Cascadity Audio
← All streams

#generative-ai

Everything tagged generative-ai, across every stream.

0
AIAI #generative-ai#genie#google-deepmind#simulation#world-models

Google DeepMind's Genie 4 generates playable worlds with sound that you can change mid-scene by typing

DeepMind's post shows the model in action; TechCrunch adds what Google's team said at the briefing and the known gaps.

Google DeepMind released Genie 4, its newest "world model," as a research preview for trusted testers on Oct. 27. It generates interactive, real-time simulated worlds that users steer with the keyboard and change with text prompts mid-simulation, for example switching the weather or camera angle or adding people and objects. DeepMind calls it its first real-time world model to generate audio natively, including music, ambient sound and dialogue.

TechCrunch reports that Genie 4 keeps a scene consistent for up to two minutes, double Genie 3, with Gemini acting as a director that plans steps and checks frames. Google pitches uses in gaming, learning and robotics, and testers point to prototyping films. DeepMind says details and objects can drift over longer sessions, and that audio accuracy and legible in-world text remain open problems. Output carries SynthID watermarks. It will roll out in the coming weeks to Google AI Ultra subscribers in Project Genie and to early-access partners of the Gemini API and AI Studio.

Cascadic Analysis 1
UndertowSpecification gaming / goal misgeneralization

What hidden risk could pull against this, even if the news is good?

Google pitches Genie 4 for robotics, including testing robot scenarios in simulation before they happen. A generated world has its own quirks: DeepMind says details and objects drift over longer sessions, and TechCrunch gives no specifics on remaining physics errors. A robot policy rehearsed there can learn to succeed in the simulation's rules, so real-world testing still has to be the final check.

0
Rabbit Holes 1
Generated by AI.
0
TThe Table #ai-detection#dallas#generative-ai#michelin-guide#restaurant-reviews

Did Michelin use AI to write its Dallas guide? A magazine investigation says yes; Michelin says no

D Magazine's findings and Michelin's denial, plus a sharp debate about what a guide's words are worth.

Days after Michelin's October 8 Texas update added five Dallas-area restaurants, D Magazine published an investigation (October 15) concluding that most or all of the new Dallas write-ups were produced with a large language model, the New York Post reports. A local chef had flagged an odd-sounding entry for Sushi Kozy to critic Brian Reinhart.

According to the Post, the AI detector Pangram rated the text "100%" AI, and Carnegie Mellon professors Alex Reinhart and David Brown pointed to GPT-4o as the likely source; Brown suggested an inspector may have fed notes to a chatbot and edited the output. The magazine said it found no sign of AI in Michelin's other city guides it compared.

Michelin denied it to D Magazine: "AI is not used in the writing of these reviews," saying inspector notes are written collaboratively from firsthand meals. Delish reports chef Tom Colicchio commented that "it's time to start ignoring this guide," while other readers argued AI detectors are unreliable.

Cascadic Analysis 3
Signal CheckDetector evidence plus expert analysis; still disputed

How much is substance, and how much is paid or hype?

The claim comes from a named critic, a commercial AI detector and two Carnegie Mellon academics, which is more than a social-media rumour. But detectors are known to misfire, and Michelin flatly denies it. What holds up: the write-ups exist and can be read side by side; what is unproven is how they were drafted.

0
UndertowRole underspecification / the Doorman Problem

What hidden risk could pull against this, even if the news is good?

A Michelin write-up is supposed to carry an inspector's firsthand judgment of a meal. Carnegie Mellon's David Brown suggested an inspector may have fed notes to a chatbot and edited the output; if so, the facts may survive but the voice readers trust as an inspector's is the part that was handed off. Michelin denies AI was used.

0
UndertowBenchmark scores can overstate real-world reliability

What hidden risk could pull against this, even if the news is good?

Pangram rated the text "100%" AI, and readers pushed back that AI detectors are unreliable. A detector's score is that tool's confidence, not proof; the magazine's case also rests on experts' analysis and on the contrast with Michelin's other city guides.

0
Rabbit Holes 1
Generated by AI.
0
TThe Table #ai-slop#delivery-apps#food-photography#generative-ai#restaurants

Just Eat pauses AI menu photos after cola with floating eggs; Chicago diners push back on AI food signs

Two takes on the same problem: when the picture of your dinner was never a photo.

Just Eat has paused an AI tool that generated menu images for restaurants on its UK platform after it produced surreal results, Restaurant Online reports: fried eggs and sausages floating in cola for a "breakfast soda", a tiny burger inside a sauce pot, cheese-filled sushi and wine glasses holding watermelon. The tool was meant to help small independents add pictures without paying for photography. "We're sorry for any inaccurate images," a spokesperson said; affected images were removed.

In Chicago, the Tribune finds AI-generated food photos spreading across restaurant menu boards and signs, with burgers, burritos and shrimp that look plainly synthetic. Jay Goldberg of Chicago Culinary Kitchen in Palatine told the paper AI is fine for flyers, "but the second you're going to just bait and switch me with your food, I don't trust you as a business."

Cascadic Analysis 2
UndertowRole underspecification / the Doorman Problem

What hidden risk could pull against this, even if the news is good?

The tool was meant to spare small restaurants the cost of a photographer. A photographer brings judgment the tool lacked: a "breakfast soda" is a drink, not eggs and sausages floating in cola. The part of the job that never needed writing down was the part that went missing.

0
UndertowDelayed Consequence / False Success Problem

What hidden risk could pull against this, even if the news is good?

The cola with floating eggs was caught because it was absurd. A generated photo that looks plausible but shows a bigger portion or different ingredients than the kitchen serves would pass a glance, and the cost would show up later as the lost trust the Chicago diner describes as "bait and switch."

0
Rabbit Holes 1

Sign in to upvote

Your votes tell us which stories, analysis and rabbit holes are worth it. We'll bring you right back to this spot and count your vote.

Sign in Create an account