OpenAI launches Presence for managed enterprise AI agents

Overview

9 stories in this issue. The first 3 are today's priorities.

Hot model dynamics

1. Top · OpenAI launches Presence for managed enterprise AI agents
2. Top · Cisco open-sources Antares security models for vulnerability localization
3. Top · Anthropic ships a Claude Economic Index connector and a $200M research fund

Global AI news 4. NVIDIA open-sources the first GPU-accelerated medical physics simulation framework 5. Substack adds Pangram to estimate how much of a newsletter is AI-written 6. monday.com runs production AI agents on Amazon Bedrock

Regional & early signals 7. Alibaba releases Qwen-Image-3.0 with 4.5× longer text input (Chinese-language source) 8. Physical AI becomes AMD Ventures' next focus beyond the data center 9. Uber cuts 10% of customer-service jobs, citing AI (Chinese-language source)

Jiufeng graphic based on the sources cited in this issue.

Hot model dynamics

OpenAI launches Presence for managed enterprise AI agents

OpenAI shifts from model supplier to deployment partner, packaging voice and chat agents into a managed service its own engineers help stand up.

On July 22nd, OpenAI launched OpenAI Presence, a managed product for deploying voice and chat agents across customer support, sales and internal workflows at large companies. Presence bundles OpenAI models with the controls needed to let an agent access corporate systems and take actions, alongside evaluations, guardrails and hands-on deployment by OpenAI engineers. OpenAI names itself as its first reference customer, saying Presence already runs its English-language AI phone support (1-888-GPT-0090).

Availability: Presence is a limited release, not general availability; deployment requires deep involvement from OpenAI engineers and systems integrators, and the phone support cited so far is English-only.

Source: RuntimeWire · OpenAI Presence · OpenAI Help Center

Cisco open-sources Antares security models for vulnerability localization

Cisco releases a family of security small language models aimed at pinpointing where vulnerabilities live in code.

Cisco introduced Antares, a family of security small language models (SLMs) purpose-built for vulnerability localization. It is releasing two of them — Antares-350M and Antares-1B — as open-weight models on Hugging Face; given limited security context, they identify the source files most likely to contain a vulnerability.

Limitations: only the two smaller models are open so far, and the task is scoped to locating suspect files rather than fixing them; beyond Cisco's own blog post, there is no independent benchmark yet.

Image source: huggingface; mirrored on Jiufeng R2.

Source: Cisco Blog · Hugging Face

Anthropic ships a Claude Economic Index connector and a $200M research fund

Anthropic lets anyone query its Economic Index through Claude and commits $200M to external research.

On July 22nd, Anthropic launched the Anthropic Economic Index connector for Claude, letting anyone ask Claude to explore the data behind the index, which measures how AI is actually being used across jobs and tasks. The same day, Anthropic said it is committing $200 million to the Economic Futures Research Fund and published the fund's research agenda, covering interventions to make the economy more flexible and resilient, ensure AI's benefits are shared, and minimize its harms.

Limitations: the connector is a data-exploration tool rather than a Claude capability update; the $200M is a commitment plus an agenda, with specific grantees not yet announced.

Source: Anthropic (connector) · Anthropic (fund)

Global AI news

NVIDIA open-sources the first GPU-accelerated medical physics simulation framework

NVIDIA open-sources the physics simulation medical robots need for training, targeting the data bottleneck.

NVIDIA open-sourced what it calls the first GPU-accelerated medical physics simulation framework for healthcare robotics. It says obtaining the large, varied datasets needed to train, test and improve robot behavior is one of the biggest bottlenecks; the framework combines classical physics simulation (modeling known rules such as device contact, friction and motion) with generative AI physics simulation (the open-source Cosmos-H Dreams). NVIDIA says the framework provides 8,192 robot-training environments in scale, turning simulation from bespoke engineering into reusable infrastructure.

Limitations: it targets robot developers, not an end product, and its real-world value depends on how developers adopt and validate it.

Source: NVIDIA Blog · Cosmos-H Dreams (GitHub)

Substack adds Pangram to estimate how much of a newsletter is AI-written

Substack lets readers scan posts for an estimate of human-vs-AI authorship.

Substack launched a feature that shows which newsletters are being written with AI, via an integration with the AI writing detection software Pangram. Users can scan posts, comments and replies in Substack's app to see an estimate of how much content was written by a human versus AI. TechCrunch frames it as part of a broader shift toward transparency around AI-assisted content.

Limitations: the result is an estimate rather than a definitive judgment of AI authorship, and it relies on Pangram's detection within Substack's app.

Source: TechCrunch · Substack (X)

monday.com runs production AI agents on Amazon Bedrock

monday.com's own production data shows AI coding tools are now routine and per-engineer PR output is up over half.

The AWS Machine Learning Blog details how monday.com runs production AI agents ("AI Teammates") on Amazon Bedrock at scale. Per monday's own internal production data, nine in ten of its Builders now use AI coding tools every month, up from roughly half a year ago (adoption has nearly doubled year over year), and per-engineer PR throughput is up by more than half. The post frames the journey in three levels; at L1, the assistant, engineers use Cursor for fast reflexive work and Claude Code for the heavy lifts.

Limitations: every figure comes from monday's own internal production data — self-reported and not independently audited.

Source: AWS ML Blog · Claude Code

Regional & early signals

Alibaba releases Qwen-Image-3.0 with 4.5× longer text input (Chinese-language source)

The third-gen Qwen image model emphasizes complex-layout and multilingual text rendering from long prompts.

On July 21st, Alibaba released Qwen-Image-3.0, the third-generation base model in its Qwen image generation and editing series. It supports up to 4.5K tokens of text input — 4.5× the previous generation — and Alibaba says it can generate knowledge diagrams and complex UIs containing formulas, geometric shapes and logic-derivation steps from longer prompts, with native rendering of 12 languages and 20-plus fonts. Cited use cases include portrait photography, math exam papers, web interfaces, product posters, film storyboards and ancient-book imagery.

Evidence boundary: these are Alibaba's own claims with no independent benchmark yet, and this item currently has only a Chinese-language source.

Source: InfoQ (Chinese)

Physical AI becomes AMD Ventures' next focus beyond the data center

AMD's ecosystem investing is tilting from the data center toward robotics and industrial automation.

SiliconANGLE reports that "physical AI" is emerging as AMD Ventures' next major focus beyond the data center, as machine intelligence moves from cloud data centers into robotics, industrial automation and other embedded, real-world systems — reshaping how chipmakers think about ecosystem investment rather than just product roadmaps. The report says AMD has built out a multi-layer investment approach around this shift.

Evidence boundary: this is SiliconANGLE's reporting and analysis of AMD Ventures' focus; specific investment amounts and targets are not given in the report.

Source: SiliconANGLE

Uber cuts 10% of customer-service jobs, citing AI (Chinese-language source)

Uber trims about a tenth of its support roles as it "embraces AI."

Per Sina Finance/36Kr, Uber said it has cut 10% of jobs in its customer-service business as part of a broader strategy to streamline headcount and embrace AI. It announced the layoff plan for its community-operations team on Wednesday, saying the change aims to simplify operations, strengthen offline collaboration and keep advancing AI adoption; under a full return-to-office policy, the team's formerly remote staff must also report to offline center offices.

Limitations: Uber attributes the cuts to streamlining and embracing AI but does not quantify AI's actual role, and this is a single Chinese-language report relaying Uber's own statement.

Source: 36Kr (Chinese)