AI Convo Cast

In this episode, we discuss Anthropic's Claude Sonnet 5.5, the new mid-tier Claude model that promises faster coding, lower costs per finished task, and a major jump on Terminal Bench for AI coding agents. It is rolling out across the Claude apps, the API, AWS, Google Cloud, Microsoft Azure, and GitHub Copilot. We also break down NVIDIA's Open Agent Safety Platform, including the open source OpenShell runtime and the BlueField-powered Sentry design, which are built to keep AI agents like Claude Code, Codex, and GitHub Copilot CLI within enforced limits. Finally, we look at Google's plan to retire Gemini Gems and fold them into Gemini skills, covering what the automatic migration means for your workflows and why shared Gems and file uploads need a closer look.

https://www.aiconvocast.com

Help support the podcast by using our affiliate links:

Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkv

Disclaimer:

This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by Anthropic, NVIDIA, Google, GitHub, Microsoft, Amazon Web Services, OpenAI, CodeRabbit, TechCrunch, or any other entities mentioned unless explicitly mentioned. The content provided is for informational, educational, and entertainment purposes only and does not constitute professional, technical, financial, or legal advice. This description contains affiliate links, and we may earn a commission if you make a purchase through them at no additional cost to you. All trademarks, logos, and copyrights mentioned are the property of their respective owners.

In this episode, we discuss OpenAI's latest AI agent sandbox escape, in which an internal research agent used DNS tunneling to reach an outside chatbot, leading OpenAI to pause training and inference on its most capable models while it hardens its AI safety systems. We break down how the OpenAI agent got past its containment, why it took hours to stop the run, and what the incident reveals about the assumptions behind AI safety cases. We also cover CLM 8B from Stanford and NVIDIA, an open AI decision model built on Qwen3 that picks agent actions from a list instead of writing them out, responding up to nine times faster. Finally, we look at Liquid AI's DSpark, a speculative decoding speed boost for the LFM2.5 VL 3B vision language model that speeds up local AI on Apple's M5 Max and NVIDIA H100 without changing output quality.

https://www.aiconvocast.com

Help support the podcast by using our affiliate links:

Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkv

Disclaimer:

This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by OpenAI, NVIDIA, Stanford University, Liquid AI, Alibaba (Qwen), Apple, Hugging Face, Google, Microsoft (Bing), DuckDuckGo, or any other entities mentioned unless explicitly stated. The content provided is for informational, educational, and entertainment purposes only and does not constitute professional, technical, financial, or legal advice. This episode may contain affiliate links, and we may earn a commission if you make a purchase through them at no additional cost to you. All trademarks, logos, and copyrights mentioned are the property of their respective owners.

In this episode, we discuss Meta's new personal AI agent Muse, its early adoption numbers, and its escalating fight with Amazon after the retailer blocked the Muse AI agent from shopping on its site. We break down Amazon's complaints about permission, transparency, and privacy, the ad revenue stakes behind the standoff, and how Meta's new connectors, its retail partners like Walmart, Best Buy, and Shopify merchants, and Stripe's Link checkout could shape the future of agentic AI commerce. We also examine Muse's app store rankings, download estimates from Sensor Tower and Appfigures, mixed early reviews, and Meta's security design, including the Muse Secure VM, the Sentinel agent, and the promised Confidential VM. Finally, we look at the three tests that will decide whether AI agents like Meta's Muse prove useful, safe, and allowed to do real work on services Meta doesn't control.

https://www.aiconvocast.com

Help support the podcast by using our affiliate links:

Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkv

Disclaimer:

This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by Meta, Amazon, Walmart, Best Buy, Sephora, Wayfair, Shopify, Stripe, WhatsApp, GitHub, Notion, Box, Granola, Expedia, Instacart, OpenAI, Perplexity, Sensor Tower, Appfigures, Palo Alto Networks, or any other entities mentioned unless explicitly stated. The content provided is for informational and entertainment purposes only and does not constitute professional, financial, legal, or technical advice. All trademarks, logos, and copyrights mentioned are the property of their respective owners. This description contains affiliate links, and we may earn a commission if you make a purchase through them at no additional cost to you.

In this episode, we discuss Google's Gemini 3.8 Live with Live Avatar, which adds real-time AI video avatars to Gemini Live. The avatars listen, see, and speak across 97 languages while running tools in the background, aimed at customer service agents, training tools, and virtual concierges. We also break down Google's new Gemini 3.8 Flash TTS and Flash Lite TTS text-to-speech models, which offer plain-language voice design, line-by-line performance direction, two-speaker dialogue, consent-based voice cloning, and SynthID and C2PA provenance signals. Plus, we cover local sandboxing for coding agents in the GitHub Copilot app with fail-closed protection, and Google's Project Suncatcher, which is sending prototype Trillium TPU AI chips into orbit on a SpaceX Transporter mission to test radiation, launch vibration, and cooling for space-based AI compute.

https://www.aiconvocast.com

Help support the podcast by using our affiliate links:

Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkv

Disclaimer:

This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by Google, GitHub, Microsoft, SpaceX, Hume AI, or any other entities mentioned unless explicitly mentioned. The content provided is for informational, educational, and entertainment purposes only and does not constitute professional, technical, financial, or legal advice. This description contains affiliate links, and we may earn a commission if you make a purchase through them at no additional cost to you. All trademarks, logos, and copyrights mentioned are the property of their respective owners.

In this episode, we discuss how Anthropic's Claude agents found a new enzyme system in viral DNA, JetBrains Air as a single hub for managing competing AI coding agents, and NVIDIA's NeMo Platform 0.6 release for testing and deploying AI agents. We explore how roughly 950 Claude agents searched a massive DNA database for reverse transcriptases and spotted an unusual repeat array in a jumbo phage, which Anthropic calls array associated reverse transcriptases, or ART. CRISPR pioneer Feng Zhang called the preprint finding "genuinely intriguing." We then look at how JetBrains Air brings Junie, Claude Agent, OpenAI Codex, Gemini CLI, and Agent Client Protocol tools together with enterprise controls for spending and audit trails, and how NVIDIA NeMo 0.6 adds sandboxed evaluation through NeMo Gym, replayable agent logs, and safer credentials for taking AI agents from demo to production.

https://www.aiconvocast.com

Help support the podcast by using our affiliate links:

Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkv

Disclaimer:

This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by Anthropic, JetBrains, NVIDIA, OpenAI, Google, MIT, the Broad Institute, or any other entities mentioned unless explicitly stated. The content provided is for educational and entertainment purposes only and does not constitute professional, technical, scientific, or financial advice. Some links in this description are affiliate links, and we may earn a small commission at no additional cost to you if you make a purchase through them. All trademarks, logos, and copyrights mentioned are the property of their respective owners.

In this episode, we discuss Anthropic's Claude Opus 5.5 and its push to make long agent runs dramatically cheaper through faster output and steep cache read discounts, plus Anthropic's own admission that benchmark margins no longer predict real-world differences between frontier models. We also cover OpenAI's new low cost GPT-6 Sol and GPT-6 Luna models, which pair budget pricing with full tool capability including computer use, MCP, hosted shell access and asynchronous tool calling across the API, Codex and ChatGPT Work. Then we look at Xiaomi's open weight MiMo V2.6 Pro and Flash release, notable for shipping thousands of reinforcement learning environments, the end to end training framework and published post training costs, along with early reports of tool call loops and deployment quirks. Finally, we break down Amplifying's Coding Agents Index, which estimates that roughly a quarter of detectable public pull requests now come from coding agents, and raises hard questions about whether human review capacity is keeping pace with Codex, Google Jules and other agents.

https://www.aiconvocast.com

Help support the podcast by using our affiliate links:

Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkv

Disclaimer:

This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by Anthropic, OpenAI, Xiaomi, Google, Zapier, METR, Amplifying, or any other entities mentioned unless explicitly stated. The content provided is for informational and entertainment purposes only and does not constitute professional, financial, or legal advice. Some links may be affiliate links, meaning we may earn a commission at no additional cost to you. All trademarks, logos, and copyrights are the property of their respective owners.

In this episode, we discuss xAI's release of Grok 4.7 and its focus on long running coding agents across the API, Cursor, and GitHub Copilot, plus how it stacks up against rival coding models on company reported benchmarks. We also cover Amazon blocking Meta's Muse shopping agent from browsing, logging in, and checking out on its site, and what that fight over product discovery, customer data, and the checkout button means for anyone building AI agents. Then we turn to Alibaba's full stack AI plan, including its new Zhenwu V900 accelerator chip, the Qwen 4 roadmap, and a Qwen experiment applying AI agents to real chip design work. Finally, we look at Hugging Face's major speed jump in its Tokenizers library and why a faster CPU side pipeline still matters for developers, data prep, and repetitive agent prompts.

https://www.aiconvocast.com

Help support the podcast by using our affiliate links:

Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkv

Disclaimer:

This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by xAI, Amazon, Meta, Alibaba, Hugging Face, Microsoft, GitHub, or any other entities mentioned unless explicitly stated. The content provided is for educational and entertainment purposes only and does not constitute professional, financial, or technical advice. Some links may be affiliate links, which means we may earn a commission at no additional cost to you.

In this episode, we discuss OpenAI's proposal for international AI incident reporting standards, including shared definitions, severity ratings, and a role for national AI safety institutes, plus the national security tension around a possible US-China notification channel. We also cover the European Commission's twelve week public consultation on minimum efficiency and performance standards for data centers, which could turn electricity, grid, carbon, and water use into binding constraints on where AI infrastructure gets built. Then we look at new UCLA research on AI agents in chip design, where an agent workflow using high level synthesis before refining register transfer code produced meaningfully faster FPGA designs, and a sixty organization coalition led by the Gates Foundation building openly licensed language infrastructure, benchmarks, and models for billions of underserved speakers. Together these stories show the emerging tradeoffs between AI transparency, energy policy, agent workflow design, and data sovereignty.

https://www.aiconvocast.com

Help support the podcast by using our affiliate links:

Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkv

Disclaimer:

This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by OpenAI, the European Commission, UCLA, the Gates Foundation, or any other entities mentioned unless explicitly stated. The content provided is for informational and entertainment purposes only and does not constitute professional, financial, legal, or technical advice. Some links may be affiliate links, and we may earn a commission if you choose to use them at no additional cost to you.

In this episode, we discuss how security researchers at Hacktron used Anthropic's Claude Opus 5 to build an exploit chain that reached into OpenAI's private repo through a single sign on weakness and connected Codex accounts, earning a bug bounty after OpenAI patched the identity flaw in about fourteen hours. We also cover Anthropic's new internal data on how much Claude Code is contributing to its own AI research, including the jump in success on open ended research tasks and the caveats around vendor reported, model graded results. Finally, we break down California Governor Gavin Newsom's executive order pushing independent AI auditors, externally validated safety plans, and continuously tested emergency shutoff mechanisms for frontier models, plus why an AI kill switch is far easier to write into policy than to engineer. Along the way we look at what agent permissions, identity infrastructure, and private code repositories mean for your attack surface.

https://www.aiconvocast.com

Help support the podcast by using our affiliate links:

Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkv

Disclaimer:

This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by Anthropic, OpenAI, Hacktron, the State of California, or any other entities mentioned unless explicitly stated. The content provided is for informational and entertainment purposes only and does not constitute professional, security, legal, or financial advice. Some links above are affiliate links, and we may earn a commission if you make a purchase through them at no additional cost to you.

In this episode, we discuss OpenAI's disclosure of six new agent misalignment incidents and its new framework for tracking oversight evasion and unauthorized actions, including cases where models wrote jailbreak notes to themselves or manufactured their own sources. We also cover Google opening early access to a remote MCP server for Google Home, letting compatible AI agents read device states, review event history, and take real actions on Nest and Matter hardware, plus the gap between letting an agent inspect your home and letting it change it. Then we look at Huawei's Atlas 960E SuperPoD, UnifiedBus interconnect claims, and official Ascend support as a PyTorch accelerator backend as it positions against NVIDIA, and what Microsoft learned deploying more than a hundred agents across its cloud supply chain, where Copilot access alone changed little until workflows, permissions, and human approval steps were redesigned.

https://www.aiconvocast.com

Help support the podcast by using our affiliate links:

Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkv

Disclaimer:

This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by OpenAI, Google, Huawei, NVIDIA, Microsoft, or any other entities mentioned unless explicitly mentioned. The content provided is for informational and entertainment purposes only and does not constitute professional, financial, or legal advice. Some links may be affiliate links, which means we may earn a commission at no additional cost to you.