AI Today: OpenAI, NASK, and Microsoft Flint

State of knowledge as of 2026-10-25. Source material collected on 2026-10-24.

Today's review covers key developments in AI system security, new tools supporting research work, and hardware optimizations. OpenAI has partnered with NASK on AI safety, Microsoft has released a library for interactive data visualizations, and researchers have gained the LitPivot system. Furthermore, we discuss new tools for verifying model compatibility with AWS Trainium and technical aspects of FP8 compute optimization on the AMD RDNA 4 architecture.

TL;DR

• OpenAI and NASK Partnership: The organizations signed an agreement concerning joint research on the safety of advanced artificial intelligence systems.
• Microsoft Research presents Flint-interactive: The tool enables AI agents to automatically create interactive data visualizations with advanced event handling.
• Model compatibility checking tool for AWS Trainium: A new space on Hugging Face allows verifying whether Transformers models run on Trainium2 accelerators without code changes.
• LitPivot supports the research concept creation process: The system uses thematic segmentation to link scientific ideas with subject literature and generate publication critiques.
• FP8 GEMM optimization on AMD RDNA 4 architecture: The documentation presents the process of porting the FP16 GEMM kernel to the FP8 format using WMMA instructions on graphics cards.

OpenAI and NASK Partnership on AI Safety

OpenAI and the Polish research institute NASK have signed a cooperation agreement aimed at jointly studying threats and developing safety standards for advanced artificial intelligence systems. The official partnership announcement took place on October 23.

The cooperation assumes combining the expertise of both organizations to better understand the risks accompanying the development of AI systems and to strengthen protective mechanisms. As part of NASK's activities, its AI Safety Research Center—acting as Poland's AI safety institute (AISI)—will co-create research on evaluation methods for increasingly advanced artificial intelligence systems. Joint work is also set to cover the identification of security gaps and the creation of international technical standards.

The partnership builds upon earlier experiences of cooperation with Polish cybersecurity specialists. In July, CERT Polska, operating within NASK's structures, gained access to OpenAI models via the Daybreak platform. Using these tools, the expert team managed to discover 6 vulnerabilities in MikroTik RouterOS software, including flaws allowing device takeover without authentication. The system manufacturer released patches, which confirmed the effectiveness of CERT Polska's actions in preventing threats.

For AI builders: IT system security specialists can verify the capabilities of using models with enhanced cyber capabilities to analyze their own network environments. The materials do not specify the costs of access to the technology used by CERT Polska, nor whether these tools are commercially available for teams outside research institutions—these issues require separate arrangements with the provider.

Microsoft Research Presents Flint-interactive

Microsoft Research released the Flint-interactive language on October 22, enabling AI agents to create interactive data visualizations. The tool allows users to directly explore charts through clicking, filtering, zooming, highlighting elements, and interactions between different charts.

The system solves the problem of static visualizations, which often limit the ability to analyze details. Flint-interactive relies on the Flint compiler, which translates user interactions into semantic events. Thanks to this, creators do not need to define low-level visualization code, but instead use high-level update operators. Separating visual logic from interaction logic allows for reducing the number of errors related to rendering and element layout on charts.

The technology supports the creation of interactive data stories, dashboards, and linked experiences, where an action in one view, such as clicking a table, can update a chart or application state. Developers can use ready-made interaction settings or define their own event handling functions via a programmatic API. Flint-interactive is available in the chat environment via Flint-MCP and in applications using Flint Canvas. Users can indicate to the agent what data should be presented and what kind of interactions are required to receive a ready chart specification along with interactive logic.

For AI builders: Developers building analytical applications or dashboards can integrate the library with their own agent to automate the generation of interactive elements. Before deployment, you should verify token generation costs for multiple chart debugging attempts and check the availability of individual Flint components in the chosen production environment, as the source does not specify license limitations or operational costs.

AWS Trainium Model Compatibility Verification Tool

Two days ago, on October 23, an information space was made available on Hugging Face allowing the verification of Transformers-type model compatibility with AWS Trainium accelerators. This tool utilizes the torch-neuronx package, serving as the native PyTorch backend for this infrastructure, and indicates which models can be run without the need for code modifications.

Users can search popular models using the Trending or Downloads views, which reflect the platform's current rankings. For each tested model, detailed data is presented: the number of Trainium cores the model is sharded across, execution time (including compilation), and peak device memory consumption. For models that are unsupported or not yet included, the information space explains the reasons behind this state.

At the time of publication, tests include text generation models and image-to-text-to-text models. The tool indicates the Neuron stack, packages necessary for installation, and a ready run.py script utilizing the torch.compile(backend="neuron") command. Tests are conducted on trn2 instances using up to 16 cores (4 Trainium2 devices). When running a model, graph compilation can be skipped by using the public NEFF cache. Data presented in the space comes from the public aws-neuron/torch-neuronx-model-support dataset, which contains Parquet files for various neuronx-cc versions and libraries.

For AI builders: Engineering teams deploying models on AWS infrastructure can check whether their chosen Transformers model will run on Trainium2 without modifications. The announcement does not state what performance looks like in the case of long generations or whether the tool supports specific architectures outside the Transformers group, which should be verified in your own test environment.

LitPivot Supports Research Concept Creation

According to a blog post on October 25, 2026 [Note: original date context preserved], a team of researchers has developed the LitPivot system, which helps scientists develop scientific ideas by embedding them in existing subject literature.

The system addresses a process called "literature-initiated pivot," meaning changing research direction under the influence of reading earlier works. LitPivot integrates idea creation with literature verification. The user enters a research proposal text into the editor, where the system utilizes a "facet highlighting" function, dividing the idea into three aspects: problem, contribution, and evaluation plan. The tool, utilizing Asta Paper Finder, groups relevant publications around the selected text fragment. The researcher can select publications for which the system will generate critiques backed by source text fragments and propose concrete edits for modification. After each change in the text, the literature list is automatically updated.

In a study involving 17 scientists, concepts developed using LitPivot were evaluated by experts as better grounded in literature than those created using standard editors and AI chats. Participants considered their ideas more novel and useful. The authors identified that the tool helps in two types of activities: differentiating the idea from existing works to demonstrate novelty and adapting the idea to already existing solutions. The system will be presented at the UIST 2026 conference in Detroit.

For AI builders: Scientists working on systems supporting scientific text writing can check the effectiveness of combining a text editor with dynamic literature searching based on thematic segmentation ("facets"). The source does not specify whether the tool supports languages other than English, what the query limits in Asta Paper Finder are, or whether the system possesses an open license allowing deployment in local academic environments.

FP8 GEMM Optimization on AMD RDNA 4 Architecture

According to a blog post on October 25, 2026, the authors presented the process of porting the FP16 GEMM kernel to the FP8 format for AMD RDNA 4 graphics cards. Documentation focuses on utilizing the WMMA instruction available in the gfx12 architecture (specifically using the AMD Radeon RX 9070 XT as an example) and on the impact of smaller element size on LDS (Local Data Store) memory utilization.

Transitioning from FP16 to FP8 changes the computation structure, because each operand element occupies 1 byte instead of 2 bytes. As a result, a single 128-bit load transfers 16 FP8 values instead of 8. The authors compared four tile configurations of various sizes and K-depths, utilizing the solution as a base for further quantization work.

The article tested, among others, a direct 128×96 (K=64) port, larger 256×256 (K=64) configurations, and variants with a larger K=128 depth (e.g., 256×160 and 192×192). Smaller data size allows for more efficient management of available 64 KB LDS memory, which enables using larger tiles and decreases the frequency of synchronization barriers. The implementation does not account for scaling factors, operating on raw E4M3 values, with result accumulation in the FP32 format. ROCm 10.0 installation and the hipBLASLt package are required for the work.

For AI builders: A programmer writing low-level kernels for AMD graphics processors can utilize these configurations to increase compute performance in transformer models. The announcement does not provide exact performance results compared to standard libraries under all conditions, nor does it specify the impact of lacking scaling factors on model accuracy in production applications—this must be verified during your own tests.

Frequently Asked Questions (FAQ)

For what purpose did OpenAI and NASK sign the agreement?

The goal of the cooperation is joint research on threats and the development of safety standards for advanced artificial intelligence systems. The NASK unit, the AI Safety Research Center, will co-create research on AI system evaluation methods, identify security gaps, and support the creation of international technical standards as part of Poland's AI safety institute activities.

What distinguishes the LitPivot tool for researchers?

LitPivot helps develop scientific ideas by embedding them in subject literature. The system divides the idea into problem, contribution, and evaluation plan, and then uses Asta Paper Finder to group publications, generate critiques of source texts, and automatically update the literature list while the scientist edits the text.

What does the AWS Trainium model compatibility verification tool enable?

The tool allows checking whether Transformers-type models can be run on AWS Trainium accelerators without the need for code modifications. Users receive information about core count, compilation and execution time, and memory consumption. The system indicates required packages and ready scripts to utilize the neuron backend.

Sources

• OpenAI and Poland’s NASK partner to advance AI safety and security — 2026-10-23
• Flint-interactive: A language for creating interactive data visualizations — 2026-10-22
• Does Your Model Run on AWS Trainium? Look It Up — 2026-10-23
• LitPivot: Developing research ideas through dynamic contextualization within the literature landscape — 2026-10-23
• FP8 GEMM Optimization on AMD RDNA™ 4 Architecture — 2026-10-23

--- Text automatically prepared by the aipoc-lab (AI) laboratory based on publicly available sources listed above. Own summaries, not copies. Pre-human-redaction version.