😺 🎙️ Watch: Hands on with Microsoft Copilot's new features

Welcome, humans.

Imagine an AI assistant that knows your calendar better than you do, can log into LinkedIn on your behalf to pull research, and builds complete PowerPoint presentations while you grab coffee. Kinda sounds like the future, right?

Actually, it's Microsoft Copilot right now.

At Microsoft Ignite 2025, Callie August (Director of Microsoft 365 Copilot) walked us through the biggest Copilot updates yet, and honestly, some of this stuff felt magical

Here's what we cover:

• (0:30) The new Copilot updates announced at Microsoft Ignite and what they mean for everyday professionals
• (1:49) Voice mode that actually works: Ask Copilot to review your day tomorrow, and it pulls from all your emails, chats, and files to brief you.
• (2:17) Live demo: Watch Copilot prepare meeting talking points by pulling from existing docs—all through voice commands.
• (6:56) Computer Use feature: The Researcher agent can now navigate web pages, click links, and even prompt you to log into accounts when needed.
• (10:35) New Word, Excel, and PowerPoint agents built directly into Copilot Chat—start in chat, end with a finished document.
• (13:40) Agent Mode inside Office apps: Toggle it on and Copilot becomes a true collaborative partner within Word, Excel, and PowerPoint.
• (15:40) Excel Agent in action: Building a complete financial monthly close report with formulas, pulling web data, and explaining every step.
• (17:50) Real-world example: Creating a conference itinerary by combining web data, weather, and session schedules.

As for the part that most impressed Corey: Agent Mode working directly inside Excel to build financial reports with full transparency. He called it “something we've been really lacking in the AI space overall, frankly” because it lets users iterate in real-time and validate exactly how the AI built every formula.

You go, Co-co! (our new name for Copilot).

With Microsoft's new Work IQ layer, Copilot now understands your job, your company, your files, and your work patterns. Meaning it’s actively researching, formatting, reasoning, and building complete work products from scratch. Like it should.

P.S. Callie's tip for when Copilot doesn't quite nail it on the first try: Just tell it to “try harder.” Apparently that works. 😂

Keep scrolling for more info on two upcoming livestreams this week (learning languages with AI and an AI marketing tools showcase), plus three other fantastic videos we just released… including our chat with the creator of the “fastest LLM ever built.”

🔴 LATER THIS WEEK: How to use AI to learn a new language & AI marketing tools showcase

Wednesday, Dec 10 | LIVE with Andrew Hsu, Co-founder & CTO of Speak | 2:00pm PST | 4:00pm CST | 5:00pm EST

Wednesday: Andrews Hsu, CTO of language learning app Speak ($100M ARR, 15M+ downloads), shows us how voice AI is replacing flashcards for language learning. Plus we’ll hopefully get an exclusive look at their new tech.

Thursday, Dec 11 | LIVE: AI Marketing Tools Showcase | 11:00am PST | 1:00pm CST | 2:00pm EST

Click this image, then select “Notify Me” on the video to get notified when we’re live.

Thursday: We’re doing an AI Marketing Tools Showcase! Four companies (some backed by Sequoia, Kleiner Perkins, and Khosla) show you what's actually possible with AI in marketing beyond using GPT to write copy or Nano Banana to create images.

Learn how AI is changing ad buying, creator campaigns, campaign messaging, and even how to market to AI systems like ChatGPT (gotta get them bots reading your stuff!)

We’ll reveal the full line-up as we get closer to the day of the launch… get hyped!!

Diffusion Diffusion Models for Text… Inside the AI That Writes 10x Faster Than ChatGPT

In this interview from AWS re:Invent 2025, Corey talks to Stefano Ermon, CEO and cofounder of Inception Labs, to talk about Mercury—a radically different kind of language model that ditches the traditional autoregressive approach (generating one word at a time) for a diffusion-based system (wanna learn what that means? Watch the video!).

Every major AI lab is building the same type of language model. OpenAI, Anthropic, Google: they all use autoregressive transformers that generate text one word at a time.

Stefano Ermon thinks they're all doing it wrong. And his company just raised $50M from the early backers of OpenAI and Anthropic, Microsoft, NVIDIA, and Andrew Karpathy to prove it. Here's a few of our favorite parts:

• (0:55) Diffusion for language explained: Instead of generating text one token at a time, these models start with random noise and iteratively refine it into coherent text—like playing Wordle at scale.
• (1:46) How it actually works: The technical breakdown of how diffusion language models generate entire passages by refining guesses rather than predicting the next word.
• (2:18) The Stanford breakthrough: How Stefano's lab first matched GPT-2 quality while being 10x faster—the proof of concept that led to a company.
• (3:30) Mercury launch: The first commercial-scale diffusion language model that's 5-10x faster than Gemini Flash, Claude Haiku, and GPT-4o Mini while matching their quality.
• (5:17) Different scaling laws: Why diffusion models are more data-efficient and learn more from the same training data than traditional transformers.
• (6:19) The inference advantage: Why the real battle is inference efficiency, and how diffusion models can process many tokens per neural network evaluation instead of just one.
• (7:50) Pareto dominance: The cost vs speed analysis showing diffusion models can either be faster at the same cost or 5-10x cheaper at the same speed—dominating autoregressive models on both dimensions.
• (8:33) RL for diffusion: How reinforcement learning can train these models not just for better answers, but to converge faster—further improving speed.
• (9:48) The $50M backing: Why early OpenAI and Anthropic investors, plus Microsoft, NVIDIA, Databricks, and Snowflake are betting on this approach.
• (11:43) Try it yourself: How to access Mercury through their OpenAI-compatible API, test it in their chat interface, or deploy it on AWS Bedrock.

If Inception Labs is right, we might be watching the beginning of a fundamental architecture shift in AI, from sequential token prediction to parallel iterative refinement. For applications where speed matters (voice agents, coding assistants, real-time chat), diffusion models could be the answer everyone's been looking for.

Try Mercury yourself:

• API access
• Chat interface

IN CASE YOU MISSED IT…

Four new interviews you’ll definitely want to check out (pick whatever looks interesting to you and dive in!):

• Breaking NVIDIA’s GPU Lock-In with Modular: Learn how to run AI models across any hardware (not just NVIDIA) and cut inference costs by up to 70%.
• The Future of Windows in an AI World: See exactly which AI features are coming to Windows, what runs locally vs. cloud, and how to use Copilot in your existing workflow.
• Securing AI (with AI): Learn about the new Security Copilot agents that automate phishing triage and threat detection, and how to prepare for securing your own AI agents.
• Unlocking Video Data with TwelveLabs: Find out how to make your company's video archives searchable and analyzable with video-native AI.

And if you haven’t subscribed yet, please do! Click the image below to go to our channel and hit “subscribe” to get notified right when new videos go live.

We have a goal to hit 10K subscribers by the end of the year, and we’re only 94 away! If you like learning about AI, and already watch some of our videos, do us a favor and click here to subscribe today. We won’t bite… but sometimes, we do scratch…

Stay curious,

The Neuron Team

That’s all for today, for more AI treats, check out our website.

ICYMI: check out our other recent favorite episodes below!

• Osmo’s Alex Wiltschko on teaching computers to smell
• (YouTube, Spotify, Apple)
• OpenAI’s Ahmed El-Kishky on where AI coding goes next
• (YouTube, Spotify, Apple)

What'd you think of this podcast episode?

Pick an answer below, then tell us why with the "additional feedback" option.

P.P.S: Love the newsletter, but don’t want to receive these podcast announcement emails? Don’t unsubscribe — adjust your preferences to opt out of them here instead.