Hi there, this is your daily ☕️ AIpresso.
In today's AIpresso:
🤖 OpenAI puts Codex agents on ChatGPT slowdowns
🖥️ Google's Gemini can now control your Mac
📵 Signal adds toggle to block Siri on iOS 27
🌐 Chrome's AI browses a fake internet to protect your real one
Plus: 💡 5 strategies & tactics, 🎁 8 more stories you might like, 🧰 6 tools, and 📚 5 papers.
OpenAI puts Codex agents on ChatGPT slowdowns LINK
- OpenAI built Performance Factory, an internal Codex-powered system that automatically investigates and mitigates performance regressions, and is already running it to keep ChatGPT fast by diagnosing slowdowns across its services.
- The task-based system ran over 1,700 investigations and pinpointed root causes in more than a third of them in about three months, using a reusable perf-forensics skill plus an adversarial reviewer that challenges each diagnosis with fresh context.
- Diagnosis is the strength, but remediation lags: of the ~600 root causes, around 400 reached a proposed fix, yet complex architectural changes still produce overcomplicated or buggy fixes that rarely ship as the exact Codex-authored pull requests.
Google's Gemini can now control your Mac LINK
- Gemini's computer use feature is rolling out in the Mac app, letting the assistant carry out multi-step tasks across local files, desktop apps and websites rather than just describing how to do them.
- Gemini works in an isolated background window with a live chat preview, keeping the main screen free, and sees only folders a user adds to the Connected folders toolbar while separate app permissions govern which programs it touches.
- It requires Apple Silicon, macOS Sequoia 15.0 or later and a Google AI Plus, Pro or Ultra subscription, and is unavailable in the EEA, UK, Switzerland and Nigeria, though Google warns it can permanently delete or share files if it errs.
Signal adds toggle to block Siri on iOS 27 LINK
- Signal is rolling out new protections that stop AI assistants like Siri and Gemini from reading or screenshotting its messages on Android and iOS 27, with the behavior toggleable in settings.
- On Android, Signal auto-enables its Screen security setting when it detects an assistant trying to read the screen, while iOS 27 gets a dedicated Block Siri AI setting that also disables in-app screenshots.
- The move echoes Signal's block of Microsoft's Recall on Windows 11 last year, though Signal admits the fix is incomplete, it can't reliably cover macOS, every Android device, or agents reaching sensitive data through local files.
Chrome's AI browses a fake internet to protect your real one LINK
- Google has built a private, full-scale replica of the web, dubbed the Mirror Dimension, to train and stress-test the Gemini-powered agents behind Chrome's AI browsing features without touching the real internet.
- The copy is stocked with working social networks, shopping sites, and bank dashboards assembled from the same components real sites use, letting agents roam freely, complete assigned tasks, and fend off the team's own security attacks.
- Because the real web follows no single protocol and agents can cause actual harm, this sandbox catches failures before they reach users, though building it required forged certificates and workarounds for Chrome's own safety checks so the browser treats agents as normal visitors.
Strategies & Tactics
> Switching LLMs mid-conversation is a ski-rental problem: Edgee's Agent Gateway decides when a coding session should switch language models by weighing one-time cache-rebuild costs against ongoing savings, avoiding wasteful mid-task swaps.
> Reducing Rollout Bottlenecks in Reinforcement Learning with NVIDIA Dynamo: NVIDIA Dynamo speeds up reinforcement learning by smartly routing requests to reuse cached prompts, sending only changed model weights, and recovering fast from worker failures.
> Build Tool-Calling Parsers in Rust with AI Agents: Use an AI agent and NVIDIA's reusable Rust components to build parsers that reliably turn messy model output into usable tool calls across different AI models.
> Uber’s 4 Claude Code defaults to rein in costs: Set cost-controlling defaults on shared coding tools so engineers don't individually configure them, cutting Uber's per-session AI spending in half.
> Google Previews On-Demand Tool Retrieval for Gemini API Agents: Gemini agents can keep hundreds of tools on standby and load full details only when needed, cutting token costs unless frequent searches outweigh the savings.
Other news you might like
- Apple may be developing its own AI agent that runs your Mac LINK
- Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost LINK
- Report: Nvidia in advanced talks to acquire Stanford-led diffusion AI startup Inception for $500M+ LINK
- Anthropic’s compute spending is increasing by 4x per year LINK
- LangChain Launches LangSmith Tool Gateway to Govern Agent Tool Calls LINK
- Genie 4 is now available LINK
- Advancing chip design with OpenAI models LINK
- Research acceleration: The jagged frontier LINK
Trending tools
Saparo: a personal shopping agent that finds items across U.S. stores, stacks promo codes, cash back, and card offers, then ranks options by total cost. LINK
Rakazo: open-source self-hosted AI agents that run locally, connect to your tools, and keep credentials and audit logs behind your firewall. LINK
readywhen: an AI assistant with its own phone number that handles tasks via calls, texts, or voice notes while tracking your projects and commitments. LINK
Frevana Multiplayer: runs no-code AI agents to find customers, create content, and automate tasks, letting teams share agents and workflows through invite links. LINK
Adside: automates paid ad campaigns across Google, Meta, LinkedIn, Reddit, and Twitter using AI agents for research, creative, and reporting, with optional human management. LINK
ChatGPT Wallet: lets AI agents complete purchases without exposing your real card details, requiring a security check per transaction while preserving card rewards and dispute protections. LINK
Trending papers & reports
Model architecture testing shows that cheap small-scale pretraining hides design differences that only surface at full scale, where a local-mixing tweak cuts training costs ~30% and standard benchmarks mask basic reasoning failures. LINK
Beam coding model delivers frontier reasoning for coding and agent tasks while needing 3-4x less computing power to run, and refuses nearly 89% of harmful requests, the best among openly available models. LINK
Four-bit AI compression runs large language models on Nvidia's next-generation Vera Rubin chips up to ~3x faster than today's standard while clawing back most of the accuracy usually lost at such low precision. LINK
Faster LLM responses come from a lightweight helper that drafts text ahead of the main model, delivering up to 3.1x speedups on servers and 2.2x on a smartphone chip with no extra setup. LINK
LLM lie detectors that compare a model's honest versus deceptive role-play beat cheaper trained detectors at spotting lies across models, but mostly track a reply's wording and often miss lies about withheld knowledge. LINK
See you tomorrow for a new dose of ☕️ AIpresso!