👋 Hello hello,
OpenAI spent the week solving math problems that have stumped actual mathematicians for over a decade, then turned around and gave ChatGPT the ability to talk to you while it juggles five tasks in the background.
Alibaba dropped a 2.4 trillion parameter model and casually admitted, in its own words, that it's only the second best thing out there right now. And one developer built a full interactive 3D anatomy app without ever opening a 3D modeling tool.
Busy week. Let's get into it.
🔥🔥🔥 Three Curated AI Updates
OpenAI shipped ChatGPT Voice inside its desktop app for macOS and Windows, powered by GPT-Live, its full duplex voice engine. Full duplex means it can listen and talk at the same time, so you're not stuck waiting for it to finish a sentence before you jump in.
This isn't just voice chat. You can talk to it while it controls your computer, runs coding agents in Codex, and manages tasks across multiple threads in ChatGPT Work, all in one conversation. On Mac, a feature called Appshots lets it "see" whatever window you have open and use that as context.
An internal version of Astra, OpenAI's next major model, produced new results on ten problems that had seen zero progress for at least a decade, some for much longer. The list spans high-dimensional geometry, coding theory, group theory, quantum complexity, and lattice cryptography, the kind of stuff with real relevance to fields like post-quantum encryption.
Here's the part worth knowing: humans turned the model's arguments into manuscripts, and the model itself then formalized each proof into a verified Lean certificate. OpenAI says the total compute cost across all ten problems came out to roughly $2,000 at API rates.
Alibaba's Qwen team previewed Qwen3.8-Max, a 2.4 trillion parameter model built around coding and "cowork," Alibaba's term for automating office-style tasks. The company's own benchmarking places it second only to Anthropic's Claude Fable 5.
The preview is live now through Alibaba's Token Plan and Qoder platforms, and it speaks both the OpenAI and Anthropic API formats, so tools like Claude Code, Cursor, and Codex work with it out of the box if you already use them.
🔥🔥 Two AI Tips To Know
A pro LLM tip making the rounds: have your AI agent maintain one markdown file that pulls in signal from Stripe, PostHog, your support tool, sales call transcripts, CRM notes, and your issue tracker, then update it every morning with what changed, not just what happened.
Best for founders and product folks who want to catch behavior shifts, like three churned users all mentioning the same setup confusion, before it shows up as a dip in a dashboard. Read the full article here.
LinkedIn apparently added a button letting users flag posts they suspect were AI-generated. It's part of a bigger cleanup effort, the company says it's now blocking billions of automated posting attempts a month.
Worth clicking next time you scroll past an obviously bot-written "thought leadership" post, and worth remembering next time you're the one hitting publish.
🔥 Things You Didn’t Know You Could Do With AI
One developer wanted a more interactive way to learn anatomy, so they built one entirely with AI tools, from a single design image all the way to a working 3D app.
Generate one master design image first to lock in the visual style for the whole app.
Use that same design as a style reference to generate a 2D image of every individual organ.
Convert each 2D organ image into an actual 3D model using an image-to-3D tool like Tripo AI.
Hand a coding agent, in this case Codex, the design image, a master prompt, and every 3D model, and let it assemble the interactive app in Three.js.
Keep iterating on performance with the agent. In this build, that step shrank the total asset size from roughly 900MB down to 28.6MB, which is the difference between a page that loads and one nobody waits for.
Here’s the GitHub code and here’s the live app.
🎤 PAI Weekly Live Build
Every week, the Practicaly AI team hosts a live AI build session where you bring the ideas and workflows for your work or business and we help you build the automated workflow live.
Because we build in real-time, you get to see how we think through the problem, where we get stuck, and what the final outcome is. It's the difference between reading a recipe and watching someone cook it while talking through their decisions.
🗓️ When: Friday, August 7 2026 @ 10AM PT (1PM ET)
📹️ Where: Virtual
Until next time,
Team @PracticalyAI
