Yesterday AI Brief: August 22 - Model Factories, Vision, and AI at Work

Yesterday's most consequential AI stories were about the layers around the model: ownership, interfaces, agent scaffolding, and the people expected to use the technology.

Bloomberg reported that Nvidia agreed to pay $6 billion for a license to Poolside's AI models, offer jobs to more than 100 employees, and make an additional $1 billion investment at a $12 billion valuation, excluding that investment. Poolside would continue operating independently. 1

A Saturday report republished by Sina Finance, citing Caixin, described the same structure and noted that Nvidia and Poolside had not immediately responded to requests for comment. That means the careful label is reported transaction, not confirmed acquisition. 2

The significance is structural: licensing, capital, and hiring can give Nvidia access to a specialized coding-model team without buying the company outright. Poolside focuses on coding automation, including government and defense software. 1

DeepSeek's release notes say its experimental V4-Flash-Vision-Exp model is live on the DeepSeek API. The model keeps V4-Flash's text capabilities, including agents, reasoning, and world knowledge, while adding mixed text-and-image input. 3

The release supports Chat Completions, Messages, and Responses. Images can be sent as base64 data, external URLs, or through the Files API, and each image is tokenized for billing at up to 384 tokens under V4-Flash pricing. DeepSeek says its multimodal-agent benchmark results come close to Opus 4.8; that is a company-reported benchmark claim, not an independent ranking. 3

For developers, the practical shift is the combination of vision, tools, and agent frameworks in one API workflow. The model is marked experimental, so real-world reliability still needs outside testing.

Nvidia research, reported by TechCrunch, used a custom system called Agentic Variation Operators, or AVO. On ARC-AGI-3, a set of 2D interactive games with no instructions, Claude Opus 5 scored 30% without the harness and 100% with a memory-aware harness plus a supervisor component, according to the report. 4

Nvidia's own research post describes AVO as a frontier-level architecture for long-horizon autonomous agents. 5 The result is not proof that the model wins every task: it comes from Nvidia's research setup, and TechCrunch notes that this is not a new Nvidia product.

The useful lesson is broader. Memory, tools, feedback, and runtime can change an agent's behavior as much as the base model can. For production buyers, a model leaderboard alone is becoming a poor proxy for what an agent will actually do.

TechCrunch reported on Saturday that AI-native accounting startup Rillet raised $100 million in a Series C at a $1 billion valuation, becoming a unicorn in 48 hours. The company said it now has 600 customers and has raised $200 million in total. 6

Rillet says customers are replacing systems from Intuit, NetSuite, Sage Intacct, Oracle, SAP, Workday, and Microsoft. The company told TechCrunch that 50% of its customers come from Intuit, 30% from NetSuite and Sage Intacct, and 20% from the other systems. Those figures come from the company, not an independent audit. 6

Rillet was built for agents working alongside human accountants. The company says its model-routing layer lets customers choose a foundation model, its harness prevents those models from training on customer data, and its governance feature lets accountants audit each agent decision. The article also notes that public-company transactions made by an AI agent still require human approval. 6

The important signal is not just the funding round. Customers are reportedly replacing core finance systems rather than running a small pilot. That suggests agentic AI is entering sensitive workflows, while also making auditability and human approval part of the product itself.