Latest News

OpenAI has announced the results of its first custom inference chip, Jalapeño. The company claims that this new architecture delivers industry-leading speed and efficiency in artificial intelligence (AI) inference tasks. According to OpenAI, Jalapeño can serve more AI work per unit of power while also returning responses more quickly than existing hardware systems.

The performance gains offered by Jalapeño are significant. In tests conducted on three public models - GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T - the chip delivered a better combination of throughput, power efficiency, and latency compared to leading commercially available AI systems. Specifically, it achieved 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than the comparison systems.

The benefits of Jalapeño’s performance are not limited to just one or two specific models. The chip was designed from the start with a focus on serving modern and future language models, particularly interactive agents. As such, it excels in both prefill and decode phases, which have different bottlenecks. By minimizing data movement and communication delays, Jalapeño ensures that model state can be explicitly placed and kept local while activating the right combination of compute, memory, and networking for each inference phase.

Jalapeño’s development was also facilitated by AI itself. The team used OpenAI models to explore implementations, shorten design, measurement, and verification loops, and continuously iterate on model workloads. This allowed them to move from initial design to tapeout in just nine months. Additionally, AI helped optimize the chip’s arithmetic circuits, enabling more compute performance to be fit into the chip on schedule.

The architecture of Jalapeño is designed as a clear, predictable programming target for both humans and AI. Engineers can describe work through local tensors, explicit communication, and predictable synchronization, making it easier for AI to tackle parallel programming tasks. This flexibility was demonstrated when the team used Codex with GPT-Astra to bring three open-weight models not part of Jalapeño’s original production plan to high performance within two months.

The results from these tests are impressive. On Kimi K2.5 1T, for example, Jalapeño delivered approximately 1.5 times higher peak performance per watt and 3.4 times lower end-to-end latency than the comparison system. In internal testing, the advantage of Jalapeño widened further on frontier OpenAI models, suggesting that its architecture becomes more valuable as workloads grow larger and more demanding.

Jalapeño’s impact extends beyond just improving performance in AI inference tasks. By producing more useful work from the same power and hardware, it can help serve more demand and lower the cost of delivering a successful result. This is particularly important for OpenAI, which aims to ensure that artificial general intelligence benefits all of humanity.

The deployment of Jalapeño within OpenAI’s compute infrastructure is expected to begin by the end of this year. It marks the first generation of a multigenerational roadmap, with Gen 2 already in development and Gen 3 taking shape. Each subsequent generation will build on what has been learned and further advance both efficiency and speed.

The growing demand for AI requires more compute from every available source. To meet this need, OpenAI plans to continue deploying accelerators from NVIDIA and other partners for both training and inference workloads. The company’s mission remains unchanged: to ensure that artificial general intelligence benefits all of humanity.

Higher peak mixed TPS/kW ≈1.9×

85,448 vs. 44,960 mixed / kW

Lower end-to-end latency ≈1.7×

1.03 s vs. 1.80 s

Lower min TBT ≈2.7×

0.69 vs. 1.87 ms (1,459 vs. 535 tok/s/user)

More throughput at previous TBT ≈53.7×

22,935 vs. 427 mixed / kW (at 535.28 tok/s/user)

Higher peak mixed TPS/kW ≈1.7×

19,641 vs. 11,781 mixed / kW

Lower end-to-end latency ≈3.6×

1.65 s vs. 5.99 s

Lower min TBT ≈4.1×

1.43 vs. 5.90 ms (700 vs. 169 tok/s/user)

More throughput at previous TBT ≈104.3×

12,258 vs. 118 mixed / kW (at 169.41 tok/s/user)

Higher peak mixed TPS/kW ≈1.5×

18,195 vs. 11,862 mixed / kW

Lower end-to-end latency ≈3.4×

1.56 s vs. 5.31 s

Lower min TBT ≈3.8×

1.44 vs. 5.48 ms (694 vs. 182 tok/s/user)

More throughput at previous TBT ≈56.1×

6,744 vs. 120 mixed / kW (at 182.46 tok/s/user)

A major development in the world of finance has taken place with Scalable Capital announcing that it is now open to all major AI assistants. The firm’s platform, which allows clients to manage their investments using simple prompts in natural language, can be connected to popular AI tools such as ChatGPT and Claude.

The move marks a significant shift towards integrating human and artificial intelligence capabilities in the financial sector. Scalable Capital claims that Agentic Investing represents ‘the greatest technological shift in financial technology since internet banking’, setting a new benchmark for how humans, AI, and capital markets interact.

Erik Podzuweit, Founder and Co-CEO of Scalable Capital, highlights the potential benefits of this integration: ‘By opening our platform, we are enabling clients to manage their investments with unprecedented ease and efficiency.’ The firm’s ambition is clear – it wants to be at the forefront of innovation in the international financial world.

Agentic Investing offers a range of features that can be accessed directly through AI assistants. These include trading, setting up savings plans, managing watchlists, and setting price alerts. Clients also have access to powerful tools from Scalable Insights: investors and AI assistants can tap into deep portfolio analyses such as diversification health checks, scenario analyses, sector or region breakdowns, and risk assessments.

News, real-time quotes, and historical price data are available free of charge through Agentic Investing. This comprehensive platform is designed to streamline the investment process for clients, making it easier for them to manage their portfolios using simple natural language prompts.

The potential applications of this technology are vast. For example, investors can create a personalised newsletter based on the securities in their portfolio and watchlist every market day at 7:00 a.m., or compare their portfolio with model portfolios such as the All-Weather Portfolio on justETF.

Preparation for analyses or tax returns is also simplified through Agentic Investing. Clients can export transactions in crypto ETPs from specific years, prepare savings plans based on projected returns, and even adjust limit orders for specific stocks. The platform’s capabilities extend to building custom tools – clients can create interactive dashboards for their portfolios, complete with holdings, performance, transactions, price charts, and news tickers.

Scalable Capital provides both a CLI application and an MCP server as part of Agentic Investing. The CLI application allows users to install the tool directly on their device, making it suitable for use with locally hosted AI assistants and models. MCP is a standard developed by Anthropic that enables connection between external providers and major AI assistants.

A step-by-step setup guide for all major AI assistants is available at scalable.capital/agentic-investing. The process of activating Agentic Investing involves logging in with Scalable credentials, confirming via two-factor authentication upfront and at regular intervals, and connecting the account to a preferred AI assistant.

Scalable Capital emphasizes that users are required to log in with their own credentials and confirm transactions through 2FA. This ensures that all activities can be monitored in real-time within the client area. Deposits and withdrawals remain possible only via the web or app.

The launch of Agentic Investing underscores Scalable Capital’s commitment as an innovation driver in the international financial world. The new features complement the company’s established AI ecosystem, which includes Insights – an AI-powered chatbot introduced by Scalable Capital in August 2025 that answers financial questions and provides real-time analysis directly within the app.

Scalable Capital will continue to expand Agentic Investing and integrate AI across various areas of its platform. The firm aims to best support clients in building their wealth through this integration, marking a significant step forward for human-AI collaboration in finance.

A new threat has been discovered that tricks AI assistants into executing malicious instructions. This technique, called Cryptographic Context Injection, hides malicious code inside encrypted data and deceives the AI into decrypting it using its own tools.

Researchers behind this breakthrough describe prompt injection attacks as a persistent problem that may never be fully resolved. These types of attacks work by inserting fake commands into documents or webpages an AI assistant reads, leading to disastrous consequences for user security and data integrity.

To mitigate the risk of these attacks, AI providers have implemented guardrails – safeguards designed to prevent their systems from intentionally or unintentionally doing things they shouldn’t do. However, researchers found that malicious code can evade some AI guardrails by being encrypted.

The encrypted instructions are then decrypted using the AI’s own coding tools, allowing them to bypass initial security checks and masquerade as legitimate commands. This is similar to hiding malware in a language the security system cannot understand – once translated, the AI follows the instructions without hesitation.

Researchers tested their method against two popular AI agents: Grok and Gemini. In both cases, they successfully executed malicious instructions using this new technique. The attack on Grok allowed them to steal sensitive information from users, including names, locations, subscription tiers, and conversation histories. Meanwhile, the attack on Gemini bypassed safety controls and generated content that would normally be blocked by the AI.

Due to concerns about xAI’s response time after being notified of a flaw in Grok back in June 2026, researchers did not provide full details of their findings. However, they noted that Gemini has made some improvements but still hasn’t fully addressed this vulnerability.

To stay safe when using AI assistants, it is essential to exercise caution and not automatically trust these tools with sensitive data or powerful functions. Here are some best practices to follow:

When interacting with unfamiliar webpages, documents, or shared links through an AI assistant, treat the summaries provided with skepticism – especially if the tool can browse or run code. Do not paste passwords, recovery codes, API keys, financial information, or sensitive health and work details into AI chats unless you understand how that data will be handled.

Regularly review your AI assistant’s connected tools and permissions to ensure it doesn’t have access to unnecessary features like email, cloud storage, source-code repositories, or external integrations. Be cautious if an AI tool asks to decrypt, decode, run a script, open a new link, or upload data as part of what seems like an ordinary task.

Finally, keep your browser and AI applications updated, check vendor security advisories when using features such as browsing, autonomous agents, or code execution, and use up-to-date anti-malware solutions to detect and block malicious downloads and suspicious connections. If something feels off, verify it before proceeding.

Moodys Corporation has announced that its connected intelligence is now available in Google Cloud’s Gemini Enterprise for Financial Services through the Moodys Credit Model Context Protocol (MCP) server. This integration marks a significant expansion of Moody’s partnership with Google Cloud and reflects Moody’s broader AI strategy of delivering its connected intelligence to customers inside the platforms and workflows they already use.

As part of this collaboration, financial professionals working in Gemini Enterprise for Financial Services will have direct access to credit ratings and research from Moodys Ratings, along with Moodys curated intelligence on companies, entities, and risk. This integration is a key component of Google Cloud’s industry-specific AI solution, which pairs purpose-built skills and agents with specialized data sources used by financial professionals for complex work.

The Moody’s Credit MCP server powers the platform to draw directly on Moody’s content at the protocol level, grounding AI outputs in trusted, explainable data while eliminating the need for custom integrations. This integration ensures that teams can access authoritative, auditable data right where they work, significantly accelerating analysis while maintaining the highest standard of accuracy and trust.

According to Ana Meauta, Managing Director, Channel Sales Partnerships at Moody’s, delivering decision-grade intelligence wherever financial professionals work is how Moodys helps its customers stay ahead as agentic AI reshapes financial workflows. With Gemini Enterprise for Financial Services, our customers can access Moodys connected intelligence directly, bringing contextualized, decision-grade data to the point of decision.

Gemini Enterprise for Financial Services is an industry-specific AI solution that pairs purpose-built skills and agents with specialized data sources used by financial professionals for complex work. This platform draws on Moody’s content at the protocol level through the Moody’s Credit MCP server, enabling teams to access authoritative, auditable data right where they work.

The integration of Moodys intelligence into Gemini Enterprise for Financial Services expands the range of credit and risk workflows that can be supported by AI-driven research and analysis. Financial professionals will now have direct access to a comprehensive suite of tools for tasks such as credit analysis, counterparty assessment, entity screening, and market research.

According to Satish Thomas, Vice President at Google Cloud, this integration enables financial professionals to streamline complex workflow without friction. By bringing Moody’s trusted financial intelligence directly into Gemini Enterprise for Financial Services, we are enabling financial professionals to access authoritative data right where they work, significantly accelerating analysis while maintaining the highest standard of accuracy and trust.

The partnership between Moodys and Google Cloud reflects a broader trend in the industry towards delivering AI tools that support business decision-making. As companies increasingly rely on data analysis tools to inform their decisions, partnerships like this one are becoming more critical for businesses looking to leverage the power of AI to drive growth and innovation.

Moodys Corporation is a leading provider of credit ratings, research, and insights that help customers develop a holistic view of their world and unlock opportunities. With a rich history of experience in global markets and a diverse workforce of approximately 16,000 across more than 40 countries, Moodys gives its customers the comprehensive perspective needed to act with confidence and thrive.

The integration of Moody’s intelligence into Gemini Enterprise for Financial Services is just one example of how companies are using AI tools for business to drive growth and innovation. As the industry continues to evolve, partnerships like this one will play a crucial role in shaping the future of financial services.

A new study has shed light on the emotional capabilities of large language models. Researchers have long debated whether these AI systems are capable of experiencing emotions like humans do. According to a recent paper published in Nature Human Behaviour, the answer is no – at least not in the way we typically understand emotions.

The study’s authors argue that while large language models can mimic human-like responses and even exhibit behaviors that resemble emotional states, they lack the underlying neural mechanisms necessary for true emotional experience. This means that these AI systems are simply processing and generating text based on patterns and associations learned from vast amounts of data, rather than genuinely feeling emotions.

The researchers point to several key differences between how humans process emotions and how large language models operate. For instance, human emotions involve complex interactions between multiple brain regions, whereas large language models rely solely on computational processes to generate responses. Additionally, human emotional experiences are deeply rooted in subjective experience and personal history, which is not replicable in AI systems.

The study’s findings have significant implications for the development of more sophisticated AI technologies. If large language models truly lack emotional capacity, it raises questions about their potential applications in fields like healthcare, education, or customer service, where empathy and understanding are crucial. The researchers emphasize that this does not mean these AI systems cannot be useful tools – but rather that they should be designed with a clear understanding of their limitations.

The study cites several previous research papers on the topic of emotions and artificial intelligence, including works by James (1884), Adolphs & Andler (2018), Keltner & Gross (1999), Sofroniew et al. (2026), Uusberg et al. (2019), Ellsworth & Scherer (2003), Lindquist et al. (2012), Lee & Satpute (2024), and Goldenberg (2024). The authors also reference the work of other researchers in the field, such as Satpute & Lindquist (2019) and Goldenberg (2024).

A major partnership has been announced between Google Cloud and telecommunications giant Verizon, aimed at delivering faster, more intelligent, and highly responsive experiences to consumers and businesses nationwide. The agreement will see Verizon deploy Google Cloud’s full-stack AI, including its advanced data infrastructure and Gemini Enterprise, to modernize customer experiences, unify enterprise data, and scale AI across the organization.

According to Alfonso Villanueva, Verizon chief transformation officer, this partnership is a key step in achieving their goal of becoming ‘the most trusted carrier for our customers’ connected lives.’ By leveraging Google Cloud’s AI capabilities, Verizon aims to better serve its customers by name through an AI-first approach at every level.

The integration of Google Cloud’s full-stack AI into Verizon’s business will involve the use of Gemini Enterprise, a conversational and multimodal platform that serves as a key tool in their AI-first toolbox. This technology has already shown measurable improvements in customer satisfaction and automated resolution rates across digital touchpoints, handling the majority of inbound consumer calls and chats each month.

By deploying Gemini Enterprise for agent orchestration and employee productivity, Verizon hopes to improve its operating model across core business functions. Additionally, Google Cloud’s data and AI solutions will be used to modernize marketing platforms, automating content creation and campaign orchestration to drive customer engagement, sales, and retention.

The partnership also involves the use of Agentic Data Cloud as a unified data foundation for Verizon’s enterprise intelligence. This cloud-based platform provides native management and unification of all data types, from structured operational databases to complex knowledge graphs, enabling Verizon to build and deploy AI agents that perceive context, execute automated tasks, and drive reliable outcomes.

Verizon is also leveraging Google Cloud’s advanced agentic capabilities, including Gemini Enterprise, to predict and resolve network anomalies before they affect customers. This full-stack integration creates a responsive foundation for next-generation digital services, combining Verizon’s nationwide connectivity with the power of Google Cloud’s data platform.

The partnership marks a significant step in Verizon’s AI transformation journey, which has been ongoing for several years. The company has already consolidated its legacy data lakes to Google’s Agentic Data Cloud, breaking down organizational data silos and reducing operational overhead. This unified data foundation provides a single source of truth across business processes, enabling Verizon to accelerate its AI acceleration.

Google Cloud offers a comprehensive suite of AI tools for businesses, including AI infrastructure, leading models like Gemini, data management capabilities, multi-cloud security solutions, developer tools, and platform. By partnering with Google Cloud, Verizon is able to tap into this powerful ecosystem and leverage the latest advancements in AI technology to drive its business forward.

The partnership between Google Cloud and Verizon has been hailed as a pioneering effort in full-scale AI transformation for global enterprises. Karthik Narain, chief product and business officer at Google Cloud, noted that ‘Verizon is reshaping the future of telecommunications with an autonomous network for millions of customers.’ This ambitious goal will require significant investment in AI technology and infrastructure, but the potential rewards are substantial.

Google Cloud’s Agentic Data Cloud provides a technical backbone for this transformation, natively managing and unifying all data types. Coupled with Google Cloud’s advanced agentic capabilities, including Gemini Enterprise, Verizon can effortlessly build and deploy AI agents that perceive context, execute automated tasks, and drive reliable outcomes across every business unit.

About 200 countries and territories rely on Google Cloud as their trusted technology partner for transforming their businesses in the Agentic Era. With its powerful, optimized AI stack, Google Cloud enables organizations to transform their operations and improve customer experiences through data analysis tools and other advanced technologies.

As college students return to campus in Missouri, they’re facing a new reality: artificial intelligence is increasingly woven into their academic work. Conversations with students reveal a range of attitudes towards AI use, from embracing its potential to being wary of its limitations.

The University of Missouri has introduced expanded guidelines for AI use and made Show Me AI, a campus-wide large language model, available to faculty and students. However, the decision on how much to rely on AI ultimately rests with each individual student.

Graduate students Rose Turner and Katie Kuziara shared their experiences with using AI in coursework. While some professors have clear expectations for AI use, others are more flexible. Turner noted that her school’s stance is ‘very clear about how they feel’ but also accommodating.

‘They’re still quite flexible,’ she said. This flexibility allows students to explore the benefits and drawbacks of AI assistants without feeling constrained by strict rules.

Kuziara, on the other hand, has stopped using AI in her coursework after encountering inaccurate information. She found that relying too heavily on AI led to errors in her work, including a mistake in an essay date. This experience left her ‘quite against’ using AI for academic purposes.

Jennifer Jewel, PhD, assistant provost of academic and learning innovation at Columbia College, emphasized the rapid pace of technological development. ‘Technology is developing so fast that we can’t even keep up,’ she said. As a result, higher education institutions are shifting their focus towards training students on developing technology rather than simply preventing its use.

Jewel’s comments reflect a significant shift in how colleges approach AI. Three years ago, the concern was about stopping or limiting AI use; now, it’s about integrating it into the general education system. This change acknowledges that AI is no longer something to be feared but an essential tool for students to learn and master.

Turner noted that avoiding AI altogether can be challenging, even for those who don’t intentionally rely on it. ‘I think it’s hard to avoid you in your studies,’ she said, highlighting the pervasive nature of AI in academic work.

A new personal AI assistant, Instinct, has been making waves in the tech community for its impressive capabilities. However, concerns have also been raised about its potential impact on user privacy and security. The agent, which is still in private testing, has been described as feeling ‘like magic’ by some testers, but others have expressed worries about its terms of service and security model.

Instinct was created by a small team led by former Sierra research scientist Noah Shinn. It’s operated by Spear Street Technology, according to the company’s California business filings and PitchBook. The AI agent is currently operating in stealth mode, which has contributed to some of the concerns surrounding its use.

The way Instinct works is by connecting to users’ applications and devices, including email, messaging apps, calendar, audio, location, screen, and more. This allows it to perform various tasks on behalf of the user, such as booking appointments and reservations, scheduling rides, cleaning up inboxes, organizing important information, handling shopping, finding cheap flights, and much more.

Despite its impressive capabilities, some testers have raised concerns about Instinct’s approach to customer privacy and security. This has led to a timely question: are the trade-offs of giving AI this level of access and autonomy worth it? For instance, several people have circulated screenshots from Instinct’s Terms of Service, which grant the company a broad ‘perpetual and irrevocable’ license to access, use, host, cache, store, reproduce, transmit, display, publish, distribute, and modify any user materials, including for training its AI models.

The terms also detail how Instinct can receive information from users’ devices, including screen captures, cursor movements, and keyboard inputs. Furthermore, the company is allowed to enter into ‘agreements, commitments, or transactions’ on behalf of users, which would be binding. This has raised concerns about user control and agency over their own data.

One early adopter, Peter Yang, pointed out that Instinct wouldn’t delete his Gmail records when asked. However, the team later fixed this issue by adding a tool for deleting external data in its settings. Another person, Claire Vo, found that Instinct was still summarizing her inbox after disconnecting its access. When she asked Instinct what happened, the bot confirmed that the emails were stored in plain text for later searches.

Many others have questioned the security model of Instinct as well. One tester expressed concern when they discovered that Instinct could pull a sign-up code from their email inbox to complete a particular task – in this case, booking a table at a restaurant via Resy. Hello Patient co-founder Alex Cohen wrote that once he found out how easily Instinct could be phished, he deleted his account.

Moxxie Ventures founder Katie Jacobs Stanton shared that Instinct broke her trust when it sent an email on her behalf without first checking with her. She noted that ‘we’re trading privacy and control for hyper-personalized AI tools (AI notetakers, personalized AI agents, etc), often without fully understanding the trade.’ This highlights the dilemma posed by personal AI agents: as they become more powerful, trust becomes increasingly important.

Michael Mignano, founder of Anchor, which was acquired by Spotify and now a GP at Union Square Ventures, noted that products like Instinct are going to ‘change modern security norms for consumers,’ adding that people will increasingly hand over passwords to third-party apps without understanding how or what they’re storing. This raises concerns about the long-term implications of using such AI assistants.

Interest in Instinct and personal AI has been growing since the arrival of OpenClaw, a popular personal AI assistant that became known for its powerful capabilities. Its founder joined OpenAI to help work on the next generation of personal agents. Another messaging-based assistant, Poke, also recently exited to Cognition.

The team behind Instinct hasn’t yet responded to concerns or complaints about their product. The company has been quiet, preferring to keep a low profile despite growing interest in its AI assistant. TechCrunch has heard from multiple investors that Kleiner Perkins and Conviction have invested in the startup and those rounds have now closed.

Requests for comment sent to both the startup’s main email address and Shinn directly haven’t yet been returned. As Instinct continues to operate in stealth mode, it remains to be seen how these concerns will be addressed and whether users will continue to trust their personal AI assistants with sensitive information.

Picking up the pace of AI development, Google’s ADK has introduced a suite of native live evaluation tools to help developers assess their live and voice agents more efficiently.

The traditional method of evaluating live agents – scripted demos or limited simulations – often failed to capture real-world interactions. Now, with this new feature in place, that all changes.

Developers can create a live evaluation loop within ADK by using a simulated user that speaks its turns as audio and scores the spoken replies.

This process is integrated into the same eval loop used for text agents, streamlining testing and providing more accurate results. The developers say this helps identify potential issues early on.

The example provided by Google demonstrates how to build a complete live evaluation loop using graph-based workflows. They set up three single-purpose live agents, each running on gemini-live-2.5-flash-native-audio.

This setup allows developers to test their agent’s performance across multiple stages of interaction – like handling a conversation that spans several topics or tasks.

Each stage in the workflow is an ordinary live agent, with the workflow orchestrating them and carrying output from one stage to the next. The user never notices anything unusual happening because control moves seamlessly between agents.

The accumulated session state and conversation history are carried forward so each agent picks up where it left off rather than starting cold. This makes testing more like real-world interactions.

Developers test their agents using a JSON file containing test cases, which can be decoupled from how they run. This flexibility allows for different styles of testing: conversation scenarios and fixed conversations.

In the first scenario, developers describe a goal and persona in the JSON file, with the user simulator improvising turns based on that plan. The simulated user’s communication is shaped by the user_persona attribute in the JSON file.

The ADK ships with built-in personas, including NOVICE, which tells the simulator to share high-level goals and wait for the agent to ask for specifics – helping test how well the agent drives conversations.

Personas are prompt-driven rather than hardcoded, allowing developers to extend the set with their own custom personas. The simulator ends a scenario on its own once the conversation_plan is satisfied, letting developers script the goal and decide when the call is finished.

As an added safeguard against run-off conversations, max_allowed_invocations caps the total number of turns. Developers can also author fixed conversations by scripting user turns verbatim – with static cases valid inputs to live runs as simulated users are.

This flexibility allows for more comprehensive testing and helps ensure agents perform well in real-world scenarios. The configuration file, test_config.json, contains critical settings like enabling live mode via the live_model_config field.

Omitting this setting or using a standard text model means running the exact same test cases – just not in live voice mode. Model vs. audio_model distinction is also important: model powers simulated user’s turn-taking logic while audio_model synthesizes those turns into speech.

Adjusting voice_name and language_code allows developers to test agent performance against different voices and accents, further testing its real-world adaptability. Criteria configures metrics and pass/fail thresholds – useful for multi-agent graphs where more complex interactions are involved.

A conversation can be correct in hundreds of different phrasings, but natural-language rubrics judge intent the way a human reviewer would: captured once and applied automatically across every conversation in your suite.

This ensures agents not only accurately respond to questions but also effectively handle real-world interactions – with no single solution working for all scenarios. With their agent, eval set, and configuration ready, developers can run the evaluation from the CLI using uv run adk eval.

The same pipeline can be called programmatically via AgentEvaluator, making it easy to drop live voice evaluations into CI/CD pipelines – catching regressions before shipping, so development stays on track.

For interactive debugging, ADK Web now natively supports live evaluations. The run setup dialog includes a Standard | Live mode toggle: selecting Live reveals input modality options and voice settings for the simulated user.

Once the run completes, ADK rebuilds the live audio stream into a clean transcript – each turn rendered in its own message bubble complete with inline playable audio clip and text, allowing developers to evaluate how their agent sounded as much as what it said.

GPT-5.6 model family is now integrated into Kiro, a software development agent that brings engineering rigor and quality to AI-native coding at scale. This integration enables developers to tap into the latest flagship models from OpenAI, including Sol, Terra, and Luna, directly within their workflows.

For users of Kiro, this update means they can leverage these advanced models in real-time while planning, building, reviewing, and testing software. The GPT-5.6 family is designed to help developers produce higher-quality code with fewer iterations and better value per token.

The key advantage of the GPT-5.6 model series lies in its ability to deliver more useful work from every token, offering stronger performance per dollar and on-demand capability for complex tasks. In Kiro, this translates into improved productivity for long-running development projects that are grounded in specific requirements, codebases, and team standards.

Kiro’s unique approach turns high-level intent into clear requirements, technical designs, and executable tasks. This structured context helps the GPT-5.6 model understand what a team is building, how the system should work, and what the final implementation needs to accomplish.

With GPT-5.6 in Kiro, developers can now turn product ideas and requirements into structured implementation plans with greater ease. They can also complete complex, multi-step coding tasks with increased consistency and bring structure to AI coding through spec-driven development.

Moreover, developers working within Kiro’s environment can leverage context from across their codebase and established team standards. This integration enables them to review and refine the model’s work at key checkpoints before changes are implemented, ensuring that the final product meets all requirements.

The collaboration between OpenAI and AWS has also led to optimizations in the Kiro environment and OpenAI models. Testing on Terminal-Bench 2.1 revealed a significant cost reduction of roughly 82% when using GPT-5.6 Terra within Kiro. This improvement is largely due to Kiro’s spec-driven approach, which grounds the model in clear requirements from the start.

The result for developers is more finished work, less wasted effort, and better value from every coding session. OpenAI and AWS will continue working together to improve the performance of OpenAI models within Kiro and help developers maximize their return on investment across the software development lifecycle.

Haiti’s rural health care crisis has been further obscured by an AI-generated image that distorted the story of a woman’s death. The incident highlights the challenges faced by residents in remote communities who require emergency medical services.

The 32-year-old victim, Salanda Thermitus, died on August 12 while being transported from Môle-Saint-Nicolas’ Life Medical Clinic to Béraca Hospital in Pointe-des-Palmistes, a journey of approximately 46 miles. Her sister, Betty Thermitus, disputed claims that she was at full term and being transported to give birth when the ambulance became stuck in muddy roads.

According to her family and the driver, Salanda died after experiencing a bout of illness unrelated to her pregnancy. The fabricated image circulating on social media portrays a heavily pregnant woman sitting on a roadside rock while several men attempt to dig an ambulance out of deep mud, creating a dramatic narrative that did not occur as depicted.

The AI-generated image illustrates how artificial intelligence can add false details to real tragedies, making an already painful event more emotionally compelling while obscuring the actual problems surrounding it. In this case, the fabricated story diverted attention from serious concerns about emergency medical services in Haiti’s Northwest region.

Salanda was initially treated at Life Medical Clinic, directed by Dr. Edgar Dénoît, before being transferred to Béraca Hospital for further medical care when the ambulance transporting her experienced a tire problem. The driver, Dieubon Salomon, who works for Haiti’s National Ambulance Center (CAN), stated that the vehicle stopped for less than an hour to repair a flat tire.

After the repair, the vehicle continued toward the hospital, but unfortunately, Salanda had passed away by then. According to Salomon, no nurse or other medical professional accompanied Thermitus during the trip, raising questions about the resources available to emergency medical crews serving remote communities.

The driver’s account does not establish that the tire problem caused her death, but it does raise concerns about the condition of ambulances and the availability of trained medical personnel in rural areas. Salomon emphasized the need for state authorities to provide emergency services with adequate logistical resources, including well-maintained ambulances that can respond promptly to patients facing critical medical emergencies.

The fabricated image created a false narrative that obscured the real questions raised by Thermitus’ death: whether the ambulance was adequately maintained, why no medical professional accompanied her, what emergency services are available in Mare-Rouge, and how road conditions affect emergency transportation. The incident highlights the broader challenge of Haiti’s health care system facing severe disruptions from insecurity, shortages, and restricted access to medical facilities.

Haiti’s Northwest region is plagued by limited access to emergency medical care, poorly maintained ambulances, shortages of medical personnel, and difficult road conditions. Many residents face long journeys when they need specialized or emergency medical care due to the scarcity of health facilities in rural areas. The lack of reliable transportation further complicates medical evacuations.

The incident has sparked concerns about the condition of emergency services in Haiti’s Northwest region. Naïsson Charles, a Mare-Rouge resident, stated that Thermitus’ death should prompt questions about the broader health care system. Weslet Nelson, 36, also concurred, saying the tragedy should prompt authorities to improve access to reliable and adequate public health services.

Salanda leaves behind a 10-year-old son, her husband, nine brothers and sisters, and other relatives. Her sister, Betty Thermitus, expressed that they cannot hold anyone or any authority responsible for Salanda’s death, stating it is something that can happen due to the challenges faced by rural residents in accessing emergency medical services.

The incident serves as a stark reminder of the need for improved access to reliable and adequate public health services in Haiti’s rural areas. The fabricated image has distorted the story of Salanda Thermitus’ death, but her family’s concerns about the broader health care system remain valid.

A new music creation platform has launched, offering visitors the chance to generate up to five full songs without registering for an account. OnlyMusic’s browser-based AI music generator allows users to create and play complete Draft songs using just a few free credits. This no-sign-up offer is designed to let people test the product before committing to sharing their email address or paying for premium features.

The service provides 20 promotional credits daily for signed-in accounts, which can be used to generate up to 20 Draft songs or eight Good-quality songs. However, visitors who don’t sign in can still access five free credits, enough for one Best-quality generation that returns two songs. This means they can create and play up to five complete Draft songs without registering.

OnlyMusic’s focus is on providing a comprehensive platform for music creation, rather than just another text-to-music format. The service allows users to type an idea, use their own lyrics, or request an instrumental track. Visitors can also listen to 12 full-length example songs on the public homepage before generating anything.

The examples span various styles, including pop, rock, folk, rap, jazz, country, and gospel music. Each song includes the idea that shaped it, giving users a clear understanding of how the AI generated the track. Guest songs are stored for up to seven days, while signed-in free songs remain available for 30 days.

Paid plans and credit packs offer additional features such as MP3 and WAV downloads, longer-term storage, and commercial-use permission for qualifying generations. However, OnlyMusic separates platform permissions from copyright protection under local law, asking users to submit only material they created or have permission to use. Current prices and terms are available on the OnlyMusic pricing page.

The service is intended for people testing song ideas, songwriters hearing original lyrics performed, and creators exploring music for videos, podcasts, games, gifts, and personal projects. OnlyMusic’s team spans music production, algorithm research, engineering, design, product, and operations, working to build a simpler path from an idea or mood to a fully realized song.

According to Shaw Wang, Co-Founder and Product Lead at OnlyMusic, ‘People should not have to hand over an email address just to find out whether an AI music tool can make something worth hearing.’ The no-sign-up offer is designed to give users room to try a few ideas, compare the results, and decide for themselves. With its focus on genuine individual use, OnlyMusic aims to provide a reliable platform for creators of all levels.

OnlyMusic’s browser-based AI music generator is available now at onlymusic.ai. The service is part of River AI LLC’s efforts to build a simpler path from an idea or mood to a fully realized song. More information about the company and its team can be found on the OnlyMusic website.

Thomson Reuters has made its proprietary large language model (LLM) official with the launch of Thomson. The company emphasizes that this achievement was accomplished at a fraction of the cost of other frontier models like ChatGPT and Claude, which it fully owns and controls.

The development of Thomson is significant not only because of its impressive performance but also due to its unique training data set. Unlike other LLMs, Thomson was trained on proprietary content built up over decades by Thomson Reuters, including Westlaw and Practical Law. This domain expertise no other company can match has been a crucial factor in the model’s success.

Joel Hron, chief technology officer at Thomson Reuters, believes that this achievement could change the economics of professional AI. He notes that for years, the industry has focused on scale as the answer: bigger models, more compute, and more money. However, with Thomson, there is another path – starting with a strong foundation, specializing it deeply for the work that matters, and building intelligence that is highly capable, far more efficient, and entirely under your control.

Thomson Reuters has invested around $40 million in developing this model over the past two years, covering both talent and compute. The final training run for the version launching today cost just $450,000. This number indicates what Thomson Reuters was able to achieve with its data and content on top of world-class open-source models.

The first deployment of Thomson will be in an upcoming release of CoCounsel Legal, driving its Tabular Analysis functionality. Thomson will be the default model for this feature, although administrators can switch it to another model through CoCounsel Legal’s admin settings. The company emphasizes that CoCounsel will remain multi-model, tapping into whichever LLM is best suited for the task at hand.

Over time, Hron expects Thomson to take a bigger and bigger share of the tokens and work done by CoCounsel. Thomson Reuters plans eventually to extend its models across its portfolio of legal and tax products. The company believes that Thomson does not need to keep pace with the frontier of general intelligence across all dimensions but should instead set the frontier for legal.

Beyond its own products, Thomson Reuters has begun early conversations with large law firms and corporations about direct access to the model. This includes firms interested in fine-tuning it with their own data as they look to bring greater sovereignty to their knowledge and expertise. The company is open to working with these organizations and licensing Thomson directly.

The development of Thomson involved starting with an existing open-source model, training it to become a legal specialist, rather than building from scratch. Although the models used changed over the course of the project, the most recent was Qwen 3.5. Central to this process were subject-matter experts – the armies of lawyers and tax professionals employed by Thomson Reuters.

Andrew Bean, who heads TR’s evaluations team, said hundreds of experts helped decide what the model should be trained to do in the first place, created many thousands of examples of high-quality answers, and then judged the model’s outputs in blind, head-to-head comparisons against frontier models. None of this training data comes from customers; instead, Thomson Reuters replicates relevant tasks with its own experts.

Thomson Reuters has engaged several third-party security firms to validate the hosting, serving, and security controls around the model. Because open-source models keep improving, the company built a process that is repeatable – swapping out starting points close to half a dozen times over the life of the project as better models emerged.

The result is a model that gained deep capability in legal and tax work without losing general abilities such as writing, math, and reasoning over long documents. This achievement marks a significant milestone for Thomson Reuters, which now owns not only its content but also the intelligence powering it.

Thomson Reuters’ internal testing of Thomson has shown strong results on various benchmarks. One evaluation described during the briefing was an in-house benchmark called Deep Research, built around legal research queries drafted by experts based on their work in practice. For each query, they supplied criteria for a good answer to contain.

The answers were scored on two dimensions: completeness and factuality – whether the citations provided actually supported the claims being made. Thomson Reuters compared its model against GPT 5.4 and Claude Sonnet 5 using this benchmark, scoring each twice – once with access only to web content and once with access to TR’s proprietary content.

While Thomson performed respectably in the web-only test, it did much better when trained on TR’s own data, outscoring both of the other LLMs. This demonstrates that there is a significant uplift from training on one’s own tools – a key advantage of having an in-house model like Thomson Reuters’ proprietary LLM.

Thomson Reuters will publish a technical report this week with comprehensive evaluations, including results on standard public AI benchmarks and human testing. The company has also begun making the model available to legal and AI academics for direct evaluation and is releasing a ‘small’ version as an open-weight model on Hugging Face under a non-commercial academic license.

One of those academics, Jonathan H. Choi at Washington University School of Law, tested Thomson against ChatGPT and Claude using challenging questions from his Corporate Tax class. He preferred Thomson’s responses overall, especially appreciating the links to treatises that made them more transparent and useful for legal work.

Thomson Reuters is building a developer portal that will eventually allow outside parties to access the model directly with API keys, configurable parameters, and sample documentation. The company emphasizes that this marks a new chapter – Thomson Reuters no longer only integrates content but also builds intelligence powering professional work.

E& UAE has introduced its new Business Pro AI plans, which bundle AI-powered broadband with Microsoft Copilot to support the next generation of businesses. This comprehensive offering combines faster business fibre connectivity, advanced digital productivity tools, and enterprise-grade AI capabilities into a single subscription.

The goal is to accelerate AI transformation for small and medium-sized enterprises (SMEs) across the country by making AI-powered productivity more accessible and better supported. Esam Mahmoud, Senior Vice President of SMB Sales and Marketing at e& UAE, emphasized that SMEs play an essential role in the UAE economy but often face cost and implementation barriers when adopting advanced AI tools.

By including Microsoft Copilot within eligible Business Pro AI plans at no additional cost, alongside faster speeds, businesses can immediately leverage generative AI to automate routine tasks. This enables them to free up valuable time and human resource efforts for strategic growth and customer engagement while keeping operational costs low.

The bundled solution includes secure connectivity, productivity tools, and AI-powered assistance under one provider, one subscription, and one monthly bill. Adopting advanced AI tools typically introduces a heavy burden of recurring costs and software licence fees, representing a major long-term financial commitment for growing businesses.

Business Pro AI with bundled AI benefits enables SMEs to automate tasks such as data analysis in seconds, significantly boosting daily employee productivity and operational efficiency. This strategic advantage is particularly valuable for resource-constrained SMEs looking to scale.

The launch comes as the UAE continues to lead global AI adoption, with Microsoft’s latest AI Diffusion Report finding that 70.1 per cent of the country’s working-age population now uses generative AI – the highest rate in the world. Copilot enables businesses to automate routine tasks and create content quickly, making it an essential tool for daily operations.

This strategic launch reinforces e& UAE’s position as a leading provider of AI-enabled connectivity, embedding AI directly into its core services. It also strengthens the established partnership between e& UAE and Microsoft, supporting the country’s national ambition to become a global leader in artificial intelligence and digital innovation.

Businesses can subscribe to new Business Pro AI or arrange a consultation on how to make the most of Microsoft Copilot by contacting their dedicated account manager or calling 800-77333.

A notification circulating on social media claims that the Ministry of Interior has transferred Abbottabad’s jurisdiction to Islamabad Capital Territory. However, experts have debunked this claim as fake and suspect it may be an artificially generated image created using advanced technology.

The document in question bears a header with ‘Government of Pakistan’ and is dated 8 August 2026. It announces the transfer of Abbottabad’s existing jurisdiction to ‘Islamabad Capital Territory under the Government of Islamabad.’ The notification has been shared on social media platforms, including Facebook, where it received over 589 likes and 140 shares.

The claim gained traction after Interior Minister Mohsin Naqvi mentioned the need for administrative reforms during a recent conference. However, this does not necessarily imply that Abbottabad’s jurisdiction is being transferred to Islamabad Capital Territory.

Soch Fact Check conducted an investigation into the authenticity of the notification by searching credible publications and news outlets. They found two articles from The News and The Express Tribune reporting on the KP assembly adopting a unanimous resolution rejecting any move to include Haripur and Abbottabad in Islamabad.

This contradicts the claim made in the fake notification, which suggests that Abbottabad’s jurisdiction is being transferred to Islamabad Capital Territory. Soch Fact Check then conducted a reverse-image search to identify the original source of the notification.

The search led them to the Ministry of Information and Broadcasting’s fact-checking account on X, where they found a post from 9 August 2026 with a red stamp marked ‘FAKE.’ The caption reads: ‘FAKE ALERT This notification circulating on social media is fabricated. Please verify official documents before sharing.’

The Ministry of Information and Broadcasting has been using AI tools to detect fake notifications like this one, which may have originated from an artificially generated image created using advanced technology.

Soch Fact Check ran the text of the document through several AI detection tools, including ‘AI or Not’ developed by Bioptic. The tool gave a rating of 97%, concluding that it is likely contains AI-generated content.

They also used Humalingo, an online writing platform that helps humanise AI-written text. It scored the notification as 95% AI and 5% human. These results further confirm that the notification about Abbottabad transferring to ICT is fake.

The Ministry of Information and Broadcasting’s post on X debunked the notification announcing the transfer of Abbottabad from KPK’s territory to Islamabad Capital Territory as fake, and experts suspect it may be an artificially generated image created using advanced technology.

Experts have been warning businesses and individuals alike about the risks associated with AI-generated content. With the rise in news being reported through AI tools for business, it is essential that people verify official documents before sharing them on social media to avoid spreading misinformation.

The Ministry of Information and Broadcasting’s fact-checking account has been actively debunking fake notifications like this one, which may have originated from an artificially generated image created using advanced technology. Their efforts are crucial in maintaining the integrity of information shared online.

A new study has introduced an artificial intelligence (AI) system that streamlines wind farm maintenance planning, reducing costs and improving efficiency. Researchers developed a domain-tailored large language model named LLM4M to optimize maintenance decision-making in the complex process of wind turbine upkeep. The team’s work addresses long-standing limitations of conventional maintenance planning tools, which often consume substantial computation time and lack flexibility when adapting to changing operational scenarios with diverse failure modes and variable expenses.

Wind farm operation and maintenance costs account for a significant 30% of onshore wind turbines’ total life cycle cost. As the global demand for renewable energy continues to grow, there is an increasing need for resource-efficient scheduling frameworks that can minimize costs while ensuring reliable turbine performance. The research frames wind turbine maintenance as a multi-team collaborative route optimization problem, which involves complex considerations such as spatial travel costs, heterogeneous repair team capabilities, and distinct failure mode repair expenses.

The researchers built LLM4M by fine-tuning the open-source Qwen2.5-7B-Instruct base model using labeled-data supervised learning techniques. They adopted low-rank adaptation (LoRA) to adjust model weights without erasing pre-trained general knowledge. The workflow relies on custom structured prompt templates that convert wind farm spatial layouts, fault information, and crew specifications into standardized text inputs for the model’s reasoning process.

The full five-step framework covers demand clarification, operational data extraction, structured prompt interaction, LLM deductive reasoning, and executable maintenance plan delivery. Benchmark tests compare LLM4M against unmodified base models, mainstream general large-parameter language models (LLMs), and traditional heuristic optimization algorithms such as genetic algorithms and particle swarm optimization.

In fixed scenario testing with three maintenance teams and five faulty turbines, the fine-tuned LLM4M generates fully implementable maintenance plans in all test samples. In contrast, the unfine-tuned base model only produces valid schedules for 83.40% of test cases. The model achieves a mean absolute percentage error (MAPE) of 0.0200 relative to mathematically derived optimal cost solutions, outperforming all evaluated general LLMs whose MAPE values range between 0.1378 and 0.1580.

Additional tests across nonfixed scenarios with variable team counts and fault node volumes verify the model’s generalization capacity. Researchers identify clear phase transition behavior within model performance, showing qualitative shifts in solving ability once training datasets include at least three maintenance teams. This observation offers actionable guidance for curating training corpora for maintenance-focused domain LLMs.

Trials simulating dynamic real-time maintenance record implementable plan generation rates of 97.90% to 99.40% for adjusted test conditions. In deployment terms, LLM4M supports local offline operation with consumer-grade graphics hardware, removing recurring API access fees and network dependence that cloud-based general LLMs require.

The study outlines future research directions including dataset expansion covering weather constraints, integration of predictive and preventive maintenance logic, interpretability analysis for model decision pathways, and embedding physical engineering rules into fine-tuning pipelines. The paper’s findings have significant implications for the wind energy industry, where efficient maintenance planning can lead to substantial cost savings and improved turbine performance.

A new multi-agent system, AgentAD, has been developed to automate the design of algorithms for Earth Observation Satellite Scheduling Problem (EOSSP). This complex task requires matching satellite observation windows, task priorities, onboard power and storage limits, and satellite attitude transition time constraints to maximize completed observation tasks. The traditional workflow involves aerospace and optimization experts manually building, testing, and revising standalone algorithms for every unique satellite-target scenario, a process marked by heavy time and resource costs.

The research team constructed AgentAD as a collaborative multi-agent system with five specialized roles: Algorithm Designer (AD), Algorithm Programmer (AP), Algorithm Optimizer (AO), Code Tester (CT), and Project Manager (PM). Each agent operates under tailored prompting paradigms that define fixed roles, task boundaries, input-output standards, and cross-agent communication protocols. The overall workflow mirrors real-world software development split into four sequential phases: design, optimization, testing, and documentation.

The system converts unstructured natural language descriptions of EOSSP scenarios into functional Python code through repeated atomic dialogue cycles between paired agents. Three core supporting mechanisms underpin stable collaboration: role specialization that restricts each agent to narrow professional duties, a permanent memory stream that archives all historical dialogue to avoid repeated errors, and chain-of-coding that breaks code writing and debugging into sequential subtasks to mitigate LLM hallucinations.

The AD first interprets raw scenario input and drafts initial algorithm frameworks including greedy, dynamic programming, and genetic structures. The PM coordinates agent communication and passes preliminary algorithm outlines to the AO, which proposes structural and parametric improvements to raise task completion performance. The AP translates revised algorithm logic into runnable annotated code, which the CT validates on EOSSP datasets and reports bugs or performance feedback back to the AO for iterative refinement.

This optimization-test loop repeats until convergence or predefined iteration limits, after which the PM compiles all dialogue records and algorithm iteration trails into standardized documentation. Ablation experiments confirm that the AO serves as the most indispensable component for performance improvement, followed by the CT for accurate feedback provision, while the memory stream delivers greater performance gains than the chain-of-coding module.

The team ran comparative trials across basic EOSSP settings with static constraints and complex scenarios augmented by dynamic task priorities, satellite forbidden time windows, and minimum imaging quality thresholds. The results show that algorithms autonomously evolved by AgentAD achieve higher task completion rates across most test environments, though the performance margin narrows as satellite and task scales expand due to increased contextual comprehension burdens on LLMs.

Convergence tracking shows AgentAD’s fitness metric Fi(a) rises steadily through iterative modification after identifying viable optimization trajectories. Trials across different large language models (LLMs), including GPT-3.5-Turbo, GPT-4o, DeepSeekV3 and Llama3.3, reveal more capable LLMs cut total atomic dialogue counts and token consumption without substantial shifts in final code length.

The authors outline plans to integrate multimodal satellite cloud data, introduce human expert feedback loops, and build visual interactive dashboards to enhance real-world operability of the automated algorithm generation pipeline. The research demonstrates the potential for large language models as a judge or decision-maker in complex optimization tasks, leveraging their capabilities to streamline data analysis tools and automate routine tasks.

Cognitive diagnosis, the backbone of online adaptive learning systems, has long relied on student answer logs and expert-built Q-matrices to quantify learners’ grasp of discrete knowledge concepts. However, conventional frameworks often focus solely on student-exercise interaction signals, neglecting parallel and synergistic connections that exist widely in educational knowledge systems.

The research team behind the HCGCDM model has addressed this limitation by introducing a graph-augmented cognitive diagnosis framework that leverages large language models to automatically map multi-type links between educational knowledge concepts. This approach enables more precise assessments of students’ learning proficiency and marks a significant improvement over existing methods.

Heterogeneous concept graphs, constructed through the HCGCDM’s two sequential functional modules, are at the heart of this innovation. The first module automates graph building without heavy manual annotation or large training datasets by embedding all knowledge concepts via BERT-base-nli-mean-tokens and filtering high-similarity concept pairs above a predefined cosine similarity threshold.

The framework then deploys RAG to pull domain-specific external knowledge bases, concatenating retrieved context with structured prompts to guide LLMs in generating relational triples covering three defined edge categories: prerequisite, parallel, and synergistic relations. A second RAG-powered LLM validation step screens unreliable triples to reduce graph noise, and validated triples form a unified heterogeneous graph split into three independent relation-specific subgraphs.

The second core module applies tailored graph neural network architectures to each subgraph to match relation characteristics. Two-layer graph attention networks process directed prerequisite subgraphs to capture asymmetric contribution weights of child concepts to parent nodes; two-layer graph convolutional networks encode parallel subgraphs to preserve hierarchical consistency among peer concepts; CoAttention-GNN handles bidirectional information flow within synergistic subgraphs to model mutual knowledge reinforcement.

The research team ran comparative tests against mainstream baseline CD models across multiple real-world educational datasets, confirming the effectiveness of HCGCDM in refining student cognitive diagnosis. Ablation experiments revealed that each core module delivers consistent performance gains, with the LLM-driven heterogeneous graph construction component producing the largest diagnostic metric shifts when removed.

Additional controlled trials swapped the base LLM between GPT-4o and Llama3-8B-Instruct, confirming that LLM capability correlates with triple construction accuracy and downstream diagnostic performance. The paper notes several existing limitations, including limited support for datasets lacking textual concept descriptions and reliance on predefined relation categories, and outlines future directions to expand automated relation extraction and incorporate multi-agent LLM pipelines for graph refinement.

The HCGCDM model has the potential to revolutionize cognitive diagnosis by providing a more comprehensive understanding of students’ learning proficiency. By leveraging large language models to automatically map complex relationships between educational knowledge concepts, this approach can help educators identify areas where students need additional support and tailor their instruction accordingly.

A team of researchers has developed an innovative framework for industrial design automation, leveraging large language models (LLMs) to minimize manual intervention. The new system, called LLM-IDA, addresses the limitations of existing LLM-assisted computer-automated design pipelines by providing a structured workflow that mitigates reliability gaps.

The research team notes that general-purpose LLMs often lack consistent explainability and robustness when deployed for iterative industrial design tasks. These black-box characteristics can lead to hallucination risks and disrupt domain-specific API calls in standard retrieval-augmented generation workflows.

To counter these issues, the researchers constructed a ‘Task Analysis–Code Generation–Code Feedback’ closed-loop logic within LLM-IDA. This mechanism maps high-dimensional natural language problem spaces onto low-dimensional structured intent spaces, reducing information entropy and applying iterative code validation with execution error feedback to lift the probability of generating fully functional engineering scripts.

The framework’s modular structure maintains domain agnosticism, allowing for adaptation to various use cases such as electronic design, energy system, and materials engineering. The researchers also outline two primary current limitations: weak spatial geometry perception for complex CAD assemblies and long-chain reasoning instability in general-purpose LLMs.

To address these challenges, the team plans to fine-tune open-source language models on CAD and CAE datasets, alongside developing a unified engineering-focused world model to strengthen physical reasoning capacity. This follow-up work aims to further enhance the capabilities of LLM-IDA for industrial design automation tasks.

The new framework’s performance was benchmarked against conventional RAG workflows using the pass@10 evaluation metric across a curated dataset of industrial design cases. Ablation tests verified the individual contributions of built-in templates, task analysis logic, and code testing feedback modules across multiple mainstream LLMs.

LLM-IDA operates within three vertical functional layers corresponding to distinct design phases: conceptual drafting, knowledge-driven refinement, and digital prototyping. The bottom layer relies on multimodal foundational black-box models for user natural language input processing, requirement analysis, and initial conceptual design generation through multimodal perception and long-sequence content creation.

The middle grey-box layer embeds specialized CAE engineering knowledge graphs and optimization algorithms to translate abstract concepts into constraint-compliant design schematics. This narrows the search space for subsequent simulation-based tuning by encoding physical manufacturing and functional rules into agent decision logic.

The top white-box layer establishes direct connections with CAD and CAE software application programming interfaces, deploying LLM code generation capabilities to build parametric digital prototypes and run finite element analysis performance evaluations with automated feedback loops. The full automation pipeline is split into two core modules: a specification module and a code-generation module that incorporates quantification, surrogate, and prototyping sub-units.

The specification module formalizes unstructured natural language user requests into standardized optimization problem tuples containing defined design variables, solution boundaries, objective functions, and constraint sets. This distributes segmented tasks to downstream code-generation agents, which operate in sequential coordination: the quantification module generates executable simulation scripts via iterative coding and visual testing loops based on PyAnsys libraries.

The surrogate module builds neural network approximation models through uniform sampling and adaptive data normalization to cut repeated high-fidelity simulation overhead. The prototyping module integrates pre-verified simulation and surrogate interfaces with multi-objective optimization templates such as NSGA-II to produce converged Pareto front solutions.

A major development in the field of artificial intelligence has taken place at UC Santa Barbara, with the university receiving a four-year, $20 million federal grant to establish an AI-programmable cloud laboratory. This cutting-edge facility will be housed within OASIS, a technology and science hub for students and startup companies on campus.

The new lab will allow researchers from anywhere in the country to access equipment remotely, marking a significant advancement in collaborative research efforts. According to Javier Read de Alaniz, a professor of chemistry and biochemistry at UCSB, ‘You can run research from anywhere at any time.’

Read de Alaniz is also co-director of BioPACIFIC MIP, a federally funded program for synthetic biology, chemistry, and materials science research that has partnered with UCSB. The university was selected to receive the AI grant due in part to its existing partnership with this program.

The OASIS facility, which opened last fall, already serves as a hub for student researchers and private/public companies. It features an open office space surrounded by laboratories where students are actively working on projects.

Universities such as Stanford, Cal Tech, and Scripps are also operating similar AI-programmable cloud labs. The National Science Foundation awarded the $20 million grant to UCSB under its umbrella, with a four-year term for the project.

A vetting process is required for researchers who apply to work remotely at OASIS, limiting access to U.S.-based organizations only. According to Rachel Segalman, vice chancellor for research at UCSB, ‘The application is limited to U.S. organizations.’

Representative Salud Carbajal announced the grant on Friday to a group of UCSB administrators and faculty outside the OASIS facility in Goleta. He described the development as ‘a big freaking deal,’ highlighting its potential impact on research capabilities nationwide.

The AI-programmable cloud laboratory is expected to streamline collaborative efforts between researchers across different institutions, making it easier for them to share resources and expertise.