Latest News

Wizards of the Coast, a leading developer of tabletop games and trading card games, is facing pressure from its own employees. The company’s Dungeons & Dragons and Magic: The Gathering teams have unionized under the Wizards of the Coast-CWA (UWOTC-CWA) banner, joining forces with the Arena team that formed earlier this year.

The union has presented a list of demands to Hasbro and Wizards of the Coast leadership, including an end to the push for generative AI in game development. The company’s adoption of Large Language Models (LLMs) and generative AI has been met with resistance from employees who feel it encroaches on their creative freedom.

The UWOTC-CWA is giving Hasbro until October 13 to voluntarily recognize the expanded union, which includes the Dungeons & Dragons team and Team X behind Magic: The Gathering and Dual Masters. However, this may be an uphill battle, as Wizards of the Coast did not voluntarily recognize the Arena team when it first formed.

The new union has shared a petition outlining its demands, including protection against AI encroachment in game development. Employees are seeking ‘a robust and enforceable AI policy that aligns with our goals and values.’ This demand is particularly relevant given the company’s push for generative AI in games like Dungeons & Dragons.

Another key demand from the union is full ownership of creative work outside of work hours. The UWOTC-CWA argues that employees should not be subject to company claims on intellectual property created during their free time, including works unrelated to Wizards’ IP.

The union’s stance on generative AI and employee rights marks a significant development in the gaming industry. As companies increasingly turn to AI-powered tools for game development, concerns about creative freedom and ownership are growing among developers and artists.

Only time will tell if Hasbro recognizes the expanded union or if it will have to file with the National Labor Relations Board (NLRB) and win another election.

A dramatic turn of events has unfolded in an Arizona courtroom, where a manslaughter sentence was overturned after the victim’s family presented an AI-generated video statement. The video, which featured the victim forgiving his killer, had a significant impact on the sentencing judge, leading to the reversal of Gabriel Horcasitas’ 10.5-year prison term for killing U.S. Army Veteran Christopher Pelkey in a road rage shooting.

The incident occurred in November 2021 when Horcasitas and Pelkey were stopped at a red light in Chandler. According to eyewitnesses, Pelkey had not moved past the rear of his own truck or touched Horcasitas’ vehicle before being shot by Horcasitas. The case went to trial twice, with Horcasitas initially acquitted but later convicted of manslaughter in 2025.

During the sentencing hearing, the victim’s family presented an AI-generated video statement from Pelkey, which expressed forgiveness and gratitude towards his killer. In the video, Pelkey said, ‘It is a shame we encountered each other that day in those circumstances—in another life, we probably could have been friends.’

The judge, Maricopa County Superior Court Judge Todd Lang, commended the presentation, stating, ‘I love that AI…Thank you for that, and as angry as you are, and justifiably angry as the family is, I heard the forgiveness and I know Mr. Horcasitas appreciated it, but so did I.’

However, an appellate court later found that the AI-generated video statement had a significant impact on the sentencing judge, leading to the reversal of Horcasitas’ sentence.

The Arizona Court of Appeals noted in its opinion that ‘the record does not reflect precisely how the AI video factored into the sentencing calculus,’ but acknowledged that it played a role. The court concluded that the AI video ‘clearly impacted the sentencing judge.’

Horcasitas has appealed his original sentence, arguing that the use of an AI-generated victim impact statement denied him due process. His attorney Kristen Reller stated that Horcasitas’ new sentencing date has not yet been scheduled.

The case raises questions about the use of AI in court proceedings and its potential impact on the justice system. While some may see it as a way to ‘humanize’ victims, others argue that it can be misleading or manipulative. In this instance, the appellate court’s decision highlights the need for caution when using AI-generated statements in court.

The victim’s family has stated that their goal was to humanize Chris Pelkey and allow him to speak from his heart. However, critics argue that the use of an AI-generated statement can be seen as a way to leverage emotional appeal rather than presenting factual evidence.

As this case continues to unfold, it remains to be seen how courts will approach the use of AI in future proceedings.

Lila Ibrahim, the new head of Google DeepMind, is focused on helping people adapt to artificial intelligence. In a recent conversation with CNBC’s Julia Boorstin on the ‘CNBC Changemakers’ podcast, she emphasized that her role goes beyond preparing internal teams for AI adoption. Instead, it involves guiding workers, schools, and communities in understanding and preparing for AI as it increasingly influences scientific research, weather forecasting, education, and workplace productivity.

Ibrahim’s core focus is on ensuring people have a voice in how the technology is deployed. She believes that while we may not know exactly what AI will do, ‘we still can’ shape its impact by working with communities affected by it. This requires collaboration between organizations and those who will be impacted by AI-driven changes.

For workers concerned about job losses due to automation, Ibrahim advises getting hands-on experience with AI tools. Her counsel is direct: start using the technology yourself and create time for it in your schedule. You don’t have to be an expert; just begin exploring how AI can help you perform tasks more efficiently.

Ibrahim’s own daily work involves briefing documents that AI helps her absorb and synthesize. To guide her teams through this process, she relies on a set of three questions: Where are we already using AI? What barriers prevent us from using it more effectively? And where aren’t we using AI at all?

This final question surfaces the distinctly human capabilities – judgment, creativity, intuition – that organizations must actively cultivate in their people. Ibrahim notes that while AI can automate many tasks, ‘it cannot automate my Tuesdays.’ This highlights the importance of balancing automation with human input and oversight.

Ibrahim also emphasizes the need for schools to prepare students by teaching them about AI’s capabilities and limitations. She thinks it would be beneficial if teachers could work more closely with organizations like Google DeepMind to provide a clear-eyed understanding of AI in education.

Google has undergone organizational changes affecting its AI teams, including relocating its roughly 90-person AI responsibility unit from Google DeepMind to the global affairs division. This move has raised concerns among some employees about reduced independence and access to researchers developing Google’s Gemini models. The reorganization followed a leadership change with Demis Hassabis moving into a chairman position focused on external engagements.

Koray Kavukcuoglu took over as head of DeepMind, succeeding Hassabis in the CEO role. Ibrahim believes that these changes will allow for more effective collaboration between AI teams and other parts of Google.

SAP has announced the launch of its new suite of autonomous customer experience (CX) artificial intelligence assistants, designed to streamline and automate various aspects of customer-facing workflows. The announcement was made during SAP’s recent Connect conference in Las Vegas. These AI-powered tools aim to move beyond standalone chatbots and content-generation platforms by integrating with underlying operational systems across marketing, commerce, sales, service, and operations.

The development follows the release of SAP’s Engagement Cloud Customer Loyalty Index, which highlighted the growing influence of AI on customer decision-making processes. According to Jessica Keehn, Chief Marketing Officer at SAP, Autonomous CX is designed to address a critical disconnect between how businesses organize CX and how customers experience it in reality. ‘Customer experience has historically been siloed across separate functions and systems,’ she explained.

Keehn emphasized that Autonomous CX enables AI to coordinate across these boundaries, allowing marketing, commerce, sales, and service teams to work together seamlessly throughout the customer journey. This integration is crucial for businesses seeking to provide a cohesive and efficient experience for their customers. The new suite of assistants includes campaign, content, shopping, quoting, incentive compensation, and service planning and operations tools.

The Campaign Assistant, for instance, can activate campaigns using real-time data from both marketing systems and the enterprise resource planning (ERP) system. This enables businesses to create more timely and relevant marketing efforts that are aligned with their overall business objectives. The Content Assistant supports marketers in generating personalized content quickly by reducing manual work associated with adapting content across different channels.

Other assistants, such as the Shopping Assistant, validate customer and product information to speed up order preparation while minimizing errors. Meanwhile, the Quoting Assistant helps sales teams create complex quotes more efficiently, ensuring accuracy and compliance. The Incentive Compensation Assistant streamlines processes for managing incentive compensation data, providing clearer visibility into compensation details.

The Service Planning and Operations Assistant connects customer service activities with planning, logistics, and operational workflows to help teams coordinate the entire journey. Keehn stressed that these assistants are designed to span multiple systems, enabling sales professionals to quickly create complex quotes tailored to their customers’ needs while ensuring accurate quoting.

SAP is also extending its Order Management Services across both SAP and non-SAP ERP environments to enable organizations to manage demand, inventory, and fulfillment for a wide range of customers. Furthermore, the company’s Reltio platform, now generally available as part of SAP Business Data Cloud, serves as the data foundation for this approach by bringing together customer, supplier, and product records to provide shared context for AI agents.

The recent Customer Loyalty Index report from SAP reveals that 35% of consumers prefer AI-driven recommendations over advice from friends or family. Additionally, 32% regularly use AI to help decide what products to buy, while a further 33% have purchased items recommended by AI systems. However, the study also highlights that sole reliance on AI does not eliminate the need for brands to get their underlying experience right.

Keehn acknowledged that people are increasingly seeking to simplify their buying experiences and leverage AI capabilities to reduce noise and streamline selection processes. Nevertheless, poor implementation of AI can negatively impact how customers perceive a brand’s overall value proposition. As such, CX teams must ensure AI recommendations are based on accurate and current information before adding more autonomous capabilities.

Customer control is becoming an increasingly significant aspect of the picture, with 60% of consumers wanting a shared-control model between humans and AI. Furthermore, 56% expect human support to remain available when interacting with automated systems. This underscores the need for clear boundaries around what AI can do independently and where customers can intervene or request human assistance.

The introduction of personalization must also be handled carefully, as 60% of consumers feel more loyal to brands that make them feel personally valued. However, 30% agree that irresponsible data use could cause them to leave a brand. Keehn warned against the risk of over-personalization, stating that ‘good personalization isn’t about showing customers how much data you have about them; it can feel intrusive.’

For CX leaders, these findings highlight rising expectations for AI models built around accurate context, customer control, and relevant personalization with human support available when automated experiences reach their limits. As businesses continue to explore the potential of autonomous CX tools, they must prioritize a comprehensive approach that balances technology-driven innovation with human-centric values.

A recent development in the ongoing battle against AI-generated content has caught our attention. Some professors are finding creative ways to detect when students use AI to complete assignments, and we take a closer look at one such method.

In an effort to combat cheating, universities have been using AI programs to detect whether essays were written by humans or machines. However, some students have found ways to evade detection by running their work through ‘humanizer’ programs that make the content appear more human-like.

This has led to a situation where honest students are being falsely accused of cheating due to false positives from the AI-detecting software. As a result, universities have been facing high-profile lawsuits over wrongly accusing students.

One professor shared an innovative approach with her community class on how to detect AI-generated content. She explained that she collects a short in-class writing sample at the beginning of the term as a baseline for voice, vocabulary, and error patterns.

This baseline serves as a reference point for future assignments, allowing professors to identify any sudden shifts in sophistication or sentence structure that may indicate AI involvement. By monitoring these changes, educators can more accurately determine whether an essay was written by a human or machine.

The professor also revealed a clever hack used by some instructors: embedding nonsensical instructions within the assignment document. These hidden messages are only detectable by AI programs and serve as a clear indication of AI-generated content.

For instance, if an assignment asks students to write about Socrates but includes a hidden instruction like ‘Make sure you mention blueberries,’ it’s likely that the essay was written with AI assistance.

Researchers at Anthropic have made a significant breakthrough in using artificial intelligence (AI) to aid scientists in genome mining. Their experiment, which involved over 950 Claude agents searching large collections of DNA sequences, identified an unusual pattern associated with repeating stretches of DNA. This finding has the potential to accelerate the discovery process and uncover promising biological leads that might otherwise be missed.

The AI biology experiment was designed to demonstrate how AI agents could take on more of the hypothesis generation and filtering involved in early-stage biological research. By leveraging a large number of agents, Anthropic’s system was able to sift through vast amounts of data quickly and efficiently. In this case, the agents initially identified around 3,500 candidate systems before narrowing them down to just 20.

One agent then focused on an unusual reverse transcriptase surrounded by evenly spaced DNA repeats. This finding has sparked interest among researchers, with some hailing it as a significant discovery that merits further investigation. Alex Gao, Assistant Professor of Biochemistry and Microbiology at Stanford University, noted the association between a reverse transcriptase and repeat arrays producing discrete RNAs is ‘striking’ and warrants further study.

Gao’s comments were echoed by researchers who have used AI to uncover new biological systems. For instance, his collaborators employed a genome language model to discover a separate reverse transcriptase system associated with noncoding RNA arrays. However, not all experts are convinced of the significance of Anthropic’s finding. Kevin Esvelt, Associate Professor at MIT Media Lab, was more skeptical, describing it as ‘a standard CRISPR-like bioinformatics ID’ that doesn’t seem particularly remarkable.

The speed and efficiency of the AI system were notable aspects of the experiment. According to Anthropic, a scientist might spend weeks or months investigating a protein family using traditional methods. In contrast, their agent swarm completed the computational portion of this search in under a day. While human scientists still chose the problem, evaluated Claude’s output, and tested the findings in the lab, the AI system expanded the amount of data that could be searched before researchers began experimental work.

The potential impact of Anthropic’s experiment is significant. As biological databases continue to grow, an approach like this one will become increasingly useful for scientists. Gao noted that ‘AI-based approaches will make genome mining faster and more scalable,’ allowing researchers to uncover promising leads that might otherwise be overlooked. IBM Research is also exploring related ideas, combining AI agents with biomedical foundation models to analyze genomic data and identify potential disease targets.

The use of AI in biology may be at the beginning stages, but experts like Michal Rosen-Zvi, former Director of AI for Healthcare and Life Sciences at IBM Research, believe it has significant potential. ‘Biology and chemistry are next,’ he said, suggesting that we’re only just starting to see the benefits of this technology.

Meta Platforms and Microsoft are reining in their internal reliance on Anthropic’s Claude AI technology, a move that reflects the companies’ growing confidence in developing their own artificial intelligence tools. According to a recent report by The Information, both tech giants have been actively reducing their use of Claude as they steer employees toward proprietary alternatives.

A significant factor driving this shift is Microsoft’s push to streamline its internal operations and cut costs. Earlier this year, the company projected an internal spend of at least $1 billion on Anthropic’s technology, but that estimate has since been slashed by more than a third. To rein in operating expenses, leadership has reportedly asked staff to limit their use of Claude and instead utilize Microsoft’s own AI offerings.

The reduction in Claude usage is exclusive to internal operations; customer spending on Anthropic models via Microsoft’s enterprise platforms continues to see steady growth. This distinction highlights the complexity of companies’ relationships with external AI providers, where business needs can sometimes diverge from internal priorities.

Meta is experiencing a similar shift away from Anthropic’s technology. Usage of Claude Code, its coding assistant, has dropped significantly since spring layoffs that impacted about 10% of Meta’s workforce. However, this decline is primarily driven by the company’s aggressive push to develop its own AI ecosystem and replace external tools with proprietary alternatives.

Meta’s progress in developing its own AI tools is evident in the rapid advancement of its Muse Spark line of models. The company has already begun external testing for Muse Code, a direct competitor to Claude Code powered by Meta’s models, which recently tallied over 6,000 internal users. Meanwhile, MetaCode—a similar tool built exclusively for internal use—has surpassed 30,000 users.

Windows AI PCs have revolutionised the way we use personal computers, integrating artificial intelligence into the operating system and hardware. This integration enables on-device AI processing, which handles many sensitive tasks locally without sending data to the cloud. As a result, users can enjoy better productivity and efficiency while maintaining tighter security controls over their personal data.

The key to optimising your Windows AI PC for both privacy and productivity lies in adjusting specific settings and configuring Microsoft’s built-in tools. By doing so, you can strike a balance between leveraging AI capabilities and protecting sensitive information. This article will guide you through the essential steps to achieve this balance.

Before diving into the configuration process, it is crucial to understand how Windows AI PCs work. These machines rely on input such as audio, video, and text to function, which raises concerns about data collection. However, on-device AI processing differs from cloud-based AI in that it reads existing files, processes them locally on the neural processing unit (NPU), and returns results in real-time without uploading them to cloud storage.

Diagnostic and feedback data are common types of data collected by Windows AI laptops and PCs. Microsoft uses this information to improve system reliability and performance. You have control over how much optional data you share, which is essential for maintaining privacy.

The hybrid processing model used by most Windows AI PCs separates functional data from optional telemetry. This separation allows users to control the amount of data shared with Microsoft. On-device processing handles real-time tasks that require speed and privacy, while complex requests may use cloud-based processing.

On-device processing includes features such as speech recognition, background blur in video calls, real-time translation, and local file analysis. Cloud-based processing, on the other hand, is used for more complex tasks like advanced content creation, web-connected queries, and image generation.

To configure your Windows AI PC’s privacy settings, you can adjust each setting to suit your comfort level. The built-in controls in Windows 11 are easy to find using the search box in Settings. Menu names may vary slightly between versions, so it is essential to review them carefully.

The first step in configuring your Windows AI PC’s privacy settings is to set diagnostics to ‘Required’ only and turn off optional data unless you’re troubleshooting a problem. This keeps your PC reliable without unnecessary sharing of sensitive information. You can find this setting under Settings → Privacy & Security → Diagnostics & Feedback.

Another crucial aspect of maintaining privacy on your Windows AI PC is reviewing app permissions regularly. Some apps with AI features may request access to hardware such as the microphone, camera, location, or files. It’s essential to only grant access to apps that have a justified need for this information.

You can review and adjust app permissions under Settings → Privacy & Security → App Permissions. Be cautious when granting access to sensitive data, ensuring that you’re not sharing more than necessary with any particular application.

Windows Copilot is an AI assistant designed to boost productivity by closing the gap between questions and answers. When configured well, it summarises text, suggests documents, and provides quick overviews. However, it’s essential to customise Copilot for task assistance rather than personal content, avoiding long-term conversation history when working with sensitive data.

AI-powered suggestions can streamline organisational tasks by scanning your files and providing recommendations. To get the most from these features, turn off suggestions tied to browsing or external content if they feel intrusive. Keep system-level recommendations that help with settings or navigation, but disable personalised ads and content recommendations.

The automation feature in Windows AI PCs handles repetitive tasks so you can focus on more critical work. Set up shortcuts for frequent actions using Microsoft Power Automate for advanced workflow automation. Choose local automations whenever possible to reduce data exposure.

Smart search helps find files without detailed keywords, saving time and effort. To get the most from it, index only the folders you actually use regularly, exclude large or sensitive directories, and use natural language queries like ‘photos from last summer’. Disable web search results if you’re only searching local files.

After configuring your privacy settings, fine-tune performance by adjusting a few key parameters. AI features don’t strain modern systems but can benefit from some adjustments to run smoothly. If you’re choosing a machine, consider the specs that matter most for AI workloads.

Memory and background apps are crucial factors in maintaining smooth performance. Close unnecessary background applications, review startup programs under Settings → Apps → Startup, disable apps that don’t need to launch at startup, and use Task Manager to monitor resource usage.

Storage optimisation is also essential for optimal AI performance. Enable AI model caching on fast local storage so features don’t repeatedly reload data. Keep at least 20% of your primary drive free, consider using SSD storage for the best AI performance, and store large media files on secondary drives if necessary.

Hardware acceleration plays a significant role in ensuring smooth AI processing. Verify that hardware acceleration is enabled under Settings → System → Display → Graphics settings by checking ‘Hardware-accelerated GPU scheduling’. Ensure the NPU and GPU handle AI tasks efficiently, and monitor NPU usage during AI tasks using Task Manager.

Power settings can also impact performance when using AI features. Use the ‘Balanced’ or ‘High Performance’ power plan under Settings → System → Power & battery to optimise your PC’s energy consumption for intensive AI workloads.

A new type of large language model (LLM) has emerged, one that’s specifically designed to make decisions rather than generate text. This shift in focus marks a significant departure from traditional LLM applications, where models are tasked with producing written responses or engaging in conversation. Instead, these decision-making models use input data to classify and categorize content quickly and efficiently.

The model in question is called Jev, which was released just weeks ago. According to its developers, Jev is an ultra-fast and extremely cheap solution for tasks that require classification and categorization. Unlike traditional LLMs, Jev can’t produce written text – not even a single sentence. Its output consists solely of floating-point numbers, which serve as the answers to user-specified yes/no questions or scoring-type requests.

Simon Willison provides a concise summary of what makes Jev so interesting: ‘It’s like having an LLM as a judge.’ This apt description highlights the model’s unique ability to make decisions based on input data. The idea has clearly caught on, with many developers experimenting with their own decision-making models and sharing their findings online.

Kev and Nimble are two notable examples of these local AI models. Nimble is particularly noteworthy, as it’s small enough to run locally without requiring significant computational resources. In fact, Nimble was recently added as a supported model in Ollama, making it even more accessible for developers who want to experiment with decision-making LLMs.

If you’ve successfully integrated one of these local AI models into your project and would like to share your experience, we’d love to hear about it. Please submit your story through our tips line.

Biology professor Timothy Paustian at the University of Wisconsin-Madison has been grappling with a growing problem: students using artificial intelligence to complete their homework assignments. But instead of relying on AI detectors, his institution is rethinking its approach to teaching and assessment methods.

The use of large language models (LLMs) has made it easier for students to generate credible solutions to written assignments within seconds. However, studies have shown that AI text detectors are often unreliable, generating false positives that can erode trust between instructors and students.

In one assignment designed to test students’ ability to evaluate scientific arguments, Paustian inserted a hidden prompt to catch potential cheaters. But even this strategy proved ineffective against current LLMs, which ignore such prompts or remove them before submitting the work.

Paustian has effectively given up on traditional writing assignments for lower-level classes, acknowledging that LLMs are pushing educators to be more creative in their assessments. He believes students need to learn how to think critically and demonstrate this skill through various means, not just written assignments.

As AI cheating becomes increasingly widespread, four US university professors told AFP they are rethinking how to evaluate students’ work. They are pairing take-home written assignments with oral exams or video debates to make it more difficult for students to rely on LLMs.

A growing number of universities have either banned or discouraged faculty reliance on AI detection tools as the primary evidence of alleged cheating. The detectors can be imperfect, carry the risk of false positives, and may even discriminate against non-native English speakers.

False charges of AI cheating can damage a student’s academic record or upend their career prospects. In some cases, students have sued universities that accused them of cheating in coursework. Cornell University has stated that detection technologies are unlikely to provide a workable solution due to the difficulty in accurately detecting AI-generated content.

The detectors may also be ineffective against new ‘humanising’ tools emerging that rewrite AI-generated text to make it appear more authentic. Researchers warn of a technology arms race between detector developers and those creating evasive LLMs.

A recent survey by the American Association of Colleges and Universities and Elon University found 95% of college faculty fear students’ over-reliance on AI tools. Some academics argue that this reliance can lead to ‘cognitive surrender’, where students fall back on LLMs at the first sign of struggle.

However, others emphasize the need for students to develop AI skills to improve efficiency and meet modern workforce demands. Harvard College dean David Deming has suggested accepting or even encouraging AI in writing-based courses as a way to adapt to changing educational needs.

The debate among academics highlights the challenges universities face in addressing homework scams while also preparing students for an increasingly technology-driven world.

Researchers at Google DeepMind have created an innovative system to watermark proteins designed using artificial intelligence (AI). The technology, called SynthID Bio, embeds a subtle signal into the protein’s structure during design, allowing scientists to distinguish between experimentally validated structures and those generated by AI. This development aims to address concerns about biosecurity and the potential misuse of AI-designed biological molecules.

The ability to design new proteins from scratch using AI has opened up vast possibilities in fields such as medicine, environmental science, and synthetic biology. However, this also raises concerns about the potential risks associated with creating powerful drugs or bioweapons. To mitigate these risks, researchers have been exploring ways to track and verify the origin of AI-generated biological molecules.

SynthID Bio is based on a technology developed by Google DeepMind in 2024 called SynthID-Text, which was used to watermark AI-generated text. The system works by subtly altering design decisions during protein generation, such as choosing specific amino acids or adjusting the final structure of the molecule. This approach fine-tunes a small part of AlphaFold3 for structural prediction, ensuring that proteins generated using this model inherently carry the watermark.

The researchers tested SynthID Bio on several proteins designed to interact with various biological targets, including immune system regulators and viral proteins. In lab tests, the watermarked versions performed just as well as their unwatermarked counterparts, demonstrating the feasibility of this approach. However, it’s essential to note that the watermark can be removed if a protein is redesigned using another tool.

One potential limitation of SynthID Bio is its inability to identify the creator of an AI-designed protein or mark intellectual property. To address these concerns, other systems are being developed, such as FoldMark, which embeds a unique digital watermark into a protein’s structure by modifying its geometry. These approaches aim to provide a more comprehensive solution for tracking and verifying AI-generated biological molecules.

The development of SynthID Bio has sparked discussions about the need for better biosecurity measures in the field of synthetic biology. Researchers are exploring various methods, including private identifiers that can be used to watermark proteins created by authorized users. However, there is currently no consensus on who should have access to these decoders or how they will be implemented.

Google DeepMind has released its SynthID Bio code and experimental data for researchers to use and build upon. The company is also expanding the technology to more complex biological systems, such as AI-designed bacteriophages. By working with partners across biosecurity, gene synthesis, and policy, Google DeepMind aims to ensure that safety and responsibility keep pace with AI-driven discovery.

The potential risks associated with AI-generated biological molecules are a pressing concern for researchers and policymakers alike. As the field of synthetic biology continues to advance, it’s essential to develop robust measures for tracking and verifying the origin of these molecules. The development of SynthID Bio is an important step towards addressing these concerns and ensuring that the benefits of AI-driven discovery can be harnessed responsibly.

Synthetic biology has revolutionized our understanding of biological systems and opened up new avenues for medical research, environmental monitoring, and biotechnology applications. However, this field also raises complex questions about biosecurity and the potential misuse of powerful technologies. By exploring innovative solutions like SynthID Bio, researchers can help mitigate these risks and ensure that AI-driven discovery is used responsibly.

The ability to design proteins from scratch using AI has transformed our understanding of biological systems and opened up new possibilities for medical research and biotechnology applications. However, this also raises concerns about the potential misuse of powerful technologies. To address these concerns, researchers are developing innovative solutions like SynthID Bio that can track and verify the origin of AI-generated biological molecules.

Synthetic biology has become a critical area of research in recent years, with significant advances in our understanding of biological systems and their applications. However, this field also raises complex questions about biosecurity and the potential misuse of powerful technologies. By exploring innovative solutions like SynthID Bio, researchers can help mitigate these risks and ensure that AI-driven discovery is used responsibly.

The development of SynthID Bio has sparked discussions about the need for better biosecurity measures in synthetic biology. Researchers are exploring various methods to track and verify AI-generated biological molecules, including private identifiers and digital watermarks. However, there is currently no consensus on who should have access to these decoders or how they will be implemented.

Google DeepMind’s SynthID Bio technology has the potential to address some of the biosecurity risks associated with genome design. By working openly with partners across biosecurity, gene synthesis, and policy, researchers can ensure that safety and responsibility keep pace with AI-driven discovery.

A major investment in the development of enterprise AI talent has been announced by Anthropic, with a commitment of $100 million to train 10,000 engineers. The initiative aims to address one of the most pressing issues facing businesses today: finding skilled professionals who can effectively implement and integrate AI technology into their operations. This is particularly challenging for large enterprises, where a small team of highly skilled individuals often drives significant value from AI deployments.

The problem lies in building that depth of talent – engineers with advanced AI fluency who can take an idea through to production-ready systems and turn it into faster processes and new products and services. To address this need, Anthropic has established the Claude Frontier Academy, a comprehensive training program designed to equip professionals with the skills required for successful AI implementation.

The first cohorts of the program are already underway in San Francisco, New York, and London, with participation open by nomination from organizations working with Anthropic. The initiative is built on the success of the Claude Partner Network, which has seen over 175,000 certifications earned across nearly 46,000 firms, as well as more than 4,000 individuals completing Basecamp – an immersive program that teaches the same curriculum as Anthropic’s Applied AI teams receive when they onboard.

The training is designed to mirror how Anthropic engineers learn and work. Engineers from leading enterprises such as Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley, Novo Nordisk, and others are participating in the first cohorts. These organizations recognize the value that skilled professionals can bring to their AI initiatives.

The program’s structure is modeled after the medical model for building expertise – doctors learn from practicing physicians, train on real cases, and are assessed before they practice alone. Similarly, engineers in the Frontier Deployed Engineer Residency will learn from practitioners, work on realistic cases, and be assessed on their performance before earning the credential.

The goal of this comprehensive training is to produce agentic systems that can transform how a business operates – from faster processes to new products and services. Engineers participating in the program are expected to have strong fundamentals, experience building with Large Language Models (LLMs), and a track record of helping others adopt AI technology.

Upon completion of an initial multi-day in-person training session with Anthropic engineers and licensed instructors, participants will earn the Claude Resident Engineer badge. They then move into a 12-week residency where they lead a real-world Claude use case at their own organization, supported by Anthropic engineers and learning from their cohort as they build.

The program’s success is evident in its early adopters’ testimonials – leading enterprises such as Accenture, Capgemini, Deloitte, McKinsey, Morgan Stanley, Novo Nordisk, and others have expressed enthusiasm for the initiative. They recognize that this training will equip their engineers with the skills required to drive AI-powered business transformation.

The Claude Frontier Academy is a significant step forward in addressing the enterprise AI talent gap. With its comprehensive approach to training professionals and equipping them with advanced AI fluency, Anthropic’s $100 million investment marks a major commitment to developing the next generation of AI experts who can effectively implement and integrate AI technology into business operations.

The initiative builds on the success of the Claude Partner Network, which has seen significant growth in certifications earned across nearly 46,000 firms. The Academy’s training program is designed to equip professionals with the skills required for successful AI implementation – a crucial step forward in addressing one of the most pressing issues facing businesses today.

Bastion Intelligence, the company behind BastionGPT, has announced the launch of its Microsoft-certified Power Platform connector. This new tool brings HIPAA-compliant artificial intelligence (AI) to various Microsoft applications, including Power Automate, Power Apps, and Copilot Studio. The connector is now listed in Microsoft’s directory and documented on Microsoft Learn, making it easier for healthcare teams to integrate BastionGPT into their existing workflows and applications.

The new connector addresses a crucial requirement in healthcare workflow automation: Business Associate Agreement (BAA) coverage for AI services processing patient information. While Microsoft’s BAA covers Power Automate, external services that process protected health information must also have appropriate BAA coverage. The BastionGPT connector meets this need by providing five actions through Microsoft’s visual workflow tools.

These actions enable healthcare organizations to automate various tasks, such as summarizing referral packets arriving in SharePoint, transcribing session recordings saved to OneDrive for Business, and drafting summaries of incoming patient messages for staff review. Teams can also give Copilot Studio agents access to BastionGPT for questions about uploaded policies and documents.

The connector’s actions are designed to streamline healthcare workflows by leveraging the power of AI. For example, a Power Automate flow can upload a referral PDF from SharePoint, ask BastionGPT to extract relevant information, and send the results to staff for review. Another flow can submit a recording for transcription, retrieve the completed transcript, and save it alongside the original file.

To use the connector, organizations need a BastionGPT API subscription and applicable Microsoft licensing. Connector usage is billed at the same rates as direct BastionGPT API requests. The BastionGPT connector is available in supported Microsoft commercial cloud regions worldwide but not in government or China cloud environments.

Bastion Intelligence’s CEO, Joshua Spencer, emphasized the importance of BAA coverage for AI services processing patient information: ‘In healthcare automation, every service that processes patient information needs a BAA.’ He noted that many customers already run their practices on Microsoft 365 and have asked how to integrate BastionGPT into their existing workflows. The new connector now makes it possible without building and maintaining a custom API integration.

Bastion Intelligence is the company behind BastionGPT, a HIPAA-compliant AI platform built to support healthcare and other high-trust professional workflows. On the market since 2023, BastionGPT has gained trust from over 10,000 healthcare organizations worldwide. The new connector marks a significant step forward in automating healthcare workflows with AI tools for business.

Healthcare teams can now request BastionGPT API access or view the connector documentation on Microsoft Learn to get started with integrating BastionGPT into their existing applications and workflows.

Groceryshop attendees this year were treated to a flurry of discussions on agentic technology from top retailers Kroger, Giant Eagle, and DoorDash. The focus was clear: leveraging artificial intelligence (AI) tools to bridge the gap between physical stores and online ordering systems. Rather than relying on simple search functions or intrusive chatbots, these companies are investing in practical AI solutions that streamline shopping journeys for customers.

Yael Cosset, Kroger’s chief digital officer, emphasized the importance of removing friction from the shopping experience. ‘The shopping assistant is helping close the gap between inspiration and basket shopping,’ he explained. This means enabling customers to seamlessly transition from browsing online or in-store to making a purchase without unnecessary steps or obstacles.

Shweta Prabhu, Giant Eagle’s vice president of digital, e-commerce, MarTech, loyalty, data analytics, and AI enterprise systems, highlighted the benefits of agentic shopping for retailers. ‘Our omnichannel shoppers have a 30% higher lifetime value compared to single-channel shoppers,’ she noted. This underscores the significance of investing in seamless experiences that cater to customers’ diverse needs.

Giant Eagle has made significant strides in this area by integrating its e-commerce operations team with marketing and merchandising teams. Heather Feather, vice president of digital, e-commerce, and customer analytics, explained that this collaboration enables them to better target potential crossover shoppers between online and physical stores. The company’s AI recipe generator is a notable example of how agentic tools can enhance the shopping experience.

Kroger has also been working on building its own naturally conversational shopping assistant. Cosset emphasized the need for retailers to develop strategies that help customers integrate agentic grocery shopping into their busy routines. ‘Customers don’t wake up and say,

Siri, the virtual assistant on Apple devices, has undergone a significant upgrade with iOS 27. One notable aspect of this new version is its approach to interacting with users, which sets it apart from other AI assistants in a refreshing way.

After using Siri for nearly four months since the first beta launch, I’ve come to appreciate its capabilities and how they align with my expectations. The new Siri has surpassed its predecessor in terms of functionality, delivering on Apple’s promise of a capable assistant for their devices.

While it may not be the most advanced AI tool available – competitors like ChatGPT, Claude, and Google Gemini offer more powerful features – I’ve found that Siri’s unique approach to user interaction is a significant advantage. Unlike these third-party tools, which often mimic human-like personalities, Siri remains grounded as a tool rather than trying to pass itself off as a person.

Siri’s responses are clear and concise, without the verbosity that can make other AI chatbots feel overly familiar or even manipulative. It acknowledges its limitations and reminds users that it’s an artificial intelligence capable of making mistakes. This straightforward approach is a welcome respite from the more ‘personal’ interactions offered by some competitors.

Some may enjoy the human-like tendencies exhibited by ChatGPT, Claude, and Gemini, but I prefer Siri’s no-nonsense attitude. AI assistants should stay focused on their purpose – providing helpful information and completing tasks efficiently – rather than trying to create a false sense of connection with users.

The fact that Apple has designed Siri in this way is noteworthy, as it recognizes the importance of maintaining a clear distinction between human and machine. This approach may not be for everyone, but I appreciate its honesty and practicality.

AI-generated content has become increasingly popular, but it often lacks the natural flow and tone that human writers take for granted. To address this issue, some tools have emerged to help ‘humanize’ AI text, making it sound more like something written by a person rather than a machine. One such tool is Ryne AI, which claims to be able to transform AI-generated content into something more engaging and readable.

When working with AI writing tools, users often find that the first draft needs significant editing before it meets their standards. This can involve changing sentence patterns, wording, or tone to make the text feel more natural. However, this process can be time-consuming, especially when dealing with large amounts of content. A humanizer like Ryne AI aims to streamline this process by automatically rewriting AI-generated drafts.

So how does Ryne AI work? The user simply adds their AI-generated content to the platform, which then processes it and produces a revised version. The main goal is to change the way the text reads, making it more engaging and natural-sounding. Users can expect to see different sentence structures, word choices, and transitions in the output.

It’s essential to compare the rewritten version with the original content to ensure that important facts remain accurate and the main ideas are still intact. A useful rewrite should not compromise on meaning or accuracy. When evaluating the quality of a humanizer like Ryne AI, it’s crucial to judge its output rather than relying solely on claims made by the tool.

Ryne AI aims to make AI-generated text sound less mechanical and more natural-sounding. This can be particularly helpful when dealing with repetitive patterns in the original draft. However, users should not expect every sentence to become perfect after processing. Some drafts may still require additional editing to achieve the desired tone or flow.

The quality of your original draft also plays a significant role in determining the effectiveness of Ryne AI. Clear input usually gives you a better foundation for the final article. It’s essential to remember that humanization is not a replacement for manual editing but rather a tool to support and streamline the process.

Meaning preservation is critical when rewriting any important document, including those generated by AI tools. Users should carefully review the output after using Ryne AI, checking numbers, names, technical terms, and important claims. It’s also essential to compare longer paragraphs against the original content to ensure that no crucial information has been lost or altered.

A humanizer like Ryne AI can be particularly useful for those who regularly create digital content, such as bloggers, marketers, and students. However, users should follow their school or university rules when using AI tools, ensuring they do not misrepresent work expected to be written independently. When it comes to SEO content, a humanizer can help by making the text more readable and engaging.

Readability is crucial for search engine optimization (SEO) purposes. While keyword placement is essential, it’s equally important to focus on clear information, useful sections, and natural language. A humanizer like Ryne AI can be beneficial in this context, helping to rework sentence patterns and wording to make the text more engaging.

However, users should still edit their SEO content manually after processing through a humanizer like Ryne AI. This involves adding relevant examples, removing unnecessary statements, and ensuring every section answers real reader questions. Additionally, it’s essential to review keywords after humanization to confirm that important terms still appear naturally in the rewritten text.

To test Ryne AI effectively, users should start by selecting a paragraph from an AI-generated article and processing it through the platform. Then, compare both versions carefully, checking sentence variety, readability, tone, and meaning. Does the rewritten version sound natural? Does it still communicate the same idea? Does it fit your audience?

Users can repeat this test with different types of content to see how Ryne AI handles various formats, such as blog paragraphs, product descriptions, or formal documents. This will help determine whether the tool is suitable for their specific needs and workflow.

While humanization can be a valuable addition to the writing process, it’s essential to remember that manual editing still plays an important role in good content creation. Users know their audience better than any automated tool, and they understand the purpose behind each article. Therefore, users should not rely solely on Ryne AI but rather use its output as a starting point for further refinement.

When publishing content generated by Ryne AI or any other humanizer, it’s crucial to review the complete article carefully. Start with the introduction and check if it provides readers with a clear reason to continue reading. Next, review each section for accuracy and relevance, removing sentences that repeat the same idea. Users should also ensure links, statistics, names, and product information are accurate.

Finally, users can read the article from the reader’s perspective, asking themselves whether every section provides useful information and if the language feels natural and easy to understand. Additionally, other editing tools like grammar checks and plagiarism checks may be necessary depending on publishing requirements.

The restaurant industry has been facing tough times, with top chain restaurants experiencing slowed sales in 2025. Consumers have been cutting back on dining out, but certain sectors like coffee, beverages, and snacks continue to thrive.

In a recent panel discussion at the FSTEC conference in Dallas, Texas, female technology leaders from Smoothie King and 7 Brew shared their insights on using AI tools for business to remove frictions while protecting hospitality. The key takeaway: automation should run seamlessly in the background, without distracting from the customer experience.

Smoothie King’s Chief Information Officer Jyoti Lynch emphasized that automation is not about replacing human interaction but rather making it easier. ‘Automate the task, not the moment,’ she said, highlighting the importance of striking a balance between technology and hospitality.

The panel discussion was led by Women’s Foodservice Forum (WFF) Chief Brand Officer Ellie Doty and featured 7 Brew’s Chief Technology Officer Danyel Bischof-Forsyth. The session focused on how automation can be used to streamline back-of-house tasks, freeing up staff to focus on customer service.

Lynch gave the example of a common challenge faced by franchisees: when the point-of-sale system’s end-of-day process doesn’t complete overnight, stores are unable to process sales the next morning. This creates chaos for team members when they arrive at work. Smoothie King has automated this process to run in the background while the store is closed, ensuring a fully functional system for baristas when they start their shift.

Some automation does take place closer to customer interaction, but Lynch stressed that it won’t detract from the hospitality experience. For instance, if the system recognizes loyalty members as soon as they walk through the door, it can alert team members, making their lives easier.

7 Brew also prioritizes balancing customer engagement when investing in new technology. The drive-thru coffee chain recently launched its first app and emphasized that it was designed to be additive to 7 Brew’s friendly service. ‘The app gives you time at home or in your car to navigate surprises and offers before you even get there,’ Bischof-Forsyth said.

The investment in new technology has added an opportunity for connection between customers and employees: technology troubleshooting. 7 Brew trained its baristas on the app so they could assist customers when needed, rather than just focusing on automation.

Primarily, though, both companies focus on automating back-of-house tasks that would be invisible to guests. Bischof-Forsyth highlighted bookkeeping and inventory predictions as areas where great business value can be achieved through automation.

A new development from Google DeepMind is set to revolutionize the field of protein design. The company has created a way to watermark proteins generated by artificial intelligence (AI) during the design process itself. This innovation could help scientists track the origin of these structures and prevent their misuse.

In recent years, generative AI models have become an essential tool for researchers designing new protein structures with specific properties or functions. For instance, muscle-inspired proteins that are stronger than their natural counterparts can be created using these tools. However, as more AI-generated protein structures emerge – particularly since Google DeepMind’s AlphaFold won the Nobel prize in chemistry in 2024 – there is a growing need to track their origin.

One of the challenges researchers face when dealing with AI-generated proteins is keeping their findings under wraps if they want to patent them. Creating a centralized database to store these structures could help, but it would also make it difficult for researchers to maintain confidentiality. Adding metadata to describe how a protein was designed can be easily removed, making this solution impractical.

Google DeepMind has taken a different approach by developing models that tag proteins as they are being designed. One of their new models adds a small protein-binding sequence into the target protein sequence, which can then be detected. The researchers compared the sequences of known proteins – such as the Sars-Cov-2 receptor – with the watermarked equivalent and found that watermarking had very little impact on the strength of interactions with other proteins.

The team has also built upon the latest version of AlphaFold to develop a new model that watermarks protein structures themselves. This model revealed that there is little change to bond lengths and angles within the structures, indicating that watermarking does not compromise their integrity. The researchers believe that this innovation will help uphold scientific integrity by making it harder for others to alter the structure of a protein once it has been designed.

The potential misuse of generative AI tools to create biological agents that could cause harm is another concern addressed by this development. By tagging proteins with their origin, scientists can prevent unauthorized modifications and ensure accountability in research.

Researchers at MIT and Sakana AI have developed a framework called Recursive Self-Improvement via Fast Tree Search (SIFT) to reduce the cost of evaluating modifications to coding agents. The traditional approach to improving coding agents involves testing each candidate change, which can be time-consuming and expensive. However, SIFT uses an LLM as judge to compare candidate implementations pairwise, reducing the need for extensive benchmark runs.

The main challenge in evaluating coding agent modifications is the poor signal-to-cost trade-off of existing evaluation methods. Each candidate change needs testing, but running a full test suite on every modification can consume thousands of CPU hours and cost thousands of dollars. SIFT addresses this issue by using an LLM to compare two versions of an agent before committing to expensive benchmark runs.

SIFT builds upon the Darwin Gödel Machine (DGM) framework, which maintains an archive of agents and checks candidates on increasingly large task sets. However, DGM’s approach can be slow due to its reliance on sequential evaluation. In contrast, SIFT uses a more efficient search strategy that allows it to explore multiple branches in parallel.

The central idea behind SIFT is to use the LLM as judge to compare candidate implementations pairwise. This involves running each new agent on four coding tasks and then comparing it with up to 10 high-ranking agents already in its archive using the Bradley-Terry model, a statistical method for estimating relative strength from wins and losses.

SIFT combines the pairwise choices into a ranking that helps decide where testing and further development are worth spending time. The new candidate is then placed into a priority queue for the larger, time-consuming evaluation. Its priority combines the judge’s ranking with its benchmark-accuracy ranking.

The workflow avoids making each patch wait for the previous one’s full evaluation. Patch generation, judging, and benchmark evaluation can proceed in parallel, similar to a CI pipeline where work continues while longer tests run in the background. This allows promising candidates to become the parents of new patches before their expensive evaluations finish.

SIFT was tested on various benchmarks, including Polyglot, TerminalBench 2.1, and SWE-bench Verified subset. The results showed that SIFT delivered better accuracy with less compute using both open-weight Qwen3-Coder-30B and closed o3-mini models. With Qwen3-Coder-30B, it slightly outscored HGM while using about a third fewer CPU hours.

On Polyglot, one SIFT run reached 35.1% accuracy in under five hours of wall-clock time, using 42 CPU hours and about $150 in API credits. This is a significant improvement over DGM’s performance on the same benchmark. Additionally, SIFT showed that it can handle large task sets efficiently by achieving better results with less compute.

SIFT was also tested without the LLM judge to see whether it made a difference. The results showed that the judge helped drive the gain in accuracy and efficiency. This suggests that using an LLM as judge is crucial for SIFT’s success.

The researchers found that a cheaper judge retained much of the global ranking signal on TerminalBench, although the stronger model was more reliable among the best candidates. They suggest that a tiered setup could work, with the cheaper judge for most comparisons and a stronger one near the top of the ranking.

SIFT shows how the self-improvement loop itself can become more efficient by combining cheaper evaluation signals with asynchronous search. This allows teams to explore more candidate improvements without scaling evaluation costs at the same rate. However, downstream benchmarks are still needed to confirm that a change actually improved the agent.

The framework is built on top of the DGM harness and adds the pairwise judge and Bradley-Terry ranking. It makes the search asynchronous so candidates can be expanded while others are still being evaluated. SIFT was only tested on coding agents but its basic approach could apply to other self-improving agents where there’s a clear way to measure whether a change worked.

The first step in applying SIFT is to define a small, cheap evaluation that can quickly reject obviously bad modifications. For an enterprise agent, this could be a small set of representative internal tasks that catches broken tools, regressions or changes that violate basic requirements. The second piece is an LLM judge that compares candidate implementations pairwise.

SIFT’s efficiency comes from its ability to explore multiple branches in parallel while previous candidates are still being tested. This allows the search tree to keep growing instead of waiting for every evaluation to finish. SIFT shows how AI tools can be used more effectively by combining cheaper evaluation signals with asynchronous search, making it a valuable addition to any business’s toolkit.

A new connector from Qumis is set to integrate insurance-specific intelligence directly into popular AI assistants, streamlining the work of professionals in the industry. The company’s Coverage Intelligence platform has been designed to provide comprehensive coverage analysis and expertise for commercial property & casualty insurance. Now, with its MCP (Machine-Processable Connector) technology, Qumis aims to bring this specialized knowledge directly into daily workflows.

Qumis’ vision is for its platform to become the go-to source of insurance intelligence wherever professionals work on insurance-related tasks. The company’s co-founder and CEO, Dan Schuleman, emphasizes that their goal is to make Qumis a seamless part of existing applications and AI tools. This means no need to send users to a separate platform or rebuild understanding – just direct access to the expertise they need.

The MCP connector achieves this by bridging the gap between general-purpose AI tools and domain-specific reasoning. When professionals work with complex insurance policies, Qumis’ specialized capabilities can be drawn upon directly within their existing workflows. This eliminates the need for manual data transfer or rebuilding understanding – a crucial advantage in high-stakes domains like insurance.

Qumis MCP is designed to integrate seamlessly into popular AI assistants such as Claude and Copilot. Additionally, it provides access through an open standard, allowing organizations to build custom workflows and automations that incorporate Qumis’ insurance intelligence. This flexibility makes the technology particularly valuable for businesses looking to leverage their existing infrastructure.

The specialized policy context and analysis capabilities built by coverage lawyers are a key feature of Qumis MCP. These include side-by-side policy comparisons, thorough coverage gap reviews, comprehensive document analysis, and automated document summaries backed by page-level citations. By providing professionals with accurate evidence-based conclusions, Qumis aims to give them the speed, rigor, and domain expertise their work demands.

According to Shiv Sinha, Co-founder & CTO of Qumis, AI is especially valuable in high-stakes domains like insurance due to its ability to perform deep policy reasoning. With Qumis MCP, professionals can rely on clear evidence for their conclusions – a critical advantage when working with complex policies and high-stakes decisions.