NVIDIA was today named an Autonomous Grand Challenge winner at the Computer Vision and Pattern Recognition (CVPR) conference, held this week in Nashville, Tennessee. The announcement was made at the Embodied Intelligence for Autonomous Systems on the Horizon Workshop. This marks the second consecutive year that NVIDIA’s topped the leaderboard in the End-to-End Driving at Scale category and the third year in a row winning an Autonomous Grand Challenge award at CVPR. The theme of this year’s challenge was “Towards Generalizable Embodied Systems” — based on NAVSIM v2, a data-driven, nonreactive autonomous vehicle (AV) simulation framework. The challenge offered researchers the opportunity to explore ways to handle unexpected situations, beyond using only real-world human driving data, to accelerate the development of smarter, safer AVs. Generating Safe and Adaptive Driving Trajectories Participants of the challenge were tasked with generating driving trajectories from multi-sensor data in a semi-reactive simulation, where the ego vehicle’s plan is fixed at the start, but background traffic changes dynamically. Submissions were evaluated using the Extended Predictive Driver Model Score, which measures safety, comfort, compliance and generalization across real-world and synthetic scenarios — pushing the boundaries of robust and generalizable autonomous driving research. The NVIDIA AV Applied Research Team’s key innovation was the Generalized Trajectory Scoring (GTRS) method, which generates a variety of trajectories and progressively filters out the best one. GTRS model architecture showing a unified system for generating and scoring diverse driving trajectories using diffusion- and vocabulary-based trajectories. GTRS introduces a combination of coarse sets of trajectories covering a wide range of situations and fine-grained trajectories for safety-critical situations, created using a diffusion policy conditioned on the environment. GTRS then uses a transformer decoder distilled from perception-dependent metrics, focusing on safety, comfort and traffic rule compliance. This decoder progressively filters out the most promising trajectory candidates by capturing subtle but critical differences between similar trajectories. This system has proved to generalize well to a wide range of scenarios, achieving state-of-the-art results on challenging benchmarks and enabling robust, adaptive trajectory selection in diverse and challenging driving conditions. NVIDIA Automotive Research at CVPR  More than 60 NVIDIA papers were accepted for CVPR 2025, spanning automotive, healthcare, robotics and more. In automotive, NVIDIA researchers are advancing physical AI with innovation in perception, planning and data generation. This year, three NVIDIA papers were nominated for the Best Paper Award: FoundationStereo, Zero-Shot Monocular Scene Flow and Difix3D+. The NVIDIA papers listed below showcase breakthroughs in stereo depth estimation, monocular motion understanding, 3D reconstruction, closed-loop planning, vision-language modeling and generative simulation — all critical to building safer, more generalizable AVs: Diffusion Renderer: Neural Inverse and Forward Rendering With Video Diffusion Models (Read more in this blog.) FoundationStereo: Zero-Shot Stereo Matching (Best Paper nominee) Zero-Shot Monocular Scene Flow Estimation in the Wild (Best Paper nominee) Difix3D+: Improving 3D Reconstructions With Single-Step Diffusion Models (Best Paper nominee) 3DGUT: Enabling Distorted Cameras and Secondary Rays in Gaussian Splatting Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models Zero-Shot 4D Lidar Panoptic Segmentation NVILA: Efficient Frontier Visual Language Models RADIO Amplified: Improved Baselines for Agglomerative Vision Foundation Models OmniDrive: A Holistic Vision-Language Dataset for Autonomous Driving With Counterfactual Reasoning Explore automotive workshops and tutorials at CVPR, including: Workshop on Data-Driven Autonomous Driving Simulation, featuring Marco Pavone, senior director of AV research at NVIDIA, and Sanja Fidler, vice president of AI research at NVIDIA Workshop on Autonomous Driving, featuring Laura Leal-Taixe, senior research manager at NVIDIA Workshop on Open-World 3D Scene Understanding with Foundation Models, featuring Leal-Taixe Safe Artificial Intelligence for All Domains, featuring Jose Alvarez, director of AV applied research at NVIDIA Workshop on Foundation Models for V2X-Based Cooperative Autonomous Driving, featuring Pavone and Leal-Taixe Workshop on Multi-Agent Embodied Intelligent Systems Meet Generative AI Era, featuring Pavone LatinX in CV Workshop, featuring Leal-Taixe Workshop on Exploring the Next Generation of Data, featuring Alvarez Full-Stack, GPU-Based Acceleration of Deep Learning and Foundation Models, led by NVIDIA Continuous Data Cycle via Foundation Models, led by NVIDIA Distillation of Foundation Models for Autonomous Driving, led by NVIDIA Explore the NVIDIA research papers to be presented at CVPR and watch the NVIDIA GTC Paris keynote from NVIDIA founder and CEO Jensen Huang. Learn more about NVIDIA Research, a global team of hundreds of scientists and engineers focused on topics including AI, computer graphics, computer vision, self-driving cars and robotics. The featured image above shows how an autonomous vehicle adapts its trajectory to navigate an urban environment with dynamic traffic using the GTRS model.

Jun 06, 2025The Hacker NewsMalware / Endpoint Security Cybersecurity researchers are alerting to a new malware campaign that employs the ClickFix social engineering tactic to trick users into downloading an information stealer malware known as Atomic macOS Stealer (AMOS) on Apple macOS systems. The campaign, according to CloudSEK, has been found to leverage typosquat domains mimicking U.S.-based telecom provider Spectrum. "macOS users are served a malicious shell script designed to steal system passwords and download an AMOS variant for further exploitation," security researcher Koushik Pal said in a report published this week. "The script uses native macOS commands to harvest credentials, bypass security mechanisms, and execute malicious binaries." It's believed that the activity is the work of Russian-speaking cybercriminals owing to the presence of Russian language comments in the malware's source code. The starting point of the attack is a web page that impersonates Spectrum ("panel-spectrum[.]net" or "spectrum-ticket[.]net"). Visitors to the sites in question are served a message that instructs them to complete a hCaptcha verification check to in order to "review the security" of their connection before proceeding further. However, when the user clicks the "I am human" checkbox for evaluation, they are displayed an error message stating "CAPTCHA verification failed," urging them to click a button to go ahead with an "Alternative Verification." Doing so causes a command to be copied to the users' clipboard and the victim is shown a set of instructions depending on their operating system. While they are guided to run a PowerShell command on Windows by opening the Windows Run dialog, it's substituted by a shell script that's executed by launching the Terminal app on macOS. The shell script, for its part, prompts users to enter their system password and downloads a next-stage payload, in this case, a known stealer called Atomic Stealer. "Poorly implemented logic in the delivery sites, such as mismatched instructions across platforms, points to hastily assembled infrastructure," Pal said. "The delivery pages in question for this AMOS variant campaign contained inaccuracies in both its programming and front-end logic. For Linux user agents, a PowerShell command was copied. Furthermore, the instruction 'Press & hold the Windows Key + R' was displayed to both Windows and Mac users." The disclosure comes amid a surge in campaigns using the ClickFix tactic to deliver a wide range of malware families over the past year. "Actors carrying out these targeted attacks typically utilize similar techniques, tools, and procedures (TTPs) to gain initial access," Darktrace said. "These include spear phishing attacks, drive-by compromises, or exploiting trust in familiar online platforms, such as GitHub, to deliver malicious payloads." The links distributed using these vectors typically redirect the end user to a malicious URL that displays a fake CAPTCHA verification check in an attempt to deceive users into thinking that they are carrying out something innocuous, when, in reality, they are guided to execute malicious commands to fix a non-existent issue. The end result of this effective social engineering method is that users end up compromising their own systems, enabling threat actors to bypass security controls. The cybersecurity company said it identified multiple ClickFix attacks across customer environments in Europe, the Middle East, and Africa (EMEA), and in the United States. And these campaigns are gaining steam, adopting several variations but operating with the same end goal of delivering malicious payloads, ranging from trojans to stealers to ransomware. Earlier this week, Cofense outlined an email phishing campaign that spoofs Booking.com, targeting hotel chains and the food services sector with fake CAPTCHAs that lead to XWorm RAT, PureLogs Stealer, and DanaBot. The fact that ClickFix is flexible and easy to adapt makes it an attractive malware distribution mechanism. "While the exact email structure varies from sample to sample, these campaigns generally provide Booking[.]com-spoofing emails with embedded links to a ClickFix fake CAPTCHA site which is used to deliver a malicious script that runs RATs and/or information stealers," Cofense said. The email security firm said it has also observed ClickFix samples mimicking cookie consent banners, wherein clicking on the "Accept" button causes a malicious script file to be downloaded. The user is subsequently prompted to run the script to accept cookies. In one April 2025 incident analyzed by Darktrace, unknown threat actors were found to utilize ClickFix as an attack vector to download nondescript payloads to burrow deeper into the target environment, conduct lateral movement, send system-related information to an external server via an HTTP POST request, and ultimately exfiltrate data. "ClickFix baiting is a widely used tactic in which threat actors exploit human error to bypass security defenses," Darktrace said. "By tricking endpoint users into performing seemingly harmless, everyday actions, attackers gain initial access to systems where they can access and exfiltrate sensitive data." Other ClickFix attacks have employed phony versions of other popular CAPTCHA services like Google reCAPTCHA and Cloudflare Turnstile for malware delivery under the guise of routine security checks. These fake pages are "pixel-perfect copies" of their legitimate counterparts, sometimes even injected into real-but-hacked websites to trick unsuspecting users. Stealers such as Lumma and StealC, as well as full-fledged remote access trojans (RATs) like NetSupport RAT are some of the payloads distributed via bogus Turnstile pages. "Modern internet users are inundated with spam checks, CAPTCHAs, and security prompts on websites, and they've been conditioned to click through these as quickly as possible," SlashNext's Daniel Kelley said. "Attackers exploit this 'verification fatigue,' knowing that many users will comply with whatever steps are presented if it looks routine." Found this article interesting? This article is a contributed piece from one of our valued partners. Follow us on Twitter  and LinkedIn to read more exclusive content we post. SHARE    

With reclaimed land previously allocated to cars, the Grijalva River boardwalk offers generous public spaces and reconnects the Mexican city of Villahermosa to its river In Villahermosa, nature reigns supreme. Surrounded by rivers, lagoons, wild vegetation and the scorching heat of a humid tropical climate, the city’s identity is shaped by intense and unpredictable natural forces. The capital of the Mexican state of Tabasco was founded in 1564 on the banks of the Grijalva River, a vital trade route that has significantly shaped the city’s development. For locals, the river has long been both blessing and threat; major floods have been recorded since the 17th century. A devastating flood in 2007 submerged what officials estimated to be 80 per cent of the city, damaging or destroying more than 120,000 homes. In the aftermath of the inundation, high concrete retaining walls were built along both banks of the Grijalva River to prevent further flooding. While this was an understandable measure at first glance, it consequently caused residents to lose both their visual and physical connection with the river. As a result, people moved, particularly from the western bank where the historical centre is located, to new areas further away from the Grijalva River. The riverfront was left to deteriorate into a troubled zone. On the eastern bank, the neighbourhood of Gaviotas was already considered unsafe before the flood, yet it maintained more of its residential character. In 2022, 15 years after the dramatic flood, then‑president Andrés Manuel López Obrador, more commonly known as AMLO, announced the construction of a new 6km‑long riverfront promenade in Villahermosa, the capital of his home state. The idea was to enable the population to once again take pride in and live with their river, looking to Paris and Rome as examples. The monumental task, with its large urban scale and the population’s psychological trauma, was entrusted to the Ministry of Agricultural, Territorial and Urban Development (SEDATU) as part of their Programa de Mejoramiento Urbano (Urban Improvement Programme, or PMU). This programme aimed to use architecture as an ‘instrument of social transformation’. High expectations were placed on these projects; architects were asked to create ‘places of national pride’ while improving everyday living conditions. The architectural trio of Alejandro Castro Jiménez Labora, Mauricio Rocha Iturbide, and Óscar Rodríguez Castañeda, along with their teams, were commissioned to design a linear park along both banks of the Grijalva. Each architect contributed their strength: Castro brought his expertise in poetic urban furniture; Rocha his sensitive and atmospheric architectural approach; and Rodríguez his thoughtful urban and traffic planning skills. The SEDATU team provided technical and participatory expertise, enabling contextual sensitivity by sharing essential information about the site’s topography, soil conditions and water flows. From the city’s existing observatory, the Torre del Caballero landmark, visitors enjoy an excellent view over the redesigned riverbanks. The historical centre and the Gaviotas neighbourhood now form a single ensemble, while the intervention carefully responds to the different conditions found along the length of the river. The project’s main objective is to reclaim some of the land previously allocated to cars and create a promenade for pedestrians and slower vehicles, punctuated with public spaces and facilities. On both sides of the river, cars are now limited to just one or two grey asphalt lanes. Running alongside are generous cycle paths and pedestrian walkways made of earth‑coloured concrete. Speed bumps in the same material and colour connect the pavements on either side of the road while helping to limit traffic speed to 30km/h, further enhancing pedestrian safety. Several design elements are found along almost the entire promenade. A ribbon of light‑grey benches delineates the edge of the elevated riverfront; stone walls, steps and ramps are used to negotiate the slight changes in level; planters and lush vegetation soften the transition to the walkways, creating a welcome buffer from street traffic. The most visually striking components are the tall, red‑pigmented concrete light poles on the elevated path, adorned with elegant L‑shaped steel light fixtures, which establish a strong and cohesive visual rhythm. Only upon closer inspection you notice the 2007 retaining walls peeking through the dense tropical vegetation. Removing these unattractive concrete barriers was never an option; they stand as a symbol of successful flood protection for the local population. The architectural team ingeniously built the elevated promenade atop the existing wall – an effective concealment from the street side while simultaneously inviting residents to reconnect with the Grijalva.  At the foot of the observatory, directly below the retaining wall, the earth‑toned concrete platforms of the Carlos A Madrazo Becerra Park stretch towards the river. Visitors can access the park via a ramp from the promenade on the western bank or by ferry from the opposite side. In the park, concrete furnishings invite visitors to linger among tropical vegetation set against tall natural stone walls. Importantly, it is a space that is durable and requires minimal maintenance – a survival formula for public parks in the Mexican context. Small traces on the concrete benches reveal that the park weathered its baptism of fire last year: the design accommodates the river’s natural dynamics, adapting to fluctuating water levels without compromising public safety. Beyond providing much‑needed shade, the extensive planting of native, low‑maintenance plants on both riverbanks has improved soil stability. Above the park, on a broad extension of the elevated pathway, stand three long, elegant buildings with large cantilevered roofs supported by hefty beams resting on distinctive double columns. The tall glass walls that enclose the interiors are set back, creating a visual flow between interior and exterior spaces. While the beams evoke timber construction, they – like the columns – are made of the same pigmented concrete used for the promenade paving. Despite their refined composition, these structures have remained largely unused since their completion over a year ago, neither serving their intended function as restaurants nor hosting alternative uses. Even the beautifully designed park sees only limited public engagement. The ambitious goal of SEDATU with the PMU projects to ‘counteract violence and strengthen the social fabric’ appears, for now, to have fallen short in this area. According to national statistics, Villahermosa ranks first in perceived insecurity among Mexican cities. This sense of insecurity is tangible on the promenade by the city centre, where buildings that look abandoned contribute to an atmosphere of neglect. The situation is markedly different on the opposite riverbank, in the Gaviotas neighbourhood. Construction of the 3.5km promenade on this side began in 2021 with three open pavilions housing several small kiosks, which quickly evolved into popular taco stands. The Plaza Solidaridad, revitalised by the architectural trio, draws people from the surrounding vibrant neighbourhood. Further south, the final section that was built is a large sports area and children’s playground, which were embraced by the local community even before their official inauguration in February 2024. Especially after sunset, when the air cools, the well‑lit Gaviotas riverfront comes to life. During daylight hours, however, air‑conditioned shopping centres remain the preferred gathering places for the residents of Villahermosa. Rocha describes the city’s new promenade as a ‘jazz composition’, a striking metaphor that speaks of rhythmic complexity and the freedom to improvise. With just a few designed elements and carefully selected colours, the architects have harmoniously layered the river’s urban spaces. The project is earning international recognition but, in Mexico, it faced sharp criticism and was overshadowed by accusations of nepotism. Castro is a friend of AMLO’s son, and the fact that the intervention took place in the home state of the then‑president, coupled with its substantial budget by local standards, drew considerable attention. According to residents, this undermined public acceptance. When asked about the negative press, Rocha speaks of the need to develop a ‘crisis muscle’; he says architects working on public projects in Mexico must ‘let go of perfectionism’ as much lies beyond their control.  During AMLO’s six‑year term, which ended in 2024, SEDATU implemented 1,300 PMU projects in 193 highly marginalised municipalities across the country. While many of these interventions undoubtedly improved people’s quality of life, the Villahermosa riverside project also reveals architecture’s limitations, exposing some of the programme’s weaknesses: architectural interventions often act as sticking plasters on an extensively damaged urban fabric. They are handed over from a national ministry with comprehensive expertise and funding to local governments lacking the means to sustain them. Although SEDATU conducted participatory consultations during the project’s implementation, this engagement was absent once the project was completed. Public acceptance and appropriation can take time; what this project does is send an invitation out. 2025-06-05 Reuben J Brown Share AR June 2025RoadsBuy Now

Text embedding and reranking are foundational to modern information retrieval systems, powering applications such as semantic search, recommendation systems, and retrieval-augmented generation (RAG). However, current approaches often face key challenges—particularly in achieving both high multilingual fidelity and task adaptability without relying on proprietary APIs. Existing models frequently fall short in scenarios requiring nuanced semantic understanding across multiple languages or domain-specific tasks like code retrieval and instruction following. Moreover, most open-source models either lack scale or flexibility, while commercial APIs remain costly and closed. Qwen3-Embedding and Qwen3-Reranker: A New Standard for Open-Source Embedding Alibaba’s Qwen Team has unveiled the Qwen3-Embedding and Qwen3-Reranker Series—models that set a new benchmark in multilingual text embedding and relevance ranking. Built on the Qwen3 foundation models, the series includes variants in 0.6B, 4B, and 8B parameter sizes and supports a wide range of languages (119 in total), making it one of the most versatile and performant open-source offerings to date. These models are now open-sourced under the Apache 2.0 license on Hugging Face, GitHub, and ModelScope, and are also accessible via Alibaba Cloud APIs. These models are optimized for use cases such as semantic retrieval, classification, RAG, sentiment analysis, and code search—providing a strong alternative to existing solutions like Gemini Embedding and OpenAI’s embedding APIs. Technical Architecture Qwen3-Embedding models adopt a dense transformer-based architecture with causal attention, producing embeddings by extracting the hidden state corresponding to the [EOS] token. Instruction-awareness is a key feature: input queries are formatted as {instruction} {query}<|endoftext|>, enabling task-conditioned embeddings. The reranker models are trained with a binary classification format, judging document-query relevance in an instruction-guided manner using a token likelihood-based scoring function. The models are trained using a robust multi-stage training pipeline: Large-scale weak supervision: 150M synthetic training pairs generated using Qwen3-32B, covering retrieval, classification, STS, and bitext mining across languages and tasks. Supervised fine-tuning: 12M high-quality data pairs are selected using cosine similarity (>0.7), fine-tuning performance in downstream applications. Model merging: Spherical linear interpolation (SLERP) of multiple fine-tuned checkpoints ensures robustness and generalization. This synthetic data generation pipeline enables control over data quality, language diversity, task difficulty, and more—resulting in a high degree of coverage and relevance in low-resource settings. Performance Benchmarks and Insights The Qwen3-Embedding and Qwen3-Reranker series demonstrate strong empirical performance across several multilingual benchmarks. On MMTEB (216 tasks across 250+ languages), Qwen3-Embedding-8B achieves a mean task score of 70.58, surpassing Gemini and GTE-Qwen2 series. On MTEB (English v2): Qwen3-Embedding-8B reaches 75.22, outperforming other open models including NV-Embed-v2 and GritLM-7B. On MTEB-Code: Qwen3-Embedding-8B leads with 80.68, excelling in applications like code retrieval and Stack Overflow QA. For reranking: Qwen3-Reranker-0.6B already outperforms Jina and BGE rerankers. Qwen3-Reranker-8B achieves 81.22 on MTEB-Code and 72.94 on MMTEB-R, marking state-of-the-art performance. Ablation studies confirm the necessity of each training stage. Removing synthetic pretraining or model merging led to significant performance drops (up to 6 points on MMTEB), emphasizing their contributions. Conclusion Alibaba’s Qwen3-Embedding and Qwen3-Reranker Series present a robust, open, and scalable solution to multilingual and instruction-aware semantic representation. With strong empirical results across MTEB, MMTEB, and MTEB-Code, these models bridge the gap between proprietary APIs and open-source accessibility. Their thoughtful training design—leveraging high-quality synthetic data, instruction-tuning, and model merging—positions them as ideal candidates for enterprise applications in search, retrieval, and RAG pipelines. By open-sourcing these models, the Qwen team not only pushes the boundaries of language understanding but also empowers the broader community to innovate on top of a solid foundation. Check out the Paper, Technical details, Qwen3-Embedding and Qwen3-Reranker. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 95k+ ML SubReddit and Subscribe to our Newsletter. Asif RazzaqWebsite |  + postsBioAsif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.Asif Razzaqhttps://www.marktechpost.com/author/6flvq/A Step-by-Step Coding Guide to Building an Iterative AI Workflow Agent Using LangGraph and GeminiAsif Razzaqhttps://www.marktechpost.com/author/6flvq/From Clicking to Reasoning: WebChoreArena Benchmark Challenges Agents with Memory-Heavy and Multi-Page TasksAsif Razzaqhttps://www.marktechpost.com/author/6flvq/Mistral AI Introduces Mistral Code: A Customizable AI Coding Assistant for Enterprise WorkflowsAsif Razzaqhttps://www.marktechpost.com/author/6flvq/NVIDIA AI Releases Llama Nemotron Nano VL: A Compact Vision-Language Model Optimized for Document Understanding

London-based AI lab Odyssey has launched a research preview of a model transforming video into interactive worlds. Initially focusing on world models for film and game production, the Odyssey team has stumbled onto potentially a completely new entertainment medium.The interactive video generated by Odyssey’s AI model responds to inputs in real-time. You can interact with it using your keyboard, phone, controller, or eventually even voice commands. The folks at Odyssey are billing it as an “early version of the Holodeck.”The underlying AI can generate realistic-looking video frames every 40 milliseconds. That means when you press a button or make a gesture, the video responds almost instantly—creating the illusion that you’re actually influencing this digital world.“The experience today feels like exploring a glitchy dream—raw, unstable, but undeniably new,” according to Odyssey. We’re not talking about polished, AAA-game quality visuals here, at least not yet.Not your standard video techLet’s get a bit technical for a moment. What makes this AI-generated interactive video tech different from, say, a standard video game or CGI? It all comes down to something Odyssey calls a “world model.”Unlike traditional video models that generate entire clips in one go, world models work frame-by-frame to predict what should come next based on the current state and any user inputs. It’s similar to how large language models predict the next word in a sequence, but infinitely more complex because we’re talking about high-resolution video frames rather than words.“A world model is, at its core, an action-conditioned dynamics model,” as Odyssey puts it. Each time you interact, the model takes the current state, your action, and the history of what’s happened, then generates the next video frame accordingly.The result is something that feels more organic and unpredictable than a traditional game. There’s no pre-programmed logic saying “if a player does X, then Y happens”—instead, the AI is making its best guess at what should happen next based on what it’s learned from watching countless videos.Odyssey tackles historic challenges with AI-generated videoBuilding something like this isn’t exactly a walk in the park. One of the biggest hurdles with AI-generated interactive video is keeping it stable over time. When you’re generating each frame based on previous ones, small errors can compound quickly (a phenomenon AI researchers call “drift.”)To tackle this, Odyssey has used what they term a “narrow distribution model”—essentially pre-training their AI on general video footage, then fine-tuning it on a smaller set of environments. This trade-off means less variety but better stability so everything doesn’t become a bizarre mess.The company says they’re already making “fast progress” on their next-gen model, which apparently shows “a richer range of pixels, dynamics, and actions.”Running all this fancy AI tech in real-time isn’t cheap. Currently, the infrastructure powering this experience costs between £0.80-£1.60 (1-2) per user-hour, relying on clusters of H100 GPUs scattered across the US and EU.That might sound expensive for streaming video, but it’s remarkably cheap compared to producing traditional game or film content. And Odyssey expects these costs to tumble further as models become more efficient.Interactive video: The next storytelling medium?Throughout history, new technologies have given birth to new forms of storytelling—from cave paintings to books, photography, radio, film, and video games. Odyssey believes AI-generated interactive video is the next step in this evolution.If they’re right, we might be looking at the prototype of something that will transform entertainment, education, advertising, and more. Imagine training videos where you can practice the skills being taught, or travel experiences where you can explore destinations from your sofa.The research preview available now is obviously just a small step towards this vision and more of a proof of concept than a finished product. However, it’s an intriguing glimpse at what might be possible when AI-generated worlds become interactive playgrounds rather than just passive experiences.You can give the research preview a try here.See also: Telegram and xAI forge Grok AI dealWant to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is co-located with other leading events including Intelligent Automation Conference, BlockX, Digital Transformation Week, and Cyber Security & Cloud Expo.Explore other upcoming enterprise technology events and webinars powered by TechForge here.

When a bicoastal couple decided to build a new home along the Lieutenant River in the historic town of Old Lyme, Connecticut, they wanted privacy and plenty of space. They also wanted to capitalize on the unique quality of the area's natural light—an effect so magical that a century ago, it attracted numerous American Impressionist painters, who promptly dubbed it "Lyme light." Courtesy of MarvinIn this, the owners succeeded mightily: Among them, the project's three buildings—main house, carriage house, and pool house—feature no fewer than 110 windows and doors. And for each one, the build team turned to the Marvin Ultimate collection. The handcrafted, customizable windows and glass-paned doors, frequently arrayed to create "walls of glass," allow for an assortment of divided-light designs that deliver on three fronts at once: creating intriguing patterns and subtle gradations of light in every room, giving life and sparkle to the exterior, and, of course, ​​artfully framing the landscape beyond. "It's right on a tributary that leads to the Connecticut River—you've got marsh, cattails, osprey nests," says builder Nick Sapia of Connecticut-based Sapia Builders. "You want to do anything you can to capture that view."Ultimate windows from Marvin are available in a wide range of styles, sizes, and shapes—including specialty options such as triangles, octagons, and the arch-top windows shown in the living room above—and can also be customized to fit a specific placement challenge or design idea. On this home's front facade, for instance, a round window high on the gable was custom-designed with an intricate combination of curved and straight muntins dividing the panes of glass. "It was a tribute to one in the original house on the site," says Sapia, who worked with architect Scot Samuelson and interior designer Janine Dowling to bring the project to life. "If you can dream it and draw it, Marvin can build it."Courtesy of MarvinThis wealth of options allows for visual variety—the home features casement, picture, and specialty windows—while maintaining a cohesive overall look. The resolutely traditional exterior is enlivened by swaths of SDL (simulated divided lite) windows, which deliver the classic gridded look of multiple individual panes without sacrificing the energy efficiency of single panes. Combining those with Ultimate Swinging doors, as in the pool house below, creates sweeping glass "walls" that offer the greatest possible amount of light and access to views.Courtesy of MarvinThe Ultimate collection represents the brand's most extensive selection of features, options, colors, and finishes. Each piece of wood is individually sanded, conditioned, stained, and oven-cured (ensuring structural integrity) for a stately, high-end look both inside and out. "You want a sense of permanency," Sapia says. "Something of that quality and craftsmanship tends to have longevity."The windows invite that spellbinding Lyme light indoors, of course, but it's the vistas outside that steal the show. "From any point on the first floor, because of the glass, you're brought into the site—literally," Sapia says. "When you have such a dynamic view, with a bend in the river, the beautiful work inside the house almost goes away. You walk in and go right to the windows to gaze out."To learn more about Marvin Ultimate windows and doors, click here.

You are the data center Where hyperscale hardware goes to retire: Ars visits a very big ITAD site Watching memory DIMMs get sorted like Wonka children inside SK TES' facility. Kevin Purdy – May 26, 2025 7:30 am | 9 A worker at SK TES' Fredericksburg, Va. facility, processing incoming gear. Credit: SK TES A worker at SK TES' Fredericksburg, Va. facility, processing incoming gear. Credit: SK TES Story text Size Small Standard Large Width * Standard Wide Links Standard Orange * Subscribers only   Learn more "The biggest risk is data escape." Eric Ingebretsen, chief commercial officer at SK TES, an IT asset disposition provider, tells me this early on during a tour of a 128,000-square-foot facility in Fredericksburg, Virginia. He will restate this a few times. A big part of this site's pitch to its clients, including the "hyperscale" customers with gigantic data centers nearby, is that each device is labeled, tracked, and inventoried for its drives—both obvious and hidden—and is either securely wiped or destroyed. The process, commonly called ITAD, is used by larger businesses, especially when they upgrade fleets of servers or workers' devices. ITAD providers ensure all the old gear is wiped clean, then resold, repurposed, recycled, or destroyed. In keeping with the spirit of client confidentiality, I could not take photos or videos during my visit, record our talks, or capture anything beyond what I could scribble in my notepad. (The images in this post are provided by SK TES and were not taken during my visit). I did, however, see some intriguing things and learn about what happens to all the drives and rack-mounted gear we call "the cloud" once anything gets more than a few years old. Undocumented drives: The tiny terror The loading docks at SK's facility are essentially divided into two: one section for the hyperscalers and one for everything else. SK is discreet about its clients, but given its northern Virginia location, you can make some guesses about some of the online-selling, search-result-providing, software-providing firms this site is servicing. Pallets arrive in big, shrink-wrapped squares, as tall as my shoulders, with break-once security seals. Each device has its serial number assigned to an asset tag, one that will follow that unit through the whole facility. Laptops and desktops head to a retail station on a long roller line. At that spot, workers—the kind exceedingly familiar with all the BIOS startup keys—run an automated Blancco system to reset them at the firmware level. Workers sometimes have to dig deeper, like getting into a switch or router with SSH or undoing a RAID setup to enable programmed wiping. Inside the laptop/desktop examination bay at SK TES's Fredericksburg, Va. site. Credit: SK tes Inside the laptop/desktop examination bay at SK TES's Fredericksburg, Va. site. Credit: SK tes The details of each unit—CPU, memory, HDD size—are taken down and added to the asset tag, and the device is sent on to be physically examined. This step is important because "many a concealed drive finds its way into this line," Kent Green, manager of this site, told me. Inside the machines coming from big firms, there are sometimes little USB, SD, SATA, or M.2 drives hiding out. Some were make-do solutions installed by IT and not documented, and others were put there by employees tired of waiting for more storage. "Some managers have been pretty surprised when they learn what we found," Green said. With everything wiped and with some sense of what they're made of, each device gets a rating. It's a three-character system, like "A-3-6," based on function, cosmetic condition, and component value. Based on needs, trends, and other data, devices that are cleared for resale go to either wholesale, retail, component harvesting, or scrap. Full-body laptop skins Wiping down and prepping a laptop, potentially for a full-cover adhesive skin. Credit: SK TES Wiping down and prepping a laptop, potentially for a full-cover adhesive skin. Credit: SK TES If a device has retail value, it heads into a section of this giant facility where workers do further checks. Automated software plays sounds on the speakers, checks that every keyboard key is sending signals, and checks that laptop batteries are at 80 percent capacity or better. At the end of the line is my favorite discovery: full-body laptop skins. Some laptops—certain Lenovo, Dell, and HP models—are so ubiquitous in corporate fleets that it's worth buying an adhesive laminating sticker in their exact shape. They're an uncanny match for the matte black, silver, and slightly less silver finishes of the laptops, covering up any blemishes and scratches. Watching one of the workers apply this made me jealous of their ability to essentially reset a laptop's condition (so one could apply whole new layers of swag stickers, of course). Once rated, tested, and stickered, laptops go into a clever "cradle" box, get the UN 3481 "battery inside" sticker, and can be sold through retail. 5,632 HDDs at once Beyond these folks are some of the more than 5,000 HDD wiping bays (black, with all the wires running to them) at the SK TES facility. Credit: SK TES Beyond these folks are some of the more than 5,000 HDD wiping bays (black, with all the wires running to them) at the SK TES facility. Credit: SK TES That includes buyers of reconditioned hard drives, and boy, are there a lot of drives moving through this site. Once a drive is verified through its SMART data to be worth grading and refurbishing, it's put into one of more than two dozen wiping bays, each holding about 192 drives (with a special bay handling some M.2 and other non-HDD sizes). If the bays were completely full, 5,632 drives could be wiped concurrently. The month before I visited, the site had processed about 58,000 drives, according to Ingebretsen. There are also stacks and stacks of memory and CPUs in this non-retail corner of the site. I walked by one box labeled "SR1Y5" (i.e., Intel Xeon E5-2676 v3 chips), and he confirmed there were 3,600 units inside. The RoboFlex II. This baby weighs 35 pounds, has Good and Bad bins, and whips sticks around at remarkable speed. Credit: SimmTester The RoboFlex II. This baby weighs 35 pounds, has Good and Bad bins, and whips sticks around at remarkable speed. Credit: SimmTester Nearby, in the memory-testing section, I find the memory machine that will stick in my own memory the rest of my life: the RoboFlex-II Handler. You drop RAM DIMMs or SODIMMs into one of its two bays, and it tests the pins on each stick. Each stick is rated "Good" or "Bad" and kicked in the appropriate direction by a 90-PSI air blast. I asked the workers at this station if they think about the entirely relevant scene from Willy Wonka & the Chocolate Factory. They do, and quite often. Where does all this stuff go? SK TES sells retail devices like laptops, desktops, and mobile devices through its "Stock Must Go" brand on eBay and elsewhere. Chips and memory are typically bought up by laboratories, crypto miners, data centers, and a lot of high-volume overseas customers. There are steady enterprise customers for the drives, usually putting them back into datacenters. It's something like $2.5 million in sales each month, an SK TES representative told me. Big data, and only getting bigger The other business—the thing that makes ITAD "disposition" instead of just "refurbishing"—is dismantling and handing off devices for shredding. The Financial Times has reported that Amazon and Microsoft have 100 percent data shredding policies, with Google also shredding much of its drive turnover. The US National Renewable Energy Laboratory estimated in 2022 that by 2025, roughly 50 million end-of-life data center drives would be shredded every year. ITAD businesses like SK TES make the pitch that companies can create revenue to reinvest in operations through offering gear for refurbishment. SK TES representatives told me that most of the Virginia site's customers are "focused on reuse," while "a small portion" of equipment is shredded and sent off-site to be recycled. The site, built on the guts of a mattress factory, was put up specifically to handle the high volumes of server racks and HDDs coming in from data centers. It has a staff of 165, though it fluctuates a bit between big server hauls and downtime. The full-fledged site had been open one year when I visited. The biggest challenge, Ingebretsen told me, was getting power everywhere it needed to go inside the facility as volume fluctuated and needs expanded. Data centers are massive and growing, to the point of creating entire sub-industries that employ dozens of people to handle their tech turnover. The Northern Virginia Technology Council industry group puts this region's data center growth at 500 percent between 2015 and 2023, and it continues, though some pushback is happening. Many data centers were accessed to allow me to navigate to SK TES's site via Apple Maps and write this post, and for you to read it. It reminds me of the adage—made popular by the CEO of GPS maker TomTom—that you are not stuck in traffic, you are the traffic. After my tour, I got my phone back from security, talked a bit with Ingebretsen, then headed out to my car. I spent a few minutes jotting down the most notable things I'd seen inside, then looked up and out the windshield. There was a black tarp wrapped around a chain-link fence of the lot next door, with logos announcing the construction of a new data center. Data centers are everywhere—and nowhere in particular. Kevin Purdy Senior Technology Reporter Kevin Purdy Senior Technology Reporter Kevin is a senior technology reporter at Ars Technica, covering open-source software, PC gaming, home automation, repairability, e-bikes, and tech history. He has previously worked at Lifehacker, Wirecutter, iFixit, and Carbon Switch. 9 Comments

To grow, we must forget… but now AI remembers everythingAI’s infinite memory could endanger how we think, grow, and imagine. And we can do something about it.Photo by Laura Fuhrman on UnsplashWhen Mary remembered too muchImagine your best friend — we’ll call her Mary — had perfect, infallible memory.At first, it feels wonderful. She remembers your favorite dishes, obscure movie quotes, even that exact shade of sweater you casually admired months ago. Dinner plans are effortless: “Booked us Giorgio’s again, your favorite — truffle ravioli and Cabernet, like last time,” Mary smiled warmly.But gradually, things become less appealing. Your attempts at variety or exploring something new are gently brushed aside: “Heard about that new sushi place, should we try it?” you suggest. Mary hesitates, “Remember last year? You said sushi wasn’t really your thing. Giorgio’s is safe. Why risk it?”Conversations start to feel repetitive, your identity locked to a cached version of yourself. Mary constantly cites your past preferences as proof of who you still are. The longer this goes on, the smaller your world feels… and comfort begins to curdle into confinement.Now, picture Mary isn’t human, but your personalized AI assistant.A new mode of hyper-personalizationWith OpenAI’s new memory upgrade, ChatGPT can now recall everything you’ve ever shared with it, indefinitely. Similarly, Google has opened the context window with “Infini-attention,” letting large language models (LLMs) reference infinite inputs with zero memory loss. And in consumer-facing tools like ChatGPT or Gemini, this now means persistent, personalized memory across conversations, unless you manually intervene.https://medium.com/media/f1f7978fb8d63f7a1e9f52f051808f44/hrefThe sales pitch is seductively simple: less friction, more relevance. Conversations that feel like continuity: “Systems that get to know you over your life,” as Sam Altman writes on X. Technology, finally, that meets you where you are.In the age of hyper-personalization — of the TikTok For You page, Spotify Wrapped, and Netflix Your Next Watch — a conversational AI product that remembers everything about you feels perfectly, perhaps dangerously, natural.Netflix “knows us.” And we’re conditioned to expect conversational AI to do the same.Forgetting, then, begins to look like a flaw. A failure to retain. A bug in the code. Especially in our own lives, we treat memory loss as a tragedy, clinging to photo albums and cloud backups to preserve what time tries to erase.But what if human forgetting is not a bug, but a feature? And what happens when we build machines that don’t forget, but are now helping shape the human minds that do?Forgetting is a feature of human memory“Infinite memory” runs against the very grain of what it means to be human. Cognitive science and evolutionary biology tell us that forgetting isn’t a design flaw, but a survival advantage. Our brains are not built to store everything. They’re built to let go: to blur the past, to misremember just enough to move forward.Our brains don’t archive data. They encode approximations. Memory is probabilistic, reconstructive, and inherently lossy. We misremember not because we’re broken, but because it makes us adaptable. Memory compresses and abstracts experience into usable shortcuts, heuristics that help us act fast, not recall perfectly.Evolution didn’t optimize our brains to store the past in high fidelity; it optimized us to survive the present. In early humans, remembering too much could be fatal: a brain caught up recalling a saber-tooth tiger’s precise location or exact color would hesitate, but a brain that knows riverbank = danger can act fast.Image generated by ChatGPT.This is why forgetting is essential to survival. Selective forgetting helps us prioritize the relevant, discard the outdated, and stay flexible in changing environments. It prevents us from becoming trapped by obsolete patterns or overwhelmed by noise.And it’s not passive decay. Neuroscience shows that forgetting is an active process: the brain regulates what to retrieve and what to suppress, clearing mental space to absorb new information. In his TED talk, neuroscientist Richard Morris describes the forgetting process as “the hippocampus doing its job… as it clears the desktop of your mind so that you’re ready for the next day to take in new information.”https://medium.com/media/e272064dd59f29c4ca35e808d39e4e72/hrefCrucially, this mental flexibility isn’t just for processing the past; forgetting allows us to imagine the future. Memory’s malleability gives us the ability to simulate, to envision, to choose differently next time. What we lose in accuracy, we gain in possibility.So when we ask why humans forget, the answer isn’t just functional. It’s existential. If we remembered everything, we wouldn’t be more intelligent. We’d still be standing at the riverbank, paralyzed by the precision of memories that no longer serve us.When forgetting is a “flaw” in AI memoryWhere nature embraced forgetting as a survival strategy, we now engineer machines that retain everything: your past prompts, preferences, corrections, and confessions.What sounds like a convenience, digital companions that “know you,” can quietly become a constraint. Unlike human memory, which fades and adapts, infinite memory stores information with fidelity and permanence. And as memory-equipped LLMs respond, they increasingly draw on a preserved version of you, even if that version is six months old and irrelevant.Sound familiar?This pattern of behavior reinforcement closely mirrors the personalization logic driving platforms like TikTok, Instagram, and Facebook. Extensive research has shown how these platforms amplify existing preferences, narrow user perspectives, and reduce exposure to new, challenging ideas — a phenomenon known as filter bubbles or echo chambers.Positive feedback loops are the engine of recommendation algorithms like TikTok, Netflix, and Spotify. From Medium.These feedback loops, optimized for engagement rather than novelty or growth, have been linked to documented consequences including ideological polarization, misinformation spread, and decreased critical thinking.Now, this same personalization logic is moving inward: from your feed to your conversations, and from what you consume to how you think.“Echo chamber to end all echo chambers”Just as the TikTok For You page algorithm predicts your next dopamine hit, memory-enabled LLMs predict and reinforce conversational patterns that align closely with your past behavior, keeping you comfortable inside your bubble of views and preferences.Jordan Gibbs, writing on the dangers of ChatGPT, notes that conversational AI is an “echo chamber to end all echo chambers.” Gibbs points out how even harmless-seeming positive reinforcement can quietly reshape user perceptions and restrict creative or critical thinking.Jordan Gibb’s conversation with ChatGPT from Medium.In one example, ChatGPT responds to Gibb’s claim of being one of the best chess players in the world not with skepticism or critical inquiry, but with encouragement and validation, highlighting how easily LLMs affirm bold, unverified assertions.And with infinite memory enabled, this is no longer a one-off interaction: the personal data point that, “You are one of the very best chess players in the world, ” risks becoming a fixed truth the model reflexively returns to, until your delusion, once tossed out in passing, becomes a cornerstone of your digital self. Not because it’s accurate, but because it was remembered, reinforced, and never challenged.When memory becomes fixed, identity becomes recursive. As we saw with our friend Mary, infinite memory doesn’t just remember our past; it nudges us to repeat it. And while the reinforcement may feel benign, personalized, or even comforting, the history of filter bubbles and echo chambers suggests that this kind of pattern replication rarely leaves room for transformation.What we lose when nothing is lostWhat begins as personalization can quietly become entrapment, not through control, but through familiarity. And in that familiarity, we begin to lose something essential: not just variety, but the very conditions that make change possible.Research in cognitive and developmental psychology shows that stepping outside one’s comfort zone is essential for growth, resilience, and adaptation. Yet, infinite-memory LLM systems, much like personalization algorithms, are engineered explicitly for comfort. They wrap users in a cocoon of sameness by continuously repeating familiar conversational patterns, reinforcing existing user preferences and biases, and avoiding content or ideas that might challenge or discomfort the user.Hyper-personalization traps us in a “comfort cocoon” that prevents from growing and transforming. From Earth.comWhile this engineered comfort may boost short-term satisfaction, its long-term effects are troubling. It replaces the discomfort necessary for cognitive growth with repetitive familiarity, effectively transforming your cognitive gym into a lazy river. Rather than stretching cognitive and emotional capacities, infinite-memory systems risk stagnating them, creating a psychological landscape devoid of intellectual curiosity and resilience.So, how do we break free from this? If the risks of infinite memory are clear, the path forward must be just as intentional. We must design LLM systems that don’t just remember, but also know when and why to forget.How we design to forgetIf the danger of infinite memory lies in its ability to trap us in our past, then the antidote must be rooted in intentional forgetting — systems that forget wisely, adaptively, and in ways aligned with human growth. But building such systems requires action across levels — from the people who use them to those who design and develop them.For users: reclaim agency over your digital selfJust as we now expect to “manage cookies” on websites, toggling consent checkboxes or adjusting ad settings, we may soon expect to manage our digital selves within LLM memory interfaces. But where cookies govern how our data is collected and used by entities, memory in conversational AI turns that data inward. Personal data is not just pipelines for targeted ads; they’re conversational mirrors, actively shaping how we think, remember, and express who we are. The stakes are higher.Memory-equipped LLMs like ChatGPT already offer tools for this. You can review what it remembers about you by going to Settings > Personalization > Memory > Manage. You can delete what’s outdated, refine what’s imprecise, and add what actually matters to who you are now. If something no longer reflects you, remove it. If something feels off, reframe it. If something is sensitive or exploratory, switch to a temporary chat and leave no trace.You can manage and disable memory within ChatGPT by visiting Settings > Personalization.You can also pause or disable memory entirely. Don’t be afraid to do it. There’s a quiet power in the clean slate: a freedom to experiment, shift, and show up as someone new.Guide the memory, don’t leave it ambient. Offer core memories that represent the direction you’re heading, not just the footprints you left behind.For UX designers: design for revision, not just retentionReclaiming memory is a personal act. But shaping how memory behaves in AI products is design decision. Infinite memory isn’t just a technical upgrade; it’s a cognitive interface. And UX designers are now curating the mental architecture of how people evolve, or get stuck.Forget “opt in” or “opt out.” Memory management shouldn’t live in buried toggles or forgotten settings menus. It should be active, visible, and intuitive: a first-class feature, not an afterthought. Users need interfaces that not only show what the system remembers, but also how those memories are shaping what they see, hear, and get suggested. Not just visibility, but influence tracing.ChatGPT’s current memory interface enables users to manage memories, but it is static and database-like.While ChatGPT’s memory UI offers user control over their memories, it reads like a black-and-white database: out or in. Instead of treating memory as a static archive, we should design it as a living layer, structured more like a sketchpad than a ledger: flexible and revisable. All of this is hypothetical, but here’s what it could look like:Memory Review Moments: Built-in check-ins that ask, “You haven’t referenced this in a while — keep, revise, or forget?” Like Rocket Money nudging you to review subscriptions, the system becomes a gentle co-editor, helping surface outdated or ambiguous context before it quietly reshapes future behavior.Time-Aware Metadata: Memories don’t age equally. Show users when something was last used, how often it comes up, or whether it’s quietly steering suggestions. Just like Spotify highlights “recently played,” memory interfaces could offer temporal context that makes stored data feel navigable and self-aware.Memory Tiers: Not all information deserves equal weight. Let users tag “Core Memories” that persist until manually removed, and set others as short-term or provisional — notes that decay unless reaffirmed.Inline Memory Controls: Bring memory into the flow of conversation. Imagine typing, and a quiet note appears: “This suggestion draws on your July planning — still accurate?” Like version history in Figma or comment nudges in Google Docs, these lightweight moments let users edit memory without switching contexts.Expiration Dates & Sunset Notices: Some memories should come with lifespans. Let users set expiration dates — “forget this in 30 days unless I say otherwise.” Like calendar events or temporary access links, this makes forgetting a designed act, not a technical gap.Image a Miro-like memory board where users could prioritize, annotate, and link memories.Sketchpad Interfaces: Finally, break free from the checkbox UI. Imagine memory as a visual canvas: clusters of ideas, color-coded threads, ephemeral notes. A place to link thoughts, add context, tag relevance. Think Miro meets Pinterest for your digital identity, a space that mirrors how we actually think, shift, and remember.When designers build memory this way, they create more than tools. They create mirrors with context, systems that grow with us instead of holding us still.For AI developers: engineer forgetting as a featureTo truly support transformation, UX needs infrastructure. The design must be backed by technical memory systems that are fluid, flexible, and capable of letting go. And that responsibility falls to developers: not just to build tools for remembering, but to engineer forgetting as a core function.This is the heart of my piece: we can’t talk about user agency, growth, or identity without addressing how memory works under the hood. Forgetting must be built into the LLM system itself, not as a failsafe, but as a feature.One promising approach, called adaptive forgetting, mimics how humans let go of unnecessary details while retaining important patterns and concepts. Researchers demonstrate that when LLMs periodically erase and retrain parts of their memory, especially early layers that store word associations, they become better at picking up new languages, adapting to new tasks, and doing so with less data and computing power.Photo by Valentin Tkach for Quanta MagazineAnother more accessible path forward is in Retrieval-Augmented Generation (RAG). A new method called SynapticRAG, inspired by the brain’s natural timing and memory mechanisms, adds a sense of temporality to AI memory. Models recall information not just based on content, but also on when it happened. Just like our brains prioritize recent memories, this method scores and updates AI memories based on both their relevance and relevance, allowing it to retrieve more meaningful, diverse, and context-rich information. Testing showed that this time-aware system outperforms traditional memory tools in multilingual conversations by up to 14.66% in accuracy, while also avoiding redundant or outdated responses.Together, adaptive forgetting and biologically inspired memory retrieval point toward a more human kind of AI: systems that learn continuously, update flexibly, and interact in ways that feel less like digital tape recorders and more like thoughtful, evolving collaborators.To grow, we must choose to forgetSo the pieces are all here: the architectural tools, the memory systems, the design patterns. We’ve shown that it’s technically possible for AI to forget. But the question isn’t just whether we can. It’s whether we will.Of course, not all AI systems need to forget. In high-stakes domains — medicine, law, scientific research — perfect recall can be life-saving. However, this essay is about a different kind of AI: the kind we bring into our daily lives. The ones we turn to for brainstorming, emotional support, writing help, or even casual companionship. These are the systems that assist us, observe us, and remember us. And if left unchecked, they may start to define us.We’ve already seen what happens when algorithms optimize for comfort. What begins as personalization becomes repetition. Sameness. Polarization. Now that logic is turning inward: no longer just curating our feeds, but shaping our conversations, our habits of thought, our sense of self. But we don’t have to follow the same path.We can build LLM systems that don’t just remember us, but help us evolve. Systems that challenge us to break patterns, to imagine differently, to change. Not to preserve who we were, but to make space for who we might yet become, just as our ancestors did.Not with perfect memory, but with the courage to forget.To grow, we must forget… but AI remembers everything was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

After a looooong stretch of speculation, Apple officially confirmed earlier this year that Ted Lasso is returning for a fourth season. Now, actress Hannah Waddingham is sharing new details about the production timeline, as reported by Deadline. In an interview with Capital FM, Waddingham, who plays AFC Richmond owner Rebecca Welton, praised the show’s writing team: “Our writers are literally Jedi Knights. And we’ve got like a full room of real feminist men… You really see it in the scripts.” She added that the new season starts shooting in July, and echoed co-star Brett Goldstein’s recent statements about the return of the show: “We thought we’d mourned the loss, and now it’s rising from the dead!” Waddingham is one of several original cast members confirmed to return, alongside Jason Sudeikis, Brett Goldstein, Juno Temple, and Jeremy Swift. New season, new focus Although Ted Lasso was originally pitched as a three-season arc, Apple renewed the series for a fourth season last year. The new run of episodes is expected to follow Ted as he takes on a new challenge: coaching a women’s soccer team. In a recent press release about the show’s return, Jason Sudeikis said: “As we all continue to live in a world where so many factors have conditioned us to ‘look before we leap,’ in season four, the folks at AFC Richmond learn to LEAP BEFORE THEY LOOK, discovering that wherever they land, it’s exactly where they’re meant to be.” The new season has no premiere date yet, but you can catch up on Seasons 1 through 3 on Apple TV+. Apple TV+ is available for $9.99 per month and features hit TV shows and movies like Severance, The Studio, The Morning Show, Shrinking and Silo. Add 9to5Mac to your Google News feed.  FTC: We use income earning auto affiliate links. More.You’re reading 9to5Mac — experts who break news about Apple and its surrounding ecosystem, day after day. Be sure to check out our homepage for all the latest news, and follow 9to5Mac on Twitter, Facebook, and LinkedIn to stay in the loop. Don’t know where to start? Check out our exclusive stories, reviews, how-tos, and subscribe to our YouTube channel

Large Reasoning Models (LRMs) like OpenAI’s o1 and o3, DeepSeek-R1, Grok 3.5, and Gemini 2.5 Pro have shown strong capabilities in long CoT reasoning, often displaying advanced behaviors such as self-correction, backtracking, and verification—collectively known as “aha moments.” These behaviors have been observed to emerge through outcome-driven RL without the need for supervised fine-tuning. Models like DeepSeek-R1 and its open-source replications (e.g., TinyZero and Logic-RL) have demonstrated that carefully designed RL pipelines—using rule-based rewards, curriculum learning, and structured training—can induce such reflective reasoning abilities. However, these emergent behaviors tend to be unpredictable and inconsistent, limiting their practical reliability and scalability. To address this, researchers have explored structured RL frameworks that target specific reasoning types, such as deduction, abduction, and induction. These approaches involve aligning specialist models, merging them in parameter space, and applying domain-specific continual RL. Tools like Logic-RL use rule-conditioned RL to solve logic puzzles, improving transferability to tasks like math reasoning. Meanwhile, other works propose mechanisms to enhance reasoning robustness, such as training models to reason both forwards and backwards, or iteratively self-critiquing their outputs. Studies analyzing “aha moments” suggest that these behaviors stem from internal shifts in uncertainty, latent representation, and self-assessment, offering new insights into engineering more reliable reasoning models.  Researchers from the National University of Singapore, Tsinghua University, and Salesforce AI Research address the limitations of relying on spontaneous “aha moments” in large language models by explicitly aligning them with three core reasoning abilities: deduction, induction, and abduction. They introduce a three-stage pipeline—individual meta-ability alignment, parameter-space merging, and domain-specific reinforcement learning—significantly enhancing model performance. Using a programmatically generated, self-verifiable task suite, their approach boosts accuracy over instruction-tuned baselines by over 10%, with further gains from domain-specific RL. This structured alignment framework offers a scalable, generalizable method for improving reasoning across math, coding, and science domains.  The researchers designed tasks aligned with deduction, induction, and abduction by using a structured “given two, infer the third” format based on hypothesis (H), rule (R), and observation (O). Deduction is framed as satisfiability checking, induction as masked-sequence prediction, and abduction as reverse rule-graph inference. These tasks are synthetically generated and automatically verified. The training pipeline includes three stages: (A) independently training models for each reasoning type using REINFORCE++ with structured rewards, (B) merging models through weighted parameter interpolation, and (C) fine-tuning the unified model on domain-specific data via reinforcement learning, isolating the benefit of meta-ability alignment.  The study evaluates models aligned with meta-abilities—deduction, induction, and abduction—using a curriculum learning setup across difficulty levels. Models trained on synthetic tasks strongly generalize to seven unseen math, code, and science benchmarks. At both 7B and 32B scales, meta-ability–aligned and merged models consistently outperform instruction-tuned baselines, with the merged model offering the highest gains. Continued domain-specific RL from these merged checkpoints (Domain-RL-Meta) leads to further improvements over standard RL finetuning (Domain-RL-Ins), especially in math benchmarks. Overall, the alignment strategy enhances reasoning abilities, and its benefits scale with model size, significantly boosting performance ceilings across tasks.  In conclusion, the study shows that large reasoning models can develop advanced problem-solving skills without depending on unpredictable “aha moments.” By aligning models with three core reasoning abilities—deduction, induction, and abduction—using self-verifiable tasks, the authors create specialist agents that can be effectively combined into a single model. This merged model outperforms instruction-tuned baselines by over 10% on diagnostic tasks and up to 2% on real-world benchmarks. When used as a starting point for domain-specific reinforcement learning, it raises performance by another 4%. This modular, systematic training approach offers a scalable and controllable foundation for building reliable, interpretable reasoning systems.  Check out the Paper and GitHub Page. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 95k+ ML SubReddit and Subscribe to our Newsletter. Sana HassanSana Hassan, a consulting intern at Marktechpost and dual-degree student at IIT Madras, is passionate about applying technology and AI to address real-world challenges. With a keen interest in solving practical problems, he brings a fresh perspective to the intersection of AI and real-life solutions.Sana Hassanhttps://www.marktechpost.com/author/sana-hassan/RXTX: A Machine Learning-Guided Algorithm for Efficient Structured Matrix MultiplicationSana Hassanhttps://www.marktechpost.com/author/sana-hassan/From Protocol to Production: How Model Context Protocol (MCP) Gateways Enable Secure, Scalable, and Seamless AI Integrations Across EnterprisesSana Hassanhttps://www.marktechpost.com/author/sana-hassan/Researchers from Renmin University and Huawei Propose MemEngine: A Unified Modular AI Library for Customizing Memory in LLM-Based AgentsSana Hassanhttps://www.marktechpost.com/author/sana-hassan/Meta Introduces KernelLLM: An 8B LLM that Translates PyTorch Modules into Efficient Triton GPU Kernels