AI News Archive: June 24, 2026 — Part 6
Sourced from 500+ daily AI sources, scored by relevance.
- HCLTech Partners with Neste for AI Efficiency
HCLTech Partners with Neste for AI Efficiency india.entrepreneur.com
Score: 59🌐 MovesJun 24, 2026https://india.entrepreneur.com/business-news/hcltech-partners-with-neste-for-ai-efficiency - May Mobility Aims to Expand Robotaxi Operations Overseas Through New Partnership
By teaming up with ride-hailing service CaoCao, May Mobility plans to bring its self-driving tech to more locations, starting in Europe.
Score: 58🌐 MovesJun 24, 2026https://www.cnet.com/roadshow/news/may-mobility-overseas-robotaxi-expansion-caocao/ - Critic of Labor’s tax changes deletes anti-immigration AI video reposted from rightwing nationalist account
Fund manager Geoff Wilson says he did not watch full video and deleted it after ‘inappropriate associations were identified’ Follow our Australia news live blog for latest updates Get our breaking news email , free app or daily news podcast The fund manager Geoff Wilson, a prominent public critic of the government’s tax changes, has deleted an inflammatory AI-generated video he reposted from a rightwing nationalist account portraying Anthony Albanese and Jim Chalmers taking money from white Australians and giving it to recently arrived migrants wearing Islamic face coverings. Wilson said he had not watched the full video before sharing it or examined other accounts, some of whose content he reposted on Wednesday morning – which included content relating to the QAnon conspiracy theory – and deleted his posts after being contacted by Guardian Australia. Continue reading...
- Figma bets on human judgment at Config 2026 while the AI powering its canvas belongs to someone else
At Config 2026, Figma turned its canvas into a full workspace with code, animation, shaders, and AI agents. But the intelligence powering all of it is rented from API providers, squeezing margins. And one of those providers is now building competing design tools. The article Figma bets on human judgment at Config 2026 while the AI powering its canvas belongs to someone else appeared first on The Decoder .
- Gulf banks want AI, but can they keep control of customer data?
As banks in the Gulf explore new AI tools, the focus is increasingly shifting from what the technology can do to how it can be used responsibly.
Score: 58🌐 MovesJun 24, 2026http://www.euronews.com/next/2026/06/24/gulf-banks-want-ai-but-can-they-keep-control-of-customer-data - HelloTwin launches ‘Digital Authority’ to bring governed AI agents to the enterprise
HelloTwin.ai GmbH today announced what it calls an accountable artificial intelligence AI twin that holds business intelligence and goals in a single source of truth. HelloTwin said it built its data model on a patent-pending compiler designed to pull answers from business context rather than generate them. This means that the agentic identity for the […] The post HelloTwin launches ‘Digital Authority’ to bring governed AI agents to the enterprise appeared first on SiliconANGLE .
Score: 58🌐 MovesJun 24, 2026https://siliconangle.com/2026/06/24/hellotwin-launches-digital-authority-bring-governed-ai-agents-enterprise/ - Why full-service partners are becoming critical to New Zealand’s cloud and AI future
New Zealand organisations are entering a new phase of cloud adoption – one where success is no longer measured simply by whether workloads move to the cloud, but by whether those environments are capable of supporting AI-driven innovation. The opportunity is significant. AI adoption is expected to add NZ$76 billion annually to NZ’s economy by 2038, growing GDP by around 1 per cent every year. That acceleration is also fuelling cloud investment. IDC forecasts indicate NZ’s public cloud spend will almost double from NZ$5 billion in 2024 to NZ$9.6 billion by 2028, as organisations increasingly modernise infrastructure and prepare for AI-enabled operations. Increasingly, cloud is no longer viewed simply as infrastructure hosting. According to IDC, Kiwi organisations are evolving from basic “lift-and-shift” migrations toward more sophisticated cloud-native and data-driven strategies, with IT leaders now treating cloud as a platform for AI-led transformation. For many organisations, however, the challenge is not whether to move – it is how to move well. According to Chris Beckett, technology strategist at Inde Technology , most NZ organisations already understand the strategic value of cloud, particularly around platforms like Microsoft Azure, but execution remains the difficult part. “Cost and budget uncertainty tend to top the list,” Beckett says. “That is not because cloud is inherently expensive, but because undisciplined adoption is. Organisations that have not built proper cost governance in from day one end up with sprawl and bill shock, and that reinforces scepticism at board level.” Governance matters more than ever One of the biggest misconceptions organisations still make is assuming cloud migration itself automatically delivers efficiency. Beckett says treating cloud as a direct infrastructure replacement means organisations lose out on the opportunity to modernise architecture and operating models. “The most common and costly mistake is lift and shift,” he says. “When you lift and shift, you take all of your existing technical debt and put it on a meter. You are now paying cloud running costs on top of architecture decisions that were made for a different world.” Instead, organisations achieving the strongest outcomes are using migration as a forcing function to modernise applications, governance and operational processes simultaneously. That includes adopting Infrastructure as Code, cloud-native platform services, embedded security controls and DevOps practices from the outset. According to Beckett, governance cannot be retrofitted later. “Visibility and control have to be built in from day one. Retrofitting them to a running environment means you are already behind, and the bill has already arrived.” This is particularly important as organisations scale AI workloads on platforms like Azure, where poorly managed environments can quickly create unpredictable consumption costs. The Azure advantage for NZ companies The launch of Microsoft’s first New Zealand hyperscale cloud region, NZ North , has also changed the conversation significantly. The region already supports tenants including Fonterra , Spark and ASB , alongside smaller organisations such as Te Tumu Paeroa. “For boards that have been asking, ‘where does the data actually live?’ the answer is now unambiguously here in NZ, with full Azure capability,” Beckett says. “That is a significant shift.” Beyond sovereignty, Beckett says Azure is delivering measurable value across scalability for customer-facing applications, improved resilience and recovery capability, Infrastructure as Code and operational consistency, and stronger cost governance through native tooling. He points to a recent engagement with a major Kiwi logistics company that migrated critical freight management systems to Azure Platform-as-a-Service infrastructure. “This was not a lift and shift, but a proper architectural rebuild. The result was scalable, always-on applications with operational visibility the team had never had before.” Why full-service partners are becoming more valuable As cloud environments become more complex – and increasingly tied to AI initiatives, governance requirements and cybersecurity obligations – many organisations are reassessing the role of their technology partners. In New Zealand, Inde works closely with Microsoft, leveraging the expertise, enablement, and partner ecosystem strength of leading IT distributor Dicker Data . Beckett argues there is a growing distinction between traditional systems integrators and full-service cloud partners. “A full-service partner stays with you through the whole lifecycle: consult, create, supply and manage,” he says. “That matters in cloud, because the technology does not stop evolving once the migration is complete.” The value, he says, lies not only in technical delivery, but in strategic guidance and long-term operational accountability. “A partner worth working with will tell you when you are not ready to migrate, not just help you move.” He cites a healthcare engagement where Inde Technology delayed migration activity in order to first address security, identity and governance readiness through the Microsoft Cloud Adoption Framework. “That was not the fastest path to a sale, but it was the right path to a good outcome,” he says. This ability to assess readiness honestly is becoming increasingly important as organisations navigate growing complexity around AI, cybersecurity, compliance and operational resilience. Planning early creates strategic advantage Beckett believes one of the biggest differentiators between successful cloud transformations and reactive migrations is timing. “The organisations that modernise well started thinking about this 18 to 36 months before the forcing function arrived.” Those forcing functions can include hardware end-of-life dates, expiring software licences or escalating operational risk. “The organisations that struggle are the ones who bring a partner in at six months. At that point, options are limited, leverage is gone, and decisions are being made under pressure.” “The infrastructure argument in NZ is resolved. Microsoft’s in-country datacentre means data sovereignty is no longer a reason to delay. The organisations building their roadmap today will be deploying with confidence in 18 months. The ones who wait will be reacting.”
- Accelerating BEV Pooling on NVIDIA GPUs for Physical AI Applications
An increasingly common design pattern for autonomous vehicles (AVs), robotics, and spatial AI systems is bird's-eye-view (BEV) perception. BEV models project...
Score: 58🌐 MovesJun 24, 2026https://developer.nvidia.com/blog/accelerating-bev-pooling-on-nvidia-gpus-for-physical-ai-applications/ - Resolve AI CEO Spiros Xanthos: AI’s impact on software production systems
Automating software’s deployment, production, and monitoring systems has been challenging. Resolve AI’s CEO explains how AI could improve reliability, pace, and workload.
- European direct lenders mull software deals as AI shapes sector
European direct lenders mull software deals as AI shapes sector PitchBook
Score: 58🌐 MovesJun 24, 2026https://pitchbook.com/news/articles/european-direct-lenders-mull-software-deals-as-ai-shapes-sector - Waymo turns to Canadian fleet manager Element to oversee part of robotaxi fleet
Waymo turns to Canadian fleet manager Element to oversee part of robotaxi fleet Automotive News
Score: 57🌐 MovesJun 24, 2026https://canada.autonews.com/technology/anc-waymo-element-fleet-deal-0622/ - Companies are scrambling to stop employees from maxing out AI budgets with small tasks
The tokenmaxxing era was brief. We now appear to be entering the era of token rationing.
- Lytho Launches AI Expert Reviewers, Embedding Instant Brand and Regulatory Compliance into Enterprise Content Workflows
Lytho Launches AI Expert Reviewers, Embedding Instant Brand and Regulatory Compliance into Enterprise Content Workflows USA Today
- Boston Dynamics plans $100 million, 1,250-job Waltham expansion
The expansion will take place over the next four years, and will see the robot maker pour millions into Massachusetts creating thousands of jobs as it seeks to expand with a new mobility and robotics center across the highway from its core facilities.
- AI is making developers more productive — and anxious about falling behind
AI is making developers more productive — and anxious about falling behind Business Insider
Score: 56🌐 MovesJun 24, 2026https://www.businessinsider.com/ai-coding-tools-software-engineers-workplace-paralysis-2026-6 - Fewer job offers for junior roles due to AI, Swiss study shows
Fewer job offers for junior roles due to AI, Swiss study shows Reuters
Score: 56🌐 MovesJun 24, 2026https://www.reuters.com/business/fewer-job-offers-junior-roles-due-ai-swiss-study-shows-2026-06-24/ - Facebook rolls out an AI companion app for creators
The new app, which is currently being tested with select creators, will have Facebook's recently launched AI creator assistant built into it.
Score: 55🌐 MovesJun 24, 2026https://techcrunch.com/2026/06/24/facebook-rolls-out-an-ai-companion-app-for-creators/ - Coca-Cola's World Cup AI Avatar Showcases Brand's Cutting-edge AI Innovation
Coca-Cola's World Cup AI Avatar Showcases Brand's Cutting-edge AI Innovation The Straits Times
- Momentic Launches Agentic Quality Platform
SAN FRANCISCO, CA — Momentic has launched a major platform update that rethinks software verification for the AI era. AI coding tools have made it possible for teams of all sizes to ship more code than ever before, but how that code gets validated has not kept up. The result is bugs reaching production faster … continue reading The post Momentic Launches Agentic Quality Platform appeared first on SD Times .
- The AI leadership gap: Even marketers who use AI fear they’ll be replaced
42.5% of marketing employees fear AI replacement. WRITER's 2026 survey reveals the leadership gap — and the four moves CMOs need to close it. The post The AI leadership gap: Even marketers who use AI fear they’ll be replaced appeared first on WRITER .
- AI literacy in Thailand is reaching office workers, not the 63% in informal employment
Thailand is both an agricultural country and an industrial manufacturing base that produces a large number of engine components. Nevertheless, workers across many industrial sectors in Thailand are now at risk of being replaced by AI technology and automation, as AI is increasingly being integrated into the Internet of Things (IoT) and across manufacturing industries. […] The post AI literacy in Thailand is reaching office workers, not the 63% in informal employment appeared first on e27 .
Score: 55🌐 MovesJun 24, 2026https://e27.co/ai-literacy-in-thailand-is-reaching-office-workers-not-the-63-in-informal-employment-20260617/ - Infosys, ANA’s Global CMO Growth Council, and LIONS Unveil the CMO AI Hub
Infosys in collaboration with ANA’s Global CMO Growth Council, global leadership body driving growth for marketers, and LIONS, the parent platform and organizer of the Cannes Lions International Festival of Creativity, today announced the launch of the CMO AI Hub. This exclusive AI-powered platform is designed to enable peer-to-peer learning among Chief Marketing Officers by bringing together insights, experiences, […] The post Infosys, ANA’s Global CMO Growth Council, and LIONS Unveil the CMO AI Hub appeared first on CXOToday.com .
- Amazon’s Zoox Shows Off Its New Toaster on Wheels
Unlike Waymos, Zoox robotaxis have no steering wheel or pedals.
Score: 55🌐 MovesJun 24, 2026https://gizmodo.com/amazons-zoox-shows-off-its-new-toaster-on-wheels-2000777065 - Pegasystems CEO and founder Alan Trefler on AI agent ‘madness’
Mainstream software suppliers are pursuing a “philosophy of madness” by persuading enterprises to deploy thousands of artificial intelligence (AI) agents , says Alan Trefler , the CEO and founder of Pegasystems. Trefler, known for his outspoken views, claims that enterprises are in danger of lining up problems for the future by letting loose en masse AI agents with the ability to make potentially business-critical decisions on their computer networks. Industry watchers say Trefler has a point when it comes to mission-critical workflows in the regulated industries Pega supplies, but that other software suppliers are getting a grip on agentic AI. Neil Ward-Dutton , research vice-president for agentic automation and AI technologies at analyst group IDC, told Computer Weekly that while there is a race by suppliers to be seen to have AI capabilities, behind the scenes, other suppliers are introducing their own guardrails for AI. “I think Pega realised pretty early on that you can use AI to help you create a deterministic system, like a Pega application. Now ... all the other vendors are starting to do that too, but Pega was one of the first,” he said. “When you get behind the headlines and speak to ServiceNow or Salesforce or Microsoft, they will all talk in their own way about this kind of hybrid approach, where yes, you can use AI, and you can use AI agents, but actually you need to combine them with traditional, deterministic, more rules-based approaches,” he said. The madness of mass AI agents Trefler told Computer Weekly that mainstream enterprise software suppliers are in danger of going down the wrong track by promoting the idea that critical systems can be safely built on AI agents. “I think they have already acknowledged they’re going to have thousands of agents running, maybe over 10,000, and I think that philosophy is madness,” he said. The big enterprise software suppliers are introducing “ control towers ” to allow enterprises to monitor and manage their AI agents. But Trefler argues that AI agents are too unpredictable to be used at scale to run mission-critical applications in enterprises. “[Software companies] have already acknowledged they’re going to have thousands of agents running, maybe over 10,000, and I think that philosophy is madness” Alan Trefler, Pegasystems “You don’t want these disaggregated, disassociated initiatives trying to run important things in the business where it might treat customers differently in ways that it shouldn’t,” he said. When AI goes wrong, it can go wrong in a spectacular fashion, as marketers who chose an AI-generated slogan for Starbucks in South Korea found out, when their “Tank Day” campaign led to national riots and customers smashing Starbucks-branded cups, followed by the resignation of its chief executive. Founded by Trefler in 1983, Pegasystems is now on track to become a $2bn revenue organisation. Based in Massachusetts in the US, the software company supplies a low-code business process platform to some of the world’s largest companies. It lists Deutsche Telekom, Lloyds Banking Group and Daimler Trucks among its customers. The AI bubble will burst Trefler sees a big future for AI in the enterprise, but he is wary of its unpredictability and the soaring costs of AI tokens , and believes the current rate of investment in datacentres to power AI is unsustainable. “It was very sobering to me to drive down Highway 101 in California about two months ago,” said Trefler. “There was billboard after billboard of AI companies. I’m in the business, and I didn’t know 80% of them.” He predicts that when the AI bubble bursts , many of these companies are unlikely to survive. “I think you’re going to get a real flash in the pan from lots of companies that may have had a good idea, but find that there’s a big step between that and sustainable business,” he said. Read more about Pegasystems Pegasystems CTO Don Schuerman on how to keep the lid on skyrocketing AI costs : Pegasystems offers an alternative take on how enterprises can use artificial intelligence to automate their business processes without burning through their budgets. Pegasystems founder and CEO Alan Trefler on the future of GenAI : The tech pioneer is optimistic about generative AI – as long as users are aware of its capacity for mistakes. Pegasystems refines Blueprint agent builder, expands marketing tools : Pegasystems emphasises ‘derisking’ agentic buildouts for its customers in regulated industries. Enterprises that succeed in agentic AI start by ‘reimagining’ business process, finds Pega research : A study of IT leaders finds businesses that succeed in deploying agentic artificial intelligence start by rethinking their business processes. Trefler has lived through tech bubbles before. The fibre optic cable bubble burst in 2000 after companies laid huge amounts of fibre optic cable that vastly exceeded demand. There was “an inevitable financial rebalancing”, with many fibre optic companies going out of business. But over the next 20 years, the excess fibre was eventually used. If and when the AI bubble bursts, enterprises will need to re-evaluate what they want to use powerful large language models (LLMs) for, he suggests. AI changing the nature of the IT profession AI has already changed the nature of the IT profession. Trefler says software platforms, such as Pega’s Blueprint and Pega’s Infinity cloud platform, are increasingly allowing people without IT skills to design and build business applications. A few years ago, the same tasks would have required experts in Pegasystems software, who were often hard to find and commanded top salaries. Today, the design work can be carried out by people with business skills rather than technical skills. Businesses will still need technically experienced people to integrate complex IT systems, but standard integrations are just going to get done, according to Trefler. “They’re already pretty much out of the box,” he said. Widespread use of AI means IT jobs will become “compressed”. There will be a need for fewer programmers, for example, as large language models churn out high volumes of code. That has put IT professionals under pressure to prove their worth by using LLMs to “ vibe code ” computer code at high volume. “The trouble is, we’ve all learned that having more and more code in the business doesn’t make that business more reliable,” said Trefler. The need for predictability Trefler describes himself as a lone “voice crying in the wilderness” and, going against perceived wisdom, warns enterprises to think twice about deploying agentic AI at scale . He does not see how letting thousands of AI agents loose to interact with customers, or make decisions that affect the business, can be done safely or cost-effectively. “What I am saying is that enterprises that want to have a certain consistency of process, consistency of soul, consistency of outcome, need to do it in some other way,” he said. AI is just too unpredictable, warns Trefler. It could treat customers differently in ways that it shouldn’t, or it might accidentally violate a law or a regulation. “You don’t want these disaggregated, disassociated initiatives trying to run important things in the business,” he said. How enterprises are using Pega’s technology Vodafone Greece automates deals for customers, saves 500 staff-days of work : Vodafone Greece hired an implementation partner for a business process management project while its own staff observed and learned how to use the technology. Wells Fargo bank turns to AI to help families settle estates after a death : Wells Fargo bank is winning customers after using business automation software and artificial intelligence to help people manage the estates of relatives following a bereavement. Citi US Personal Banking turns to AI to ‘delight’ customers with personalised services: Citigroup’s US Personal Banking business has created a repository of customer data and is rolling out a decision engine. Bupa turns to data to provide personalised health services : Private healthcare provider Bupa says a project to deploy business process automation is bringing it closer to APAC customers. Over the past two months, the economic impact of AI has become a growing issue, with CEOs starting to worry about the rising cost of AI tokens. Trefler, only half joking, compares current token price hikes to a drug dealer winning over clients by offering free or cheap drugs, only to ramp up the price when they are hooked. The tendency to measure the impact of AI by the number of AI tokens employees consume is misplaced, in Trefler’s view, with the only credible metric being whether AI is helping businesses save money or make money. The case for deterministic workflows Neil Ward-Dutton, research vice-president for agentic automation and AI technologies at IDC, says Trefler is “right on the money” when it comes to dealing with the high-volume business-critical workflows, such as checking the eligibility of people with insurance claims, that Pega specialises in. “It would be complete madness to try to replace what Pega does with fleets of agents running around, kind of guessing at what to do, trying to make a fist of it. That would be an absolute disaster,” he added. But he said there are other, less critical applications where fleets of AI agents would be just fine, such as building marketing campaigns, translating advertising copy, or checking an online product catalogue is up to date. Software agents, such as Claude Cowork and Microsoft Scout , sit on the desktop and can help people schedule emails and work through tasks relatively easily. He said more software suppliers are coming around to a similar combination of deterministic workflows and AI agents , and are using their own technologies to execute agentic AI safely. This includes using their own platforms to provide plug-in skills for AI. Salesforce , for example, has introduced AgentForce agents, and is using workflows and templates. A German supplier, Cumunda , which competes with Pega, is also using a mixture of deterministic workflows, with AI dynamic decision-making. How Pega’s Blueprint is solving the problems of agentic AI Pegasystems’ solution to the unpredictability of artificial intelligence (AI) as a tool to run enterprises is to use it to design their business processes first. The company introduced its AI-powered design studio, Blueprint , two years ago, as a tool that is intuitive enough to allow people without IT skills to design business processes and workflows. Pega’s customers use Blueprint to build business workflows in the supplier’s Infinity cloud software platform, with the latest version expected to be released in autumn 2026. For example, a bank could develop workflows to validate an application from a customer for a loan, a mobile phone company could build a workflow for managing offers when it’s time for a customer to renew their contract, or a healthcare company could create a workflow to automatically send a follow-up email after a call with a patient. In the Pega model, businesses can use AI agents, such as a chatbot, to choose the right workflow, for example, when a retail customer orders a product, or a bank user requests the balance on their current account. Time savings that make a difference In one example, insurance group MetLife has used Pega’s software to redesign its underwriting processes in Mexico. The company built a structured workflow, controlled by a web portal, that minimises manual data entry and can automatically validate policy data. This has helped the insurer reduce the time it takes to issue a policy by 40%, down from a 45- to 60-day turnaround to between 30 and 40 days. It has also reduced the number of applications that are held up because of missing information by 40% to 50%. In another example, NHS 24, Scotland’s National Digital Health Service, turned to Pega to replace its legacy IT systems, which were becoming difficult for staff to use and to keep patched with security updates. Working with Pega, Amazon Web Services (AWS) and Coforge, NHS 24 built an intuitive system to guide staff when people call for medical advice. The project has reduced the average time to answer queries by 10 minutes, freeing up staff to deal with more enquiries. AI value is in redesigning business processes Alan Trefler, CEO and founder of Pegasystems, argues that using AI to do the hard thinking required to design standard business processes makes more sense than using AI agents that have to rethink the same problem every time they are asked. “What makes AI agents very expensive is that they are figuring out what the next step is, looking at the inputs they have and what the goal is … they need to rethink from the beginning,” he said. Pega’s technology “snaps” to pre-written workflows, but still has the capability to stop and ask for more information if it needs to complete a business process. Trefler says Blueprint has reduced the time taken for enterprises to redesign their business processes , with 80% of projects going live within 90 days. In the past, Pega would have sent in a team of people with design and programming skills to spend one or two weeks at a company mapping out their business processes with a whiteboard and Post-it notes. Blueprint can do the same thing in a morning. Pega has demonstrated how Blueprint can ingest and analyse technical documents and videos of people describing how they use existing IT systems to perform tasks and turn that into business workflows. The technology will allow Pega to offer its software to a wider range of companies. Until now, it has limited itself to 1,000 customers, mainly large Fortune 500 companies, because of the resources needed to support them. Trefler believes that working with partners – which include Accenture, EY, Infosys and AWS – he can extend Pega’s coverage to what he calls the “Gartner Universe” – the 12,000 organisations that turn to the analyst group for advice on IT and technology.
Score: 55🌐 MovesJun 24, 2026https://www.computerweekly.com/news/366644888/Pegasystems-CEO-and-founder-Alan-Trefler-on-AI-agent-madness - Agentic AI Predictions — What CMOs Need to Know
Explore upcoming trends in agentic AI and how CMOs can prepare for the evolving marketing landscape.
- The Cost Of AI Productivity Is Less Creativity
AI is now standard across marketing agencies — but its impact remains constrained. Forrester’s “The State Of AI Inside US Marketing Agencies, 2026” finds that CMOs must shift their AI ambitions toward creativity, differentiation, and growth to unlock real value.
Score: 55🌐 MovesJun 24, 2026https://www.forrester.com/blogs/the-cost-of-ai-productivity-is-less-creativity/ - Schroders Capital CIO: AI is about returns, not headcounts
Schroders Capital CIO: AI is about returns, not headcounts PitchBook
Score: 55🌐 MovesJun 24, 2026https://pitchbook.com/news/articles/schroders-capital-cio-ai-is-about-returns-not-headcounts - Achieve state-of-the-art inference latencies with speculative decoding
How Modal and Decagon worked together to cut inference latency - and you can too.
- Europe’s Power Grids Are Not Ready for AI, Envision CEO Says
Europe’s Power Grids Are Not Ready for AI, Envision CEO Says Caixin Global
- STT, Tata Delhi data centre fire leaves clients fearing decades of data lost; Google hit
One client says it lost two decades of data
- India Wants Its AI Talent Back; But What’s The Incentive?
For decades, India has exported some of its brightest minds to the world’s leading research labs. Researchers of Indian origin…
Score: 55🌐 MovesJun 24, 2026https://inc42.com/features/india-wants-its-ai-talent-back-but-whats-the-incentive/ - What Ambani’s deep-tech ambitions reveal about India’s AI limits
Country still needs to play catch-up if it is to become self-reliant in the use of the technology
- A battery-free wireless intelligent aligner for spatially resolved, closed-loop theranostics of chronic oral diseases
Science Advances, Volume 12, Issue 26, June 2026.
- DeepMind Chief Demis Hassabis says Google’s still winning AI talent
Despite the departure of two top AI leaders that sent the company’s stock price down, Hassabis says Google is still on top.
Score: 55🌐 MovesJun 24, 2026https://www.semafor.com/article/06/23/2026/deepmind-chief-demis-hassabis-says-googles-still-winning-ai-talent - AI Slop Is Coming for Your Kids
AI Slop Is Coming for Your Kids Business Insider
- The AI readiness gap: Why networks matter more than ever
Ask enterprise leaders about AI and you’re likely to get a wave of excited responses. BCG research found that two-thirds of global CEOs put accelerating AI among their top three priorities , with CIOs under pressure to turn that ambition into business value. But there’s a problem. Many enterprise AI initiatives are struggling to move beyond pilots into production. Despite near-universal adoption, McKinsey finds that 88% of organizations now use AI in at least one business function , while almost two-thirds remain stuck in pilots and experimentation. “When it comes to AI readiness, most organizations are still trying to figure it out,” says industry expert Bill Burns. “We’re all asking the same questions: where should workloads live, how will traffic move, what does security look like, and where are the bottlenecks going to appear?” The reasons are well documented, and most have nothing to do with infrastructure: unclear ROI, poor data quality, governance gaps, change-management fatigue, and a shortage of talent. Any honest account of why pilots stall has to start there. But there is a common thread why these problems keep surfacing at the same companies, and it sits underneath all of them. Businesses can fix their data strategy, governance model, and talent pipeline, and still find that workloads won’t move where they need to, when they need to, at the cost they need. That constraint is the network – the one layer that gates whether the rest can actually run in production. Why AI traffic is different and legacy networks can’t cope Enterprise networks have always evolved to reflect changes in technology and working patterns. The rise of cloud computing and mobile devices in the mid-2000s, for example, shifted enterprise applications from the data center to public clouds and made the internet the network of choice. AI is triggering the next major shift. It changes the shape, speed and economics of data movement, creating new traffic patterns that legacy infrastructure was never designed to handle. Unless networks adapt, AI will struggle to move beyond pilots into production. The first challenge comes from training AI models. Unlike traditional enterprise traffic, AI workloads are persistent and continuous, creating demands that can overwhelm existing networks. “The problem is that many of us are trying to modernize while still keeping the lights on,” says Burns. “It’s a pendulum every day between operational stability and preparing for what comes next.” Training AI models requires data centers with high bandwidth, ultra-low latency and near-zero packet loss. Networks previously handling 100Gb may now need 400Gb or even 800Gb capacity. In distributed GPU clusters, one delayed packet can stall synchronization across thousands of dollars of compute resources in real-time. The inference challenge The second challenge comes from inference, where users interact with AI systems and AI agents talk to each other. This shifts traffic from north-south flows to far greater volumes of east-west machine-to-machine traffic, potentially increasing network demands by as much as 100x. Furthermore, AI agents operate far faster than humans, meaning millisecond-level delays can become critical bottlenecks. As devices are increasingly used by both people and agents, enterprise networks will need to operate at machine speed. “The network is no longer a foster child in the AI era,” says Murali Krishnan, associate vice president and head of the strategic products group for the Americas at Tata Communications. “It is the fabric – the epicenter around which performance, ROI and experience will be measured. CIOs need to unlearn what they knew about networks of the past, because how you design and deploy the network has changed from the ground up.” What AI-ready networks look like After the physical networks of the 1990s and the software-defined networks of the 2010s, we’re moving into the era of cognitive and contextual networks, fit for the unique requirements of AI. Static, best-effort infrastructure is giving way to networks that can observe, prioritize and adapt in real-time. We believe this new infrastructure must be built on three principles. Unlike today’s enterprise networks, AI-ready networks will be natively intelligent and autonomous, with deep observability built in as standard . In AI environments, one delayed flow can ripple across an entire workload. “Most networks can move AI traffic. The difference is whether they understand it,” says Rajat Gopal, vice president, cloud networking and security solutions at Tata Communications. “That means application awareness – knowing which workload a flow serves – consistency you can measure in jitter, not just an uptime number, and sovereignty enforced in the path itself, so data is geofenced by default.” Given enterprises’ hunger for data, IT leaders will need to architect their future networks with elasticity and scalability in mind – not just increased link capacity, but also more effective congestion domain boundaries and more controlled interconnect paths between clouds. Because the old perimeter-based security model is defunct in an era of AI-powered threats, when data moves continuously across domains, security and control have to be embedded into routing logic, not bolted on. Those guiding principles start to map out a way for enterprises to prepare for AI at a foundational level. The network is becoming an active control plane for AI performance, cost and compliance. It also helps address some of the biggest headaches facing IT leaders currently, such as data sovereignty compliance (through visibility into data paths and metadata) and cost optimization (via lowering egress fees). “We didn’t set out with AI in mind,” says Thor Wallace, CIO at NETSCOUT. “But as it turns out, the decisions we made through our digital transformation have put us in a position where we’re ready for it. The biggest driver was ensuring we had pervasive visibility across the network.” The time to act As AI agents spread, the network is becoming a critical – yet frequently overlooked – enabler of enterprise AI success. The opportunity is significant. As Seth Goodman, CRO at Boost Payment Solutions, argues: “To view AI as primarily a cost saver is missing the point entirely.” The organizations seeing the greatest value are using AI to increase productivity, accelerate decision-making and unlock entirely new capabilities. With industry leaders already benefiting from AI’s productivity gains, CIOs have no time to waste. Fixing the foundations should be the key first step for IT leaders looking to get ready for AI. AI-powered enterprises are being built today. It’s time to get real about your AI readiness. Discover how to evolve your network for the next era in Tata Communications latest whitepaper .
Score: 55🌐 MovesJun 24, 2026https://www.cio.com/article/4188336/the-ai-readiness-gap-why-networks-matter-more-than-ever.html - Getting past the pilot: Why so many AI test projects have trouble scaling
Getting past the pilot: Why so many AI test projects have trouble scaling Fortune
Score: 55🌐 MovesJun 24, 2026https://fortune.com/2026/06/24/getting-past-the-pilot-why-so-many-ai-test-projects-have-trouble-scaling/ - See How AI Giants Are Using AI in Their Own Offices
The activities of AI agents inside OpenAI, Google and Anthropic offer a glimpse of what white-collar workers can expect in their own jobs.
Score: 55🌐 MovesJun 24, 2026https://www.wsj.com/tech/ai/see-how-ai-giants-are-using-ai-in-their-own-offices-1ac00a70?mod=rss_Technology - How Shopify built an AI stack that doesn't care which models survive
Shopify built an LLM proxy that gives every engineer access to multiple AI providers — with automatic failover when any one of them goes down, changes, or disappears. When Claude Fable 5 shut down , Shopify's engineers didn't go into panic mode. The proxy shifted them to Claude Opus or GPT 5.5 automatically, without interrupting their workflows. “Fable looks amazing; we used it of course,” Farhan Thawar, Shopify’s head of engineering, says in a new VentureBeat Beyond the Pilot podcast . “When a model comes and then it goes, or it could be as innocuous as an update, the proxy allows us to spray across the different providers,” Thawar says. Shopify buys tokens in bulk and all users connect to models through its proxy, Thawar says. This gives his team access to reporting and failover; when there’s an availability issue with one provider, users can be “automatically, seamlessly” transferred to another. Enterprises can learn from this example and consider how a disruption might affect their business, Thawar says. At the very least, they should establish a solid backup plan. It’s important to have a system that allows for movement across models so enterprises are not “super tied” to a specific provider. Distillation is another important strategy. With distillation, a student model learns from a teacher model and typically becomes specialized in a narrower task. These small language models (SLMs) can be more beneficial than generalized, off-the-shelf models in some circumstances. For instance, Shopify’s flagship AI assistant, Sidekick, which performs numerous specialized subtasks for merchants so they can “remove toil” from their day-to-day. Using smaller distilled models can be faster and cheaper than more generalized models, Thawar says. In some cases they have proven to be 2x cheaper and faster; in more extreme cases 30x cheaper and faster, he says. But “it isn’t just about cost and latency, which are big; it’s about accuracy,” Thawar says. Engineers feed the UDP their teacher model, training data, evals, and a target model — say, Opus 4.8 distilling down to Qwen 3.5. The pipeline runs for about a day, then returns an evaluation showing what the fine-tuned model actually achieved on speed, cost, and accuracy for that subtask. If the tradeoff looks good, the engineer deploys it — no approval process required. Shopify's internal platform, Tangle, lets anyone visualize the pipeline as it runs. Thawar says his “dream” is to eventually not give the distillation pipeline a target model at all. Instead, users could provide the teacher model with data and evals and the directive: ‘Based on your learnings over time, I want you to look at a different class of model, different sizes, different types, and you tell me what the right distillation target is.’ “Maybe we'll get surprised. Maybe it'll be such a small model it could run on a phone,” Thawar says. “Other times, maybe it comes back and says, ‘There isn't a way to distill this down to anything better than what we have at the frontier.’” Moving away from "AI reflexivity" to "AI leverage" Shopify users can apply whatever harness they want: Claude Code, Codex, Cursor, GitHub Copilot for VS Code. “We expose everyone to the different harnesses so they can get a feel for what may or may not work in their workflow.” But the company also implemented a usage dashboard; this allows Thawar’s team to ask interesting questions around not just token spend, but: Who’s using the most expensive tokens? Who's spending more time on reasoning? What types of models are being used, and what disciplines and levels? Regarding the " tokenmaxxing " question, Shopify does have “circuit breakers” in place. If a user has a model running for a long time (say, 10 hours) and it’s consuming a lot of tokens, they will get pinged, “Did you mean to spend this?” As Thawar explains, sometimes the reply is “Oh, absolutely.” Other times it’s: ‘Whoa, I didn't know that was running in the background. I totally forgot about it. I'd rather stop it now.’ The ultimate goal, as Thawar describes it, is to move from “AI reflexivity” to “AI leverage,” and get people to really think deeply about where they can benefit most from AI in their workflows. Listen to the full podcast to hear more about: Shopify’s philosophy of building infrastructure before features. As Thawar puts it: “We've always built more infra. We will continue to always build more infra.” How Shopify’s internal AI agent, River, creates a “substrate of information” across the company. How Thawar's OpenClaw agent figured out he was traveling from his calendar — and what that moment told him about where agents are actually headed. You can also listen and subscribe to Beyond the Pilot on Spotify , Apple or wherever you get your podcasts.
Score: 55🌐 MovesJun 24, 2026https://venturebeat.com/orchestration/how-shopify-built-an-ai-stack-that-doesnt-care-which-models-survive - Midjourney Wants to Scan Your Body in 60 Seconds
Midjourney is developing a 60-second full-body ultrasonic scanner and spa-like scanning locations, raising questions about health data and regulation. The post Midjourney Wants to Scan Your Body in 60 Seconds appeared first on TechRepublic .
Score: 54🌐 MovesJun 24, 2026https://www.techrepublic.com/article/news-midjourney-full-body-ultrasonic-scanner/ - Talk of a bubble is 'blasphemy against AI' says SoftBank's Son
Talk of a bubble is 'blasphemy against AI' says SoftBank's Son Reuters
Score: 53🌐 MovesJun 24, 2026https://www.reuters.com/business/media-telecom/talk-bubble-is-an-insult-ai-says-softbanks-son-2026-06-24/ - What’s My SERP Ranking Today? BHMarketer.ai Introduces AI-Powered SERP Analysis with SERP.fyi
What’s My SERP Ranking Today? BHMarketer.ai Introduces AI-Powered SERP Analysis with SERP.fyi USA Today
- Lakewood startup CentSight launches AI CFO platform for small businesses
The founder built his AI-powered tool in just four months with backing from a Detroit venture capital firm, drawing on lessons learned from decisions he wishes he'd made differently at his previous companies.
- Tecno’s EllaClaw AI agent wants to clean up your phone and run your errands
Tecno's EllaClaw agent moves past chatbot replies and starts doing real work, from clearing RAM and saving battery to hailing rides and shopping inside your apps, all while keeping you in control.
Score: 52🌐 MovesJun 24, 2026https://www.digitaltrends.com/phones/tecnos-ellaclaw-ai-agent-wants-to-clean-up-your-phone-and-run-your-errands/ - Avid Launches Fundraising Operating System 4, Adding AI Features That Build Audiences and Draft Campaigns
Avid Launches Fundraising Operating System 4, Adding AI Features That Build Audiences and Draft Campaigns USA Today
- Sustaining Intellect in the AI Economy
Sustaining Intellect in the AI Economy Tepperspectives
Score: 52🌐 MovesJun 24, 2026https://tepperspectives.cmu.edu/all-articles/sustaining-intellect-in-the-ai-economy/ - In Small-Data Medical Imaging, Variance Is the Enemy
Building a pediatric tuberculosis screening tool that generalizes across ages and centers, from 2,000 chest X-rays and 65 models. Photo by CDC on Unsplash Pediatric tuberculosis is one of the harder problems in clinical imaging. In children the radiological signs are subtle and non-specific, microbiological confirmation is unreliable because the disease is often paucibacillary, and many of the settings with the highest burden have very few radiologists. A model that can reliably triage chest X-rays, flagging the children who need confirmatory testing and ruling out those who do not, has real value. This post is a technical case study of the engineering behind one such model, developed in collaboration with a partner hospital and a public TB program in Brazil. My role was the machine learning engineering: the preprocessing, the modeling, and the deployment. Clinical direction and label definitions came from the radiology team. A note on anonymization: institutional names and a few specifics are withheld here because the underlying results are still under peer review. Publishing them in full beforehand could count as prior publication and jeopardize the journal submission. I want to focus on the three problems that took the most work: A preprocessing pipeline that generalizes across children of different ages and across imaging centers. A modeling approach that stays robust with only about 2,000 images. What the results mean and how the model is meant to be used. The data, and why it is hard The development set was 2,011 pediatric chest X-rays collected across several clinical sites in Brazil. Two characteristics shaped every decision that followed. The first is class imbalance. TB prevalence in the development cohort was about 9.8 percent (197 positives). With a positive class that small, plain accuracy is meaningless and even AUROC can flatter a model, so the metrics that mattered were AUPRC and balanced accuracy alongside AUROC. AUPRC is the honest one here, because its no-skill baseline is the prevalence itself, around 0.10. The second is heterogeneity. Pediatric thoraxes change shape dramatically between an infant and an adolescent, so a fixed crop or a fixed normalization does not transfer across ages. On top of that, each center used different equipment, exposure settings, and file conventions. Some of that heterogeneity showed up in ways that are easy to underestimate until they corrupt your labels, which I will come back to. A word on what “positive” means, because pediatric TB makes this harder than it sounds. In children the disease is often paucibacillary, so microbiological confirmation frequently fails even when the child genuinely has TB. Our positive class therefore combined two kinds of label: cases confirmed by laboratory PCR, and presumptive cases where the clinical team diagnosed TB from the full clinical picture together with the radiograph. Folding presumptive cases into the positive class is the deliberate choice for a screening tool, where the cost of missing a case is high and waiting for microbiological proof that often never comes would systematically drop real positives. Part 1: Preprocessing that generalizes The goal here was a single pipeline that takes a raw chest X-ray from any age and any center and produces a consistent input for the network. A few steps did most of the work. Detecting the actual X-ray field. Many radiographs arrive with collimation borders, black margins, or burned-in console text that are not part of the image. The pipeline locates the exposed field in two passes: a coarse pass that reads the transition out of the near-black background to bracket the film, and a refinement pass that scores candidate rectangles by area, fill, and how centered and chest-shaped they are. The result is a bounding box around the real radiograph, and that box is what gets cropped and fed forward. On a clean full-frame image the box hugs the edges and does little; on an image with borders or a console strip it does real work. That is where most of the cross-center variance lived. Why this mattered for our data specifically. The clinical images came from several centers, on different equipment, at genuinely different quality levels: some sites produced clean, well-exposed films, others did not. That heterogeneity is the whole reason field detection and contrast normalization earned their place. Without a step that standardizes the field and the contrast, the network would be learning the scanner and the center as much as the disease. Standardizing the signal before it reached the CNNs was vital, not cosmetic. Aspect-preserving resize to 512. The field crop is letterboxed to 512 by 512, so the thorax keeps its true proportions instead of being stretched to a square. This is the step that lets a toddler’s image and an adolescent’s image be compared as network inputs without distorting either. Contrast normalization with CLAHE. Exposure varied across machines, so I applied CLAHE (Contrast Limited Adaptive Histogram Equalization) to the grayscale crop. CLAHE equalizes contrast locally, which brings out lung detail consistently across sources without blowing up noise the way global equalization can. The output is single-channel grayscale replicated across three channels, which is what the ImageNet-pretrained backbones expect. Figure 1. The pipeline applied to public Montgomery and Shenzhen chest X-rays. Left, the original with the detected field outlined in green. Right, the CLAHE-normalized 512 by 512 output. One pipeline produces consistent inputs across two different sources and both classes. Note on the images: every figure here was generated by running the real production pipeline on public data, the Montgomery and Shenzhen TB datasets. No images from the clinical training set appear anywhere in this post, since those are pediatric patient radiographs from clinical sites and are not mine to publish. Part 2: Modeling with only 2,000 images With a small, imbalanced dataset, the central risk is variance. A single deep network trained on 2,000 images will overfit, and its performance will swing depending on which images happen to land in the validation split. Two design choices addressed this. Five-fold cross-validation, used for everything. Rather than carving off a single fixed validation set and wasting a fifth of an already small dataset, I used five-fold cross-validation and reported out-of-fold (OOF) predictions. Every image gets a prediction from a model that never saw it in training, which gives an honest estimate while using all the data. Because the study reports OOF predictions, it is more accurate to describe a single development cohort than to pretend there is a separate held-out validation split. An ensemble across many architectures. Instead of betting on one network, I trained 13 architectures, each across the five folds, for 65 models in total. Each was initialized from ImageNet weights, with the backbone frozen to train the classifier head first and the last stage then unfrozen for fine-tuning. The final score for an image is the mean predicted probability across the ensemble. Averaging diverse architectures reduces variance far more reliably than tuning a single model, and in a small-data regime that variance reduction is most of the battle. The 13 were ConvNeXt-Tiny, ConvNeXt-Small, EfficientNet-B0, EfficientNet-V2-S, MobileNetV3-Small, MobileNetV3-Large, ShuffleNetV2-x1.0, ResNet18, ResNet50, DenseNet121, RegNet-Y-8GF, ViT-B/16, and a DenseNet121 initialized from CheXNet chest-X-ray weights. Most are convolutional, with one vision transformer and one backbone pretrained on chest radiographs instead of natural images, so the ensemble spans different inductive biases and different pretraining, not just different depths. The payoff shows up in the numbers. Individually, these models landed between about 0.63 and 0.72 AUROC on the internal OOF cohort, with ResNet50 the strongest single architecture at 0.718. The ensemble mean reached 0.744, above every individual member. That gap is the variance reduction doing its job, and it is the whole reason for training thirteen models instead of crowning one. Worth noting against intuition: the chest-X-ray-pretrained DenseNet was the weakest single model here at 0.632, a reminder that in-domain pretraining is not automatically better than ImageNet on a given task. I also tested whether selecting the top few architectures by validation AUPRC would help. It looked better on paper but it was selection bias dressed up as an improvement, so I dropped it. The plain mean of all architectures is the cleaner, more defensible primary metric, and that is what the paper reports. One more negative result is worth stating plainly, because negative results rarely get published. I tested adding lateral views through cascade and fusion approaches. At this dataset scale, frontal-only consistently matched or beat the multi-view approaches on AUPRC. The lateral views added noise rather than signal, so the production model is frontal-only. Figure 2. Architecture of the ensemble: 13 architectures, each trained across 5 folds, averaged into a single mean probability. Part 3: Results The headline numbers come from two cohorts: the internal development cohort (OOF predictions, n = 2,011) and an external validation cohort (n = 199), drawn from the pediatric images within the Montgomery and Shenzhen datasets (the main public TB chest X-ray sets, otherwise adult) plus independent clinical data. On the internal cohort the ensemble mean reached an AUROC of 0.744 (95% CI 0.705 to 0.781) and an AUPRC of 0.279 (0.226 to 0.342). On the external cohort it reached an AUROC of 0.860 (0.785 to 0.924) and an AUPRC of 0.412 (0.237 to 0.650). Against a no-skill baseline equal to the prevalence, around 0.10, that external AUPRC is roughly four times chance, which says more about real performance on the rare positive class than the AUROC does. The external number being higher than the internal one is not a contradiction. The two cohorts have different case mixes, and the only test that counts for a screening tool meant to leave its home institution is whether it generalizes to data it never trained on. It does, at least on this cohort. One honest caveat on that external number. The public images are, on average, cleaner and acquired on better-controlled equipment than the heterogeneous multi-center clinical data the model trained on. So part of the high external AUROC reflects an easier, higher-quality cohort, not pure generalization skill. What that caveat does not touch is the version-over-version gain: 0.751 to 0.860 is measured on the same external cohort, so the improvement is real and not an artifact of cohort quality. The bigger story is that jump from the previous model version, and what drove it. The previous version was trained on about 1,700 images; the new one adds roughly 300 more, bringing the development cohort to 2,011. That is the only substantive change between the two, so the gain is attributable to more data rather than to architecture or hyperparameter tuning. Internally the two versions are essentially tied, AUROC 0.741 to 0.744, a difference of 0.002. Externally the new version went from 0.751 to 0.860, a gain of 0.109 with a DeLong p of 0.0015 and a 95% CI on the difference of 0.044 to 0.183. The improvement is specifically in cross-site generalization, which is exactly where it counts, and it came from adding a few hundred well-curated images to a small dataset, a reminder of how much each image is worth in this regime. Figure 3. ROC curves for the ensemble mean across the development (OOF) and external validation cohorts. The operating point For a screening tool the threshold matters as much as the curve. At the Youden’s J operating point on the external cohort the model showed a clear screening profile: sensitivity 0.95, specificity 0.61, NPV 0.991, PPV 0.22, and balanced accuracy 0.78. It missed one of the twenty positives. That is the shape you want for triage, because a negative result is highly trustworthy, so the model can rule children out and concentrate confirmatory testing where it is needed. The honest caveat: twenty positives is a small denominator, so the sensitivity estimate is encouraging but wide, and this threshold was chosen on the same cohort it is reported on, so treat the operating point as illustrative rather than locked in. We are actively expanding the external validation cohort with pediatric images from other countries, to test the model on a wider range of scanners, populations, and disease patterns than the current set covers. Once that larger, more diverse external cohort is in place, it becomes the right basis for selecting and locking a production threshold, rather than tuning the operating point on the same handful of cases it is reported on. Until then, the numbers above are a credible screening profile, not a fixed deployment setting. Figure 4. Operating point on the external ROC curve, with sensitivity, specificity, and NPV annotated. From model to deployment A model that only runs in a notebook does not help anyone. Training ran on AWS SageMaker, on an ml.g4dn.xlarge GPU instance, but the deployment target was a different world: a server inside the partner program rather than a cloud environment I controlled. The production model was packaged as a container with DICOM handling built in, so it ingests the format hospitals actually produce rather than expecting pre-converted images. It runs frontal-only inference as the production configuration: as Part 2 showed, frontal-only matched or beat the multi-view approaches at this dataset scale, so there was no reason to ship anything heavier. The deployment had a deliberate constraint: it had to run inside the institution, so no patient image leaves it. The ensemble was exported to ONNX, DICOM decoding was handled in-container, and the model was exposed through a local REST endpoint so the equipment calls it over the internal network. That on-prem, data-stays-local design is what keeps it compliant and deployable in a public-health setting. It was validated end to end on the program’s own hardware, a single NVIDIA RTX A4000 16 GB, a modest prosumer GPU rather than a datacenter cluster, which matters because the sites with the highest TB burden are the least likely to have serious GPU infrastructure to spare. The point I would make to anyone doing similar work: budget real time for the deployment surface. DICOM decoding, codec support, and getting a heavy container to run reliably on someone else’s hardware is its own project, separate from the modeling. What it means, and what it does not Used as intended, this is a triage and rule-out tool. The high sensitivity and very high negative predictive value let it prioritize confirmatory testing, directing limited expert attention to the children most likely to need it. It is not a replacement for a radiologist or for microbiological confirmation, and it was never designed to be. The limitations are real and worth naming plainly. The training data came entirely from clinical sites in Brazil, so good performance elsewhere has to be earned, not assumed. There is no dedicated open pediatric TB imaging set. The two main public TB datasets, Montgomery and Shenzhen, are otherwise adult, and the groups that hold large pediatric cohorts generally do not release them. That scarcity is itself one of the hardest constraints of this problem, and it is why the external cohort had to be assembled from the few pediatric cases sitting inside those public sets plus independent clinical data, at 199 images and 20 positives. Good numbers on a problem this data-starved are hard-won, and the same scarcity is exactly what keeps the external validation small. The model did well on it, which is encouraging. Even so, a couple of hundred images pulled from a handful of sources is not broad, multi-site, independent pediatric validation, and I will not pretend it is. The risk that follows is distribution drift. A new deployment site brings its own scanners, exposure settings, patient mix, and disease patterns, none of which have to match the Brazilian development distribution. This is exactly why the planned cross-country external validation is not a formality: images from other countries will carry their own distributions, and the screening profile that looks clean here could move in either direction on data that differs from what the model has seen. The honest next step is prospective validation on independent pediatric cohorts in deployment, with monitoring for drift, rather than treating the external AUROC as a settled property of the model. Where this is meant to go. The goal was never a benchmark number; it was a tool that reaches the children who need it. The direction we are working toward is a deployable screening aid integrated into the public diagnostic pathway for pediatric TB in Brazil, through the SUS, the country’s universal public health system that covers the entire population at no point of care. The case for it is concrete: many towns in the interior of Brazil have no radiologist on site at all, so a chest X-ray taken there may wait days to be read, or be read by a non-specialist. In that gap, a preliminary screening tool that runs locally and flags the films that need urgent expert attention has real triage value, and it is most useful precisely where specialists are scarcest. The framing matters: this is one more instrument in the diagnostic process for pediatric TB, sitting alongside the clinician and the confirmatory test to help triage and prioritize, not a system that decides on its own. One could imagine it entering the network first as exactly that preliminary screen and evolving from there as it earns broader validation. That is a direction, not a roadmap, and it is the kind of future that building for accessible, in-facility hardware from the start is meant to keep open. Takeaways If I had to compress this into a few lessons for the next person building a medical imaging model on a small, multi-center dataset: Preprocessing that generalizes across patients and sites is most of the work, and detecting the true image field and normalizing contrast consistently did more for generalization than any modeling trick. With limited data, variance is the enemy, and the ensemble is not a nice-to-have, it is the result. A cross-validated ensemble of diverse architectures beat every single model it was built from, and that variance reduction, not any one clever backbone, is what produced a number worth reporting. On a small, imbalanced dataset, averaging across architectures is the single highest-leverage decision. More views did not mean more signal. Across extensive experimentation, frontal-only consistently matched or beat the multi-view cascade and fusion approaches at this scale. What actually moved performance was not more projections but better inputs: standardized, high-quality frontal images. Spend the effort on input quality and consistency before reaching for extra views. Report the negative results. The selection-bias trap in top-k architecture picking, where filtering to the best validation scores looked like an improvement but was just leakage in disguise, is as useful to readers as the headline AUROC. None of these are exotic. They are the unglamorous discipline of small-data medical imaging: respect the variance, distrust the easy win, and put the work into the inputs and the validation rather than the cleverness of the model. It comes down to this. When the positive class is fewer than two hundred cases, every single one is precious, and there is no volume of data to paper over a sloppy pipeline. That is what makes the preprocessing almost artisanal: each image has to be found, verified, field-detected, and normalized with care, because one silently dropped or distorted positive is a meaningful fraction of everything the model has to learn from. And it is why the internal cross-validation is not a box to tick but the backbone of the whole effort, the only way to get an honest read on a dataset too small to spare a held-out split. Get the careful, per-image work and the rigorous internal validation right, and the reward is not a leaderboard score. It is a model that still behaves when it leaves the building it was trained in, which, for a tool meant to help find sick children where specialists are scarce, is the only result that matters. In Small-Data Medical Imaging, Variance Is the Enemy was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
- Vishal Sikka’s AI startup Hang Ten Systems raises $32 million led by Mayfield
Former Infosys CEO Vishal Sikka has launched Hang Ten Systems, an AI startup focused on enterprise solutions, securing $32 million in seed funding led by Mayfield. The Palo Alto-based company aims to help large corporations translate AI investments into tangible business results.
- MoEngage acquires AI startup Aampe in first acquisition; eyes more deals in US, Europe
Customer engagement platform MoEngage has acquired AI startup Aampe, bolstering its artificial intelligence capabilities. MoEngage cofounder and chief executive Raviteja Dodda said the company is evaluating more inorganic growth opportunities, particularly in markets such as the US and Europe.
- Investors Are Using AI for Financial Decisions. They Still Want a Human Advisor.
Investors Are Using AI for Financial Decisions. They Still Want a Human Advisor. Barron's