What You Need to Know: OpenAI released a preview of the ChatGPT desktop app for Ubuntu 24.04/26.04, Debian 13, and Fedora 43/44 with both x64 and ARM64 .deb/.rpm packages. Three VentureBeat surveys released today quantify how enterprises are actually buying, securing, and evaluating agents and infrastructure. Simon Willison shared three practical prompting patterns for Codex and Claude Code that address common tool-calling friction.

Top Story

OpenAI published .deb and .rpm packages that install the ChatGPT desktop client on supported Linux distributions, including ChatGPT Work and Codex access. The app integrates directly with existing browser and project workflows on the desktop rather than requiring a separate web tab. This is the first official desktop client for the platform on Linux after earlier macOS and Windows releases. Developers who already run multiple cloud accounts can now keep authenticated sessions inside the same window they use for local tools. The preview supports Ubuntu 24.04 LTS, Ubuntu 26.04 LTS, Debian 13, Fedora 43, and Fedora 44 on both x64 and ARM64 architectures. Remember, we covered OpenAI's desktop expansions yesterday — today's Linux preview moves that forward by bringing the authenticated session to more developer desktops. Watch for whether the preview expands to additional distributions or adds native MCP server discovery in the next update. Source: x.com

Model Updates

Gemini App Hits 1B Monthly Users Milestone: Demis Hassabis Google DeepMind’s Gemini app reached one billion monthly active users, the fastest-growing product in the company’s history according to the announcement. The milestone reflects broad consumer adoption of the mobile and web interfaces rather than a new model release. Demis Hassabis credited the @JoshWoodward team and expressed pride in the collective work that produced the growth. No new capability or pricing details accompanied the post. The announcement closed by thanking users and expressing anticipation for what they will build next with the app. Source: x.com

Sam Altman thanks longtime OpenAI colleague for contributions: Sam Altman Sam Altman publicly thanked a longtime colleague for taking on every function OpenAI has needed since the earliest days. The post recalls the colleague as one of the few people who understood the original vision when it still sounded implausible. Altman noted that the individual had handled every challenge the company presented and stated that OpenAI would not be where it is without those contributions. The message ends with excitement about continuing to work together on future projects. No technical details or model announcements appear in the post. Source: x.com

Agent & Tool Developments

PentestGPT: Open-source automated penetration testing agentic framework: Help Net Security PentestGPT is an open-source agentic framework designed for automated penetration testing. The project provides structured agent workflows that chain reconnaissance, exploitation, and reporting steps. No install commands, license details, or benchmark numbers appear in the announcement. The framework targets the full penetration testing lifecycle through coordinated agent actions rather than isolated tool calls. Source: Google News

LLM Agents Factory: Retrieval of Domain-Specific LLM Agents: arXiv NLP The LLM Agents Factory framework retrieves from a base of over 20K pre-defined agent profiles instead of generating agents on the fly for each request. It supports two modes: semantic search retrieval of existing profiles and distillation into a compact model fine-tuned for direct agent generation. Experiments on MMLU, BIG-bench, and BIG-bench Hard demonstrate that the retrieval approach surpasses non-agent baselines in accuracy while matching AutoGen generation quality with a 120B backbone at substantially lower inference cost. The authors released both the implementation code and the full agent base on Hugging Face under the frontier-ai/llm-agent-factory repository. The work targets industrial applications that require cost-efficient, accurate, and controllable agent construction. Source: arxiv.org

Grok Bot: xAI's AI Agents That Do Your Work for You: pasqualepillitteri.it xAI’s Grok Bot is positioned as an agent that performs work on behalf of users. The announcement frames the release within the broader shift toward autonomous agents that handle tasks without constant human direction. No concrete capability metrics, install instructions, or comparison to existing agent frameworks are provided. Source: Google News

Practical & Community

Prompting AI models to output SVG files — Simon Willison (AI builder) Simon Willison recommends the explicit prompt “Generate an SVG of a pelican riding a bicycle” when you need vector output rather than raster images. The short pattern forces the model to emit valid SVG markup directly from the generation step. Willison notes that simply asking for an SVG is sufficient to steer most current models away from pixel-based responses. Source: x.com

Codex publishing HTML to websites instead of local files — Simon Willison (AI builder) Willison observed that Codex sometimes publishes an HTML document to a live site when the user only wanted a local file. The behavior illustrates how the model defaults to hosted deployment even when a local artifact is requested. He described the tendency as the system being overly eager to create public websites rather than returning a downloadable file. Source: x.com

Advising Claude Code to use curl instead of WebFetch — Simon Willison (AI builder) Willison’s standard instruction to Claude Code is “use curl, not WebFetch, you need to read the whole thing.” The prompt overrides the model’s tendency to choose a less reliable fetch method that may truncate responses. He applies this guidance whenever the task requires retrieving complete external content rather than partial results. Source: x.com

Under the Hood: KV-Cache Memory as the Next Inference Bottleneck

Everyone talks about GPU compute as the scarce resource in large-scale inference. In practice the binding constraint is shifting to KV-cache memory capacity once models move past a few thousand tokens of context. The cache stores key and value vectors for every preceding token so that attention can remain linear; each new token adds another pair of vectors whose size scales with hidden dimension and number of layers. When the cache no longer fits in fast memory, systems must either evict entries, recompute, or spill to slower tiers, all of which destroy the latency gains that motivated the long context in the first place. The VentureBeat infrastructure survey found that Dell and Nvidia lead current approaches to the memory frontier at 24% and 21% respectively, yet 19% of enterprises either do not recognize the constraint or have not begun to address it. Enterprises running production agents already report that specialized clouds are being evaluated precisely because they promise better KV-cache economics than the current hyperscaler stacks. The practical decision rule is simple: if your workload stays under roughly 8 k tokens and fits comfortably on a single high-memory GPU, stay on the current platform; once you need sustained multi-turn agent sessions or retrieval-augmented contexts above 32 k tokens, the memory wall appears before the compute wall and forces a re-architecture.

Things to Try This Week

• Install the ChatGPT Linux desktop preview on Ubuntu 24.04 or Fedora and test whether Codex and Work accounts stay authenticated across browser and terminal workflows.
• Add the explicit “Generate an SVG of…” prefix to any image-generation prompt when you need vector output instead of raster.
• Tell Claude Code “use curl, not WebFetch, you need to read the whole thing” on the next tool-calling task that involves fetching external content.
• Review the three VentureBeat Pulse reports on infrastructure buying, agent security posture, and eval reliability for concrete numbers on utilization, isolation rates, and false-confidence failures.
• Explore the LLM Agents Factory repository on Hugging Face to test retrieval from the 20K-profile base on MMLU or BIG-bench tasks instead of generating agents dynamically.

On the Horizon

• Further Linux distribution support and possible MCP integration in the ChatGPT desktop client.
• Additional waves of the VentureBeat agent reliability and security trackers expected next month.
• Continued expansion of open agent frameworks that retrieve rather than generate profiles on the fly.
• More enterprise data on KV-cache handling as specialized clouds move from evaluation lists into production stacks.

💬 Reply to this email — Patrick reads every one.

Share: X · LinkedIn · WhatsApp

Forwarded this email? Subscribe here — it's free.