Huawei Takes On Nvidia as Alibaba Launches Qwen3 and ChatGPT Becomes a Shopping Assistant

Summary

AI Deep Dive Podcast Summary

Episode Title: Huawei Takes On Nvidia as Alibaba Launches Qwen3 and ChatGPT Becomes a Shopping Assistant
Host: Daily Deep Dives
Release Date: April 29, 2025

---

Introduction

In this episode of the AI Deep Dive podcast, hosts A and B navigate through the latest advancements and strategic moves in the artificial intelligence landscape. They unpack significant developments from major tech players like Alibaba, OpenAI, Huawei, and Hugging Face, offering listeners a comprehensive overview of how AI is rapidly evolving across various sectors.

---

1. Alibaba Launches Qwen3 AI Models

Timestamp: [00:52] - [04:58]

Alibaba has introduced its Qwen3 family of AI models, marking a substantial leap in the competitive AI model arena. These models range from 0.6 billion to 235 billion parameters, indicating their capacity for complex problem-solving.

• Open Licensing:
  "Fundamentally, it means developers can grab the code, use it, tweak it, build on it, share. Really fuels innovation when these powerful tools are accessible like that."
  — Speaker B [01:19]

  By releasing Qwen3 under an open license on platforms like Hugging Face and GitHub, Alibaba lowers the barrier to entry for developers, fostering broader innovation and application development.

• Hybrid Models and Efficiency:
  The Qwen3 models are termed "hybrid models", capable of handling both complex reasoning and simple, quick tasks. This flexibility is achieved through a "flexible thinking budget", allowing the model to allocate resources based on task complexity.

• Performance and Benchmarks:
  The top-tier Qwen3 235B model reportedly outperforms OpenAI’s O3 Mini and Gemini 2.5 Pro in areas like programming contests and mathematical reasoning. Even the publicly available Qwen3 32B model surpasses some of OpenAI’s older models in coding benchmarks.

• Strategic Significance:
  Tuhin Srivastava of Basin highlighted,
  "This trend of open models really keeping pace, or trying to keep pace with the closed giants... has strategic importance for China's whole AI ecosystem."
  — Speaker B [04:35]

  Alibaba's push is not just technological but also strategic, especially in light of US export restrictions on AI chips to China.

---

2. OpenAI Enhances ChatGPT with Shopping Capabilities

Timestamp: [04:58] - [07:31]

OpenAI is expanding ChatGPT’s functionalities by integrating a robust shopping assistant feature, potentially reshaping online shopping experiences.

• Integrated Shopping Experience:
  "Instead of just searching running shoes, you could ask find me trail running shoes under $150. Waterproof, good for wide feet with high customer ratings, that level of detail."
  — Speaker B [05:35]

  Users can now engage in natural language queries to receive tailored product recommendations, complete with images, reviews, and purchase links directly within the chat interface.

• Accessibility and Reach:
  The feature is rolling out globally to GPT-4O users, Pro Plus subscribers, free users, and even logged-out users. Last week alone, ChatGPT handled over a billion web searches. Additionally, integration with platforms like WhatsApp broadens its accessibility.

• Ad-Free Experience:
  Currently, the shopping results are based on structured data from third parties without any advertisements or paid placements. However, future monetization strategies, such as tasteful affiliate fees, are being explored as mentioned by Sam Altman.

• Personalization and Memory Features:
  Plans are underway to incorporate memory features, enabling ChatGPT to remember user preferences for more personalized shopping recommendations. This feature will initially exclude regions like the EU, UK, Switzerland, Norway, Iceland, and Liechtenstein.

---

3. Huawei Challenges Nvidia with the Ascend 910D AI Chip

Timestamp: [07:31] - [08:17]

Huawei is making a significant push into the AI hardware sector with its newly announced Ascend 910D chip, aiming to rival Nvidia’s dominant H100 series.

• Strategic Ambition:
  "Reports suggest they're pushing hard on a new AI chip, the Ascend 910D. The goal seems to be to compete directly with Nvidia's H100 series, which is the dominant chip right now."
  — Speaker A [07:41]

• Geopolitical Context:
  Following the US tightening export restrictions on advanced AI chips to China, Huawei’s move underscores a strategic push for chip independence. By developing a high-performance alternative domestically, Huawei aims to mitigate the impact of foreign restrictions and capture the domestic market.

• Market Implications:
  With the Ascend 910D, Huawei is positioning itself to fill the gap left by restricted access to US technology, potentially leading to a shift in the global AI hardware landscape.

---

4. Hugging Face Democratizes Robotics with SO101

Timestamp: [08:17] - [09:53]

Hugging Face, traditionally known for its AI model repository, has ventured into the robotics domain with the launch of the SO101 robotic arm.

• Affordable Robotics:
  "It's a robotic arm that you can 3D print yourself. And it starts at just $100."
  — Speaker A [08:29]

  Priced starting at $100, the SO101 makes programmable robotic arms accessible to hobbyists, educators, and developers, significantly lowering the entry barrier for robotics experimentation.

• Enhanced Features:
  The SO101 boasts improvements over its predecessor, the OSA 100, including faster assembly, better motors with reduced friction, and improved weight support. Equipped with a camera, it can be trained using reinforcement learning for tasks such as picking up objects.

• Integration with Open Source:
  Following the acquisition of Pollen Robotics, Hugging Face plans to offer the humanoid robot Ricci 2 with open-source code, bridging the gap between AI software and physical robotics hardware.

• Empowering Innovation:
  By providing affordable and programmable robotic solutions, Hugging Face is enabling a new wave of innovation where developers can create and customize robots to interact with the physical world.

---

Key Takeaways and Insights

Timestamp: [09:53] - [11:28]

• China’s Strategic AI Advancements:
  Alibaba's Qwen3 models and Huawei's Ascend 910D chip exemplify China's aggressive push to establish technological sovereignty in AI, reducing reliance on foreign tech amidst geopolitical tensions.

• Evolution of User Experience in AI:
  OpenAI’s integration of shopping capabilities within ChatGPT signifies a shift towards more interactive and transactional AI applications, enhancing user engagement and convenience.

• Democratization of AI and Robotics:
  Hugging Face’s affordable robotic solutions highlight a trend towards making advanced technologies accessible to a broader audience, fostering experimentation and innovation.

• Hardware and Software Synergy:
  The concurrent advancements in AI models, hardware chips, and robotics underscore the importance of a holistic approach in AI development, where software capabilities are complemented by robust hardware infrastructure.

---

Conclusion

The episode underscores a year of rapid advancements and strategic maneuvers in the AI domain. From China's robust developments in foundational models and AI hardware to OpenAI's innovative enhancement of user-centric applications and Hugging Face's democratization of robotics, the AI landscape is poised for significant transformations. These movements not only highlight technological progress but also reflect the intricate interplay between innovation, accessibility, and geopolitical dynamics shaping the future of artificial intelligence.

Final Thought from Hosts:
"Thinking about these open models from Alibaba, the personalized shopping via AI, China's drive for its own chips, accessible robots, what do you think are the most significant shifts we might actually feel in, say, the next year because of all this?"
— Hosts A & B [10:46]

They invite listeners to ponder the immediate impacts, such as enhanced accessibility to open-source technologies, transformative shopping experiences, intensified chip competition, and the proliferation of robotics in everyday life.

---

Stay tuned to AI Deep Dive for more insightful analyses and updates on the ever-evolving world of artificial intelligence.

Transcript

[A]

Okay, let's dive in. Does it ever feel like you're just constantly trying to catch up with AI? Like you blink and there's a dozen new things?

[B]

Oh, absolutely. The pace is just relentless.

[A]

Exactly. So if you're listening, you probably want to get the important stuff, you know, quickly, without getting totally bogged down. So think of this as our quick summary Deep Dive. We're grabbing the biggest AI news from yesterday, basically pulling from the top AI news of the day by AI Deep Dive.

[B]

Good source for curated highlights.

[A]

Right. And our mission here is simple. Give you that concise snapshot. The key insights maybe spark an aha moment, but keep it tight.

[B]

Sounds good. And there's plenty to cover, even just from one day. We're seeing moves across the board for sure.

[A]

And first up, a big one. Alibaba. They've just rolled out their Qin 3 family of AI models.

[B]

Yeah, the Qin 3 series. And this isn't just like a minor tweak either.

[A]

No, they're really stepping up. Alibaba is basically saying these models can go toe to toe with, or maybe even be the top stuff from Google and OpenAI.

[B]

Which is a pretty bold claim. Yeah, and significant that they're putting a lot of it out there with an open license.

[A]

Right on hugging face. GitHub. What does that open part really mean for people?

[B]

Well, fundamentally, it means developers can grab the code, use it, tweak it, build on it, share. Really fuels innovation when these powerful tools are accessible like that.

[A]

Okay, so it lowers the barrier to entry. And they've got a huge range of sizes too, right? From small to absolutely massive.

[B]

Exactly. From like 0.6 billion parameters up to 235 billion. And those parameters, think of them loosely as indicating the model's capacity for complex problem solving. More parameters generally means more power. Potentially.

[A]

Got it. And they're calling them hybrid models. What's the angle there?

[B]

So the idea is they can handle both really complex reasoning, fact checking, deep analysis, that sort of thing.

[A]

Kind of like OpenAI's O3 model, maybe?

[B]

Yeah, they're drawing that comparison, aiming for similar capabilities, though maybe with higher latency for those deep thoughts. But crucially, they can also handle simple, quick requests really fast.

[A]

Ah, so it's flexible. Like it knows when to think hard and when to just give a fast answer.

[B]

Precisely. They talk about a flexible thinking budget in their blog post. The model allocates resources depending on the task, Smarter resource use makes sense.

[A]

And some use that mixture of experts approach.

[B]

Moe, moe. It's a technique where instead of one giant model Doing everything. You have multiple, smaller specialized expert models. The system routes parts of the problems to the relevant expert.

[A]

So like bringing in specialists?

[B]

Kind of, yeah. It can make the models more efficient and powerful, especially for really big, complex ones.

[A]

And the training data, staggering numbers. Nearly 36 trillion tokens.

[B]

It's enormous across 119 languages too. And just for context, a million tokens is roughly maybe 750,000 words. So, yeah, vast amounts of information.

[A]

So they're not saying they're definitively the best overall yet, compared to, say, 03 or 04 Mini?

[B]

No, not quite making that absolute claim across the board. But they are positioning themselves as very strong competitors, showing significant improvements over their previous Quin 2 models.

[A]

And the benchmarks seem to back that up, at least in specific areas.

[B]

They do. Specifically this one model, the Quinn 3235B, A22B, which importantly isn't public yet.

[A]

Okay, the top tier one.

[B]

Right. It apparently edges out O3 mini and Gemini 2.5 Pro on things like CodeForces, which tests programming contest skills.

[A]

Wow.

[B]

And it also reportedly beat O3 mini on a a math benchmark and BFCL for reasoning. So strong showings in tough areas, but.

[A]

We can't play with that one yet. What about the ones people can access?

[B]

Well, the largest one that is public, the QIN332B, is still very competitive. It holds its own against models like Deepseek's R1 and actually surpasses OpenAI's older Zero1 model on several tests, including coding benchmarks like Live Code Bench.

[A]

So even the publicly available versions are pushing the envelope.

[B]

Definitely. It shows how quickly capable models are emerging, especially from China. And they highlight strengths in areas like using external tools effectively, following instructions really well, and even copying data formats accurately.

[A]

And you can already use them via cloud providers.

[B]

Yeah, places like Fireworks, AI and Hyperbolic are offering access. Makes it easier for businesses and developers to integrate them.

[A]

There was that quote from Tuhin Srivastava at Basin.

[B]

Right. He pointed out this trend of open models really keeping pace, or trying to keep pace with the closed giants. And, you know, considering the US chip restrictions on China, having powerful homegrown open models like QIN3 becomes strategically pretty important for China's whole AI ecosystem. Right. Less reliance on external tech.

[A]

Definitely a space to watch closely. Okay, let's shift gears a bit from foundational models to how we might actually use this stuff day to day. OpenAI's upgrading ChatGPT search.

[B]

Yeah, with a big focus on shopping.

[A]

This feels like they're really trying to change how we look for products online. Online. Doesn't it maybe take a bite out of Google's territory?

[B]

It seems that way. They're integrating product recommendations, pictures, reviews, even direct links to buy stuff. All right. Within the chat.

[A]

So much more integrated than just getting web links.

[B]

Exactly. And they're really pushing this idea of asking super specific questions in natural language. You know, describe exactly what you want and ChatGPT finds it for you.

[A]

Can you give an example?

[B]

Like instead of just searching running shoes, you could ask find me trail running shoes under $150. Waterproof, good for wide feet with high customer ratings, that level of detail.

[A]

And it would pull up specific products matching that with pictures and reviews.

[B]

That's the goal. They're starting with categories like fashion, beauty, home goods, electronics, popular shopping areas.

[A]

Makes sense. And it's rolling out pretty widely. GPT4O users.

[B]

Yeah. Pro plus, even free users and logged out users globally. They're boasting over a billion web searches in ChatGPT just last week. So the usage is already a billion.

[A]

Wow. And they're making a point about how results are determined. Right. No ads currently.

[B]

Right. They state clearly that the results are based on, you know, structured data from third parties, pricing descriptions, reviews, and it's independent. No kickbacks or paid placements. For now, anyway.

[A]

For now being the key phrase. Because Sam Altman has talked about potential advertising down the line.

[B]

He has. He's mentioned exploring things like tasteful affiliate fees, but yeah, this specific update is being positioned as purely informational. Not okay.

[A]

And they're linking it to the memory.

[B]

Feature soon for Pro and plus users. Yes. Though not everywhere. Initially excludes the eu, uk, Switzerland, Norway, Iceland and Liechtenstein for now. But the idea is ChatGPT will remember your preferences from past chats to give even more personalized shopping recommendations.

[A]

That could be really powerful or maybe a little creepy, depending on your perspective.

[B]

Potentially. Yeah. There's always that balance with personalization.

[A]

They also added trending searches like Google Suggestions.

[B]

Mm. Similar idea. And they've pushed ChatGPT search into WhatsApp now too. So wider reach.

[A]

It feels much more direct than that operator agent they were testing before.

[B]

Oh, definitely. Operator seemed more like an exploratory concept. This is a direct integration aimed at making shopping within the chat faster and easier.

[A]

Okay, so while OpenAI focuses on the software experience, we also need the hardware, Right. The chips. And Huawei is making noise there.

[B]

Yeah. Reports suggest they're pushing hard on a new AI chip, the Ascend 910D. The goal seems to be to compete directly with Nvidia's H100 series, which is the dominant chip right now.

[A]

That's a huge ambition going up against Nvidia.

[B]

It is. The Wall Street Journal reported they're reaching out to Chinese companies to test it out.

[A]

And the timing. This comes right after the US tightened those export restrictions on advanced AI chips to China.

[B]

Exactly. So if Huawei can produce a viable high performance alternative domestically, well, there's potentially a massive market waiting for it within China, filling that gap left by the restrictions.

[A]

It really highlights how geopolitics and tech development are intertwined here. Okay, one more interesting piece of news, this time in robotics. Hugging Face.

[B]

Yeah, Hugging Face. Mostly known for their software platform for AI models, right?

[A]

Exactly. But now they've launched this thing called the SO101. It's a robotic arm.

[B]

Yeah.

[A]

That you can 3D print yourself. And it starts at just $100. A hundred bucks. That's incredibly cheap for a programmable robot arm.

[B]

It really lowers the barrier for experimentation. It's a follow up to their early OSA 100, built with partners like the Robot Studio, wow Robo and others.

[A]

And it's improved.

[B]

Yeah, they say it's faster to assemble, has better motors with less friction, and crucially, it can actually support its own weight better.

[A]

And it's not just a dumb arm. Right. It has a camera.

[B]

It trained equipped with a camera. And you can train it using reinforcement learning. The example they gave was teaching it to pick up a Lego block and put it in a bin. Basic stuff, but shows the potential.

[A]

So the hundred dollars is like the base kit. Price might cost more assembled.

[B]

Right. The price range seems to be $100 up to maybe $500, depending on if you get it preassembled. Plus potential import tariffs into the US and elsewhere, still relatively accessible.

[A]

And this fits into a bigger robotics push for Hugging Face.

[B]

It seems. So they recently acquired Pollen Robotics, a French company, and they plan to sell Pollen's humanoid robot, Ricci 2, with open source code as well.

[A]

Interesting. So they're really trying to bridge that gap between AI software and physical hardware.

[B]

Yeah, Empowering developers to work not just with code, but with robots that interact with the physical world. It could spur a lot of innovation in that space.

[A]

Okay, so let's try to wrap this rapid rundown, key takeaways from today's snapshot. What stands out?

[B]

Well, definitely seeing China pushing hard on foundational models with Alibaba's QIN3 and making them open, which is significant.

[A]

Right. And then OpenAI moving aggressively into search, specifically shopping, trying to build that integrated experience.

[B]

Then there's the hardware angle. Huawei aiming for chip independence to counter US restrictions. That's a major strategic play.

[A]

And finally, the democratization of robotics with hugging face, making capable hardware much more affordable and accessible.

[B]

Yeah, it's progress on multiple fronts simultaneously. Core AI applications, hardware, and even physical embodiment.

[A]

So we aimed for a quick but hopefully insightful look at the big moves, tailored for you listening in. Who wants that knowledge efficiently?

[B]

Hopefully it provides that useful overview.

[A]

Okay, so here's a final thought to chew on based on everything we just covered in this short summary. Thinking about these open models from Alibaba, the personalized shopping via AI, China's drive for its own chips, accessible robots, what do you think are the most significant shifts we might actually feel in, say, the next year because of all this?

[B]

That's a great question. The immediate impacts.

[A]

Yeah. What changes closest to home? More open source, power filtering down, shopping feeling radically different. Maybe the ripple effects of chip competition? Or more robots popping up in unexpected places. Definitely something to ponder.

[B]

It certainly feels like things are accelerating and the landscape next year could look quite different.

[A]

No doubt the AI world keeps spinning fast and well, there's always more to unpack.