AI news: OpenAI’s pause, an AI-powered cancer vaccine, and Qwen 3.8 goes local

AI had another busy week, with big moves from OpenAI, a major milestone in AI-assisted cancer treatment, and a powerful new open model you can actually run on a high-end PC. On top of that, we saw new tools for 3D creation, agents, productivity, and even some record-breaking robot athletics.

Trippo 2.0: text and image to 3D models you can actually use

A standout tool this week is Trippo 2.0, an AI model that generates full 3D objects from images or prompts. These aren’t just pretty renders — you can export the meshes, pull them into Blender or Unreal Engine, and even 3D print them.

Trippo can take a photo of a person and generate a 3D mesh with textures. Clothing, poses, and overall proportions are often impressively accurate, though faces can still look a bit uncanny, with distorted lips, noses, or sunglasses. From the neck down, however, the results can be remarkably close to the original photo.

Where Trippo really shines is with fantasy and AI-generated characters. For example, feeding it an image of a humanoid wolf with horns, wings, and torn jeans produced a detailed 3D mesh and textured model that closely matched the concept art. For game devs, hobbyists, and 3D artists, this is already close to “game-ready” asset generation.

Trippo also lets you export different visual styles of the same model:

• Unlit – a flat, lighting-free version that can emphasize structure (sometimes dropping elements like wings down to their underlying bones).
• Cartoon – adds bold outlines for a cel-shaded look.
• Sketch – a more hand-drawn, line-art style.
• Hologram – semi-transparent, sci-fi style that could be great for VFX or in-game ghost/hologram effects.

As of now, Trippo offers two free generations, which is enough to experiment with a couple of characters or props and see if it fits into your 3D or game dev workflow.

AI-assisted mRNA cancer vaccine passes phase 3 melanoma trial

Among all the AI doom-and-gloom headlines, there was a genuinely hopeful story: an mRNA cancer vaccine for melanoma has succeeded in a phase 3 clinical trial.

The trial included 1,137 patients with stage 2B to stage 4 melanoma whose tumors had already been surgically removed. The vaccine, known as mRNA-4157, extended the time patients remained cancer-free compared to standard treatment alone.

While some coverage skipped the AI angle, Moderna’s own materials make it clear that AI and machine learning are central to how this vaccine is designed. The system uses a series of integrated AI algorithms to analyze next-generation sequencing data from tumor and blood samples, identify genetic mutations, and then predict up to 34 neoantigens (tumor-specific targets) that are most likely to trigger a strong immune response.

These models learn over time by pairing clinical data with immunogenicity outcomes, gradually improving their ability to pick the most effective targets. This isn’t generative AI in the ChatGPT sense, but it’s still using neural networks and predictive modeling to design personalized cancer treatments.

It’s a concrete example of AI contributing to real-world medical breakthroughs — the kind of story many people hoped for when AI first entered the healthcare conversation.

Hyper Agent: teams of AI agents you can share with your coworkers

Another trend gaining momentum is multi-agent systems — instead of one general-purpose AI, you spin up a team of specialized agents that collaborate on tasks.

Hyper Agent is a platform built around this idea. You can create agents with specific roles (research, writing, video planning, coding, etc.) and have them work in parallel. Crucially, these agents are shared: once you build one, your entire team can use it in the Hyper Agent app or even talk to it via Slack as if it were a coworker.

One practical example is a B-roll generator agent: you paste in a section of a video script, and the agent outputs a list of photorealistic B-roll ideas, motion graphics concepts, and other visuals tailored to that segment. Because it lives in Hyper Agent, anyone on the team can reuse that workflow instead of rebuilding it from scratch.

Hyper Agent also has a marketplace where you can try prebuilt skills like this, or publish your own internal tools for your team.

Qwen 3.8 27B: a powerful open model you can run locally

Alibaba released a new variant of its Qwen 3.8 family: a 27 billion parameter model called Qwen 3.8 27B. Unlike the massive 2.4 trillion parameter Qwen 3.8 2.4T, this one is actually designed to be run on high-end consumer hardware.

According to independent benchmarks like Artificial Analysis, Qwen 3.8 27B performs better than several popular closed models in agentic tasks, while remaining small enough to run on a GPU with roughly 24–32 GB of VRAM. That means top-tier consumer cards like an RTX 3090, 4090, or 5090, or Apple Silicon machines with large unified memory (e.g., Mac Studio with a high-memory M-series chip) can realistically handle it.

On the Artificial Analysis intelligence index — which combines many benchmarks into a single score — Qwen 3.8 27B lands around the middle of the current frontier pack with a score of 52. It sits just below models like GPT-4.8 Luna Max and DeepSeek V4, and importantly, it’s the only model near that level that’s truly practical to run locally on consumer hardware.

How to run Qwen 3.8 locally with LM Studio

If you want to try Qwen 3.8 27B on your own machine, LM Studio is one of the easiest entry points. It’s a free desktop app for running local LLMs.

The basic workflow:

1. Install LM Studio and open the model search.
2. Search for “Qwen 3.8” and pick the official Hugging Face release.
3. Choose a quantization level (4-bit, 5-bit, 6-bit, or 8-bit). Higher bit = better quality, but more VRAM required.
4. LM Studio will tell you whether full GPU offload is possible on your hardware for each variant.

If your GPU can handle it, aim for the highest-bit version that still fits in memory. There are also “uncensored” variants, which remove safety filters. These will answer almost any question, including harmful or unsafe ones, so they should be handled with extreme caution.

Once downloaded, you can start a chat directly in LM Studio. It behaves like a standard chat model: answer questions, write code, and reason through problems — all fully offline.

Turning Qwen 3.8 into an agent with Hermes

LM Studio itself is just a chat interface. If you want agent-like behavior (breaking tasks into steps, writing and running code, iterating on a project), you can pair LM Studio with an agentic harness like Hermes.

Hermes can detect models exposed by LM Studio as a local server. You start the server in LM Studio, then select that model inside Hermes. From there, Hermes can orchestrate multi-step workflows similar to tools like Claude Code or OpenAI’s Codeex: creating task lists, writing code, and iterating on a project.

In testing, Qwen 3.8 27B was able to work through a full “build a simple game” brief using Hermes, generating code, assets, and a UI. Performance was slower than cloud models — multi-hour runs aren’t unusual — and the final game didn’t fully work, but the direction is clear: local models are rapidly closing the gap with hosted frontier systems.

OpenAI pauses reinforcement learning on its next frontier model

OpenAI published an update titled “Pacing Model Development in an Era of Cyber-Critical Capabilities,” outlining a temporary pause in reinforcement learning (RL) on its latest frontier models.

According to the update, OpenAI has:

• Paused RL training on its newest models intended for deployment, while it strengthens red-teaming and security testing.
• Put its largest planned frontier RL run on hold, instead focusing on smaller-scale training and evaluation to better understand model behavior.

RL is typically a post-training phase where a base model is fine-tuned using human feedback and safety constraints. The fact that RL is paused suggests the underlying next-gen model (often referred to in rumors as “GPT-6” or “Astra”) may already be trained, but OpenAI is holding back deployment until it’s more confident in its cybersecurity and containment properties.

This cautious stance fits into a broader debate about how fast frontier models should be released, and how to handle risks like autonomous hacking, jailbreaks, and large-scale misuse. For more context on the arguments for slowing down frontier AI, see this deep dive on calls to pause advanced AI development.

New OpenAI features: computer history and Apple Messages integration

While the next big model is on hold, OpenAI is still shipping features around its existing stack, especially for its desktop ChatGPT and Codeex apps.

Computer history: a recall layer for your Mac

OpenAI introduced a “computer history” feature that records your activity across apps and websites into a searchable timeline. ChatGPT and Codeex can then reference this history to:

• Answer questions about what you worked on earlier in the day.
• Help you pick up where you left off on a project.
• Spot repeated workflows and suggest turning them into reusable automations or “skills.”

You must explicitly turn this feature on in the ChatGPT desktop app, and you can choose which apps and websites are tracked. You can also pause tracking from the menu bar and delete entries from your history.

The idea is similar to Microsoft’s controversial recall features: your AI assistant watches what you do so it can automate and summarize your work. It’s powerful, but it also raises obvious privacy and data-governance questions.

Apple Messages plugin for ChatGPT

OpenAI also added an Apple Messages plugin to the Mac ChatGPT app. Once installed, ChatGPT can:

• Search your recent messages.
• Summarize or “catch up” on long threads.
• Draft and even send replies directly from within the ChatGPT app.

To enable it, you install the Messages plugin from the ChatGPT app’s plugin section, then grant it access. From there, you can ask things like “Summarize my last five conversations” or “Draft a reply to the last message from [contact].”

Meta, Perplexity, Google, and Claude roll out new productivity tools

Several other major players shipped updates aimed at turning AI into more of a daily work companion.

Meta’s new Mac app

Meta launched a dedicated Mac app for its Meta AI assistant. The interface looks similar to modern AI workspaces like Codeex or Claude’s coworking environment, with:

• Chat history and media generated over time.
• Artifacts and scheduled tasks.
• Image generation and text chat powered by Meta’s models.

It’s still lighter on features than some competitors, but clearly aimed at becoming a central hub for Meta’s AI tools on desktop.

Perplexity Brain: a built-in second brain

Perplexity introduced “Brain,” a second-brain-style knowledge base built into its Perplexity Computer product. It can store and organize information, build a knowledge graph between topics, and help you recall or connect ideas later — similar in spirit to tools like Obsidian.

The concept is very close to the “AI wiki” idea that’s been circulating in the community: a personal knowledge base where AI helps you link, summarize, and resurface information over time.

Google removes visible watermarks on AI media

Google quietly rolled out the ability to remove visible watermarks from AI-generated images, videos, and music. Its invisible SynthID watermark remains embedded, so tools like Gemini can still detect whether content was AI-generated, but the visible “AI” labels are now optional.

Critics worry this will make it easier to spread AI-generated fakes that look like authentic photos or videos, especially since most people won’t run every image through a detector. It’s another example of how usability and transparency can pull in opposite directions.

Claude adds Gmail and Google Drive actions

Anthropic’s Claude now integrates directly with Gmail and Google Drive. You can:

• Ask Claude to reply to email threads and send responses (with configurable approval steps).
• Browse, summarize, and work with files stored in Google Drive.

This brings Claude closer to tools like Codeex that already act as a central hub for email and document workflows. For a broader look at how OpenAI, Anthropic, and Google are positioning their assistants, see our earlier roundup of major assistant updates.

AI for teens and students: ChatGPT and Gemini

OpenAI and Google both made moves aimed at younger users and students.

ChatGPT’s teen mode

OpenAI is rolling out a teen-focused version of ChatGPT. The system tries to infer whether a user is a teenager based on what they say or how they chat. If you explicitly state your age, it will use that; otherwise, it may classify you as a teen based on your language and behavior.

Teen accounts get a more restricted version of the model, with tighter safety and content filters. This raises interesting questions about how much AI students should rely on for homework and learning vs. doing their own research and problem-solving.

Free Google Gemini AI plan for college students

Google announced that eligible college students around the world can get one year of its Google AI plan for free. This includes enhanced Gemini access and new study tools designed to help with coursework, research, and exam prep.

It’s part of a broader trend: AI assistants are quickly becoming standard tools in education, even as schools and universities are still figuring out policies around their use.

New music model: Alibaba’s Happy Shrimp 1.0

Alibaba also released a music generation model with a memorable name: Happy Shrimp 1.0. It’s positioned as a competitor to tools like Suno, letting users generate full tracks from text prompts.

In early tests, Happy Shrimp can produce basic genre tracks — for example, a “future garage” piece or pop-punk songs about specific topics. The output is often generic and not yet at the quality level of the best commercial tools, but it’s another sign that open and regional players are pushing into AI music, especially as some users look for alternatives to more restrictive platforms.

Developer tools: Cursor Origin and agentic coding in Slack

On the coding side, several tools are trying to integrate AI more deeply into the development stack.

Cursor Origin: a GitHub-style hub

Cursor, the AI-first code editor, announced Origin — a GitHub-like platform for hosting code, sending pull requests, and browsing repositories. It can sync directly with GitHub, letting you import existing repos while layering Cursor’s AI features on top.

The move positions Cursor as not just an editor, but a full development environment and collaboration hub, with AI baked into the entire workflow.

Slack adds agentic coding workflows

Salesforce is bringing agentic coding into Slack with a feature called Slack Code. The idea is to let human developers and AI agents collaborate directly inside Slack channels.

Agents can write and review code in-channel, while humans discuss, refine, and approve changes. Under the hood, Slack still relies on external agent harnesses (like Codeex or Claude Code) to run the heavy lifting; Slack becomes the shared interface where everyone — bots and humans — interacts.

The internet is increasingly written by AI

A new Pew-style analysis suggests that a growing share of the web is AI-generated. In early 2022, almost none of the content on the internet showed signs of AI authorship or heavy AI editing. Today, roughly 10% of all pages show AI involvement.

If you filter out older pages published before ChatGPT’s release, the numbers are even starker: about 35% of newer pages appear to be AI-generated or heavily AI-assisted. It’s a strong data point for the “dead internet theory” crowd — and a reminder that a significant portion of what we read online is now produced or shaped by models.

Robots that jump higher and run faster than humans

To cap off the week, we got a glimpse of the future of robotics. A new bipedal robot from Unite Robotics, nicknamed “Superman,” can jump from a standstill to a height of about 2 meters (around 6 feet) and reach running speeds of 12.66 meters per second.

That jump height and speed surpass human world records for both standing high jump and sprinting. Video demos show the robot easily clearing a person’s height and sprinting at speeds no human could match.

Combined with advances in AI control systems, robots like this hint at a near future where autonomous or semi-autonomous machines can outperform humans not just mentally, but physically — with obvious implications for logistics, security, and more.

Where things are heading

This week’s updates paint a clear picture: frontier labs are slowing down visible model releases to grapple with safety and cybersecurity, while the rest of the ecosystem races ahead with tools, agents, local models, and domain-specific systems.

Local models like Qwen 3.8 27B are getting surprisingly strong. AI is quietly powering breakthroughs in medicine. Productivity tools are weaving AI into every corner of our workflows. And even the physical world — from music to robots — is being reshaped by these systems.

With events like Meta Connect and OpenAI’s Dev Day on the horizon, the next big model drops may still be a few weeks away. In the meantime, the incremental upgrades we’re seeing each week are steadily changing how we work, create, and interact with technology.