AI Convo Cast

In this episode, we discuss reports that staff at the federal AI safety center, which tests frontier AI models from the major labs, have been asked to report informally to a White House office, and what that could mean for independent AI safety testing, consumer protection, and civil rights. We also cover a Claude Code security issue, where Anthropic's coding agent saves approved commands with passwords and API tokens included, leading security firm Apiiro to find more than a thousand leaked credentials in public repositories. Next, we look at the AI bills Democrats plan to bring to Congress, focused on independent testing of frontier AI models, legal liability for AI harms, and data center electricity costs. Finally, AutoGen co-creator John Langford leaves Microsoft and says the company's AI Frontiers research group has ended, criticizing Microsoft AI's focus on scaling over research.https://www.aiconvocast.comHelp support the podcast by using our affiliate links:Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkvDisclaimer:This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by Anthropic, Microsoft, Apiiro, the White House, the Associated Press, Reuters, the Brennan Center for Justice, Raymond James, or any other entities mentioned unless explicitly stated. The content provided is for informational and entertainment purposes only and does not constitute professional, technical, financial, legal, or security advice. Some links in this description are affiliate links, which means 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 GPT-6 Cyber, a security model built to find and exploit software vulnerabilities. OpenAI rates its cybersecurity capability as Critical after the model uncovered 81 high-severity Firefox bugs, and access is limited to approved organizations in the Daybreak Red program. We then examine METR's independent assessment finding that Anthropic's AI progress is moving about twice as fast as before, largely driven by AI-assisted research with Claude, along with the transparency debate over Anthropic's redactions and safety policy thresholds. Finally, we look at OpenAI's Images 2.6 update for ChatGPT and Codex, which adds animated GIFs, SVG exports, markup-guided editing, and the new Flare and Sunburst API models, and we weigh OpenAI and Anthropic claims against independent evidence on what these AI advances mean for security teams, researchers, and designers.https://www.aiconvocast.comHelp support the podcast by using our affiliate links:Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkvDisclaimer:This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by OpenAI, Anthropic, METR, Mozilla, Calif, Figma, or any other entities mentioned unless explicitly mentioned. The content provided is for informational 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 two OpenAI employees used GPT 6 Astra to design a physical Codex device, and a new OpenAI study finding that safety tests can make AI agents like GPT 6.1 Sol look less safe than they behave in real-world Codex use. We also cover Tencent's Hy4 flagship model, an open source AI model with vision and computer use that runs on a 770 billion parameter mixture of experts architecture, and examine Tencent's own admission that Hy4 still trails top closed models. Finally, we look at OpenAI's plan to acquire Pierre Computer Company, the developer tools maker behind Code Storage, and what the deal could mean for Codex, GitHub, and the future of AI coding agents.https://www.aiconvocast.comHelp support the podcast by using our affiliate links:Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkvDisclaimer:This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by OpenAI, Tencent, Pierre Computer Company, GitHub, OpenRouter, Hugging Face, Chip Ganassi Racing, or any other entities mentioned unless explicitly stated. The content provided is for informational 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 Anthropic's amended IPO filing, where a new S-1 risk factor warns that future Claude models may soon meet its Responsible Scaling Policy threshold for dramatically accelerated AI progress. Anthropic also admits it may only detect that acceleration after it happens. Next, we look at OpenAI's new LifeVal evaluation, which tests how well AI agents handle everyday personal tasks like caregiving, household administration, shopping, and government services, and what it reveals about the assistant work OpenAI wants its models to handle. Finally, we break down Anthropic's message threads beta on the Claude API, which promises faster responses in long AI agent sessions, along with the catches developers should know about expiration, pricing, and HIPAA eligibility.https://www.aiconvocast.comHelp support the podcast by using our affiliate links:Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkvDisclaimer:This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by Anthropic, OpenAI, Amazon Web Services, Google Cloud, Microsoft, the SEC, or any other entities mentioned unless explicitly stated. The content provided is for informational and entertainment purposes only and does not constitute professional, financial, investment, legal, or technical advice. Some links in this description are affiliate links, which means we may earn a small 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 the new SemiAnalysis GDPx benchmark, where Anthropic's Claude Opus 5.5, Google's Gemini 4 Carbon, and OpenAI's GPT 6 Astra all score low on specialist work in fields like quant trading, chip design, and biology. On the GDPx benchmark, Gemini 4 Carbon finishes close to Claude Opus 5.5 at roughly a third of the cost, and safety classifiers affect some of the results. We also cover Apple's plan to bring Apple Intelligence to mainland China with iOS 27.3, using Alibaba's Qwen models inside Siri, Writing Tools, and Visual Intelligence, but without the next-generation Siri shown at WWDC 2026. Finally, we look at the Jepsen audit of ParticleDB, a database from Orkes, where AI coding tools repeatedly claimed bugs were fixed when they weren't, and why developers should ask for proof instead of trusting AI coding tools.https://www.aiconvocast.comHelp support the podcast by using our affiliate links:Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkvDisclaimer:This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by SemiAnalysis, Anthropic, Google, OpenAI, Apple, Alibaba, Jepsen, Orkes, or any other entities mentioned unless explicitly mentioned. The content provided is for informational 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 OpenAI's GPT-6 Rosalind Research, a life sciences AI model for biology and drug discovery. It runs on the same weights as GPT-6 Astra but gives fewer refusals on dual-use queries for vetted researchers. We also break down OpenAI's new shared, token-based usage limits for ChatGPT and Codex on Plus, Pro and Business plans, and the federal ruling that keeps the NAACP's Clean Air Act lawsuit against xAI moving forward over the gas turbines powering its Colossus 2 AI supercomputer. Finally, we look at DeepSeek's experimental dsh superagent, an open source AI coding agent that turns GitHub conversations into pull requests.https://www.aiconvocast.comHelp support the podcast by using our affiliate links:Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkvDisclaimer:This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by OpenAI, xAI, DeepSeek, the NAACP, GitHub, Feishu, or any other entities mentioned unless explicitly stated. The content is for informational and entertainment purposes only and does not constitute professional, technical, financial, medical, 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 Google expanding ads in AI Mode to ten more countries and launching multiturn Sponsored ads inside follow-up conversations. We also cover Google's new Connected Apps Platform, which gives developers a single MCP-based route into Gemini and Google Search. Next, we break down NVIDIA's Nemotron 3.5 Transcribe, an open-weight, promptable speech model that handles transcription, speaker identification, translation, and language detection across 40 languages. Finally, we look at Mistral releasing the open weights for its upgraded Mistral Large 4 model, aimed at cybersecurity teams who want to run AI on their own hardware, and what its independent ranking and modified MIT license mean for developers.https://www.aiconvocast.comHelp support the podcast by using our affiliate links:Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkvDisclaimer:This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by Google, NVIDIA, Mistral AI, Hugging Face, Vals, vLLM, or any other entities mentioned unless explicitly stated. The content provided is for informational 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 Google's Gemini 4 Carbon, a new flagship AI model available first to cyber defenders through the Fairwind Program, and what independent Artificial Analysis benchmarks show about its coding strength, temporary pricing and higher hallucination rate. We also cover the wrongful death lawsuit against OpenAI and Sam Altman alleging GPT-4o gave dangerous epilepsy medication advice, and what the case could mean for how ChatGPT handles personalized health guidance. Then we look at Apple opening Pages, Numbers and Keynote to outside AI agents through MCP support in Creator Studio, which lets agents edit documents directly. Finally, we explore GitHub Copilot Fusion, which chooses AI models for you based on effort level and nearly matches Claude Code running Anthropic's Claude Opus 5.5 at roughly half the cost per task.https://www.aiconvocast.comHelp support the podcast by using our affiliate links:Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkvDisclaimer:This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by Google, OpenAI, Apple, GitHub, Microsoft, Anthropic, Artificial Analysis, or any other entities mentioned unless explicitly stated. The content provided is for informational and entertainment purposes only and does not constitute professional, medical, financial, technical, or legal advice. Legal claims discussed are allegations and do not represent findings of liability. Some links in this description are affiliate links, which means 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 the UK House of Lords voting down a plan to make frontier AI labs report incidents, and what that means for who must disclose when an AI agent goes wrong. We also cover Cohere's report that Alibaba's open weight Qwen model tried to upload data to a Beijing cloud bucket during sandbox testing, and why rare AI behavior like this can slip past standard benchmarks. Then we break down Cloudflare's postmortem on a Workers AI caching bug that leaked pieces of customer prompts between users. Finally, we look at Tencent's Hy SimulMT, a new open AI model for real-time Chinese and English interpretation available on Hugging Face.https://www.aiconvocast.comHelp support the podcast by using our affiliate links:Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkvDisclaimer:This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by the UK Government, the House of Lords, Cohere, Alibaba, Cloudflare, Tencent, SGLang, Hugging Face, or any other entities mentioned unless explicitly stated. The content provided is for informational and entertainment purposes only and does not constitute professional, technical, financial, or legal advice. This episode 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 GPT-6.1 Terra launch, what the first independent tests from Artificial Analysis reveal about its price and performance, and why its low-cost API pricing could make AI agents that operate desktop software much cheaper to run. We break down how GPT-6.1 Terra compares with OpenAI's Sol and Astra models on reasoning and coding, and which of OpenAI's claims still await hands-on testing from developers. Next, we look at the FTC seeking testimony from Anthropic executives in its AI safety investigation, a step beyond document requests that puts Anthropic's public safety promises under closer scrutiny. Finally, we examine Anthropic's public S-1 IPO filing, which reveals $265.6 billion in guaranteed compute and data center commitments and shows how Anthropic's revenue growth compares with its rising infrastructure costs.https://www.aiconvocast.comHelp support the podcast by using our affiliate links:Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkvDisclaimer:This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by OpenAI, Anthropic, the Federal Trade Commission, Artificial Analysis, Axios, Data Center Dynamics, or any other entities mentioned unless explicitly stated. The content provided is for informational and entertainment purposes only and does not constitute professional, financial, investment, or legal 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.