. I talked to Greene, who is known as PlayerUnknown, about it in an exclusive interview. It’s down at the bottom of this introduction and I hope you like it. At the end, I asked him about ambition. Greene got the idea from the movie that he could stage a battle where 100 people would compete with each other. With each player eliminated, the battle space would get smaller until the last two were battling it out in a very small circle. The last one standing was the winner. Greene first created a “mod” called DayZ in the Arma universe. Then he teamed up with South Korea’s Krafton to make PUBG. The game debuted in 2017, disrupted shooter games like Call of Duty. On the strength of PUBG’s 80 million in sales, Krafton went public and Greene became wealthy from that. That gave him the money to work on something even more ambitious. Brendan Greene is the creator of PUBG and he is on to his next survival project. I had a front row seat to this plan. Greene went off on his own to create a new startup, PlayerUnknown Productions, in 2021 to make a gaming survival world that was a lot like a metaverse. Then he gave me a scoop on his ambitions. Without anything to show me except a screenshot at the time, Greene said was creating a world called Prologue that had a huge amount of terrain — about 100 square kilometers. That world, bigger than just about any existing game world, would be a test where players would drop into the world and try to survive until they exited the world in a given spot. It would be different every time they dropped into it. Now Greene has released a video that describes his intentions more concretely. Prologue now has a real preview in the video and the world looks very realistic, with trees and grasses swaying in the wind. And it’s still a huge world, fashioned with machine learning and AI tools. The aim is to release it sometime in the middle of next year as a single-player game for people to try to survive. AI will generate the terrain of Prologue. The challenge is that the open-world of Prologue will be an emergent place, where anything can happen and the weather will get progressively worse. It may seem simple to get to the exit point on the map, but it’s likely going to be hell getting there. Then there will be something else. The company will do a shadow drop of the company’s free tech demo, called Preface: Undiscovered World, showcasing its in-house game engine called Melba. Preface will be able to generate terrain for an Earth-size virtual world, using very little in the way of computing resources. This demo aims to provide users with an early look at the innovative technology that will power the subsequent titles in the series, and eventually a third game called Project Artemis. Project Artemis is the large-scale end goal project of the series. As described in the past, Greene sees this as an Earth-size world where players can drop in and create their own gaming experiences in different sections of the world. We don’t use the word metaverse so much anymore, but that’s what it seems like to me. The journey to get there could take another five or ten years. In the video, Greene said he embarked on Prologue three years ago and “then life happened” and it has taken three years to get it into a solid and breakthrough shape. Now the company can start sharing it and getting feedback “to make it into really something different.” In our interview, Greene said that the team started pulling together when Laurent Gorga joined as CTO. About a year ago, Gorga started putting in motion a process that enabled the team to make a lot more process. While they were making the tech, the team would now create frequent builds to test the tech on a granular level. They started making enough progress so that they started scheduling the timelines for Prologue and Preface. And they talked about it in a video stream on December 6, during the PC Gaming Show. It made a lot of jaws drop. Prologue is expected to drop into early access on the second quarter of 2025. Here’s a view of Preface, another test of technology from PlayerUnknown Productions. “When I started this I was trying to make a larger open world experience than most people made, and we tried to provide a couple of years and we found a way to do that,” Greene said. “We essentially reinvented how you create these worlds using machine learning technology, using natural earth data to generate” the terrain. Now the company is ready to test this terrain, which will form the basis for the larger worlds. He said the team broke the journey into three stages. The first job was to fill out the terrain of the world. The second was to fill that terrain with lots of interaction when scaling up. And then third, the goal was to pull a bunch of those players onto the world, Greene said. The company will keep enhancing Prologue with its current game engine and then it will move it over to the next version of its game

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More For the 10th day of OpenAI’s “12 Days of Shipmas” event, the company decided to go a bit old-school by launching a phone number for people to call and speak with ChatGPT. U.S.-based users can call 1-800-ChatGPT (1-800-242-8478) on any device that can make calls — including a rotary phone — as OpenAI demonstrated in its live stream. If you have an international number, you can message ChatGPT using WhatsApp. It might seem odd that ChatGPT, which has a web version, mobile apps on iOS and Android, and desktop applications for MacOS and Windows, will need a phone number to ask ChatGPT a question. But OpenAI explained that the feature is ideal for those without consistent data connections.  Deploying AI wherever it can OpenAI said the feature enables wider access to ChatGPT. After all, calling a phone number is usually free and accessible for people who may not have unlimited data or who are not close to a Wi-Fi connection.  However, anyone who wants to call ChatGPT can only do so for 15 minutes a month.  This isn’t the first time a tech company has utilized calling or texting to expand its user base. In countries like the Philippines, Facebook offers a way for people to use SMS to post on the social media platform. This proved a boon in expanding Facebook’s reach in the Philippines, where most people have a cellphone, although not necessarily a smartphone. Filipinos are now some of the largest users of the social network.  Not that ChatGPT is hurting for users. It is still one of the most used AI platforms out there, even with more competition in the market. OpenAI said its business offerings, ChatGPT Enterprise, Team and Edu alone, logged more than 1 million users as of September. A month earlier, ChatGPT itself reached 200 million users. OpenAI offers ChatGPT for free, but it gives its more advanced models and other powerful features to subscribers at the Plus, Teams, Pro, Enterprise and Edu levels.  “This is an experimental way to talk to ChatGPT, so availability and limits may change,” the company said. “For a fuller experience with more tools, higher limits and more personalization, existing users should continue using ChatGPT directly through their accounts.” The drawbacks Calling ChatGPT feels like using OpenAI’s Voice and Advanced Voice Model. I called ChatGPT and asked for ideas on what to do during a layover in Tokyo. As expected, it gave me some suggestions.  The difference though, is I did not get a readout of my conversation with ChatGPT. If I use Advanced Voice Mode, anything we discussed is transposed into a written chat that I can look back at when I log into my account.  I also highly doubt any developers will call ChatGPT and ask it to code anything.  But the feature is interesting and it does work. Not bad for something OpenAI developers created just a few weeks ago for a Hack Week.  source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More In its latest push to redefine the AI landscape, Google has announced Gemini 2.0 Flash Thinking, a multimodal reasoning model capable of tackling complex problems with both speed and transparency. In a post on the social network X, Google CEO Sundar Pichai wrote that it was: “Our most thoughtful model yet:)” And on the developer documentation, Google explains, “Thinking Mode is capable of stronger reasoning capabilities in its responses than the base Gemini 2.0 Flash model,” which was previously Google’s latest and greatest, released only eight days ago. The new model supports just 32,000 tokens of input (about 50-60 pages worth of text) and can produce 8,000 tokens per output response. In a side panel on Google AI Studio, the company claims it is best for “multimodal understanding, reasoning” and “coding.” Full details of the model’s training process, architecture, licensing, and costs have yet to be released. Right now, it shows zero cost per token in the Google AI Studio. Accessible and more transparent reasoning Unlike competitor reasoning models o1 and o1 mini from OpenAI, Gemini 2.0 enables users to access its step-by-step reasoning through a dropdown menu, offering clearer, more transparent insight into how the model arrives at its conclusions. By allowing users to see how decisions are made, Gemini 2.0 addresses longstanding concerns about AI functioning as a “black box,” and brings this model — licensing terms still unclear — to parity with other open-source models fielded by competitors. My early simple tests of the model showed it correctly and speedily (within one to three seconds) answered some questions that have been notoriously tricky for other AI models, such as counting the number of Rs in the word “Strawberry.” (See screenshot above). In another test, when comparing two decimal numbers (9.9 and 9.11), the model systematically broke the problem into smaller steps, from analyzing whole numbers to comparing decimal places. These results are backed up by independent third-party analysis from LM Arena, which named Gemini 2.0 Flash Thinking the number one performing model across all LLM categories. Native support for image uploads and analysis In a further improvement over the rival OpenAI o1 family, Gemini 2.0 Flash Thinking is designed to process images from the jump. o1 launched as a text-only model, but has since expanded to include image and file upload analysis. Both models can also only return text, at this time. Gemini 2.0 Flash Thinking also does not currently support grounding with Google Search, or integration with other Google apps and external third-party tools, according to the developer documentation. Gemini 2.0 Flash Thinking’s multimodal capability expands its potential use cases, enabling it to tackle scenarios that combine different types of data. For example, in one test, the model solved a puzzle that required analyzing textual and visual elements, demonstrating its versatility in integrating and reasoning across formats. Developers can leverage these features via Google AI Studio and Vertex AI, where the model is available for experimentation. As the AI landscape grows increasingly competitive, Gemini 2.0 Flash Thinking could mark the beginning of a new era for problem-solving models. Its ability to handle diverse data types, offer visible reasoning, and perform at scale positions it as a serious contender in the reasoning AI market, rivaling OpenAI’s o1 family and beyond. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More The messaging app where millions of office workers share memes and coordinate projects is quietly transforming into something far more ambitious: A platform where AI agents work alongside humans as digital coworkers. As part of Salesforce’s sweeping AI initiative announced Tuesday, Slack is evolving from a communication tool into what company officials call a “work operating system” — one where AI agents can attend your meetings, summarize your conversations, create presentations and even negotiate with other AI agents on your behalf. “We’ve been on a journey to become what we call a work operating system — one that simplifies the complexity of all the systems you use daily,” Slack CPO Rob Seaman, who oversees the company’s AI integration, said in an interview with VentureBeat. “While messaging and human interaction remain at its core, the system now provides access to automation and all the apps you need to do your job, plus AI agents, which we believe will be crucial players in the workplace.” The new digital workplace: AI Agents as your always-on colleagues The transformation is already visible at companies like Accenture, where client executives are using AI agents within Slack to dramatically reduce time spent on administrative tasks. These agents can prepare for meetings, summarize discussions and even draft proposals — all within the familiar Slack interface where employees already spend their workday. “To ensure these AI agents are both widely adopted and continuously improving, it’s critical to integrate them where people are already working,” Seaman explained. Inside Slack’s game-changing AI integration: What you need to know Unlike traditional chatbots or AI assistants that require users to visit separate websites or apps, Slack’s AI agents will be integrated directly into existing workflows. They appear in channels alongside human colleagues and can be called upon through natural conversation. The system is designed to be accessible to non-technical users. “There’s actually no code involved,” said Seaman. “From an end user perspective, there really is no technical work for you.” But the implications for workplace dynamics are profound. In demonstrations, AI agents showed they could independently schedule meetings, analyze documents, create visualizations and even collaborate with other AI agents on complex tasks. For example, during Tuesday’s presentation, an AI agent helped an Accenture executive returning from vacation quickly get up to speed on client activities, prepare for upcoming meetings and draft a proposal — tasks that would typically take hours of human effort. Building trust and control: How Slack’s AI safeguards protect your data While Salesforce will provide template agents for common business tasks, Seaman expects most organizations will customize their AI workers for specific needs. “We’ll provide ready-to-use templates that will meet about 80% of most needs, but we expect organizations will customize their AI agents for specific purposes,” he said. “We’ve seen this pattern with Slack already — companies tend to significantly customize the platform to fit their needs.” The company has built extensive safety measures into the system. “The agents execute on your behalf with no greater permissions than what you have,” Seaman explained. “They don’t have God permissions or admin permissions… we don’t create any holes for the AI to see things that it should not be able to.” The human-AI partnership: Redefining workplace collaboration For many employees, the prospect of AI colleagues raises concerns about job displacement. But Salesforce executives frame it as an augmentation of human capabilities rather than a replacement. “It’s not about handing over the data,” said Claire Cheng, VP of machine learning and engineering at Salesforce. “It’s about unlocking the full potential in the data to enable Agentforce to deeply understand your business and your customer and empower the agents to take more effective actions.” Looking ahead, Salesforce envisions even more sophisticated collaboration between human and AI workers. Future versions will enable multiple AI agents to work together on complex tasks, with specialized agents handling different aspects of projects. “Right now there is a human evoking different agents,” Silvio Savarese, who leads Salesforce’s AI research, told VentureBeat. “The future will have an orchestrator agent which will be calling out different specialized agents that will be talking, working together, performing tasks.” This vision of the workplace — where humans and AI agents collaborate seamlessly through platforms like Slack — represents a fundamental shift in how office work gets done. And while the full implications remain to be seen, one thing is clear: The office chat app where you’ve been sharing cat GIFs is about to become much more powerful. “We’re at the beginning of the beginning,” said Marc Benioff, Salesforce’s chief executive. “When you’re at the beginning of the beginning, you see these little things, and then you try to extrapolate what this is going to be. This is an incredible moment.” source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More A startup founded by former Meta AI researchers has developed a lightweight AI model that can evaluate other AI systems as effectively as much larger models, while providing detailed explanations for its decisions. Patronus AI today released Glider, an open-source 3.8 billion-parameter language model that outperforms OpenAI’s GPT-4o-mini on several key benchmarks for judging AI outputs. The model is designed to serve as an automated evaluator that can assess AI systems’ responses across hundreds of different criteria while explaining its reasoning. “Everything we do at Patronus is focused on bringing powerful and reliable AI evaluation to developers and anyone using language models or developing new LM systems,” said Anand Kannappan, CEO and cofounder of Patronus AI, in an exclusive interview with VentureBeat. Small but mighty: How Glider matches GPT-4’s performance The development represents a significant breakthrough in AI evaluation technology. Most companies currently rely on large proprietary models like GPT-4 to evaluate their AI systems, a process that can be expensive and opaque. Glider is not only more cost-effective due to its smaller size, but also provides detailed explanations for its judgments through bullet-point reasoning and highlighted text spans showing exactly what influenced its decisions. “Currently we have many LLMs serving as judges, but we don’t know which one is best for our task,” explained Darshan Deshpande, research engineer at Patronus AI who led the project. “In this paper, we demonstrate several advances: We’ve trained a model that can run on-device, uses just 3.8 billion parameters, and provides high-quality reasoning chains.” Real-time evaluation: Speed meets accuracy The new model demonstrates that smaller language models can match or exceed the capabilities of much larger ones for specialized tasks. Glider achieves comparable performance to models 17 times its size while running with just one second of latency. This makes it practical for real-time applications where companies need to evaluate AI outputs as they’re being generated. A key innovation is Glider’s ability to evaluate multiple aspects of AI outputs simultaneously. The model can assess factors like accuracy, safety, coherence and tone all at once, rather than requiring separate evaluation passes. It also retains strong multilingual capabilities despite being trained primarily on English data. “When you’re dealing with real-time environments, you need latency to be as low as possible,” Kannappan explained. “This model typically responds in under a second, especially when used through our product.” Privacy first: On-device AI evaluation becomes reality For companies developing AI systems, Glider offers several practical advantages. Its small size means it can run directly on consumer hardware, addressing privacy concerns about sending data to external APIs. Its open-source nature allows organizations to deploy it on their own infrastructure while customizing it for their specific needs. The model was trained on 183 different evaluation metrics across 685 domains, from basic factors like accuracy and coherence to more nuanced aspects like creativity and ethical considerations. This broad training helps it generalize to many different types of evaluation tasks. “Customers need on-device models because they can’t send their private data to OpenAI or Anthropic,” Deshpande explained. “We also want to demonstrate that small language models can be effective evaluators.” The release comes at a time when companies are increasingly focused on ensuring responsible AI development through robust evaluation and oversight. Glider’s ability to provide detailed explanations for its judgments could help organizations better understand and improve their AI systems’ behaviors. The future of AI evaluation: Smaller, faster, smarter Patronus AI, founded by machine learning experts from Meta AI and Meta Reality Labs, has positioned itself as a leader in AI evaluation technology. The company offers a platform for automated testing and security of large language models, with Glider its latest advance in making sophisticated AI evaluation more accessible. The company plans to publish detailed technical research about Glider on arxiv.org today, demonstrating its performance across various benchmarks. Early testing shows it achieving state-of-the-art results on several standard metrics while providing more transparent explanations than existing solutions do. “We’re in the early innings,” said Kannappan. “Over time, we expect more developers and companies will push the boundaries in these areas.” The development of Glider suggests that the future of AI systems may not necessarily require ever-larger models, but rather more specialized and efficient ones optimized for specific tasks. Its success in matching larger models’ performance while providing better explainability could influence how companies approach AI evaluation and development going forward. source