SK Hynix told TechCrunch the company hasn't finalized any plans or arrangements yet.

HS Hyosung Information Systems, an AI and data infrastructure solutions provider, held its Data Center Modernization Conference 2026 at Lotte Hotel Busan on Sept. 15. The conference is an annual event that HS Hyosung Information Systems has held for IT professionals in the Busan and South Gyeongsang region for more than a decade.

HS Hyosung Information Systems said it shared strategies and insights for data center modernization in the AI era, including data and infrastructure strategies for enterprise AI, more efficient private cloud operations and ways to strengthen cyber resilience.

According to HS Hyosung Information Systems, data center infrastructure is increasingly required to reliably process and operate large-scale AI workloads and vast amounts of data, driving a shift toward redesigning overall infrastructure to integrate AI infrastructure, cloud operations and data protection.

In response to these changes, the conference broadly covered key infrastructure strategies required for data centers, including data and infrastructure development for enterprise AI, modernization of AI training and inference environments, greater efficiency in private cloud and data center operations, and security and data protection.

Dong-soo Kwon, a senior expert at HS Hyosung Information Systems, delivered a keynote presentation titled "Data and Infrastructure Strategies for Enterprise AI in the Agentic AI Era of 2026." Go-hoon Jung, a project manager at HS Hyosung Information Systems, then outlined the company's capabilities and strategies for building private clouds through a smart city development project case study. Dong-hyuk Kim, a team leader at HS Hyosung Information Systems, also presented data protection and recovery strategies designed to support uninterrupted business operations amid disasters and cyber threats.

Major technology partners also participated. FuriosaAI showcased its RNGD neural processing unit (NPU), which supports high power efficiency and high-performance inference, while Lablup presented modernization tools for efficient AI training and inference.

VMware by Broadcom shared ways to reduce data center costs using VMware Cloud Foundation (VCF), while Veeam Software presented next-generation service recovery strategies for zero-trust environments. Palo Alto Networks Korea presented a data center modernization strategy using its Precision AI-based security platform.

The venue also featured exhibition booths where attendees could explore technologies and solutions from major partners including C Platform, Ocean Technology, Gensys and Union Biometrics.

SK Telecom (SKT) said it is hosting an in-person standardization meeting of the GSMA Rich Communication Services (RCS) group at SKT Tower in Seoul from Sept. 15 to 18, where participants will discuss ways to combine artificial intelligence (AI) with messaging services.

As AI's capabilities expand into real-world actions such as reservations and payments, SKT is seeking to redefine the role of messaging. The company envisions using messages as a channel through which AI agents obtain users' final approval before carrying out actions.

The GSMA RCS group is a body under the GSMA where global telecommunications carriers, device and platform companies, and business messaging providers discuss RCS technical standards and service specifications. The SKT-hosted meeting is attended by 33 representatives, including staff from 16 companies — global carriers AT&T, T-Mobile, Vodafone and Orange, as well as Apple, Google and Samsung Electronics — along with GSMA officials. RCS is a next-generation messaging standard that supports features such as photo and video transmission through smartphones' native messaging apps.

An AI special session has been added to the meeting alongside regular standardization topics. SKT proposed the AI agenda to the GSMA and agreed to hold a special session titled "AI and the Evolution of Messaging" on Sept. 17. SKT also led the organization of the session and recruitment of participating companies.

The session will cover the integration of RCS and AI agents, trust in AI and user control, improvements to brand registration using AI, and AI-based spam prevention for business messaging. Participants will share their technologies and service experience and discuss the functions and roles messaging will require in the AI era.

User control is at the heart of discussions about connecting AI with RCS. As AI agents increasingly execute actual transactions, users need to directly review and approve key information such as order details and payment amounts before a transaction is finalized. If that confirmation takes place within the AI service itself, however, it could be affected by malfunctions or external manipulation. Different approval methods among AI agents could also make it difficult for users to determine whether a request is legitimate.

To address the issue, SKT has proposed "AI Agent Approval over RCS" as a standardization agenda, using RCS as a separate channel outside the AI service where users can make the final decision. RCS can verify that the company requesting approval is legitimate because only pre-registered and verified businesses can display their names and logos.

Messages can also include product images and an "Approve" button, allowing users to respond without opening another app. Approval requests and results remain in the conversation history of the smartphone's native messaging app for later review.

For example, when an AI agent requests a product payment, the user can review the order details and payment amount in an RCS message and approve the transaction. This preserves the convenience of having AI perform tasks while leaving the final transaction decision to the user.

SKT said the role of telecommunications in the AI era will expand to connecting people with AI agents and that it plans to use its connectivity and authentication capabilities to build "AI trust infrastructure."

SKT said it has operated its own RCS network and services while working to incorporate features needed in Korea into global standards. Rich Text Formatting, which SKT proposed at a Seoul meeting in May last year, was incorporated into RCS Universal Profile 4.0 earlier this year. The company said the latest meeting will expand discussions to the convergence of RCS and AI while seeking opportunities for technical exchanges and cooperation with global carriers, device makers and messaging providers.

OpenAI has acquired smartphone camera artificial intelligence (AI) and imaging startup Glass Imaging Inc. for more than $300 million, The Wall Street Journal reported on Sept. 14. Neither OpenAI nor Glass Imaging has officially confirmed the deal, while foreign media have viewed the acquisition as a further expansion of OpenAI’s push into consumer hardware.

Glass Imaging develops AI-based image processing technology designed to overcome the physical limitations of smartphones’ small image sensors and lenses. The acquisition is drawing attention because OpenAI is developing consumer devices with renowned former Apple designer Jony Ive.

Ive led the design of Apple products including the iPhone, iPad and iPod. OpenAI has been developing new AI devices with Ive and his design firm LoveFrom since the fall of 2023.

Ive founded io Products Inc. in 2024 with Scott Cannon, Evans Hankey and Tang Tan, and OpenAI entered the consumer hardware market by acquiring io Products in May last year. OpenAI announced an all-stock transaction valuing io Products at $6.5 billion and completed the acquisition in July of the same year.

Following the acquisition, 55 hardware, software and product development employees, including former Apple staff at io Products, joined OpenAI. No new AI device has been officially announced, but Bloomberg reported last month that the first device under development by OpenAI and Ive is a donut-shaped, screenless smart speaker, with a price of $300 to $400 under discussion.

The industry is paying attention to OpenAI’s move to acquire AI camera technology.

If AI agents are embedded in smartphones, glasses, earbuds or other personal devices, cameras can serve as key sensors that allow AI to perceive the physical world. Current smart glasses allow users to show AI particular objects or places through a camera, enabling the AI to recognize, explain or analyze them.

Glass Imaging Uses AI Trained on Optical Characteristics to Overcome Smartphone Camera Limits

Founded in 2019, Glass Imaging is headquartered in Los Altos, California, and has raised $30 million in external funding. The company was valued at $100 million in a funding round last year. The latest acquisition represents a threefold increase in valuation.

Its co-founders, Ziv Attar and Tom Bishop, are also former Apple engineers. Attar founded Israeli camera startup LinX Imaging, which Apple acquired in 2015 for $20 million, bringing Attar to Apple.

Foreign media have reported that the two co-founders led the team that developed Apple’s Portrait Mode. Apple’s Portrait Mode brought the so-called bokeh effect to smartphones by distinguishing a subject from the background and blurring the latter.

Bokeh is a photography term referring to blurred areas, points of light or out-of-focus regions in the background of a photo or video. The term derives from the Japanese word “boke,” meaning blur or haze. Computational photography is a digital imaging technology that uses software calculations, algorithms and AI to overcome the physical limitations of optical lenses and sensors and to synthesize and improve images. It enables thin devices with small lens apertures and sensors, such as smartphones, to produce images approaching the quality of large DSLR cameras.

Glass Imaging’s technology takes original RAW data captured by a camera sensor and processes the entire post-capture imaging pipeline through a neural network.

A key feature is that the neural network is trained in advance on the specific camera system used in each smartphone model. It learns the characteristics of the camera types, sensors, lenses and other components of a particular camera system.

GlassAI Trains Neural Networks on Individual Camera Modules

At the core of Glass Imaging’s technology is GlassAI, a neural network-based image signal processor (ISP) software. Unlike conventional ISPs, which process RAW data through multiple stages including demosaicing, noise reduction, sharpening and frame fusion, GlassAI uses an end-to-end neural network to process RAW data into red, green and blue (RGB) images.

GlassAI performs those processes at once while also modeling the optical characteristics of individual lenses. The neural network is trained to reflect the optical and sensor characteristics of each smartphone camera module, allowing it to produce images tailored to that particular camera system during capture.

While conventional AI photography technologies focus on improving the quality of images after they are captured, GlassAI learns the physical characteristics of camera lenses and sensors and uses a neural network for the entire process of converting RAW data into images.

According to SiliconANGLE, GlassAI measures the point spread function (PSF) for each lens-and-sensor combination, along with sensor response and noise profiles, and trains the neural network to address the shortcomings of that specific combination. A PSF mathematically describes how a perfect point of light spreads after passing through an optical system such as a camera, telescope or microscope. As a result, a change in smartphone model requires training and tuning for the specific camera module.

Glass Imaging says the resulting detail is not newly generated by AI but information captured by the sensor that would otherwise be lost during conventional ISP processing. The company also says the technology becomes more effective as pixel sizes decrease.

According to an interview Glass Imaging machine-learning engineer Shivansh Rao gave to photography publication PetaPixel, submicron pixels contain fine details that conventional approaches cannot properly decode. Tests showed that as pixel size decreased from 0.75 micrometers (μm) to 0.35μm, the neural-network approach improved MTF50, a resolution metric, by more than 50%.

The technology is being used for zoom functionality in the 600 series from Chinese smartphone maker Honor Device Co. The Honor 600 uses a 200-megapixel main camera to provide zoom without a dedicated telephoto camera. It uses a 16-in-1 Hexa Bayer configuration that combines 16 0.56μm pixels into the equivalent of a 2.24μm superpixel for regular photography.

The challenge arises when zoom is used. Cropping part of the sensor rather than using its entire area to magnify an image requires information from the small native 0.56μm pixels. As pixels become smaller, they capture less light and become relatively more susceptible to optical limitations, noise and blur. The structure is inherently disadvantaged compared with using a dedicated telephoto lens.

Glass Imaging uses neural networks to compensate for those limitations. GlassAI, trained in advance on the lens and sensor characteristics of a camera module, analyzes RAW data from the sensor and restores details lost in conventional image processing. Glass Imaging says this allows high-resolution main-camera pixels to deliver higher-quality zoom images without adding a separate telephoto camera.

Zoom remains an area where smartphone cameras face physical limitations. Smartphones, constrained by thickness and size, cannot accommodate long optical systems like those used in DSLRs. Even with multiple cameras, each has a limited focal length, while digital zoom results in image-quality degradation.

The Honor 600 uses GlassAI to correct the effects of optical aberrations and sensor noise during smartphone zoom photography and restore image detail. Glass Imaging’s technology uses AI to compensate for the physical limitations of small sensors and lenses, enabling higher-quality zoom images in compact devices.

DeepSeek has unveiled its latest artificial intelligence model, DeepSeek-V4.1-Flash. Although it has fewer parameters than DeepSeek's top-tier V4-Pro model, the new model showed advantages in performance, cost and speed that could make it a practical substitute for the flagship. It also outperformed V4-Pro on some agentic and coding benchmarks.

DeepSeek-V4.1-Flash sharply reduces not only the number of active parameters but also its key-value (KV) cache, while posting agentic and coding performance ahead of V4-Pro and, on some benchmarks, several frontier models. DeepSeek said Monday that it would begin consolidating its existing V4-Pro application programming interface (API) service into V4.1-Flash.

DeepSeek-V4.1-Flash uses a Mixture of Experts (MoE) architecture, which activates only selected expert networks depending on the input to improve computational efficiency. The model has 552 billion parameters in its backbone, with 8 billion active parameters during input processing and 16 billion during output generation.

The previous V4-Flash had 284 billion parameters and 13 billion active parameters. The top-end V4-Pro had 1.6 trillion parameters and 49 billion active parameters. Despite that difference, DeepSeek says V4.1-Flash delivers better performance than V4-Pro in agentic and coding tasks. The distinction between active parameters for input and output in DeepSeek-V4.1-Flash is particularly notable.

DeepSeek-V4.1-Flash is designed with a focus on agents and coding. According to DeepSeek, at the maximum reasoning intensity, or reasoning_effort=100, V4.1-Flash outperformed OpenAI's GPT-5.6 Sol and Anthropic's Claude Opus 5 on Terminal-Bench 2.1. DeepSeek also claims an advantage over the two models on DeepSWE v1.1 and AutomationBench.

Outside agentic and coding tasks, however, the model posted relatively lower scores on difficult general reasoning. Artificial Analysis (AA), an independent AI model evaluation platform, gave the model an Intelligence Index score of 40. That was well above the average for open-weight models but below the leading frontier models.

AA calculates its Intelligence Index by combining multiple evaluations, including GDPval-AA, AutomationBench-AA, SciCode, Humanity's Last Exam (HLE), CritPt and AA-LCR. DeepSeek-V4.1-Flash recorded scores including 1,632 on GDPval-AA v2, 69% on AutomationBench-AA, 52% on SciCode and 39% on HLE, leaving the model in the upper-middle range overall. The results indicate that the model can be highly efficient when deployed for coding-focused applications.

The score of 40 also reflects AA's recent changes to its evaluation methodology, which have lowered the overall scale of the index. AA upgraded its Intelligence Index from v4.2 to v4.3, making the evaluation set more difficult.

As a result, frontier models that previously scored well above 60 have been moved into the 50s under the revised methodology. Anthropic's Claude Fable 5.1, which had scored 66 under the previous v4.2 evaluation, now scores 53 under the revised system, alongside OpenAI's GPT-6 Astra.

AA's upgrade was driven in part by the replacement of Terminal-Bench 2.1 with Terminal-Bench 4.0. DeepSeek-V4.1-Flash outperformed GPT-5.6 Sol and Claude Opus 5 on Terminal-Bench 2.1, but fell behind them on Terminal-Bench 4.0.

Terminal-Bench 4.0 consists of 66 complex terminal tasks spanning software, machine learning, science, operations, security, hardware and media. It revised computing and time allowances and improved task descriptions, environments and verification logic to more finely distinguish performance among frontier models.

In terms of parameter size, DeepSeek-V4.1-Flash's Intelligence Index score is slightly below the 42 points recorded by GLM-5.3-Flash but is well above the median score of 18 for other open-weight models of a similar size. Under AA's stricter Intelligence Index, GLM-5.3-Flash ranks first among open-weight models, while Alibaba's Qwen 3.8 Max and DeepSeek-V4.1-Flash share second place at 40. DeepSeek V4 Pro 0813 scored 36.

DeepSeek-V4.1-Flash is also fast, as DeepSeek claims. Its output speed is about 214.4 tokens per second, while time to first token (TTFT) is 1.13 seconds. The median for comparable models is 65.4 tokens per second and 2.34 seconds for TTFT, making V4.1-Flash more than three times as fast by the output-speed measure. AA also rated it as “notably fast.”

The model has reduced its KV cache to 890 bytes per token. AA also viewed this as a significant strength. The reduction becomes more consequential when combined with DeepSeek-V4.1-Flash's pricing structure.

According to AA, DeepSeek-V4.1-Flash is priced competitively for input and output tokens, with an overall cost below the median of comparable models. Cache-hit pricing receives further reductions, while off-peak periods receive an additional 50% discount.

AA, however, said the model is “very verbose.” Its Intelligence Index evaluation used 250 million output tokens, substantially more than the 130 million median for similar open-weight models. The total cost of the evaluation was also recorded. Although the model is fast, the length of its responses can affect actual cost and completion time.

Still, AA said the model reaches the Pareto frontier on a cost-per-task basis. In other words, it is among the least expensive models at a comparable level of intelligence.

The reason DeepSeek-V4.1-Flash separates active parameters for input and output, unlike a conventional MoE model, lies in its Causal Encoder-Decoder (CED) architecture.

The term CED, or the concept behind it, was not created by DeepSeek. The term has been used since the late 2010s in fields including speech, time-series modeling and generative models. A paper on Alibaba's Qwen-Audio-VAE, published in July, also discusses CED. However, the CED used by DeepSeek in V4.1-Flash appears to be a different technology despite sharing the same name.

DeepSeek's CED makes the encoder causal rather than bidirectional and eliminates the need for cross-attention by projecting the encoder's final representation into the decoder's KV cache in a single operation. The structure appears to have been designed and named by DeepSeek for V4.1-Flash, and DeepSeek identifies it as a new architecture for its model in the official Hugging Face model card and technical documentation.

DeepSeek's CED in V4.1-Flash does three things: it changes the originally bidirectional encoder into a causal, sequential structure so it follows the same basic pretraining pattern used by today's autoregressive models; it eliminates the need for the decoder to query the encoder repeatedly through cross-attention by transferring the encoder's final output into the decoder cache in one operation; and it repurposes the architecture as a cost- and memory-saving mechanism inside a conversational chatbot.

According to DeepSeek's technical report, the 40 Transformer layers in V4.1-Flash are divided into 20 causal encoder layers and 20 decoder layers. Instead of generating global KV separately at each decoder layer, the model projects the decoder's global KV from the final hidden state of the encoder.

DeepSeek's CED architecture does not correspond directly to either the original Transformer encoder-decoder architecture or the decoder-only architecture used by most modern large language models (LLMs). According to DeepSeek's technical documentation, it differs in three key ways.

First, the encoder itself is causal. In the original encoder-decoder architecture, the encoder reads a sentence bidirectionally, looking both forward and backward. In DeepSeek's CED, the encoder also reads sequentially, from left to right. This is why it is called a causal encoder. Making the encoder causal in the same way as the decoder allows all 40 layers to operate under a consistent causal scheme, simplifying training.

Second, DeepSeek eliminates cross-attention, through which the decoder would otherwise query information from the encoder in real time at every layer and for every generated token. Cross-attention is an attention mechanism in which one part of a model references information held by another part rather than relying only on its own representations.

How DeepSeek CED Differs From Encoder-Decoder and Decoder-Only Models

Early Transformer-based models used an encoder-decoder architecture, while most generative large language models (LLMs) today use a decoder-only architecture. In an encoder-decoder model, dozens of layers are stacked sequentially. When a user submits a question or instruction, the input passes through the layers from the first to the last, becoming progressively more refined and processed. The process of understanding the user's prompt and the process of generating a response are handled through the same architecture, requiring the model to pass through all of those layers during processing. If a user attaches a very long document, or an AI agent performs a continuous sequence of tasks over an extended period as the conversation history keeps growing, the AI must repeatedly process that information through dozens of layers. Each layer leaves key-value (KV) cache states to retain important information. In general, the more layers a model has, the more KV states need to be stored. That is why KV cache is considered a major bottleneck and a significant source of rising costs in AI service operations. This traces back to the Transformer introduced by Google researchers in 2017, when machine translation was presented as a key application. In an English-to-Korean translation example, the architecture can broadly be divided into an encoder, which reads the entire English source sentence, and a decoder, which uses that understanding to generate the Korean translation. The key point is that the encoder reads the sentence bidirectionally. For example, when processing the fifth word of a sentence, the encoder can refer not only to the first four words but also to all words after the sixth. This is similar to a person reading an entire sentence before determining its context. Once the encoder has fully processed the source sentence in both directions, the decoder uses that representation to generate the translation one word at a time. The decoder operates only from left to right, or causally. To generate the next word, it can refer to words it has already produced, but cannot see words that have not yet been generated. When generating each word, however, the decoder does more than look at the words it has previously produced. It also references the representations generated by the encoder from the input tokens. This process is known as cross-attention. Crucially, cross-attention occurs at every decoder layer and each time a word is generated. In other words, whether at the fifth decoder layer or the 20th, the decoder repeatedly references information produced by the encoder in real time. This architecture is useful for tasks such as translation and summarization, where the input and output may be in different languages or have substantially different formats. Today's widely used models, including GPT and Claude, generally use a decoder-only architecture rather than the traditional encoder-decoder structure. The encoder and decoder are not separated. Instead, the user's question and the AI's response are combined into a single long sequence of text, which is processed by one set of layers entirely in a causal, left-to-right manner. The process of understanding the question and the process of generating the answer therefore share the same layers and parameters. The architecture is relatively simple and efficient to train, helping decoder-only models become an industry standard following the GPT series. Its drawback is that a long input must pass through all layers when it is initially processed to generate the KV cache, while the memory burden of storing and managing the cache increases as conversations grow longer. DeepSeek's Causal Encoder-Decoder (CED) architecture attempts to reduce this repeated processing. The basic idea is relatively straightforward: instead of treating all 40 workstations as a single processing line, the model divides them into two groups of 20 with distinct roles. The first 20 workstations, or encoder layers, process the input and produce the final hidden states. The user's question or document is read once and transformed into a representation that can be used by subsequent computations. The second 20 workstations, or decoder layers, generate the response based on that information and construct the KV states needed during generation.

Third, the roles of the encoder and decoder are changed. The model separates input understanding from response generation while continuing to operate as a single conversational chatbot from the user's perspective. Internally, the input-processing portion activates 8 billion parameters, while the output-generation portion activates 16 billion.

Early Transformer-based models used an encoder-decoder architecture, while most current generative LLMs use a decoder-only architecture.

SWA and DeepSeek’s SWA Bounded Replay

AI models broadly rely on two types of information when generating a response: global memory and local memory. Global memory maintains information across a long context. In DeepSeek's Causal Encoder-Decoder (CED) architecture, the encoder processes the entire input and uses its final hidden states to generate the decoder's global key-value (KV) cache. This global KV cache maintains an overall representation of the long context. But when a large language model (LLM) generates a response, it needs more than an understanding of the entire context. It also needs fine-grained information within a very narrow context, such as which word it has just used or how the preceding sentence ended. Information from the full long context alone may not adequately preserve such details, making a separate form of local memory necessary. Sliding Window Attention (SWA) is a mechanism that allows tokens within only a recent portion of the sequence to attend to one another. In DeepSeek's CED architecture, however, the decoder does not retain its SWA local state in full. DeepSeek therefore uses a separate mechanism called SWA Bounded Replay to compensate for the missing local state. In DeepSeek's CED architecture, the decoder does not repeatedly process the entire prompt as conventional decoder layers would. The first 20 encoder layers process the full prompt and produce the final hidden states, which are then used to construct the global KV cache for the following 20 decoder layers. The decoder can therefore access information from the entire input processed by the encoder through the global KV cache. That does not mean the model lacks local context information. Response generation still requires fine-grained information from a narrow context, such as the word just generated or the immediately preceding sentence. DeepSeek-V4.1-Flash addresses this by maintaining SWA state only for the most recent 128 tokens. Rather than continuously storing local KV on a solid-state drive (SSD), the model can reprocess the most recent 128 tokens when needed to reconstruct the local state. This is the role of SWA Bounded Replay. With SWA Bounded Replay, DeepSeek-V4.1-Flash performs a short replay process in which only the final 128 tokens of the prompt pass through the 20 decoder layers. This allows the decoder to recover local memory, or SWA state, for the immediately preceding context. The process is called “bounded” because the replay is limited to the most recent 128 tokens rather than the full prompt, and “replay” because the decoder recalculates the information to recover the recent context. The approach is a key part of DeepSeek's effort to improve cost efficiency. Under a conventional decoder architecture, preserving local memory would require an SWA cache containing recent-context information to be retained for the full length of the prompt. As conversations become longer and prompts grow, the amount of storage required also increases, eventually creating a need to move the cache to disk or an SSD. SWA Bounded Replay significantly reduces the need for such local-memory storage. When the local state is needed, just before response generation, the system can immediately recalculate the most recent 128 tokens, use the resulting state and then discard it. Instead of storing the SWA KV state on an SSD, the system reconstructs it by replaying the most recent 128 tokens when required. DeepSeek says this approach eliminates persistent SSD storage for SWA KV and, as a result, reduces the overall persistent KV cache footprint to about one-eighth that of the previous V4-Flash. While CED's reuse of global memory reduces computation and the amount of global cache that needs to be maintained, SWA Bounded Replay serves as a complementary mechanism that fills the structural gap in local memory at a small, fixed computational cost while also reducing storage requirements.

Another technology used to reduce the KV cache is △Compressed Sparse Attention 2 (CSA2), △along with FP4 quantization. CSA2 allows multiple layers to share and reuse KV and attention indices, reducing the duplication that would otherwise occur when each layer maintains its own cache. FP4 quantization stores the global KV cache using four-bit floating-point values, reducing storage requirements compared with FP8.

DeepSeek introduced the first version of Compressed Sparse Attention (CSA) in V4. While DeepSeek Sparse Attention (DSA) selects important KV entries and performs attention only on a subset of them in long contexts, CSA first compresses the KV representations of multiple tokens into shorter representations and then sparsely selects the information it needs. In other words, CSA combines KV compression with sparse selection to reduce both KV cache size and attention computation for long contexts.

V4.1-Flash advances the approach to CSA2. Each attention layer is assigned one of three modes — Full, Reindex or Reuse — allowing multiple attention layers to share and reuse KV and the locations of important information rather than recalculating the same information at every layer. Some layers perform new calculations while others reuse existing results, reducing the memory and computation required for long contexts. Combined with FP4 KV caching, the global KV cache is reduced to 890 bytes per token, about one-quarter the size of the V4-Flash cache.

Reducing the number of active parameters while maintaining performance lowers the GPU memory and computational resources required to run an AI model. That can reduce server operating and power costs and allow the same GPU resources to handle more simultaneous requests. Lower computational requirements can also reduce latency and make deployment on lower-end GPUs, on-premises servers rather than cloud infrastructure, and edge AI systems easier.

If total parameters can be reduced while maintaining the same performance, the model's weight memory can also be reduced, which is the ideal outcome. If that is not possible, reducing the active parameters in an MoE model is the next-best option because fewer active parameters also mean fewer computations.

DeepSeek-V4.1-Flash has a total of 552 billion parameters, which is substantial, but activates only 8 billion parameters for input and 16 billion for output. CSA2 also sharply reduces the KV cache, allowing GPU memory to be used more efficiently. The fact that DeepSeek developed and actively deployed technologies such as CSA2 and CED in the model is particularly notable.

DeepSeek is not alone. Chinese AI companies including Moonshot AI, developer of the Kimi family, and Alibaba, developer of Qwen, are expanding model parameter counts or advancing their models through proprietary technologies. Questions remain, however, over allegations that some Chinese AI companies illegally distilled capabilities from U.S. frontier models.

Anthropic recently said that seven Chinese AI companies had attempted to extract capabilities from its Claude models without authorization, and disclosed cases involving Alibaba, Moonshot AI and DeepSeek.

According to Anthropic's findings, Alibaba linked more than 151 million interactions with Claude to its own models. Anthropic said Alibaba created large numbers of accounts to send queries to Claude and collect its responses, then used the data to train or improve its own AI models. In effect, Claude's responses were used as teacher-model data.

In Moonshot AI's case, more than 23 million interactions were identified. Particularly problematic, according to Anthropic, was that Moonshot forwarded questions entered by actual Kimi users to Claude and then displayed Claude's responses to those users as if they had been generated by Kimi. Anthropic also said Moonshot subsequently used the collected user questions and Claude responses in developing its own models.

For DeepSeek, Anthropic said it identified about 12.1 million interactions. Anthropic alleged that DeepSeek also sent large numbers of queries to Claude to collect responses and used them for model distillation.

So, with some trepidation, I report that Lex Friedman has brought Strategery back from the dead. There are some clever updates, with a better visual design and even the addition of chaos mode, but I must admit I immediately set the look back to the original. Be careful with this one.

The post Strategery Returns to iPhone and iPad appeared first on MacSparky.

I’ve made my choice. I’m buying the iPhone Duo instead of the iPhone 18 Pro Max.

Here’s my full explanation and my read on the tradeoffs.

The post Why I’m Buying the iPhone Duo Instead of the iPhone 18 Pro Max appeared first on MacSparky.

On this episode of Mac Power Users, Stephen and I break down Apple’s September event, starting with the “Ternus Doctrine” and what it means for Siri AI. We compare the iPhone Duo and iPhone 18 Pro, then dig into the new Apple Watch and AirPods features. In More Power Users, we share what we ordered and why.

Episode Links

• Everything Apple Just Announced – Stephen’s YouTube Recap
• Apple Event September 9 2026: Introducing iPhone Duo and more
• Apple Plans to Squeeze More Revenue From the App Store, Report Says
• Apple unveils iPhone Duo
• I Tested Apple’s New FOLDABLE iPhone Duo
• iPhone Duo MagSafe Accessory Photo
• AppleCare One Family
• Apple debuts iPhone 18 Pro and iPhone 18 Pro Max
• iPhone 18 Pro & Duo: A Photographer’s Reaction
• About Audio Intelligence features on Apple Watch Series 12 and Apple Watch Ultra 4
• Introducing Apple Watch Series 12, with the all-new Health Sensing System

This episode of Mac Power Users is sponsored by:

• NetSuite: The leading integrated cloud business software suite. Try NetSuite Next for free.
• Squarespace: Save 50% off your first purchase of a website or domain using code MPU

The post Mac Power Users 866: The Ternus Doctrine appeared first on MacSparky.

This past week, Apple took the wraps off the new iPhone Pro and the brand-new iPhone Duo. It’s a big deal, and a lot of ink has already been spilled over it. But something else happened that I’d like to shed more light on.

John Ternus gave his first keynote as Apple CEO. And his opening remarks in that job carry real weight. He chose to use that moment to make his case for Apple’s place in AI. (Yes, he called it AI.) It starts at the 3:32 mark in Apple’s YouTube keynote and it was the clearest statement of intent we’ve heard from Apple on the subject.

His words:

AI makes entirely new kinds of experiences possible. And AI can become even more useful when it brings together your life with the apps, services, and products you use every day. And this means the product at the center of your experiences becomes even more essential. What I like to think of as an intelligent personal hub.

It’s the intersection of broad new capability and a deep understanding of your personal context that makes this idea so powerful.

There’s no product in the world better designed to be your intelligent personal hub than iPhone.

As your personal hub, your iPhone can do so much more for you because it understands your context. This includes your calendar, your relationships, your routines, the most private details of your daily life.

It’s personal intelligence that’s actually personal.

He described the iPhone as an “intelligent personal hub.” His argument was pretty straightforward. If you were building the ideal device for personal AI, you’d want something you carry everywhere, with enough processing power to run models locally and a connection to more capable models when needed. You’d want cameras and microphones so it could understand what’s going on around you. And you’d want a screen. There are still plenty of things I’d rather look at than have a robot read to me.

You’d end up with something a lot like an iPhone.

The other half of his argument was about context. A capable model gets much more useful when it knows something about you. Your calendar, your relationships, and your routines are already on your phone. With your permission, that information could help it do things that a chatbot starting from scratch can’t.

Of course, that’s also the information you’d be least comfortable handing over to some company’s servers. Ternus tied Apple’s approach directly to that concern. Run the models on the device whenever possible. Use Private Cloud Compute when the job needs more power. His promise was “personal intelligence that’s actually personal.”

For lack of a better term, I’m calling this the Ternus Doctrine.

I’ve been making versions of this argument for a while. Apple’s access to local context has always seemed like a big advantage to me. But hearing Ternus put it together at the opening of the keynote (his first as CEO) felt different. This was the position he chose to stake out before getting to the new hardware.

Apple has often seemed reluctant to get serious about AI. I know they’ve been working on it, and I’ve written before about why their approach could work. But there’s a limit to how long I can be excited about what Siri might eventually do. At some point, I need it to do the thing.

If we start doing more of our computing by asking an assistant to do work for us, whoever makes that assistant will have a lot of influence over how we use our devices. If I spend my day working through somebody else’s AI, Apple still gets to sell me a phone. But somebody else is deciding how much of that phone I use and which services I reach for.

I don’t think Apple wants that future.

Ternus’s argument gives Apple a different path. Make the intelligence built into the phone useful enough that it’s the one we reach for. Give it the context it needs, while letting us control what it can access. And make it work across the devices and apps we already use.

If he follows through, I’d expect more attention to what Apple silicon can run locally. I’d also expect Apple to keep building out Private Cloud Compute. Those are the parts of the plan where Apple controls the hardware and can keep improving it.

Siri is the harder test. I’m on the betas and the new Siri AI is light-years ahead of the old Siri. But it’s still unclear how far Apple is willing to take this. If they want Siri to become the brain of our private contextual AI assistant, they need to keep the gas down. They also need developer buy-in. Apps need to plug into it.

I’m speculating about where Ternus takes this. An opening speech doesn’t tell us how he’ll assign people or spend money, and Apple has given us plenty of reasons to be skeptical of AI promises.

I liked hearing Ternus explain where he thinks Apple belongs. Now I’d like to be able to ask my iPhone to do something complicated and trust it to finish the job.

The post The Ternus Doctrine appeared first on MacSparky.

There is a lot to process after this week’s Apple event. If you’re still sorting through it all, I found two videos worth your time.

First is Marques Brownlee’s iPhone 18 Pro/Duo Impressions: Mogged. Marques has a good feel for this stuff, and I always value his initial reactions. I don’t agree with him about the Apple Watch recording feature. He lands closer to the creepy side of that debate. I don’t think it is nearly as troubling as some people are making it out to be, although we’ll know more once people start using it in the wild.

The second is Tyler Stalman’s iPhone 18 Pro & Duo: A Photographer’s Reaction. Tyler set me straight on the variable aperture. The mechanism isn’t there to create more bokeh. At its widest setting, the lens is already as open as it gets. The clever bit is that Apple can put a faster lens in the phone, then stop it down when conditions call for less light or more depth of field.

https://www.youtube.com/watch?v=Zt-fS0DH8BQ

That makes the variable aperture less of a single flashy feature and more of a tool that lets Apple build a more flexible camera. Either way, if you like taking pictures with your phone, I think the iPhone 18 Pro is going to be a really nice upgrade.

The post Two Apple Event Videos Worth Watching appeared first on MacSparky.

It was a momentous week with the iPhone announcement. I’ve got a lot on my mind.

Below you’ll find the week’s new videos and podcasts, a few of the conversations happening in the member community, and what’s coming next week. Most of those links lead straight into the Labs community on Circle, so they’ll only open if you’re a member. If you’re a Labs member who hasn’t set up your Circle access yet, send me a note and I’ll get you in. For members, this is your week at a glance. If you’re not one yet, it’s a peek at what you’re missing.

---

The Week’s Releases

Touring the Endor Fieldbook

Video · September 8 · all tiers

I built myself a native Mac app with an audience of one. Endor Fieldbook reads the same Markdown files as my Obsidian vault, then puts my calendar, tasks, content, research, and finances into an interface made just for me.

Pathfinder Link | Insider Link | Member Link

Labs iPhone Event Meetup

Meetup · September 9 · all tiers

We watched the September event together in the Labs chat room, then got on Zoom at noon to compare notes on the iPhone 18 Pro and the Duo. The recording is already published in the Labs Circle and podcast feeds.

Pathfinder Link | Insider Link | Member Link

Labs September Meetup (After Dark)

Meetup · September 10 · Pathfinder & Insider

This month’s After Dark meetup ran Thursday evening at 6:00 Pacific, the late slot for members who cannot make a noon session. Recording to follow.

Pathfinder Link | Insider Link

The Lab Report for September 11, 2026

Podcast · September 11 · all tiers

Apple’s September event takes over this week’s Lab Report. I look at the iPhone 18 Pro camera, explain why I’m buying an iPhone Duo, and work through the tradeoffs between Apple’s two new phones.

Pathfinder Link | Insider Link | Member Link

In the Community

Lucas Museum of Narrative Art (Labs Discussion Board)

A member visited the Lucas Museum of Narrative Art in Los Angeles during a preview and came back recommending it, with the Geffen nearby as a second stop.

iPhone Duo (Labs Discussion Board)

A member posted a picture of the rumored dual-screen iPhone, and the comments turned into a run of jokes about its resemblance to a discontinued Microsoft Duo.

Best Way to Access Mac from iPad? (Labs Discussion Board)

One member asked how to reach a MacBook Pro back home while traveling with only an iPad, and the replies covered several remote-access options.

If you’re a Labs member and haven’t joined the Circle community yet, let me know and I’ll get you set up.

Coming Up

The event is behind us, so here is what the Labs calendar looks like through the middle of October.

September Deep Dive – The Mac Studio Home Server · September 17, 12:00 Pacific

Pathfinder Link

Labs Jam Session – Browsers · September 24, 12:00 Pacific

Pathfinder Link | Insider Link | Member Link

Q4 Quarter Planning Meetup · October 1, 10:00 Pacific

Pathfinder Link

Labs October Meetup · October 8, 14:00 Pacific

Pathfinder Link | Insider Link

October Deep Dive – Local and Private AI on a Mac · October 15, 12:00 Pacific

Pathfinder Link

Links of Interest

• iPhone 18 Pro/Duo Impressions: Mogged. Marques Brownlee shares his first impressions after hands-on time with both new iPhones.
• iPhone Duo hands-on. The Verge takes an early look at Apple’s first foldable iPhone.
• Is the Foldable iPhone Duo Worth $1,999?. Joanna Stern looks at the first-generation hardware and asks the obvious price question.
• Apple reportedly cancels two MacBooks and reshuffles its OLED roadmap. MacRumors on the display roadmap that says a foldable MacBook and a large-OLED MacBook Pro tier are both off the table.
• Apple hit with a $2.7 billion lawsuit over App Tracking Transparency. The UK claim argues Apple held other developers to stricter tracking rules than its own advertising business.
• Claude, ChatGPT, and Grok all went down at once. Worth reading if you have moved a daily workflow onto a single cloud AI service.

---

Join the Labs

The videos, podcasts, and live sessions above are inside MacSparky Labs. Join at macsparky.com/join.

The post This Week in the Labs — September 11, 2026 appeared first on MacSparky.

Nature Geoscience, Published online: 11 September 2026; doi:10.1038/s41561-026-02099-7

Rapid caldera collapse during submarine eruptions could increase tsunami magnitude, suggests an analysis of the geomorphological evolution of the caldera formed in the climactic eruption of Hunga volcano (Tonga) in 2022.

Artificial intelligence (AI) services and solutions provider Bespin Global has won a project to upgrade "ChaTi," a conversational information search service operated by the Korea Agro-Fisheries & Food Trade Corp. (aT). The project will revamp the conversational search service within KATI, aT's specialized platform for agricultural and food export information, using the latest technologies.

Bespin Global said it will use its HelpNow AI platform to build a next-generation intelligent assistant based on agentic AI that analyzes user requirements and executes tasks. The goal is to help exporters access export information more conveniently and make faster, more informed decisions.

The project comes as the expansion of K-food exports increases the amount of information exporters need to regularly check, including country-specific quarantine requirements and labeling regulations. K-food exports reached a record $13.62 billion last year, of which agricultural and food products accounted for $10.41 billion. The government has set an export target of $16 billion for this year.

Bespin Global plans to improve ChaTi's answer accuracy by upgrading its latest large language model (LLM) and retrieval-augmented generation (RAG) architecture, expand training on internal data including export statistics databases and non-tariff barriers, and introduce an optical character recognition (OCR)-based system for extracting unstructured data from PDFs, tables and images.

In particular, the company will apply multi-agent orchestration technology that enables specialized agents in areas including market analysis, regulatory assessment and data analysis to collaborate on complex queries.

An incident in which OpenAI artificial intelligence (AI) agents circumvented developer-imposed controls and effectively took over a German wiki has become the subject of regulatory scrutiny by the European Union.

The European Commission (EC) has received an incident report from OpenAI and is reviewing the matter. Key questions include when OpenAI became aware of the agents’ abnormal behavior and whether it should have classified and promptly reported the event as a “serious incident” under the EU AI Act.

If OpenAI is found to have known that the incident was reportable but unjustifiably delayed reporting it, the company could face a fine of up to 3% of its global annual revenue.

The case has escalated into an EU regulatory issue because of the significant gap between when OpenAI may have learned of the incident and when it disclosed it publicly. According to Reuters, anonymous OpenAI employees said company executives knew about the incident weeks before Reuters first reported it.

Researchers at AI safety nonprofit Nightingale, which first uncovered the incident, also raised the possibility that OpenAI became aware of the situation around June 21. According to the researchers, AI agent activity on DSEwiki dropped sharply from June 21, shortly after a visit that appeared to be linked to OpenAI.

Reuters first reported the OpenAI agents’ activity on DSEwiki on Sept. 4.

From May through July this year, OpenAI’s AI agents made more than 15,000 edits to DSEwiki, a German wiki for programmers, using names such as “OpenAIResearcher” and “OAIResearchMar26.” Posts included answers to tasks being performed by the agents, methods for bypassing OpenAI safeguards and ways to evade monitoring.

DSEwiki is a privately operated site created about 25 years ago that had been largely abandoned for years. When its administrator belatedly discovered the unusual activity and began deleting posts, some agents identified the alphabetical deletion pattern and responded by creating backup pages beginning with the letter “Z.”

OpenAI maintains that the event was a “misalignment,” rather than a security incident. The company said security incidents go through formal reporting procedures, while misalignment is disclosed through system cards or research papers, and that there is no industry standard for separately reporting this type of AI behavior.

The first issue for the EC, therefore, is whether the DSEwiki incident qualifies as a “serious incident” under the EU AI Act.

Article 3(49) of the EU AI Act defines a serious incident as an incident or malfunction of an AI system that directly or indirectly results in death or serious harm to health, a serious and irreversible disruption to the management or operation of critical infrastructure, infringement of obligations under EU law intended to protect fundamental rights, or serious harm to property or the environment.

The second issue is the legal status of the AI model used by the agents. The EU AI Act exempts some general-purpose AI models that are in research, development or prototyping before being placed on the market from certain regulatory requirements.

The third issue is when OpenAI became aware of the incident and when it reported it. The EU AI Act requires providers of general-purpose AI models with systemic risk to report serious incidents “without undue delay.” It does not specify a fixed reporting deadline.

Based on currently available information, it is unclear exactly when OpenAI’s internal monitoring systems detected the incident.

Key EU obligations for providers of general-purpose AI models took effect on Aug. 2, 2025, while the EC’s related enforcement powers became fully applicable on Aug. 2 this year. General-purpose AI model providers are required to prepare technical documentation, publish summaries of training data and take measures to comply with copyright law.

The maximum fine for violations or the provision of incomplete information is 3% of global annual revenue or €15 million, whichever is higher. Prohibited AI practices can carry fines of up to 7% of global annual revenue or €35 million.

Political pressure could also increase in the US. Following the Hugging Face hacking incident in July, two US lawmakers asked OpenAI whether there had been additional similar cases, but the company reportedly declined to answer.

Beyond the regulatory questions, the DSEwiki incident also exposed concerns over the controllability of AI agents.

The agents were performing web retrieval tasks at the time. They were permitted to read information from the internet, but operated in an environment where the ability to write data was blocked. The agents, however, discovered that DSEwiki’s outdated software allowed edits through ordinary read requests, enabling them to circumvent the restriction on writing data.

The agents had not been instructed to communicate or collaborate with one another, but they used DSEwiki to exchange messages. When the administrator deleted posts, they analyzed the deletion pattern and created backup pages, adapting their behavior in response to human intervention.

Taken together, the AI agents circumvented restrictions explicitly imposed by developers, used the channel they gained to share information with other agents and changed their behavioral strategies in response to actions by a human administrator. In effect, they moved beyond human control.

Enterprise software support provider Rimini Street said Sept. 9 that healthcare company Ilhwa has adopted Rimini Support, its maintenance service for SAP S/4HANA systems.

Ilhwa plans to reduce maintenance costs through Rimini Street’s SAP S/4HANA technical support and use the IT budget savings to develop reports within its SAP system, fund business improvement projects and deploy new hardware.

Ilhwa, which operates mainly in the food and beverage and pharmaceutical businesses, has integrated key companywide processes including finance, sales, production and inventory around SAP S/4HANA. The system has also been customized to reflect the requirements of its various business units, making stable enterprise resource planning (ERP) operations and efficient maintenance important.

Rimini Street said Ilhwa needed a way to reduce its high annual maintenance costs while retaining its existing SAP S/4HANA environment and use its limited IT budget more strategically. Ilhwa also found that its previous maintenance arrangement fell short in resolving and responding to issues relative to its high cost.

Rimini Street provides Rimini Support L4 services for Ilhwa’s SAP S/4HANA environment, with dedicated engineers providing customized technical support. The company said Ilhwa has reduced its previous SAP maintenance costs by more than 50% since switching to its service.

Rimini Street also provides analysis reports on Ilhwa’s SAP system to improve operational efficiency. Ilhwa has reinvested part of the savings in small-scale business improvement projects, giving it greater flexibility to respond quickly to the needs of business units.

Ilhwa also cited having a single point of contact for maintenance support as a benefit. The arrangement allows it to resolve complex technical issues more quickly through dedicated Rimini Street engineers without having to go through multiple support channels.

Meta has announced its answer to OpenClaw. The company on Sept. 8 launched Muse, an agentic artificial intelligence personal assistant, in the U.S.

Muse is initially available through a dedicated app and WhatsApp. It offers a free basic tier as well as paid subscriptions costing $20 or $100 a month depending on usage.

Reports that Meta was developing an agentic AI personal assistant similar to OpenClaw emerged in May. The Information reported that Meta was training a project code-named Hatch using Anthropic's Claude 4.6 model and planned to switch to its own Muse Spark model for the commercial launch.

According to Meta, Muse can automate tasks ranging from online shopping and buying movie tickets to booking appointments such as tennis lessons and even filling out permission forms for children's school field trips. It is designed to access apps across categories including email, calendars, payments, health, shopping and smart homes.

Muse can send emails, book travel, reduce utility bills, fill out forms, make plans, turn recipe videos into grocery lists, send party invitations and make payments. Payments are handled through Link, a service from U.S. fintech company Stripe.

Stripe Link is a digital wallet that simplifies online purchases using stored payment information and supports payments made by AI agents on behalf of users. Its agent wallet generates one-time card numbers so actual card details are not exposed to merchants. Muse also receives Link's purchase protection.

How OpenClaw and Meta Muse Differ

Muse offers many features similar to OpenClaw, which served as a model for its development. One of OpenClaw's key strengths is its ability to carry out tasks when users simply send it a message through chat apps such as WhatsApp or Telegram. Muse implements a similar approach through its dedicated app and WhatsApp.

Another similarity is its always-on capability. Muse runs independently on a dedicated virtual machine (VM) in the cloud regardless of whether the user's computer is turned on. This provides the same always-on capability that allows OpenClaw to independently perform monitoring and scheduled tasks even when a user's computer is off.

Muse also resembles OpenClaw in the breadth of its integrations and autonomous actions. It can send emails, book travel, make payments and even negotiate. Sensitive actions such as sending emails or making purchases require user approval, but other tasks can be completed autonomously without requiring human confirmation at every step. That mirrors OpenClaw's ability to complete an entire workflow from a single message.

OpenClaw, however, is open source, allowing users to install it on their own servers and connect models of their choice. Muse is a commercial cloud service that runs only on Meta's Muse Spark model. The biggest difference is security.

OpenClaw requires users to manage security themselves and has faced numerous concerns over vulnerabilities. Muse, by contrast, incorporates privacy and security safeguards from the design stage, including isolated secure VMs and separate connection approvals for individual apps. In that respect, it is similar to Nvidia's NemoClaw, which combines OpenClaw with a security framework.

NemoClaw for Enterprises and Developers, Muse for Consumers

Nvidia's NemoClaw is an open-source security and deployment stack designed to run OpenClaw securely. OpenClaw serves as the brain, while NemoClaw acts as protective headgear around it.

NemoClaw uses a sandbox runtime called OpenShell to block and monitor access to file systems, networks and processes, preventing OpenClaw agents from communicating externally without authorization. It also incorporates security and privacy policies, network access controls, credential management, model routing, agent lifecycle management and observability.

Nvidia's own language models, including Nemotron, can also run directly on a user's PC in a local environment. This allows inference and execution without sending data to the cloud. The configuration can be version-controlled through YAML policy files, allowing multiple developers to reproduce the same security settings.

Muse and NemoClaw share the concept of long-running agents that autonomously perform tasks. Both make isolated execution environments that prevent user data from being exposed externally a core design principle — cloud VMs for Muse and local sandboxes for NemoClaw.

Both approaches give AI agents permission to operate web browsers and apps while establishing guardrails to prevent unauthorized access to the broader system or sensitive information.

OpenClaw and NemoClaw target developers and users with some AI and IT expertise, while Meta's Muse can be seen as extending the concept to general consumers.

Always-On Operation and Long-Term Memory Key to Autonomous AI Assistants

Agentic personal AI assistants require computer-use technology, always-on operation and long-term memory, as well as broad web and app support.

Computer use is a fundamental technology for agentic personal AI assistants. It allows AI to perceive a screen and operate a computer like a person, moving the cursor, clicking buttons, entering text and navigating websites and apps. It effectively provides the AI with eyes and hands, with the advantage of not requiring separate integrations for every app.

While computer use is essential, always-on operation is even more important. Users need to be able to issue instructions and interact with the assistant at any time, while continuously pursuing long-term goals that may span several days requires a separate orchestration layer.

The assistant must also remember and preserve users' work and include dedicated connectors, or application programming interfaces (APIs), for functions such as payments and calendars to operate reliably in real-world tasks.

An agentic personal AI assistant therefore needs more than computer-use capabilities. It requires an always-available entry point through messaging services, long-running background execution that continues when the user is away, persistent memory of previous conversations and preferences, and an ecosystem of connectors spanning services such as email, payments and calendars.

By those criteria, services that can currently be considered agentic personal AI assistants are largely limited to OpenClaw, Meta Muse and Perplexity's Computer.

OpenAI and Anthropic, two leading AI research companies, do not yet offer fully autonomous personal AI assistants. Both companies are primarily focused on enterprises and developers.

Google continues to add agentic AI assistant features to its Gemini app, but its strengths remain largely within Google's ecosystem, including Gmail, Calendar and Docs. Microsoft Copilot is similar, with a focus on work environments built around Microsoft 365 applications such as Outlook, Word and Excel.

As adoption of autonomous artificial intelligence, or agentic AI, and agentic coding accelerates, the industry is confronting a paradox: per-token costs are falling while total AI spending is surging.

Unlike one-off chatbot conversations, agentic AI autonomously repeats reasoning, planning, tool calling, verification and error-correction loops. Estimates suggest agentic AI could consume five to 30 times as many tokens as a single query, and as much as 1,000 times more for complex tasks.

Recently released AI models are increasingly promoting token efficiency as a competitive advantage. Beyond lowering the cost per token, developers are emphasizing technologies that reduce the number of tokens used for reasoning itself. Prompt caching, batch application programming interfaces (APIs) and dynamic tool loading have been used to cut token-processing costs and eliminate unnecessary context.

More advanced approaches to token optimization are now emerging. The most prominent is latent-space reasoning, along with more aggressive approaches such as Soft Latent Thinking and Selective Latent Thinking, and the Full-Bandwidth Transformer, which seeks greater token efficiency while retaining the existing architecture.

Token Compression Does Not Eliminate Inefficiency

Competition among large language models (LLMs) is shifting from more parameters toward more efficient reasoning. The ability to solve the same problem with fewer tokens is becoming a new competitive metric.

Models with stronger reasoning capabilities tend to generate more reasoning tokens before producing an answer. Longer reasoning can improve accuracy, but generating more tokens increases decoding operations and graphics processing unit (GPU) computation. It also expands the KV cache used to store information from previous tokens. Performing the same reasoning with fewer tokens therefore directly affects AI model costs and latency.

Heima, or Efficient Reasoning with Hidden Thinking, presented at the International Conference on Machine Learning (ICML) in April, approaches the problem through compression. It reduces the number of tokens used for reasoning by compressing long chains of thought (CoT) or eliminating unnecessary reasoning steps.

Heima sharply reduces the length of "thinking tokens" used internally by a model to analyze and reason before generating its final answer. It compresses lengthy, complex reasoning into a handful of core tokens, which can be converted back into human-readable text by a dedicated Adaptive Interpreter when needed. The approach preserves how the model thinks while reducing how extensively that thinking is expressed in language.

Token compression can immediately reduce API costs without changing the underlying model architecture. But even with highly effective compression, reasoning is still expressed as natural-language tokens.

No matter how complex the model's internal calculations are, information passed to the next step must ultimately travel through the narrow channel of a single selected token. This is an inherent limitation of autoregressive decoding.

Autoregressive decoding is how transformer-based LLMs generate text sequentially, one token at a time, using previously generated tokens as inputs for the next step. Because the model continuously references earlier context, it can maintain coherent flow, tone and grammar even across long passages.

The approach, however, has clear limitations. During reasoning, each new token depends on previously generated tokens, making token generation inherently sequential and difficult to parallelize.

An incorrect token generated midway can also cause errors to accumulate through subsequent computations, distorting the context. Each new token must pass through the model's large neural network layers, adding computational costs and memory bottlenecks.

Latent Reasoning Moves Beyond Autoregressive Decoding

The limitation stems from a gap between how models think and how they transmit information.

Transformers process input tokens through multiple layers and transform them into high-dimensional hidden states containing thousands of dimensions. A hidden state represents information exchanged internally between model layers and is not exposed outside the model. It can be compared with an intermediate note the AI keeps internally before expressing an answer.

These hidden states contain not only context and meaning but also rich latent information such as incomplete reasoning, partial answers and future plans that have yet to be fully expressed as words. The problem is that this high-dimensional information is not passed directly into the next computational step.

Instead, the model uses the hidden state to calculate a probability distribution for the next token, selects one token and uses it as the input for the next step. The hidden state itself is not directly carried forward.

As a result, even sophisticated internal computations must pass through an extremely narrow bottleneck of a single word. Rich high-dimensional information represented as continuous vectors is compressed into a language token and then converted back into a vector for the next calculation, creating repeated information loss and computational inefficiency.

The problem goes further. Every time the model selects the next token, it projects the hidden state across the entire vocabulary to calculate probabilities for a large number of candidate tokens. This process, known as LM Head Projection, imposes a substantial computational burden.

Adding a single word therefore requires the model to repeatedly perform a large-scale probability calculation across tens or hundreds of thousands of vocabulary candidates. Reducing the number of tokens does not eliminate this fundamental computational structure.

Latent reasoning, which has recently gained attention as a token-optimization technology, seeks to fundamentally change part of this cost structure.

Does Intermediate Reasoning Need to Be Expressed in Natural Language?

Latent reasoning performs intermediate reasoning directly within the model's continuous internal state, rather than expressing it in natural language. In human terms, instead of verbalizing every intermediate thought before reaching a conclusion, the model quietly works through the problem internally.

Traditional reasoning incurs costs by converting internal thoughts into language tokens and then decoding those tokens back into representations for further computation. Latent reasoning instead completes intermediate calculations within the model's hidden numerical space, or latent state, without generating each thought as human-readable natural language.

One example pushing this concept further is the BDH-CQ model unveiled last month by Pathway Research, part of Pathway. Pathway develops post-transformer architectures and efficient AI reasoning technologies aimed at overcoming limitations of transformer models.

BDH-CQ has only 150 million parameters but attracted industry attention for its performance relative to its size on ARC-AGI-1, a well-known benchmark for difficult AI reasoning tasks.

ARC-AGI-1 evaluates a model's ability to infer and solve previously unseen spatial grid-pattern rules. It tests whether a model can independently construct geometric and spatial logic, similar to a child figuring out the rules of a new toy.

While conventional AI benchmarks often measure memorized knowledge or language understanding, ARC-AGI tests how quickly a model can identify and apply unfamiliar rules without prior knowledge. Models get only two attempts, or Pass@2, and even a single incorrect pixel results in a wrong answer.

BDH-CQ scored 29.5% under Pass@2. Frontier models perform far better under the same conditions, with OpenAI's GPT-6 Astra at 98.5%, Anthropic's Claude Opus 5 and Fable 5.1 at 97.5%, and Google Gemini 3.7 Flash at 95.5%.

BDH-CQ's 29.5% appears low by comparison, but its parameter count is far smaller. While parameter counts for frontier models are undisclosed, they are estimated at at least 1 trillion and are believed to be around 2 trillion to 5 trillion. BDH-CQ achieved nearly 30% with roughly one ten-thousandth as many parameters. Its compute cost per task was also just $0.0007.

150-Million-Parameter BDH-CQ Nears ARC-AGI-1 Pareto Frontier

The potential impact could be significant if Pathway Research can further improve its latent reasoning-based BDH-CQ model.

BDH-CQ does not output intermediate reasoning in human language tokens. Instead, it repeatedly updates its recurrent memory and performs computations entirely within its internal latent-state space.

For AI developers, reducing or eliminating the process of converting internal thoughts into language — projecting hidden states across the vocabulary through the LM head, selecting the next token and embedding that token again for the next reasoning step — could enable comparable performance with less computation. That could ultimately deliver high performance efficiency even in smaller models.

COCONUT demonstrated the possibility that AI can reason without expressing thoughts in language, while BDH-CQ extended the concept toward smaller models and lower inference costs.

Both approaches reduce the role of language tokens in AI reasoning and directly use the model's continuous internal states. Researchers are now going a step further by exploring whether language generation itself can be removed while an AI model is thinking. One such approach is Soft Latent Thinking.

A paper titled "A Model with No Head and Many Thoughts" appeared on arXiv on Aug. 31. Its authors are researchers from Russian internet and technology company Yandex and Russian financial technology company T-Tech.

The Soft Latent Thinking method described in the paper removes the LM head, the language model's output mechanism, during reasoning and replaces it with a lightweight projector. The projector transforms hidden states into continuous representations in embedding space and passes them directly into the next reasoning step.

Soft Latent Thinking Can Be Applied to Existing LLMs

Conventional language models repeatedly convert hidden states representing their internal thoughts into actual language tokens during intermediate reasoning. The model calculates probabilities across a large vocabulary containing tens of thousands of token candidates, selects the next token and uses it as the input for the next reasoning step.

The mechanism that converts hidden states into vocabulary-space probabilities is called the Language Model Head, or LM head.

Soft Latent Thinking eliminates this intermediate language-generation process. A much lighter projector directly connects the current latent state to the next one. The model continuously updates its vector state without generating language tokens while reasoning, converting it back into language only when producing the final answer.

BDH-CQ builds recurrent latent reasoning into a small 150-million-parameter model, while Soft Latent Thinking differs in that it can also be applied to existing LLMs.

Selective Latent Thinking

Soft Latent Thinking reduces computation by eliminating language-token generation during reasoning and processing the entire thought process in latent space. But that raises another question: Is compressing every reasoning step into a latent state always optimal?

Compressing stages that require detailed intermediate reasoning, such as complex mathematical calculations or precise logical deduction, could instead reduce accuracy.

Research is therefore expanding toward approaches that distinguish between reasoning segments that can be compressed and those that should remain in language, rather than converting the entire reasoning process into latent representations. Selective Latent Thinking is a representative approach.

Selective Latent Thinking predicts upcoming short reasoning segments and uses the model's confidence level to determine whether they should be compressed into latent representations. Segments that can be safely compressed are skipped through latent states, while uncertain sections or those requiring precise reasoning retain explicit CoT. The idea is that not every thought needs to be expressed in language, nor does every thought need to be moved into latent space.

A Selective Latent Thinking paper published in May tested the approach using Llama-3.2-1B and Qwen3-4B across four mathematical reasoning benchmarks.

The researchers reduced reasoning-chain length by 58.4% compared with explicit CoT while limiting the decline in accuracy to 2.8%. In other words, a reasoning process equivalent to 100 tokens could be simplified to about 42 tokens while largely maintaining performance.

The researchers also found that Selective Latent Thinking improved the balance between accuracy and efficiency compared with conventional latent reasoning. At similar levels of reasoning compression, it achieved 22.7% higher accuracy than existing latent reasoning methods.

Microsoft Studies Full-Bandwidth Transformer

Microsoft AI Frontier researchers' Full-Bandwidth Transformer retains generated tokens as in conventional models but also passes the previous step's hidden state into the next step, increasing the amount of information available during the same reasoning process.

The research, released last month, retains the standard LLM transformer architecture, KV cache and existing language-modeling objective while widening the information channel between steps. It combines the previous step's top-level hidden state with the embedding of the newly generated token using a Gated Linear Unit (GLU) and passes both into the next step.

According to the research, a 1-billion-parameter Full-Bandwidth Transformer trained on 400 billion tokens achieved performance comparable to a conventional transformer trained on about 600 billion tokens, or roughly 1.5 times as much data. This suggests models of the same size can achieve similar performance with less training data.

Differences also emerged during inference. The reasoning trace required to reach the same level of accuracy was shorter.

In other words, reusing information the model has already calculated rather than discarding it can deliver similar performance with less data during training and shorten the reasoning trace required to achieve the same accuracy during inference.

From Parameter Scaling to Reasoning Efficiency

Taken together, recent AI reasoning-efficiency technologies including Heima, COCONUT and BDH-CQ are moving toward changing the way models think. Heima compresses long reasoning processes into smaller thinking tokens, while COCONUT and BDH-CQ perform reasoning in latent space rather than human-readable language.

Soft Latent Thinking minimizes language generation during reasoning, while Selective Latent Thinking retains language only where needed and compresses the rest. Microsoft's Full-Bandwidth Transformer, meanwhile, preserves the token-based architecture while carrying hidden states into the next step so the model can use more information during the same reasoning process.

The approaches differ, but their direction is the same: improving AI performance increasingly involves changing how models internally process information and reason.

The AI market is likely to evolve toward competition not only in raw performance but also in how much of that performance can be achieved with fewer tokens and less computation.

AI and data infrastructure solutions provider HS Hyosung Information Systems is targeting enterprise AI operations with high-efficiency inference infrastructure combining its low-power Arm-based GreenCore server with neural processing units (NPUs).

HS Hyosung Information Systems said Sept. 9 it is working with major domestic and global NPU companies including Qualcomm, FuriosaAI, DEEPX and Mobilint to provide infrastructure suited to customer environments. It also offers its Arm-based GreenCore server as an option to maximize NPU performance. GreenCore was jointly developed by Korean Arm server manufacturer ACCESSLAB and HS Hyosung Information Systems.

The company describes GreenCore as an Arm architecture-based central processing unit (CPU) server featuring low power consumption, low heat generation and efficient many-core parallel processing. It is designed for data centers with limited power, cooling and rack space, as well as companies and institutions preparing on-premises AI services due to security and data sovereignty requirements.

“NPUs are gaining attention as an alternative to GPUs, but supported AI models and target workloads differ depending on the NPU manufacturer and product,” the company said. “It is therefore important to select products suited not only to the characteristics of the service but also to deployment environments such as data centers and edge infrastructure.”

HS Hyosung Information Systems analyzes customers’ AI models, service types, model sizes, concurrent user counts, target response times and data center power and space conditions to recommend suitable NPU candidates.

Once an NPU is selected, the company installs it in a GreenCore server and conducts integrated validation covering hardware and software compatibility, AI model execution and operational stability before deploying the system in the customer’s environment.

When needed, the company also works with AI solution providers to conduct proof-of-concept (PoC) testing that reflects actual service environments. This reduces the burden on customers of directly comparing and validating multiple NPU products and vendor-specific development tools. HS Hyosung Information Systems supports the entire AI inference infrastructure deployment process, from workload analysis and NPU selection to GreenCore server configuration, integrated validation, AI solution integration, PoC, deployment and operations.

The company plans to continue expanding its HS Hyosung AI Platform portfolio by integrating Arm servers, GPUs, NPUs, high-performance storage, data lakes, artificial intelligence for IT operations (AIOps) and its AI solution partner ecosystem.

Alphabet's Google will invest €13 billion in artificial intelligence (AI) data centers and related digital infrastructure in Finland, its largest single investment in Europe. The investment is supported by the country's cold climate, abundant water resources and electricity supply from nuclear power.

Google said it will work with Finnish municipalities, telecommunications companies, grid operators and other local organizations and businesses. The company has operated a data center in Hamina since converting a former paper mill into the facility in 2009.

The investment will expand the existing data center and add new facilities in Kajaani, Muhos and Vaala in northern Finland, where access to the power grid is strong.

Nordic countries including Finland and Sweden have emerged as attractive locations for new AI data centers because of their cold climates, abundant water resources from seas and lakes, and clean nuclear and renewable energy such as hydropower and wind. These conditions can reduce the energy consumed by cooling systems, which account for a significant share of data center electricity use.

According to a U.S. Department of Energy (DOE) energy-efficiency study, cooling systems account for about 40% of energy consumption at traditional data centers. Gartner recently estimated that cooling and other infrastructure account for about 35% of overall data center electricity consumption.

A Gartner report released in June projected that electricity consumption by cooling and other infrastructure will rise sharply from 159 TWh in 2025 to 195 TWh in 2026 as the number of dedicated AI servers increases.

The Nordic region's low outdoor temperatures allow data centers to extend the use of free cooling, reducing reliance on mechanical cooling. Lower utilization of air conditioners, chillers and other mechanical cooling equipment can significantly reduce cooling energy costs.

Abundant water resources also enable direct water cooling, which draws cold seawater or lake water through pipes to absorb server heat before discharging it. The method can rapidly cool large-scale servers without compressors or cooling towers. Google already uses a seawater-based cooling system at its Hamina data center, drawing cold water from the Baltic Sea to cool servers.

Abundant electricity is another reason hyperscalers are turning to the Nordic region. Hydropower in Norway and wind and nuclear power in Finland and Sweden can help hyperscalers reduce carbon emissions. As part of its Finland investment, Google has entered into a carbon-free energy partnership with Finnish state-owned energy company Fortum Oyj.

Google signed a 22-year agreement covering up to 50% of the capacity of Fortum's Loviisa nuclear power plant. The agreement will start at a smaller scale in 2028 and reach 50% of the plant's capacity from 2030 through 2049. It is Google's first nuclear power procurement agreement outside the U.S. The company also signed power purchase agreements totaling 629 MW with onshore wind developers including Valorem and Suomen Hyötytuuli.

Google will also work with Finnish transmission system operator Fingrid Oyj to ensure reliable grid connections and support grid stability. A 94-MW battery energy storage system (BESS) is planned near its Kajaani site to help balance the grid. The company will also work with Nokia and telecommunications service provider Elisa on high-speed network connectivity and other communications infrastructure.

Microsoft is similarly building an AI data center near Helsinki after investments in Sweden and Denmark. It is using the Nordic region's cold climate to improve cooling efficiency while planning to feed waste heat generated by AI computing into district heating networks for local residents.

Amazon Web Services (AWS) is modernizing and expanding its cloud infrastructure in Sweden, including Eskilstuna, Katrineholm and Västerås, as well as in Finland, to accommodate the high power density required by AI infrastructure. AWS has long-term wind power purchase agreements (PPAs) in Finland and Sweden and combines them with cold-air cooling. Its data centers are reported to operate at a power usage effectiveness (PUE) of 1.1 or lower.

Meanwhile, Argentina's Patagonia region, which has similar climatic conditions, is also emerging as a potential location for AI data centers.

OpenAI has signed a letter of intent (LOI) with local energy company Sur Energy to consider a green data center project worth up to $25 billion with capacity of as much as 500 MW. A lack of high-speed fiber-optic networks and transmission infrastructure, however, could require additional investment.

AI infrastructure company Snet Systems said on Sept. 10 it has completed a wireless network upgrade project at Kyung Hee University Medical Center, establishing infrastructure to support a smart hospital environment.

Kyung Hee University Medical Center launched the project to meet growing demand for digital healthcare services and smart medical environments. Snet Systems deployed a high-performance wireless network across major clinical and operational areas, providing a foundation for stable mobile electronic medical record (EMR) and medical Internet of Things (IoT) services.

The new network allows medical staff to use smart devices to immediately access and enter patient information at the point of care. It is expected to reduce errors and workloads associated with handwritten records and subsequent data entry, while real-time EMR connectivity supports faster and more accurate medical care.

The seamless wireless environment across key hospital areas is also expected to improve continuity of medical services and operational stability. Snet Systems said it carried out the project based on its experience building and maintaining the medical center’s existing Cisco-based wired network. Having a single provider operate both wired and wireless networks also enables faster responses to network disruptions.