ThursdAI - The top AI news from the past week
ThursdAI - Mar 6, 2025 - Alibaba's R1 Killer QwQ, Exclusive Google AI Mode Chat, and MCP fever sweeping the community!
Episode guests
Podcast summary created with Snipd AI
Quick takeaways
- Alibaba's new QW32 billion reasoning model, QwQ, marks a significant advancement in AI technologies through extensive development and reinforcement learning.
- Google's upcoming AI mode in Search, powered by Gemini 2.0, aims to enhance user experience with advanced reasoning capabilities and real-time information access.
- The discussion on Model Context Protocols (MCPs) highlights their potential to revolutionize AI interactions by enabling seamless communication between diverse models and APIs.
- Dina explains the user-friendly process of creating MCP servers, emphasizing how it can enhance productivity and accessibility for developers and non-developers alike.
Deep dives
Host Introduction and Guests
The episode begins with host Alex Volka welcoming listeners and introducing co-hosts Wolfram Ravenwolf and Jun Yang, both of whom have significant experience in AI. Wolfram rejoins after a brief absence due to illness, while Jun is back after being busy leading the Qun team at Alibaba and working on AGI-related projects. The host expresses excitement about discussing the latest developments from Jun regarding the QW32 billion model, a significant release in AI reasoning models. This introduction sets the stage for a deep dive into AI advancements and the impact of newly released models.
Announcement of QW Model
Jun Yang unveils the QW32 billion model, highlighting it as a notable advancement that stems from Alibaba's efforts in building innovative AI technologies. Previously, the QWQ32 billion model was mentioned as merely a play tool, but this new iteration is positioned as a serious product after extensive development work. Through reinforcement learning and experimentation, this model aims to exceed performance expectations, reaching high accuracy on benchmarks such as AIME24. The discussion emphasizes that, while the model is still in a rough state, the results are promising, showcasing comparative performance against competitors like R1, and indicating a strong focus on reasoning capabilities in AI development.
Challenges and Insights in Model Development
During the episode, Jun elaborates on the challenges faced in scaling reinforcement learning for the QW32 billion model, particularly underlining the obstacles encountered while tuning and improving hyperparameters. The use of math and code rewards in reinforcement learning was explored, as the team claimed to have made significant progress in model optimization. Jun revealed that the development team had successful experiments that showed consistent performance improvements over time, indicating the potential for better future iterations. This willingness to acknowledge both the progress and the ongoing journey of model refinement illustrates a transparent approach to AI development.
Collaborative Enhancements to AI Models
The discussion shifts to the broader implications of recent developments in AI tooling and infrastructure, emphasizing collaboration among developers. Both Jun and Alex address the notion of combining different AI models and capabilities to create a more unified and powerful user experience. The desire is to merge reasoning and non-reasoning modes within a single model for versatility. Jun elaborates on their goal to develop multi-modal models that can process diverse input types and perform various tasks, ultimately enhancing the user interaction with AI.
Google's AI Mode and New Features
The conversation transitions to an important segment featuring Robbie Stein from Google, who discusses innovations in AI overviews and the newly introduced AI mode for Google Search. This mode enables Google to leverage AI-powered responses to a wider range of user queries, significantly enhancing the search experience. Robbie highlights that the mode utilizes Gemini 2.0 models to provide advanced reasoning capabilities that can fetch real-time information, thus improving user satisfaction. The introduction of this feature reflects Google's efforts to compete in the AI space by transforming how users interact with search engines and obtain information.
Exploration of MCPs (Model Context Protocols)
The podcast delves into the concept of Model Context Protocols (MCPs) with insights from Jason and Dina, who describe their potential to revolutionize AI interfaces and agents. They explain MCPs as a connectivity protocol that allows different models and APIs to communicate seamlessly, emphasizing the ease of integrating various tools and services. They illustrate how MCPs can lead to more effective and efficient workflows for developers by enabling agents to interact with multiple APIs without requiring extensive customization. The discussion underscores the transformative potential of MCPs in creating collaborative AI environments that enhance productivity.
Creating and Using MCPs
Dina provides a breakdown of the steps involved in creating and deploying MCP servers, highlighting Cloudflare's worker tools that simplify the process for developers. She mentions how basic functions can be wrapped into MCP formats, enabling easy tooling deployment for applications. Dina illustrates the user-friendly nature of MCP setup, stating that users can quickly integrate and customize functionalities without heavy technical knowledge. This accessibility stands to enhance adoption rates among developers and non-developers alike in utilizing advanced AI tools and processes.
Live Demonstrations and Application Examples
Towards the end of the episode, there are live demonstrations showcasing how MCPs can be integrated into different tools and environments, such as Cursor and Windsurf. By demonstrating the ease of installing and configuring an MCP server, the hosts show how applications can quickly leverage external APIs for AI functionalities. This segment illustrates not only the flexibility of MCPs but also reinforces the collaborative nature of modern AI development. By making interfaces more accessible, the podcast highlights the potential for greater integration and innovative use cases as MCP technology continues to evolve.
What is UP folks! Alex here from Weights & Biases (yeah, still, but check this weeks buzz section below for some news!)
I really really enjoyed today's episode, I feel like I can post it unedited it was so so good. We started the show with our good friend Junyang Lin from Alibaba Qwen, where he told us about their new 32B reasoner QwQ. Then we interviewed Google's VP of the search product, Robby Stein, who came and told us about their upcoming AI mode in Google! I got access and played with it, and it made me switch back from PPXL as my main.
And lastly, I recently became fully MCP-pilled, since we covered it when it came out over thanksgiving, I saw this acronym everywhere on my timeline but only recently "got it" and so I wanted to have an MCP deep dive, and boy... did I get what I wished for! You absolutely should tune in to the show as there's no way for me to cover everything we covered about MCP with Dina and Jason! ok without, further adieu.. let's dive in (and the TL;DR, links and show notes in the end as always!)
ThursdAI - Recaps of the most high signal AI weekly spaces is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
🤯 Alibaba's QwQ-32B: Small But Shocking Everyone!
The open-source LLM segment started strong, chatting with friend of the show Junyang Justin Lin from Alibaba’s esteemed Qwen team. They've cooked up something quite special: QwQ-32B, a reasoning-focused, reinforcement-learning-boosted beast that punches remarkably above its weight. We're talking about a mere 32B parameters model holding its ground on tough evaluations against DeepSeek R1, a 671B behemoth!
Here’s how wild this is: You can literally run QwQ on your Mac! Junyang shared that they applied two solid rounds of RL to amp its reasoning, coding, and math capabilities, integrating agents into the model to fully unlock its abilities. When I called out how insane it was that we’ve gone from "LLMs can't do math" to basically acing competitive math benchmarks like AIME24, Junyang calmly hinted that they're already aiming for unified thinking/non-thinking models. Sounds wild, doesn’t it?
Check out the full QwQ release here, or dive into their blog post.
🚀 Google Launches AI Mode: Search Goes Next-Level (X, Blog, My Live Reaction).
For the past two years, on this very show, we've been asking, "Where's Google?" in the Gen AI race. Well, folks, they're back. And they're back in a big way.
Next, we were thrilled to have Google’s own Robby Stein, VP of Product for Google Search, drop by ThursdAI after their massive launch of AI Mode and expanded AI Overviews leveraging Gemini 2.0. Robby walked us through this massive shift, which essentially brings advanced conversational AI capabilities straight into Google. Seriously — Gemini 2.0 is now out here doing complex reasoning while performing fan-out queries behind the scenes in Google's infrastructure.
Google search is literally Googling itself. No joke. "We actually have the model generating fan-out queries — Google searches within searches — to collect accurate, fresh, and verified data," explained Robby during our chat. And I gotta admit, after playing with AI Mode, Google is definitely back in the game—real-time restaurant closures, stock analyses, product comparisons, and it’s conversational to boot. You can check my blind reaction first impression video here. (also, while you're there, why not subscribe to my YT?)
Google has some huge plans, but right now AI Mode is rolling out slowly via Google Labs for Google One AI Premium subscribers first. More soon though!
🐝 This Week's Buzz: Weights & Biases Joins CoreWeave Family!
Huge buzz (in every sense of the word) from Weights & Biases, who made waves with their announcement this week: We've joined forces with CoreWeave! Yeah, that's big news as CoreWeave, the AI hyperscaler known for delivering critical AI infrastructure, has now acquired Weights & Biases to build the ultimate end-to-end AI platform. It's early days of this exciting journey, and more details are emerging, but safe to say: the future of Weights & Biases just got even more exciting. Congrats to the whole team at Weights & Biases and our new colleagues at CoreWeave!
We're committed to all users of WandB so you will be able to keep using Weights & Biases, and we'll continuously improve our offerings going forward! Personally, also nothing changes for ThursdAI! 🎉
MCP Takes Over: Giving AI agents super powers via standardized protocol
Then things got insanely exciting. Why? Because MCP is blowing up and I had to find out why everyone's timeline (mine included) just got invaded.
Welcoming Cloudflare’s amazing product manager Dina Kozlov and Jason Kneen—MCP master and creator—things quickly got mind-blowing. MCP servers, Jason explained, are essentially tool wrappers that effortlessly empower agents with capabilities like API access and even calling other LLMs—completely seamlessly and securely. According to Jason, "we haven't even touched the surface yet of what MCP can do—these things are Lego bricks ready to form swarms and even self-evolve."
Dina broke down just how easy it is to launch MCP servers on Cloudflare Workers while teasing exciting upcoming enhancements. Both Dina and Jason shared jaw-dropping examples, including composing complex workflows connecting Git, Jira, Gmail, and even smart home controls—practically instantaneously! Seriously, my mind is still spinning.
The MCP train is picking up steam, and something tells me we'll be talking about this revolutionary agent technology a lot more soon. Check out two great MCP directories that popped up this recently: Smithery, Cursor Directory and Composio.
This show was one of the best ones we recorded, honestly, I barely need to edit it. It was also a really really fun livestream, so if you prefer seeing to listening, here's the lightly edited live stream
Thank you for being a ThursdAI subscriber, as always here's the TL:DR and shownotes for everything that happened in AI this week and the things we mentioned (and hosts we had)
ThursdAI - Recaps of the most high signal AI weekly spaces is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
TL;DR and Show Notes
* Show Notes & Guests
* Alex Volkov - AI Eveangelist & Weights & Biases (@altryne)
* Co Hosts - @WolframRvnwlf @ldjconfirmed @nisten
* Junyang Justin Lin - Head of Qwen Team, Alibaba - @JustinLin610
* Robby Stein - VP of Product, Google Search - @rmstein
* Dina Kozlov - Product Manager, Cloudflare - @dinasaur_404
* Jason Kneen - MCP Wiz - @jasonkneen
* My Google AI Mode Blind Reaction Video (Youtube)
* Sesame Maya Conversation Demo - (Youtube)
* Cloudflare MCP docs (Blog)
* Weights & Biases Agents Course Pre-signup - https://wandb.me/agents
* Open Source LLMs
* Qwen's latest reasoning model QwQ-32B - matches R1 on some evals (X, Blog, HF, Chat)
* Cohere4ai - Aya Vision - 8B & 32B (X, HF)
* AI21 - Jamba 1.6 Large & Jamba 1.6 Mini (X, HF)
* Big CO LLMs + APIs
* Google announces AI Mode & AI Overviews Gemini 2.0 (X, Blog, My Live Reaction)
* OpenAI rolls out GPT 4.5 to plus users - #1 on LM Arena 🔥 (X)
* Grok Voice is available for free users as well (X)
* Elysian Labs launches Auren ios app (X, App Store)
* Mistral announces SOTA OCR (Blog)
* This weeks Buzz
* Weights & Biases is acquired by CoreWeave 🎉 (Blog)
* Vision & Video
* Tencent HYVideo img2vid is finally here (X, HF, Try It)
* Voice & Audio
* NotaGen - symbolic music generation model high-quality classical sheet music Github, Demo, HF
* Sesame takes the world by storm with their amazing voice model (My Reaction)
* AI Art & Diffusion & 3D
* MiniMax__AI - Image-01: A Versatile Text-to-Image Model at 1/10 the Cost (X, Try it)
* Zhipu AI - CogView 4 6B - (X, Github)
* Tools
* Google - DataScience agent in GoogleColab Blog
* Baidu Miaoda - nocode AI build tool
This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit sub.thursdai.news/subscribe
Remember Everything You Learn from Podcasts
