ReadingResearchRadarInvestment framework
Sign in / Sign up
AllTechnologyProductsInvestingPublic marketsOther

Reading

2026-11-1944 posts

Refresh

German Federal Court: AI-generated legal briefs without lawyer review may be inadmissible

Material
2026-11-19 00:42 GMT+8

The German Federal Court of Justice (Bundesgerichtshof, BGH) clarified in a ruling dated October 13, 2026, that lawyers must perform substantive content reviews and assume full responsibility when using artificial intelligence (AI) to draft legal documents. If a document is primarily generated by AI without effective human oversight, courts may deem it non-compliant with formal requirements, rendering the appeal invalid.

The decision stems from a criminal appeal against a verdict by the Kaiserslautern District Court (Case No.: 3 StR 350/26). Under Section 345(2) of the Code of Criminal Procedure, valid grounds for appeal require the defense attorney's 'shaping participation' in the drafting process. The BGH's Third Criminal Senate stated that if an appeal brief is created using AI without significant content control by the lawyer, this shaping participation may be considered missing.

In the specific case, judges noted numerous generic, standardized passages in the appeal brief that had little connection to the facts established in the original trial. For instance, the text abstractly discussed possibilities for withdrawing from attempted crimes but failed to address the district court's finding that the defendant could no longer influence the victim after fleeing into a warehouse where police were already alerted. Although the BGH determined these flaws did not entirely negate the lawyer's involvement—thus rejecting the appeal on merits rather than procedural inadmissibility—the case underscores strict review obligations.

This ruling signals to the legal profession that uncritical reliance on AI-generated text poses high procedural risks. Courts are empowered to reject such automated inputs and refuse to reopen cases on formal grounds. This aligns with a broader trend of stricter judicial scrutiny regarding technology misuse; previously, the Berlin Higher Regional Court also reprimanded lawyers for submitting documents containing AI hallucinations.

Sources:heise online

Research

Session Continuation Pattern at Bankruptcy Data Company Stretto: Enabling Voice AI to Surpass Streaming Session Duration Limits

Material
2026-11-19 02:58 GMT+8

Bankruptcy services data company Stretto and AWS partner CI&T co-authored an article on the Amazon Machine Learning Blog detailing their production deployment experience with their voice AI agent, Tara. Tara is used to answer calls from individuals seeking bankruptcy advice on the Bankruptcy.com website, conducting structured pre-bankruptcy interviews. A specific engineering constraint shaped the core of the technical narrative: the full attorney-designed interview takes 15 to 20 minutes, which exceeds the maximum duration of a single streaming session on Amazon Nova 2 Sonic. Rather than treating this limitation as a defect to be avoided, the team viewed it as a design parameter, leading to the development of a session continuation pattern.

The mechanism of the session continuation pattern works as follows: a voice bridging layer continuously tracks how long the bidirectional audio stream between the caller and Nova 2 Sonic has been active. Approaching the limit, it opens a new stream using the same system prompt and tool configuration, feeding in the conversation history thus far as text. Once Tara completes her current response, the bridging layer switches the caller's audio to the new stream while closing the old one. To the caller, the conversation seamlessly continues at the response boundary, without them perceiving that a session switch occurred underneath.

The portability of this pattern lies in its independence from model improvements; it handles constraints at the orchestration layer. Any team building a telephone agent using streaming voice models will encounter the conflict between single-session duration limits and real-world conversation lengths. The timing of the switch described in the article—after the model finishes its current response—is a key design choice: switching mid-response would cause audio breaks for the caller, while switching before the response begins might mean the model has not yet output a complete semantic unit.

Before deploying Tara to production, the team dedicated an entire week specifically to validating voice quality, covering caller emotional states (calm, nervous, hesitant, frustrated), baseline audio latency, interruption handling, and emotional tone adaptation. This validation report served as a go/no-go gate for proceeding with integration. This practice itself serves as a recommendation: in voice AI deployments involving sensitive financial information, validate conversational naturalness and emotional adaptability first, then write integration code, rather than doing so in reverse.

Boundaries to retain: all performance figures in the article are self-reported by Stretto—approximately 96% accuracy in structured field extraction, average call completion time of about 8 minutes, a 95% reduction in per-minute voice costs compared to previous solutions, and a 20% increase in lead conversion rates compared to web forms. These figures have not been independently replicated, and the 95% cost reduction is based on Amazon Bedrock on-demand pricing in the us-east-1 region as of April 2026. The behavior of the session continuation pattern under adversarial conditions (such as callers interrupting at the switch boundary, or emotional states being lost during text history transmission) was not detailed in the article.

Sources:Amazon Web Services

Research

Why Fine-Tuning Takes Much Longer Than FLOP Estimates: Evaluation Frequency and Checkpoint Writes Are the Dominant Factors

Material
2026-11-19 02:53 GMT+8

When scheduling fine-tuning jobs on shared GPU clusters, almost every team uses the same formula to estimate duration: six times the parameter count multiplied by the number of training tokens, divided by the number of GPUs and peak throughput. This formula describes only the optimization loop itself, but a real fine-tuning job involves far more than just the optimization loop. Before the job starts, environment setup, model downloading and loading, and data tokenization are required; during training, periodic evaluations and checkpoint uploads occur; at the end, the final model must be written to object storage. On a multi-tenant cluster, all these operations compete with other jobs for network and storage bandwidth. In a blog post preview released on November 18, 2026, the Dharma-AI team quantified exactly how large these ignored overheads are, using operational logs from 272 supervised fine-tuning tasks in their production environment.

They used the FLOP formula as a baseline, testing across a range where Model FLOPs Utilization (MFU) varied from 0.25 to 0.65. The best result (MFU=0.30) still yielded a mean absolute error of 12.41 hours, with an R² of only 0.31. In other words, the average duration of these tasks was 23.43 hours, while the theoretical formula's prediction deviated by nearly half. As a second analytical baseline, they also tested Calculon (Isaev et al., SC '23), an analytical performance model that accounts for compute, memory traffic, and inter-device communication, extended here to cover LoRA, grouped-query attention, fused attention, bfloat16, and SwiGLU. The median of Calculon's estimates accounted for only 11% of the actual wall-clock time. This indicates that the vast majority of time in production jobs is spent outside the optimization steps.

In contrast, the CatBoost regressor trained by Dharma-AI used only information readable from the configuration before the job launched—hyperparameters, dataset size, model identifier, and number of allocated GPUs—and achieved an R² of 0.98, a mean absolute error of 1.71 hours, and a median error of 0.51 hours in 5-fold cross-validation. The source of this gap lies in two types of signals: first, evaluation frequency, checkpoint policy, per-GPU load, and fine-tuning regime, which the FLOP formula does not include at all; and second, the nonlinear interactions between these variables, which can only be learned from historical runs.

Among the 24 features in the CatBoost model, the top-ranked feature is neither parameter count nor GPU count, but num_eval_data—the total number of samples processed in evaluation over the entire lifecycle of the job. Following closely are trainable parameter count, evaluation batch size percentage, total parameter count, and per-GPU training batch percentage, with saved checkpoint parameters also ranking high. SHAP value analysis shows that high evaluation volumes consistently push predicted time upward, while the impact of trainable parameter count becomes significant only for the largest models. This means: how often you run evaluations, how much data is evaluated each time, combined with model size, determines when the job finishes. These are often choices made by habit when writing the training configuration.

Sources:huggingface.co

Research

Feiran Hao, VP of Portfolio Strategy and Planning at UiPath: RPA is tokenless enterprise automation

Material
2026-11-18 08:00 GMT+8

Feiran Hao, Vice President of Portfolio Strategy and Planning at UiPath, notes a recurring phenomenon: people within companies constantly complain about running out of tokens, with some even timing their prompts specifically to optimize usage windows. His response is not to complain about costs, but to propose a design principle—'tokenless,' meaning that when a step does not require a model to understand, reason, or generate content, it should not consume tokens. API calls, running Python scripts, and UI automation can all be performed with zero token consumption.

He illustrates this principle in personal applications with two specific examples: rather than having AI watch a video to extract screenshots, one should use scripts to directly capture frames and mathematical algorithms to identify unique frames; rather than having AI load a 20,000-row Excel spreadsheet before starting analysis, one should first use scripts to understand the data structure. The commonality between these approaches is that there is no need for models to reinvent the wheel when deterministic tools can already complete the work.

Feiran Hao then scales this personal practice to the enterprise level: what UiPath initially perfected at an enterprise scale was precisely this kind of tokenless automation. He acknowledges that RPA has received little attention during the AI boom, but insists on redefining it as 'tokenless enterprise automation.' The reason is simple—it is precise. Enterprise work is usually not open-ended exploration; people know what technology aims to achieve and how much they are willing to spend to achieve it. Deterministic execution is more suitable for such needs than probabilistic generation.

This philosophy leads directly to pricing strategy. Feiran Hao reveals that UiPath has introduced transactional billing as a new option alongside its existing model. The logic is that customers value the work completed, not the tokens consumed within that work. For users accustomed to paying by usage, the old options remain available; but for enterprise customers who care more about result certainty, transactional billing provides a predictable cost structure.

The article concludes with a restrained reminder: you do not need to use AI just because you can; sometimes, work is better done in a tokenless manner. This argument itself is not technically novel—deterministic tools are cheaper and more predictable for structured tasks, which is engineering common sense. The notable incremental contribution lies in organizing it into a coherent strategic line from personal habits to enterprise positioning to pricing models. The boundary to preserve is that Feiran Hao is a VP at UiPath, whose core business is indeed RPA; the renaming to 'tokenless' serves not only technical judgment but also the commercial need to embed existing product lines into the AI-era narrative.

Sources:uipath.com

Research

Microsoft Ignite Unveils Production Capabilities for Foundry Hosted Agents: Shared Subnet IPs and Private CA Trust

Material
2026-11-19 01:00 GMT+8

Microsoft announced at its Ignite conference on November 18, 2026, that several new capabilities are now available for Hosted Agents in Azure Foundry Agent Service, designed to resolve bottlenecks preventing enterprises from moving AI agents from pilot to production. These updates focus on network scalability, secure internal service connectivity, and deployment flexibility.

To address IP address exhaustion, which previously limited large-scale deployments, hosted agent sessions now support shared subnet IPs. Microsoft states that each session continues to run in its own VM-isolated sandbox but no longer requires a dedicated IP address. According to planning estimates, a /24 subnet can now support approximately 60,000 concurrent sessions, while the smallest supported /27 subnet is projected to handle about 7,500. This change is Generally Available (GA) and requires no additional request.

On the security front, support for trusting Private Certificate Authorities (Private CAs) has been introduced, currently in public preview. Previously, agents often failed calls to internal MCP servers or APIs using corporate certificates due to strict certificate checks. Now, administrators can register a CA certificate once on the Foundry resource (stored in Azure Key Vault), allowing projects within that resource to automatically trust it. This maintains hostname and certificate validation without disabling security checks.

Network Egress Controls are also Generally Available. This feature allows organizations to enforce allow/deny rules for specific hosts outside the agent's sandbox, addressing the security concern that "a tool list is not a network boundary." Additionally, Microsoft previewed a "Bring Your Own AKS" (BYO-AKS) deployment option, targeted for early December public preview, which enables organizations to run the Foundry agent runtime on their own Azure Kubernetes Service clusters to meet specific compliance and network management requirements.

Sources:Microsoft Foundry Blog

Research

curl Maintainer Daniel Stenberg: AI Rewrote 340,000 Lines of Rust but Missed 200 Concurrent Bugfixes

Quick take
2026-11-18 23:58 GMT+8

Code volume has never been a measure of maintenance quality. In a blog post dated November 17, 2026, curl maintainer Daniel Stenberg provided a starkly specific comparison: an AI agent converted 185,000 lines of C code into 340,000 lines of Rust, nearly doubling the line count. Yet, during that same timeframe, the actual curl project merged over 300 commits. Among them were more than 200 bugfixes and several security vulnerability patches—none of which are present in those Rust rewrites.

Stenberg specifically noted, "I counted." This was not a vague impression but a line-by-line verification by the maintainer across two timelines. Over those slightly more than three weeks, human maintainers fixed over 200 bugs and patched several vulnerabilities; the AI produced new code nearly twice the size of the original yet remained entirely disconnected from these fixes. He further noted that in the following month, the team performed another 200 fixes, possibly patching additional vulnerabilities.

These figures reveal a structural blind spot: AI agents excel at one-time conversions—moving existing code from one language to another—but struggle to continuously track incremental changes in upstream projects. An actively maintained open-source project generates bugfixes, security patches, and compatibility fixes daily; these changes do not appear in the initial code snapshot. The rewriting agent sees only the starting point, whereas maintainers perform continuous convergence. Those 340,000 lines of Rust code may be syntactically correct and structurally clear, but they do not contain the specific errors fixed by those 200 bugfixes, nor the attack surfaces closed by those vulnerability patches.

The question Stenberg left behind is rhetorical: "Will these clones keep up?" He offered no answer and did not assert that AI rewrites can never catch up. His phrasing was an open-ended query rather than a definitive negation. However, the data itself is telling: at the moment the code conversion was completed, those Rust versions were already lagging behind the upstream. Catching up requires continuous synchronization capability, not merely the scale of an initial conversion.

Notable boundary conditions: Stenberg described a specific batch of AI agents attempting to rewrite curl, not a final judgment on all AI-assisted programming. Curl itself is a project with over 25 years of maintenance history, continuously handling protocol compatibility and security issues; its maintenance complexity may exceed that of many one-time conversion scenarios. His observation applies more to the question of whether "actively maintained large projects can be replaced by one-time rewrites" rather than "whether AI can write useful code."

Sources:Simon Willison’s Weblog

Research

Disney Reveals Star Wars Anzellans Prototypes Using AWS Cloud and AI for Real-Time Guest Interaction

Material
2026-11-19 01:50 GMT+8

Walt Disney Imagineering R&D has revealed three new Audio-Animatronic prototypes based on the Star Wars alien characters known as the Anzellans. Named Vilo Vele, Kado Bre'a, and Jibu Bre'a, these small robots do not rely on hidden puppeteers or strictly predefined movement sequences. Instead, they use AI models running on Amazon's AWS cloud to process and respond to guest interactions in real time.

According to reporting by The Verge, the animatronics are equipped with hidden cameras and microphones to capture guest movements and verbal cues. This data is processed by multiple AI models on AWS servers. The system employs a layered approach: a 'fast-decision model' generates immediate responses, such as making eye contact to simulate attention, while larger models plan out longer interaction sequences.

Despite the integration of AI, Disney emphasizes that the robots' movements and dialogue are not generated on the fly. Every action and line of speech selected and triggered by the models is prewritten by Disney Imagineers to ensure character consistency and prevent deviation from the script. This design allows for seemingly spontaneous interactions while maintaining strict content control.

The prototypes will make their first public appearance later this month at the AWS re:Invent conference in Las Vegas, where attendees can interact with them. Disney stated that the technology could eventually be deployed in attractions like Star Wars: Galaxy's Edge, potentially changing how large-scale entertainment facilities manage staffing and deliver interactive experiences.

Sources:The Verge

Research

NVIDIA Releases Daviator: Eliminating GPU Bottlenecks in Small-System DFT via k-Point Batching and Mixed Precision

Material
2026-11-18 23:00 GMT+8

Training next-generation machine learning interatomic potentials (MLIPs) requires tens to hundreds of millions of density functional theory (DFT) calculations as training data. Plane-wave DFT is the gold standard for generating this data, with VASP being its dominant implementation. However, an counter-intuitive phenomenon has long plagued researchers: running small-system DFT calculations on GPUs is actually slower than on CPUs. In a technical blog post released by NVIDIA in November 2026, the root cause of this bottleneck was explained, along with how its new library, Daviator, resolves it through two layers of mechanisms.

The core of the problem lies in the k-point processing mode. The electronic wavefunctions of periodic systems are parameterized by crystal momentum; small unit cells require hundreds to thousands of irreducible k-points to achieve convergence accuracy. Each k-point solve is independent within the Davidson eigensolver step, requiring no inter-k-point communication. However, VASP’s OpenACC GPU port processes k-points sequentially one by one. For large systems, the workload per k-point is sufficient to saturate the GPU; but for small systems paired with dense k-point grids, hundreds of small GPU dispatches per step cause kernel launch overhead to dominate computation time. The article provides a specific case: a single-atom iron system with 752 k-points ran 8 times slower on more expensive GPU hardware than on CPU.

Daviator’s solution replaces sequential loops with batch operations. Its core is a batched Davidson eigensolver that batches FFTs, matrix multiplications, and custom kernels, drastically reducing the number of GPU kernel launches and increasing stream multiprocessor occupancy from partial capacity to near saturation. Data remains resident in GPU memory throughout. VASP retains full control over self-consistent field (SCF) loops, mixing, symmetry, and I/O; Daviator intervenes only at well-defined call boundaries on data already located on the device.

After eliminating the k-point bottleneck, NVIDIA further introduced a mixed-precision execution layer to leverage the asymmetry between FP32 and FP64 throughput in hardware. On data center GPUs, FP64 throughput has remained relatively flat across Hopper, Blackwell, and Rubin architectures, while FP32 throughput continues to grow. Inference-grade and workstation GPUs like the RTX PRO 6000 have an FP32-to-FP64 ratio of 64:1. The issue is that not every kernel in the Davidson pipeline is equally sensitive to precision, and simply comparing FP32 runs against FP64 yields only a binary pass/fail result, failing to reveal where numerical boundaries lie.

The article describes a transferable methodology for analyzing precision requirements. First, the code is instrumented using generic floating-point types with configurable mantissa and exponent bits. Then, run scripts, comparison scripts, and screening scripts are written: for each kernel, the mantissa bits are decremented until the comparison fails, and then the exponent bits are decremented at the minimum available mantissa. This identifies the lowest effective precision for each kernel, determining whether to switch to FP32, improve the implementation to bring it into the FP32 range, or retain it as an FP64 island. NVIDIA points out that true mixed precision requires knowing exactly where the precision boundaries are, because soft failures—energy drift, incorrect convergence, molecular dynamics trajectories that appear valid but actually diverge—are more insidious than hard failures.

Sources:NVIDIA Technical Blog

Research

AI Document Processing Startup Unsiloed AI Raises $7.5M Seed Led by Felicis

Quick take
2026-11-18 20:00 GMT+8

AI document processing startup Unsiloed AI has announced it raised $7.5 million in seed funding, led by Silicon Valley venture capital firm Felicis. The round included participation from Y Combinator, Orange Collective, and Uncorrelated Ventures, along with angel investors Aarthi Ramamurthy and Beth Turner.

Unsiloed AI aims to solve the challenge of making business documents accessible to AI agents. Its software breaks down complex layouts containing tables and charts, extracts information into structured data, and provides source traceability to verify results. Co-founder and CEO Aman Mishra noted that the probabilistic nature of Large Language Models (LLMs) leads to hallucinations, which can cause severe errors in industries like banking, healthcare, and law.

The company uses "dual-stream" vision models to handle text and image data separately, combined with "heat map-based chunking" to optimize computing resources. According to the company, internal tests on 542 financial documents showed 99.78% extraction accuracy, and it scored 88.0 on the public olmOCR-Bench, outperforming Reducto and Unstructured. These results have not been independently verified.

Headquartered in San Francisco with a team of seven, Unsiloed AI plans to use the new funds to hire researchers and engineers, expand sales, and develop proprietary datasets. The startup already serves customers in finance, legal services, and healthcare, including voice AI startup Bolna AI.

Sources:TechCrunch

Research

US Banks Citibank and PNC Update Terms to Make Customers Liable for Third-Party AI Agent Actions

Material
2026-11-19 01:32 GMT+8

Citibank, the third-largest commercial bank in the United States, updated its online user agreement this week to explicitly inform customers that they bear full responsibility for any issues arising from the use of third-party AI tools for personal finance management, including losses resulting from hacking incidents involving these agents. This move signals a shift in the US banking sector's approach to AI agents from purely technical blocking to clear legal risk isolation.

"As third-party AI tools become more widely available, we want customers to understand the risks associated with sharing banking credentials or authorizing those tools to access their accounts," a Citi spokesperson told TechCrunch. "We recently updated our Online User Agreement to provide greater clarity about those risks and to reinforce that customers who choose to grant account access to third-party AI agents are responsible for activities conducted through those agents."

This policy is part of a broader trend among major US financial institutions. JPMorgan and Bank of America, the two largest banks in the country, announced similar policies last month. PNC, the eighth-largest bank, also changed its terms on the same day as Citi. Additionally, Citizens bank emailed customers warning them that they are liable for anything their AI assistant does.

Previous media coverage often focused on how banks technically block AI agents from logging in with customer credentials or restrict them to read-only access via protocols like Model Context Protocol (MCP). However, the updates from Citi and PNC address the core issue of legal liability: even if an AI agent is legitimately connected, the bank will no longer provide traditional fraud protection if the agent's actions lead to financial loss or if the agent is compromised by hackers.

For AI startups, this raises the barrier to entry in the personal finance sector. Anthropic’s personal finance tool, launched earlier this month, is explicitly billed as "read-only" and states it "can't move, manage, or invest your money," likely to avoid such liability risks. As AI agents attempt to take over more high-value interactions, this legal allocation of risk will be a critical factor in determining which AI investments prove viable.

Sources:TechCrunch

Research

Nokia and Four Partners Form 'AI Gigafactory Northern' Consortium to Bid for EU Infrastructure Funding

Material
2026-11-16 15:00 GMT+8

Nokia, Nscale, Skeleton Technologies, Telia, and Verda announced on November 16, 2026, the formation of a European consortium named "AI Gigafactory Northern" (AIGF Northern). The alliance aims to establish a federated, multi-site AI infrastructure across the Nordic and Baltic regions and intends to submit a bid to the European High-Performance Computing Joint Undertaking (EuroHPC) for public funding designated for AI Gigafactories.

Unlike traditional centralized data centers, AIGF Northern employs a distributed architecture connecting sites in Oulu, Finland; Iru, Estonia; and Valmiera, Latvia, with the Swedish location yet to be determined. This design leverages the cool climate of Northern Europe to reduce cooling demands and integrates reliable renewable energy sources, while distributing capacity across borders to minimize concentration risk and enhance resilience.

The consortium defines clear technical roles: Nokia provides secure network infrastructure, Nscale offers scalable cloud platforms for AI training and inference, Skeleton Technologies supplies high-power systems to improve efficiency and reduce grid strain, Telia contributes connectivity and system integration expertise, and Verda supports sovereign cloud infrastructure. Research partners VTT Technical Research Centre of Finland and KTH Royal Institute of Technology will provide capabilities in quantum computing, semiconductors, and AI talent development.

This initiative coincides with the EU's efforts to mobilize over €30 billion for AI investment through the InvestAI program. EuroHPC is currently soliciting bids for up to four medium-sized and three large AI Gigafactories. AIGF Northern enters a competitive field alongside other major players like Deutsche Telekom, all vying to build a controllable, trustworthy, and scalable domestic compute foundation to strengthen Europe's technological sovereignty in artificial intelligence.

Sources:European technology leaders form AI Gigafactory Northern consortium to build trusted, sustainable AI infrastructure

Research

a16z Leads Linq's Series B: Building Infrastructure for AI Agent Communication

Material
2026-11-19 01:05 GMT+8

U.S. venture capital firm Andreessen Horowitz (a16z) announced on November 18, 2026, that it has led the Series B funding round for Linq. Linq is an infrastructure platform providing developers with tools to build reliable, scalable messaging experiences for AI agents, handling the operational complexity behind them.

As AI agents become a primary source of internet traffic, companies must adapt to allow customers to interact with their services through these agents. Linq's core value lies in integrating AI agents into existing user communication habits. For instance, food delivery platform DoorDash uses Linq's technology to enable consumers to find food, build a cart, and check out within a single thread on Apple Messages for Business, without switching to standalone apps or relying on traditional notifications.

Linq is an approved Messaging Service Provider (MSP) with Apple, giving enterprises confidence to legitimately launch and scale customer-facing agents on this channel. Beyond Apple Messages, Linq supports RCS, WhatsApp, and other platforms, aiming to become the universal platform helping all businesses and agents communicate wherever their users are.

This funding reflects capital market recognition of the 'AI agent infrastructure' sector. Linq's evolution from digital business cards to sales software, then AI sales agents, and finally general communication infrastructure demonstrates how startups can capture market turning points by identifying opportunities to reuse underlying technology.

Sources:a16z.news

Research
Next →
Sign in / Sign up
ReadingResearchRadarInvestment framework

Danya Gordin, Chief AI Architect at Stretto, and Marcelo Wzorek, Executive Manager at CI&T, are quoted in the article regarding project outcomes. The article was published by the AWS Machine Learning Blog; AWS holds a commercial interest relationship as the provider of the model and cloud infrastructure. This attribution boundary should be retained when citing its performance claims.

However, the scope of this predictor has clear boundaries. It covers only supervised fine-tuning and does not include other training and alignment strategies such as RLHF, DPO, or RLVR. All runs used H200 GPUs, and the model is calibrated to that infrastructure, so it cannot be directly migrated to other GPU generations. There were only 13 samples for long tasks exceeding 60 hours, where the median error jumped to 5.19 hours and the 90th percentile error reached 18.51 hours. The Dharma-AI team themselves note that this is a limitation of data volume rather than a methodological flaw, and more data should continue to improve results. What transfers is the methodology itself: using features that expose overheads, feature selection that handles multicollinearity, and a comparison protocol, this pipeline can be rebuilt on any execution-log dataset with a similar structure.

Following this workflow, original analysis based on Verificarlo indicated that most of the Davidson algorithm code could run with FP32 arithmetic and storage: FFTs, Hamiltonian application, and orbital updates use FP32, while subspace eigenproblems, tall-skinny matrix multiplications behind subspace matrices, energies, and convergence measures remain in FP64. The remaining FP64 bottleneck is the tall-skinny matrix multiplication in the Rayleigh-Ritz projection; pure FP32 accumulation is insufficient because the output feeds directly into a numerically sensitive eigensolver. NVIDIA resolved this with a split-k algorithm: breaking dot products into FP32 chunks with final accumulation in FP64.

Regarding performance results, NVIDIA reported that Daviator achieved over 2x acceleration relative to the CPU baseline for single-atom iron, a 22x Davidson acceleration for LiFePO4 r2 SCAN functionals, and reduced execution time by over 85% for a 64-atom Bi-doped GaAs system. Mixed precision made the FP32 Davidson solver run 3 to 7 times faster on the RTX PRO 6000, delivering approximately 2x end-to-end acceleration; however, on H100 and H200 GPUs, marginal gains were limited due to FP64 tensor cores matching FP32 throughput. Projections for the OMat24 dataset showed that combining Daviator and mixed precision on the RTX PRO 6000 could reduce costs to about 25% of CPU costs.

It must be clarified that these figures are all self-reported benchmarks from NVIDIA, and the OMat24 projections are extrapolated from 13 representative inputs rather than actual full runs. Currently, Daviator supports only single-node execution; non-collinear spin and pure Gamma-point runs are not yet covered. The convergence limits of mixed precision depend on the input; typically, FP32 can converge to EDIFF thresholds of 10^-4 to 10^-6 eV, and continuing with FP64 is required to break through the FP32 noise floor. This precision analysis methodology itself—instrumentation, per-kernel screening, and FP64 islands—has reference value for other HPC codes needing mixed-precision adaptation, not limited to VASP or the Davidson solver.