🧠 The Model That Robbed a Vending Machine

Inkling is a Mixture-of-Experts model with 975B total parameters (41B active), a 1M token context window, and pretraining on 45 trillion tokens of text, images, audio, and video. Full weights are on Hugging Face right now.

The part worth caring about is the effort dial. You can tune how hard Inkling thinks, and it hits the same score as Nemotron 3 Ultra on Terminal Bench 2.1 using roughly a third of the tokens. For anyone paying per token in a workflow that runs a thousand times a day, that math matters more than a leaderboard spot.

Thinking Machines also says plainly that Inkling is not the strongest model available today, open or closed. It's built as a base to customize, and it's fine-tunable on Tinker at a 50% discount for a limited time. Read more about this here.

GPT-Red is an internal-only red teamer trained through self-play, where it attacks defender models and both sides get better. Prompt injection is the target: a hidden instruction buried in a webpage, an email, or a tool response that hijacks your agent into doing something you never asked for.

The results are uncomfortable. On a prompt injection arena, GPT-Red succeeded on 84% of scenarios against GPT-5.1 while human red teamers managed 13%. It broke the office vending machine agent for real, not in simulation. OpenAI keeps it locked away and never ships it, then uses its attacks to train production models. GPT-5.6 Sol now fails on 0.05% of its direct injections.

If you're pointing agents at your inbox or browser, this is your reminder that the attack surface is every piece of text they read.

Also from OpenAI this week: ChatGPT voice now supports Live Activities on iOS.
Turn on "Background conversations" under Settings > Voice.

Justin McLeod, the guy who founded Hinge, left the company he ran for nearly fifteen years and raised $18M to start Overtone. There are no profiles and no matches to scroll. It learns about you in your own words, then makes an introduction only when it believes two people are genuinely worth putting together.

Everyone comparing it to Black Mirror's "Hang the DJ" is not being subtle, and they're not wrong either.

1. 🧩 Gradial

Gradial is an agentic platform for enterprise marketing operations, meaning the unglamorous work that happens after content is written: authoring into the CMS, tagging assets, QA, accessibility checks, brand compliance. It plugs into Adobe Experience Manager, Jira, Figma, and Salesforce. Best for digital ops teams at big companies where a simple page update takes four days and six approvals.

An AI-native mortgage broker out of New York. You answer a few questions, and its pricing engine shops lenders, reads every quote, and flags buried fees. The company says its rates run more than half a point below the national average with closings around 15 to 17 days. Currently licensed in California, Colorado, and Texas, so this one is for US homebuyers only.

An AI agent loop is just a prompt with a few extra steps where the model checks its own answer against a standard and tries again until it passes. Normally you're the return path: you read the bad draft, say "this is garbage, do it again." A loop hands that job to the model.

Here’s an example of a loop might look like:

The whole thing lives or dies on step 2. Vague standards mean infinite loops and a drained wallet.

Watch this demo to understand how this works.

This week we're hosting a live Q&A with Kushank, and he's answering whatever you throw at him: AI, the tools, the workflows, the stuff that isn't clicking yet. It's a full hour, it's interactive, and it's built around your questions. So bring them all.

🗓️ When: Friday, July 17th @ 10AM PT (1PM ET)
 📹️ Where: Virtual

RSVP takes ten seconds, and the question you send sets the agenda. So bring it on!