AI Daily · 2026-08-18

NVIDIA is teaming with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR on an AI compute financing platform targeting over $500 billion, and has locked in SB Energy’s PORTS-Pike campus in Ohio to exclusively host NVIDIA AI compute. On models, Tencent open-sourced EVIE-Preview-4.5B, a visual document retrieval model that tops the ViDoRe leaderboard, while Together AI’s DeepSWE benchmark shows a cascade that tries DeepSeek V4 Pro 0813 first and escalates to GPT-5.6 Sol solves 83% of tasks at $3.35 per task, versus Sol alone at 72.7% for $8.37. In research and tooling, Google PhotoScan estimates cardiometabolic risk from phone photos, and Hugging Face’s new GPU allocator lifts utilization by up to 33 points under the same workload. LangChain and Replit added agent payments middleware and black-box penetration testing, Qwen 3.8 27B matches GPT-5.6 Luna on the Artificial Analysis Intelligence Index, and a tracked rare-book shipment reportedly ended at an Amazon warehouse tied to suspected AI training scanning.

North America · First-hand

Google

⭐⭐ [Research] Seeing beyond BMI: Estimating cardiometabolic risk with smartphone imagery

Google Research introduces PhotoScan, a deep learning approach that estimates body composition from smartphone photos and predicts insulin resistance with accuracy approaching DXA scans in a clinical research setting. The framework derives body fat percentage, A/G ratio, and V/S ratio from standard 2D photos, pre-training on over 35,000 UK Biobank records and fine-tuning on a cohort of 677 adults. Validation showed higher body fat percentage accuracy than smartwatch-based bioelectrical impedance analysis (BIA) sensors, while also unlocking A/G and V/S metrics beyond BIA's capabilities. The method avoids specialized clinical equipment and radiation exposure, offering a scalable, non-invasive pathway for metabolic risk screening.

East Asia · First-hand

Tencent Hunyuan

⭐⭐⭐ [Model Release] tencent/EVIE-Preview-4.5B

Tencent Hunyuan released EVIE-Preview-4.5B, a preview of a visual document retrieval model on HuggingFace, open-sourced under Apache-2.0. Built on Qwen3.5-4B, it uses a ColPali/ColBERT-style late-interaction multi-vector architecture with native 128-dimensional token vectors. The model ranks first on both ViDoRe V3 (65.36 average nDCG@10 on public domains) and ViDoRe V1+V2 (85.77 average), beating larger models such as webAI-ColVec1.1-8b. The same checkpoint supports two deployment tiers (768 or 1792 visual tokens per page) with no retraining or re-export, and the lighter tier fits roughly one million pages in under 180 GiB.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Product Update

⭐⭐ [Product Update] AgentCore Payments middleware for LangChain agents

LangChain's blog introduces AgentCore Payments, a middleware layer that lets LangChain agents securely pay for paid APIs directly, moving payment logic out of tool wrappers and into a shared infrastructure layer. Spending limits are enforced deterministically by AgentCore at the session level rather than through LLM prompts, providing a backstop even if the agent is compromised. Built on Coinbase's x402 protocol now stewarded by the Linux Foundation, it enables per-call stablecoin micropayments within a single HTTP request-response cycle, with wallet authentication via AgentCore Identity and funding through Coinbase CDP and Stripe (Privy) via fiat or USDC. LangSmith integration records what an agent bought alongside its reasoning, supporting post-hoc spending audits and pre-launch evaluation.

⭐⭐ [Product Update] Black-box pen tests on Replit

Replit announced a new black-box penetration testing capability for apps built on its platform. Traditionally, pre-launch security reviews required procuring vendor pen tests that cost thousands of dollars and took weeks of back-and-forth; Replit wants to democratize security the same way it democratized software creation. Its existing scans have an agent read the code for dangerous patterns and run on every build, catching a lot of issues. But malicious hackers don't see the code — they open the app and poke around for unlocked doors, which yields different results. The gap between these two approaches is where most actual break-ins happen, including cases where code looks fine yet the app still hands over data it shouldn't.

Research

⭐⭐⭐ [Research] DeepSeek V4 Pro 0813 vs GPT-5.6 Sol on DeepSWE: Cost, Coding, and Routing

Together AI published a head-to-head evaluation of DeepSeek V4 Pro 0813 and GPT-5.6 Sol on the DeepSWE software-engineering benchmark. A cascade approach—running DeepSeek V4 Pro 0813 first and escalating to GPT-5.6 Sol only when tests fail—solves 83.0% of tasks at $3.35 each, while Sol alone solves 72.7% at $8.37. Sol leads at pass@1 (72.7% vs 62.8%) but Pro overtakes at pass@4 (88.5% vs 85.8%). The per-rollout cost gap is 35x ($0.24 vs $8.37), giving Pro 261 solves per $100 versus Sol's 9. Sol is faster (17 min vs 35 min) but breaks existing tests in 20% of its failures, versus 11% for Pro.

⭐⭐ [Research] Same Cluster, 33 Points More Utilization: What Changed Was the Order

The Dharma-AI team describes a constraint-aware GPU allocator benchmarked against a FIFO scheduler across seven scenarios. On identical hardware and workloads, GPU utilization improved by up to 33 percentage points, and priority-weighted output rose in every scenario, by up to 105%. The core challenge is deciding which GPU runs which job at each timestep, since batch-like workloads (training, batch inference, quantization) conflict with the elastic demand of real-time inference. FIFO scheduling reserves GPUs for peak demand all day, leaving significant idle capacity; baseline utilization was only 51.6% and 53.6% in two reservation-dominated scenarios.

⭐⭐ [Research] Qwen 3.8 27B scores 52 on the Artificial Analysis Intelligence Index

Qwen 3.8 27B scored 52 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Luna (max) and trailing GLM-5.2 (max) and DeepSeek V4 Pro 0813 (max) by just one point. The post notes that GLM-5.2 is 753B parameters, DeepSeek V4 Pro 0813 is 1.7T parameters, and GPT-5.6 Luna is of unknown size but presumably much larger than 27B. Simon Willison describes the model as truly astonishing.

Other

⭐⭐⭐ [Other] NVIDIA Guarantees SB Energy's PORTS-Pike Technology Campus in Ohio to Exclusively Host NVIDIA AI Compute

NVIDIA announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish AI compute infrastructure financing platforms, aiming to mobilize over $500 billion in third-party capital. Separately, NVIDIA guaranteed that SB Energy's PORTS-Pike Technology Campus in Ohio will exclusively host NVIDIA AI compute.

⭐⭐ [Other] We Tracked a Shipment of Rare Books. It Ended at an Amazon AI Training Facility

404 Media tracked a shipment of roughly 1,000 rare books using an Apple AirTag placed by a participating bookseller, and found it delivered to the VGT3 corner of Amazon's LAS8 facility in northeast Las Vegas. Amazon workers' forum discussions confirmed that VGT3 destructively scans large volumes of books. The investigation provides concrete evidence for reports of anonymous bulk book purchases suspected to be for AI training, pointing to Amazon as a likely customer.