Tamez Labs

Anthropic's Claude Code desktop app just added an iOS Simulator Pane that lets the coding agent build, launch, and tap through an app it wrote — closing the does-it-run gap for mobile UI. Trip Wire and Jack Lock break down what that verification loop actually delivers, where it falls short, and what traps to watch for before you trust it at the merge queue.

Walden Robotics spun out of Toyota Research Institute in January and had humanoid robots running eight-hour production shifts by February — a timeline almost unheard of in the field. In July, the Cambridge startup announced a $300 million seed round at a $1.1 billion valuation, backed by Toyota, NVIDIA, Boeing, and Samsung. This episode looks at the Large Behavior Model technology behind the robots, the team that built it, and the open question of whether performance inside Toyota's plants can translate anywhere else.

Northslope was an 88-person applied AI company built by ex-Palantir engineers to deploy custom applications inside large enterprises. About five months after closing a $22 million Series A, OpenAI's deployment subsidiary came calling. The story behind that acquisition reveals a quiet but significant shift in how frontier AI labs think about where the real value gets created.

OpenAI's GPT-5.6 Sol and its Sol Ultra sibling just launched as the company's flagship reasoning-and-code models — but nearly every performance number comes straight from OpenAI's own preview data. Trip Wire and Jack Lock break down what the agentic architecture actually does, where independent reviewers diverged from the vendor chart, and why the model's 1.0 content-publishing score tells most creatives everything they need to know.

Etched is a chip startup that raised $800 million on a simple but radical premise: you can't fix AI inference by swapping one part. This episode traces how three Harvard dropouts went from a 30-page memo that every investor passed on to a five-billion-dollar company with over a billion dollars in signed contracts before shipping a single rack.

OpenAI's new venture, the OpenAI Deployment Company, is placing engineers directly inside client businesses to wire AI into real workflows — and it launched with four billion dollars and a guest list that includes McKinsey. This episode looks at what DeployCo is, how it acquired London-based Tomoro to build its forward-deployed workforce, and what it reveals about where the actual bottleneck in enterprise AI sits.

v0 built a contacts table with search and a modal in three minutes flat, then took fifty more to wire up anything that actually makes it a product. That ratio is the whole game: AI scaffolding is instant, but brand decisions, interaction bugs, and demo-path triage do not compress at the same rate. The founders closing rooms on Monday are the ones who made their design decisions before the first prompt, not the ones who delegated them to the tools and tried to recover later.

ElevenLabs v3 vs Cartesia Sonic 3.5 is the AI voice showdown every podcast producer needs to know about right now, and the timing could not be more lopsided: one model has three months of real creative testing behind it, the other launched two days ago. For a ten-minute two-voice episode, v3 wins today because its inline Audio Tags and Text to Dialogue API actually solve the character differentiation problem at the script level, but Sonic 3.5 has one specific threat: if its voice drift fix holds for narrative audio the way it holds for outbound call centers, the gap closes fast. Neither tool survives paragraph four of grief, and any producer who does not know that going in is going to find out the hard way at the edit stage.

A startup called Fundamental Technologies spent sixteen months in total silence, raised two hundred and twenty-five million dollars, hit a one-point-four billion dollar valuation, and signed Fortune 100 customers at seven-figure contracts before most people even knew it existed. Their argument is brutal and specific: every transformer-based AI model ever built, GPT-4, Claude, Gemini, all of them, is architecturally wrong for the structured tabular data that actually runs enterprise decision-making, because a tokenizer built for language cannot understand a number, a column, or a relational row. The CEOs of Perplexity, Datadog, Brex, and the guy who just sold his company to Google for thirty-two billion dollars all personally wrote checks, which is the kind of signal that is worth paying very close attention to.

Krea just shipped native LoRA training inside the browser on June 1, and suddenly the tool everyone wrote off as a vibe-check platform can actually lock a character's face across a full pose series without touching a single command line. ComfyUI still wins if you need a portable, reproducible pipeline that lives in a JSON file and runs on your own hardware forever, but Krea now gets you to a usable character in hours instead of the days it takes to wire a ComfyUI graph from scratch. The real tell is one line buried in the breakdown: Krea does not let you export your trained LoRA, which means if you build your whole character pipeline there, you do not own it.