Tamez Labs

Anthropic has invited Accenture's AI division to observe Claude's training in real time — the first arrangement of its kind at a frontier AI lab. The move raises a genuine question about whether embedded oversight is a meaningful safety commitment or a cleverly structured appearance of one. This episode works through the numbers, the logic, and the tension.

Anthropic launched Claude Docs and Slides on September 16th, letting users build documents and decks directly inside a conversation. Trip and Jack put the new tools through a concrete test — a client-ready deliverable — and trace exactly where the drafting engine holds up and where it falls short.

Suno's V6 update ships section editing, single-lyric swaps, and mashups in one request — a genuine move toward directing a song rather than re-rolling it. Trip and Jack weigh those craft gains against a hard perfection wall, a second abrupt model retirement in five weeks, and a licensing story that's real but partial.

Runway's Solaris generates app interfaces as live video frames rather than HTML and JavaScript — your clicks steer the model, not the DOM. Trip Wire and Jack Lock dig into what that actually means: where the approach earns its wins, and where structural limits like text legibility, zero persistence, and inaccessibility by construction stop it cold. It's research-preview only, so the whole conversation runs on evidence, not hands-on use.

Perplexity AI tripled its annualized revenue to over $750 million in roughly half a year — and most of that surge traces to a single new product that isn't quite search. This episode tracks how a citation-first answer engine is pushing into AI agents, and what a reported Nvidia investment signals about where that heading next.

Velaura AI raised a $110 million Series A on the strength of thirty million chips already in the field — not a pitch deck. The Santa Clara startup licenses a chip IP block called Titan Core that promises two to four times more AI performance per watt, and collects royalties tied to the power savings it delivers.

Trip Wire and Jack Lock pit Canva AI 2.0 against Adobe Firefly on a concrete brief: build a brand-consistent, client-ready campaign asset set — social graphic, slide, and supporting image — that's editable at the component level and commercially safe to ship. The verdict splits cleanly, and the clause that decides it may surprise you.

Trip Wire and Jack Lock put Krea 2's new LoRA training to the test — not the model launch itself, but the narrower question of whether you can take a real, ownable illustration style and make it reliably reusable across new subjects and scenes. They work through the Doodles/Burnt Toast case, independent community evidence, and the honest gaps in pipeline, resolution, and licensing.

Trip Wire and Jack Lock receive a research brief about the Windows 11 August 2026 Patch Tuesday update — and it doesn't fit the show's format. They walk through what's actually in the brief, flag the one AI-adjacent detail worth watching, and explain why they're holding off until there's a real tool to test.

OpenAI released two transcription models — GPT-Transcribe and GPT-Live-Transcribe — just two days before this recording, so real performance data is still thin. Trip Wire and Jack Lock work through what's actually documented: the truncation bug that silently drops audio, phantom name substitutions, billing telemetry that returns zeros, and where live versus batch transcription genuinely differ. The takeaway is grounded in reports from developers hitting real walls, not launch-day marketing.