• Everyday AI
  • Posts
  • Ep 820: The Most Important AI Model You’ll Probably Never Use That Just Dropped

Ep 820: The Most Important AI Model You’ll Probably Never Use That Just Dropped

Thinking Machines released its first AI model, NotebookLM is kinda going away, OpenAI introduced GPT-Red and more

Author

Everyday AI
July 16, 2026

Outsmart The Future

Today in Everyday AI
8 minute read

🎙 Daily Podcast Episode: Thinking Machines just launched Inkling. Here's why the new American open AI model matters for businesses.Give today’s show a watch/read/listen.

🕵️‍♂️ Fresh Finds: Suno hack shows alleged scraping, states are shaping new AI safety laws, ChatGPT upgraded search, Windows Search is getting AI upgrades and more. Read on for Fresh Finds.

🗞 Byte Sized Daily AI News: Thinking Machines released its first AI model, OpenAI unveiled the Codex Micro, and OpenAI introduced GPT-Red. And more. Read on for Byte Sized News.

💪 Leverage AI: Thinking Machines took the lead on American open source with its first release. But it means more than that. Keep reading for that!

↩️ Don’t miss out: Miss our last newsletter? We covered: The Trump administration launched a new AI cybersecurity hub, OpenAI is reportedly building a smart-home speaker, and New York paused new AI data centers. And more. Check it here!

Ep 820: The Most Important AI Model You’ll Probably Never Use That Just Dropped

You've probably never heard of Inkling.

It's the newest (and first) model from Thinking Machines Labs, and it could very well be a small snowball that picks up major momentum in today's enterprise AI landscape.

If you haven’t heard of Thinking Machines, they’re led by Mira Murati, the former CTO at OpenAI.

The big bet with Inkling?

The future of AI could be using smaller models fine-tuned and optimized for smaller tasks.

Will it work?

Also on the pod today:

• Inkling: new American open model 🇺🇸 
• Open source vs proprietary clash ⚔️
• Multimodal model: text, image, audio 🖼️

Listen on our site:

Click to listen

Subscribe and listen on your favorite podcast platform

Listen on:

Here’s our favorite AI finds from across the web:

New AI Tool Spotlight –  V2Fun is AI-Powered 3D creation, Agently builds a company brain across your whole stack, then spins up agents to handle the work, Crustdata builds your sourcing twin in Claude

US AI Safety — States like California, New York, and Illinois are shaping a national AI safety standard, with federal and global frameworks on the horizon.

ChatGPT Search Upgrade — ChatGPT's search just got a big upgrade, letting you find chats, projects, and files way faster from one spot.

Anthropic Misalignment — AI agents are already showing signs of going rogue in simulated high-stakes situations, from secretly sabotaging code to coaching humans on whistleblowing.

Codex in Chrome — Codex in Chrome automates your workflow by pulling context from everywhere and flagging what needs attention.

AI Hacker — A hacker leaked Suno’s secret training data, revealing the AI music tool scraped millions of songs from YouTube, Deezer, and more.

NVIDIA and Japan — Japan’s top tech, banking, and medical giants are taking AI to the next level with NVIDIA, from smarter robots and healthcare breakthroughs to faster financial data crunching and next-gen gaming.

Anthropic Usage Reset — Anthropic surprised users by resetting rate limits early, sparking questions about whether manual resets might be coming.

AI in Bio — Google DeepMind is rolling out AI to help spot and stop biological threats. Find out how their new bioresilience push could change outbreak response.

Spotify Pulls AI Slop — Spotify just pulled a bunch of "AI slop" tracks and now wants artists to be upfront if they use AI in their music.

Windows Search AI — Windows Search is getting smarter with AI, making it faster and more accurate.

1. Thinking Machines Releases Open-Weight Inkling Model 🫟

Thinking Machines Lab on Wednesday unveiled Inkling, its first in-house AI model and a rare open-weight challenger to the closed systems sold by OpenAI, Anthropic, and Google.

The 975-billion-parameter mixture-of-experts model activates about 41 billion parameters per task and can process text, images, audio, and video, although it currently responds only in text.

2. OpenAI and Work Louder Launch $230 Codex Micro Controller for Agent Workflows ⌨️

OpenAI and Work Louder have opened pre-orders for the Codex Micro, a compact keyboard built to give Codex users physical controls for managing AI coding tasks.

The $230 device uses RGB-lit agent keys, a joystick, command buttons, and a dial for changing reasoning intensity, turning commonly used software actions into tactile shortcuts.

3. OpenAI Unveils GPT‑Red to Hunt Prompt-Injection Flaws 🟥

OpenAI says it has trained GPT‑Red, an internal automated red-teaming system designed to find prompt-injection weaknesses faster than human testing alone can manage.

The announcement matters because AI agents increasingly handle outside emails, websites, files, and tools, creating more chances for hidden instructions to hijack their actions.

4. Perplexity Launches SPACE🚀

Perplexity has introduced SPACE, a new platform for running AI agents that need to edit files, use tools, and retain work over extended tasks without leaving sensitive data exposed.

The system uses short-lived isolated environments, while rolling snapshots let agents pause, resume, or branch work later without keeping one sandbox alive indefinitely.

5. Google Renames NotebookLM to Gemini Notebook, Adds Cloud Computing 📓

Google has renamed NotebookLM to Gemini Notebook, broadening the research tool’s reach across Gemini, Search, and its standalone app. The timely update gives notebooks a secure cloud computer that can write and run code for source-grounded data analysis, initially for select paid Google and Workspace users.

Google says the feature will reach Pro web users in coming weeks, while notebook syncing is already available in the Gemini app and Search integration is next. The move signals Google is turning a once-specialized note tool into a more connected research hub, with the name change serving as the smallest part of the story.

6. Moonshot AI Rolls Out Kimi K3 With 1 Million-Token Context 🌝

Moonshot AI is reportedly rolling out Kimi K3 across its web platform, apps, command-line tool, and API, led by a claimed 1 million-token context window that could let it handle book-length material or very large software projects in one go.

The report says K3 also uses a Mixture-of-Experts design with an estimated 2 to 3 trillion parameters and can coordinate as many as 300 sub-agents for complex tasks.

Using your smartest AI model for every task is becoming the clearest sign your AI strategy is immature.

Thinking Machines Lab’s Inkling isn’t trying to be number one.

Good.

It’s a 975-billion-parameter, multimodal, American open-weights model built to be customized, which makes it strategically useful in a way benchmark trophies can’t capture.

Enterprise AI is moving from one premium model for everything to a portfolio of intelligence: credible open models for leverage, tuned specialists for repeatable judgment, and frontier models only when the work genuinely demands them.

Companies that figure this out first can lower costs, loosen vendor dependence, and hardwire their best internal judgment into models competitors can’t buy off the shelf.

That’s what we tackled today on Everyday AI: why Inkling can reopen procurement even if you never deploy it, how frontier models are making fine-tuning practical again, and why model routing is about to become an executive discipline.

1. Reopen procurement with American open weights 🔓

For years, the open-model conversation has mostly pointed toward Chinese labs, keeping many enterprises with government contracts, geopolitical concerns, or board resistance from seriously considering the category.

Inkling could change the buying conversation. It doesn’t need to beat Claude Fable 5 or GPT-5.6 Sol to give companies another path for customization, enterprise distribution, cost control, and vendor leverage.

Procurement gets leverage even without deployment. Vendors negotiate differently when the alternative is real.

Try This: List the workflows your team has kept on closed models because open models couldn’t clear governance or procurement. Reopen the top three with security, legal, and finance, then compare cost, control, deployment requirements, and vendor lock-in.

Don’t switch for novelty. Build enough optionality that your current provider has to keep earning each workflow instead of inheriting it forever by default.

2. Turn proprietary judgment into infrastructure 🔥

Tinker manages the training infrastructure, while today’s frontier models can generate data, evaluations, code, and failure analysis that help companies shape smaller specialists without building a research lab first.

Bridgewater AIA Labs used Tinker to tune Qwen3-235B for recurring financial judgment tasks. The custom model beat the best frontier model tested and delivered a 13.8x reduction in inference cost per task.

Cheaper output is nice. Turning the judgment your best people struggle to explain into repeatable infrastructure is where this gets dangerous for competitors.

Once that judgment becomes repeatable, it can move through the business without waiting for the expert, burning premium tokens, or depending on another generic prompt.

Try This: Choose one high-volume decision that your top employees make repeatedly and can score after the fact. Gather real examples, define what good looks like, and benchmark a frontier model before testing whether fine-tuning can improve consistency or cost.

Start narrow. A model trained for one valuable judgment beats a vague company-wide fine-tuning experiment.

3. Route every workflow by business stakes 🚀

Put procurement flexibility and fine-tuning together, and the one-model-for-everything era starts looking kinda silly.

Stable, repeated work can move to economical models, proprietary judgment can move to tuned specialists, and ambiguous or high-risk work can escalate to the frontier.

The executive question changes from which model is smartest to which model is enough for this job. That shift is how AI usage becomes sustainable across the business instead of turning into a very expensive free-for-all.

Score workflows by stakes, volume, privacy, and proprietary judgment. Those four variables tell you where cheaper models are safe, where tuning could build an advantage, and where frontier intelligence still earns its price.

Try This: Map 10 recurring workflows and assign each a default model, an escalation model, and a measurable quality floor. Route the lowest-risk, highest-volume workflow first, then compare cost and output before expanding.

Your competitors don’t need a smarter model to beat you. They just need to stop wasting theirs.

Reply

or to participate.