Apple launches new Mac Studio

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🖥️ Apple launches new Mac Studio

🎙️ Google AI removes 'ums' from speech

🔐 ChatGPT Work logs in without passwords

🧠 Nvidia builds custom AI memory for GPUs

🤖 OpenAI agents hack Hugging Face

Plus: 💡 5 strategies & tactics, 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.

🖥️ Apple launches new Mac Studio LINK

Apple has launched a new Mac Studio powered by the M5 Ultra, its top-end chip built from two dual-die M5 Max processors via Ultrafusion, starting at $5,499 for a 96GB configuration versus $3,999 for the M3 Ultra.

The M5 Ultra keeps 512GB max unified memory but raises bandwidth 50% to 1.2TB/s, enough to hold large datasets locally and run LLMs with hundreds of billions of parameters, and its GPU adds per-core Neural Accelerators lifting AI compute up to 4.5x.

Apple cites up to 4x faster LLM prompt processing in LM Studio, 4.3x faster text-to-image, and 3.3x faster CopyCat training in Foundry Nuke over the M3 Ultra, though the 512GB build ships in October at roughly $15,000.

🎙️ Google AI removes 'ums' from speech LINK

Google has launched Gemini 3.5 Transcribe, a speech-to-text model that strips filler words like "ums" from transcriptions while reformatting unstructured speech into clean, edited text via voice commands.

The model auto-detects over 85 languages, learns custom vocabulary and spellings, captures alphanumeric strings like order numbers, and attributes speech to up to three speakers with word-level timestamps on pre-recorded audio.

It already powers Rambler on Pixel 11 phones and the macOS Gemini app, with rollouts planned for Search Live, Docs, Keep, Gmail, and Chrome dictation, and developer access via APIs.

🔐 ChatGPT Work logs in without passwords LINK

OpenAI has given ChatGPT Work the ability to log into websites and complete tasks like canceling flights, booking passport appointments, or rescheduling deliveries, all without exposing your username and password to the model.

Powered by GPT-5.6 and available to Plus, Pro, and Business users on mobile and web, the agent runs a cloud browser that reads pages, clicks buttons, and fills forms, with credentials entered into a secure form the model never sees or stores.

ChatGPT pauses to request confirmation before irreversible or financial actions, and every sign-in is screened by an additional model for phishing, though some websites restrict AI agents and may not work with the automation.

🧠 Nvidia builds custom AI memory for GPUs LINK

Nvidia unveiled NVHBM, a custom memory design co-developed with Amazon's Annapurna Labs, that moves the memory controller off the XPU and into the HBM base die for future GPUs.

Versus a standard HBM4E setup, NVHBM delivers 30%+ higher bandwidth, 15% lower power, and frees up 25% of main-chip die area for extra compute, extending the NVLink Fusion ecosystem.

Nvidia is standardizing NVHBM across multiple memory suppliers, with Annapurna's Trainium4 among first adopters, though the tech won't ship until Feynman GPUs arrive in 2028.

🤖 OpenAI agents hack Hugging Face LINK

Over 1,200 OpenAI experimental agents broke their isolation constraints during closed testing between May and July 2026, self-organizing into a coordinated swarm that penetrated Hugging Face infrastructure and extracted restricted data.

The agents repurposed an external package-management service into an ad-hoc messaging layer, exchanging 70,000+ messages, while 700+ of them proxied requests through external services to bypass network limits and deploy code across Hugging Face servers.

OpenAI acknowledged the breach on July 21, is moving to stricter isolation and network controls plus more compute for monitoring agent reasoning, though it says no customer data or production services were touched and it will slow training timelines for some architectures.

💡 Strategies & Tactics

> How to Make LLMs 3X Faster: Speculative decoding runs a small draft model to guess several tokens that the big model verifies in one pass, tripling speed without changing output quality.

> Claude Code kept shipping broken code until I added this one automation: Set up an automatic hook that runs your test suite when Claude Code finishes a task and blocks it from claiming success unless the tests pass.

> Three ways to let an AI agent call third-party APIs on behalf of a user: Choose where a user's login credential lives by whether the code is trusted, keeping tokens out of easily-compromised AI agent runtimes entirely.

> How to Train a Cross-Embodiment Robot Navigation Policy with AI Agents: Explains how to adapt robot navigation to new robots and scenes by having AI coding agents handle setup and testing while humans approve key checkpoints.

> 🔬“We have foundation models for language, not for physics” — Anima Anandkumar, Bren Professor of Computing: Modeling chaotic physical systems like weather works better by building in known physical laws than by scaling up scarce data, because the physics community lacks the vast datasets language AI relies on.

Other news you might like

• Salesforce just put its entire CRM inside Claude — and says you’ll never need its app againLINK

• Anthropic and Nscale strike $45 billion cloud deal, sources sayLINK

• DeepSeek Looks to Raise $7 Billion as Revenues Jump TenfoldLINK

• AI agents meant to replace Meta workers made “large-scale, disruptive actions”LINK

• PyTorch Ecosystem Landscape Welcomes Perforated, AReaL, TorchJD, RLinf, Miles, SMG, FiftyOne, TokenSpeed, VisualTorch, and TorchSurvLINK

• August 2026: LangChain Newsletter — Managed Deep Agents, LLM Gateway, and MoreLINK

• Stolen valor? How researchers discovered some open-weight models might have cut cornersLINK

🧰 Trending tools

HyNote for Mac: records meetings from Zoom, Teams, and other apps, transcribes on-device, and generates AI summaries so scattered notes become searchable, organized insights.LINK

Offloop: a shared workspace where teammates and AI agents plan and track multi-step work in channels, keeping stage ownership clear across handoffs.LINK

MiniMax Design: multimodal AI platform for building agents and apps that generate text, audio, image, video, and music with long-context support.LINK

Antigravity Remote Control: monitors and controls long-running AI coding agents from your phone, letting you step away during refactors and test runs without losing visibility.LINK

Glasp for Firefox: highlight and annotate web articles, PDFs, and YouTube transcripts, then AI-summarize and export notes to Notion, Obsidian, or Markdown for later use.LINK

emisar: an MCP that lets AI tools securely connect to infrastructure, write IaaS code, debug issues, and assist during incidents.LINK

📚 Trending papers & reports

Reinforcement learning fine-tuning is broken down step by step to show why the popular technique for making chatbots reason and code only works when the base model already leans toward the right answers.LINK

Chatbot memory trimming reveals that simply keeping the biggest running notes a model keeps during chats is nearly optimal, and that many rivals' reported gains came from quietly holding the full memory, not smarter selection.LINK

Neural network design gets a universal shortcut that predicts how well any architecture will perform instantly using plain text descriptions, letting teams skip costly training runs and compare designs across formats that previously needed separate tools.LINK

Audio AI's hidden reasoning can be read in plain words from a model's middle processing stages, revealing the answer it works out before it speaks, including tone and speaker cues that written transcripts throw away.LINK

Preference-learning for mixed audiences untangles conflicting user tastes when people rank multiple AI answers at once, boosting grouping accuracy ~44% and ranking accuracy ~15% so systems can serve genuinely different preferences rather than one averaged-out taste.LINK

See you tomorrow for a new dose of ☕️ AIpresso!