The Polylog AI Intelligence Brief

Morning Edition · Friday, August 28, 2026

Tech

Moonshot AI Seeks Up to 30% of Revenue to Put Kimi K3 on Microsoft, Amazon and Google Clouds

The talks are early, with data access and token auditing still unresolved, and they continue despite a warning from the US Treasury that Moonshot could be added to a trade blacklist.

3 sources
Corroborated

Tech

Nvidia Begins Volume Shipments of Vera, an 88-Core Server CPU Aimed at Agent Workloads

Oracle Cloud Infrastructure plans to deploy hundreds of thousands of the processors starting this year, and early units went to Anthropic, OpenAI and Amazon Web Services.

3 sources

Tech

Anthropic Opens a Research Preview of a Standard for AI Agents to Operate Lab and Factory Machines

Amazon Web Services will support the specification through its Strands Robots library, and Anthropic says it plans to open source the standard later.

3 sources

Tech

Amazon Commits to Two Million More Nvidia GPUs as Nvidia Moves the Memory Controller Into the Memory Stack

The added Blackwell Ultra, Rubin and Rubin Ultra capacity lands in 2027 and 2028, and Amazon's Annapurna Labs is the first partner for Nvidia's custom NVHBM memory.

3 sources

Tech

Google Ships Gemini Omni 1.1 Flash With 40-Second Video Sequences and a Cheap Draft Tier

Developers can extend a clip in 10-second increments, set first and last frames, and preview at 360p for about a third of the 720p price before committing to a full render.

3 sources

Tech

Altman Says OpenAI Will Have an Internal System He Calls AGI Before the End of 2026

The claim rests on OpenAI's own definition of the term, and the model it points to, Astra, has not been released or independently evaluated.

3 sources
Plausible

Tech

Google DeepMind Runs Model Evaluations Inside an Encrypted Enclave So Neither Side Sees the Other's Data

The pilot tests a Gemini Flash Lite model against confidential benchmarks using Confidential Space, an Nvidia H100 confidential GPU and Intel memory encryption, with the Singapore AI Safety Institute and MLCommons among the partners.

3 sources

Tech

Tencent's WeChat Team Open-Sources a Multimodal Embedding Family Already Running in Production

The 2B, 4B and 9B models handle text, images, video and visual documents, and Tencent reports the 9B version scoring 80.6 overall on the MMEB-v3 evaluation.

3 sources

Tech

Researchers Argue Capture-the-Flag Scores Hide How AI Agents Actually Break Into Systems

A paper posted Friday says pass or fail grading on security benchmarks discards the agent trajectory, which is the part that reveals whether a solve was reasoning or a lucky tool call.

2 sources

Tech

Small Models and Compression Research Push More Inference Off the Cloud

A widely circulated developer essay argues small models have become good enough for real products, while new work on binarized network pruning and multi-drafter speculative decoding attacks the cost from both ends.

3 sources

Tech

New Benchmark Tests Whether Computer-Use Agents Refuse Traps and Ask When Instructions Are Vague

ADeptS-Bench pairs benign and malicious versions of the same task with threats embedded in the visual interface, and separately scores whether an agent seeks clarification instead of guessing.

1 source

Tech

Berkeley Lab Cuts Beamline Image Analysis From a Month to About 15 Minutes Using Meta's Open Vision Models

The SYNAPS-I project runs Meta's SAM 3 and DINOv3 across 300 Nvidia A100 accelerators, one of the first concrete results from the US Department of Energy's Genesis Mission.

2 sources