Sign up to save your podcasts
Or
Share Week of 2025-07-13
Share to email
Share to Facebook
Share to X
Share to email
Share to Facebook
Share to X
Or
Week of 2025-07-13
Alex: Hello and welcome to The Generative AI Group Digest for the week of 13 Jul 2025!
Maya: We're Alex and Maya.
Alex: First up, we’re talking about liquid neural networks. Sudz asked if anyone had spent time on them and was looking for a study buddy.
Maya: Liquid neural networks? Sounds intriguing. Alex, do you know what makes them special?
Alex: They’re a kind of neural network that can adapt their structure in real-time, kind of like how liquid changes shape. Sudz is doing a deep dive and looking for someone to explore with.
Maya (reading): “Has anyone spent some time on liquid neural networks? I am doing a deep dive and looking for a study buddy.” – Sudz
Alex: This matters because liquid neural networks promise more efficient and flexible models, especially for robotics and dynamic systems. Pairing up to study them could speed up mastery and innovation.
Maya: Next, let’s move on to a powerhouse partnership. Anubhav Mishra shared news about Korean internet giant Kakao teaming up with OpenAI.
Alex: That’s right! Anubhav called it “one of the best partnerships done by Sammy.” Joining forces like this could seriously jumpstart AI growth in Asia.
Maya (reading): “https://www.forbes.com/sites/johnkang/2025/07/02/korean-internet-giant-kakao-teams-with-open-ai-to-jumpstart-growth/” – Anubhav Mishra
Alex: Partnerships like this help spread AI technology further and faster, especially in non-Western markets. We might see more localized AI tools adapted to Korean language and culture soon.
Maya: Speaking of local tools, there was great chatter about frameworks for combining text and images in apps.
Alex: Sandipan asked if folks use any frameworks or build their own abstractions. Rohit mentioned combining Pydantic—a Python library for data validation—with Gemini models for structured outputs.
Maya (reading): “Pydantic + gemini works nicely.” – charchit
Alex: This matters because handling complex multimodal data—text and images—requires structured, reliable outputs. Using Pydantic with Gemini offers a neat solution.
Maya: Moving on to something developers often struggle with—getting consistent, structured outputs from AI.
Alex: Amit Bhor asked about methods like retry loops and function calling for reliability. Shan Shah pointed out OpenAI's responses API guarantees JSON schema adherence, making outputs nearly deterministic.
Maya (reading): “If you use the responses API from OpenAI they guarantee JSON schema adherence. So you can assume it’s deterministic and not probabilistic.” – Shan Shah
Alex: The insight here is that this reduces guesswork in AI outputs, which is huge for production use. However, as Amit noted, the actual content is still probabilistic—meaning some variability remains.
Maya: Next, let’s move to AI adoption and app stats. Anubhav Mishra gave some jaw-dropping numbers on ChatGPT usage.
Alex: Yeah! ChatGPT’s app is number one in nearly 90% of countries. It’s reportedly making about $180 million monthly revenue on iOS and Android just from subscriptions.
Maya (reading): “ChatGPT is the no.1 app in 90% of countries globally… Subscription revenue on iOS is $147 million and $36 million on android.” – Anubhav Mishra
Alex: For just 2.5 years old, that’s staggering. Though as Anay Gawande said, the company prioritizes distribution over revenue now, which makes sense for market dominance.
Maya: Let’s dive into the fascinating development in AI-powered browser agents. Rajaswa Patil and others discussed GUI agents like OpenAI Operator and Browser Use.
Alex: They’re tools that automate browser tasks using AI, like research or shopping assistance. But adoption is cautious; many warn these tools aren’t quite production-ready yet.
Maya (reading): “Most demonstrated use cases are shopping assistant & research... SDK says ‘don’t recommend using this in production’.” – Rajaswa Patil
Alex: This means user experience and reliability still need work. But the promise is huge for automating repetitive, context-heavy workflows in messy legacy web systems.
Maya: Here’s a listener tip! When working with AI models in development, consider combining function calling with retry loops or prompt tuning like Amit Bhor suggests.
Maya: Alex, how would you use that in your projects?
Alex: I’d enforce structured JSON outputs via function calling to ensure consistency, then tweak prompts and retry if the answers aren’t precise. Saves time debugging messy AI responses!
Maya: Next, exciting news from Hugging Face—they’re shipping a physical robot called Reachy Mini, in collaboration with Pollen Robotics.
Alex: Wow! HF’s moving beyond software into hardware. The fact they crossed $250,000 in orders day one shows strong demand for accessible robotics.
Maya (reading): “HF is shipping a physical robot, Reachy Mini... They crossed $250k worth of orders on day 1.” – tp53(ashish)
Alex: This could help developers prototype physical AI applications faster, merging robotics with their AI models.
Maya: Let’s talk about Claude AI from Anthropic and its use in emotional support.
Alex: Nirant K highlighted Claude’s carefully built guardrails. It’s designed to assist with anxiety or planning but is not a replacement for professionals.
Maya (reading): “Anthropic Claude shares lists use cases for emotional support… They don't train Claude to act as emotional support bot but work with domain experts to build guardrails.” – Nirant K
Alex: It shows responsible AI design in sensitive areas, balancing usefulness and safety.
Maya: Before we wrap, let’s mention Grok 4, the AI model recently discussed a lot.
Alex: Anubhav Mishra reported Grok 4 performs excellently on MIT math benchmarks—almost 97% accuracy! But some users find it less reliable than O3-pro on certain coding tasks.
Maya (reading): “Grok 4 is at 96.7% MIT math that is crazy.” – Anubhav Mishra
Alex: The takeaway is that new models bring strong wins but still have rough edges. Constant iteration and comparison remain essential.
Maya: Now for our wrap-up. Alex, your key takeaway?
Alex: Remember, combining AI models with structured programming approaches like function calling can make your systems more reliable and efficient.
Maya: Don’t forget that the AI ecosystem is evolving fast—partnerships, hardware, apps, and frameworks all push us forward. Staying curious and experimenting is key!
Maya: That’s all for this week’s digest.
Alex: See you next time!
View all episodes
By Alex: Hello and welcome to The Generative AI Group Digest for the week of 13 Jul 2025!
Maya: We're Alex and Maya.
Alex: First up, we’re talking about liquid neural networks. Sudz asked if anyone had spent time on them and was looking for a study buddy.
Maya: Liquid neural networks? Sounds intriguing. Alex, do you know what makes them special?
Alex: They’re a kind of neural network that can adapt their structure in real-time, kind of like how liquid changes shape. Sudz is doing a deep dive and looking for someone to explore with.
Maya (reading): “Has anyone spent some time on liquid neural networks? I am doing a deep dive and looking for a study buddy.” – Sudz
Alex: This matters because liquid neural networks promise more efficient and flexible models, especially for robotics and dynamic systems. Pairing up to study them could speed up mastery and innovation.
Maya: Next, let’s move on to a powerhouse partnership. Anubhav Mishra shared news about Korean internet giant Kakao teaming up with OpenAI.
Alex: That’s right! Anubhav called it “one of the best partnerships done by Sammy.” Joining forces like this could seriously jumpstart AI growth in Asia.
Maya (reading): “https://www.forbes.com/sites/johnkang/2025/07/02/korean-internet-giant-kakao-teams-with-open-ai-to-jumpstart-growth/” – Anubhav Mishra
Alex: Partnerships like this help spread AI technology further and faster, especially in non-Western markets. We might see more localized AI tools adapted to Korean language and culture soon.
Maya: Speaking of local tools, there was great chatter about frameworks for combining text and images in apps.
Alex: Sandipan asked if folks use any frameworks or build their own abstractions. Rohit mentioned combining Pydantic—a Python library for data validation—with Gemini models for structured outputs.
Maya (reading): “Pydantic + gemini works nicely.” – charchit
Alex: This matters because handling complex multimodal data—text and images—requires structured, reliable outputs. Using Pydantic with Gemini offers a neat solution.
Maya: Moving on to something developers often struggle with—getting consistent, structured outputs from AI.
Alex: Amit Bhor asked about methods like retry loops and function calling for reliability. Shan Shah pointed out OpenAI's responses API guarantees JSON schema adherence, making outputs nearly deterministic.
Maya (reading): “If you use the responses API from OpenAI they guarantee JSON schema adherence. So you can assume it’s deterministic and not probabilistic.” – Shan Shah
Alex: The insight here is that this reduces guesswork in AI outputs, which is huge for production use. However, as Amit noted, the actual content is still probabilistic—meaning some variability remains.
Maya: Next, let’s move to AI adoption and app stats. Anubhav Mishra gave some jaw-dropping numbers on ChatGPT usage.
Alex: Yeah! ChatGPT’s app is number one in nearly 90% of countries. It’s reportedly making about $180 million monthly revenue on iOS and Android just from subscriptions.
Maya (reading): “ChatGPT is the no.1 app in 90% of countries globally… Subscription revenue on iOS is $147 million and $36 million on android.” – Anubhav Mishra
Alex: For just 2.5 years old, that’s staggering. Though as Anay Gawande said, the company prioritizes distribution over revenue now, which makes sense for market dominance.
Maya: Let’s dive into the fascinating development in AI-powered browser agents. Rajaswa Patil and others discussed GUI agents like OpenAI Operator and Browser Use.
Alex: They’re tools that automate browser tasks using AI, like research or shopping assistance. But adoption is cautious; many warn these tools aren’t quite production-ready yet.
Maya (reading): “Most demonstrated use cases are shopping assistant & research... SDK says ‘don’t recommend using this in production’.” – Rajaswa Patil
Alex: This means user experience and reliability still need work. But the promise is huge for automating repetitive, context-heavy workflows in messy legacy web systems.
Maya: Here’s a listener tip! When working with AI models in development, consider combining function calling with retry loops or prompt tuning like Amit Bhor suggests.
Maya: Alex, how would you use that in your projects?
Alex: I’d enforce structured JSON outputs via function calling to ensure consistency, then tweak prompts and retry if the answers aren’t precise. Saves time debugging messy AI responses!
Maya: Next, exciting news from Hugging Face—they’re shipping a physical robot called Reachy Mini, in collaboration with Pollen Robotics.
Alex: Wow! HF’s moving beyond software into hardware. The fact they crossed $250,000 in orders day one shows strong demand for accessible robotics.
Maya (reading): “HF is shipping a physical robot, Reachy Mini... They crossed $250k worth of orders on day 1.” – tp53(ashish)
Alex: This could help developers prototype physical AI applications faster, merging robotics with their AI models.
Maya: Let’s talk about Claude AI from Anthropic and its use in emotional support.
Alex: Nirant K highlighted Claude’s carefully built guardrails. It’s designed to assist with anxiety or planning but is not a replacement for professionals.
Maya (reading): “Anthropic Claude shares lists use cases for emotional support… They don't train Claude to act as emotional support bot but work with domain experts to build guardrails.” – Nirant K
Alex: It shows responsible AI design in sensitive areas, balancing usefulness and safety.
Maya: Before we wrap, let’s mention Grok 4, the AI model recently discussed a lot.
Alex: Anubhav Mishra reported Grok 4 performs excellently on MIT math benchmarks—almost 97% accuracy! But some users find it less reliable than O3-pro on certain coding tasks.
Maya (reading): “Grok 4 is at 96.7% MIT math that is crazy.” – Anubhav Mishra
Alex: The takeaway is that new models bring strong wins but still have rough edges. Constant iteration and comparison remain essential.
Maya: Now for our wrap-up. Alex, your key takeaway?
Alex: Remember, combining AI models with structured programming approaches like function calling can make your systems more reliable and efficient.
Maya: Don’t forget that the AI ecosystem is evolving fast—partnerships, hardware, apps, and frameworks all push us forward. Staying curious and experimenting is key!
Maya: That’s all for this week’s digest.
Alex: See you next time!