Welcome, humans.

Every major AI model you know of (think of any from ChatGPT, Claude, and Gemini) works the same way: they predict one word at a time, left to right, like the world's most expensive typewriter. This is called being auto-regressive. It works, but it's painfully inefficient.

Meanwhile, Stefano Ermon's lab at Stanford invented a technology called diffusion — the technology behind Midjourney, Stable Diffusion, and Sora.

Now, his company Inception Labs just launched Mercury 2, a reasoning model that applies that same approach to text — and it might be a preview of where the whole industry is headed.

Instead of generating one token at a time, it produces the whole answer at once and then refines it. The result? 1,000 tokens per second on NVIDIA Blackwell GPUs, roughly 10x the throughput of Claude 4.5 Haiku and GPT 5.2 Mini at comparable quality.

The price tag is wild, too:

• $0.25 per million input tokens (half of Gemini 3 Flash, 4x cheaper than Claude Haiku).

• $0.75 per million output tokens (4x cheaper than Flash, 6.5x cheaper than Haiku).

• 128K context window, full tool use, and JSON output support.

In our latest podcast episode, we sat down with Stefano to understand how this process works, why every other lab hasn't switched yet, and what this means for AI agents, coding tools, voice assistants, and even NVIDIA GPUs.

Here's some of our favorite parts:

• (5:20) How diffusion models are trained to fix mistakes instead of predict the next word — and why that matters for inference (running the AI model in production).

• (8:31) The GPT-2 experiment that convinced Stefano to leave Stanford and start a company: same quality, 10x faster.

• (12:35) The GPU bottleneck explained: why every autoregressive model wastes most of its compute — and why throwing more hardware at it can't fix a sequential problem.

• (15:41) At $1 per million output tokens, how Mercury 1 (previous version) made the math work when other models can't (if you didn’t notice, Mercury 2 is cheaper).

• (29:16) The use cases where Mercury dominates right now: coding IDEs, voice agents, customer support — anything where a human can't wait.

• (36:28) Does diffusion fix hallucinations? Stefano's honest answer might surprise you.

• (46:35) What's next: Mercury's new reasoning model (which literally launched today).

Why this matters: Mercury 2 is the first commercial diffusion language model with reasoning capabilities. The key insight = today's LLMs are memory-bound, spending most of their time shuffling data around instead of actually computing.

Diffusion models flip that equation by processing tokens in parallel, making far better use of existing GPUs. As Ermon put it in our interview: for the same hardware, you produce more tokens, so the cost per token goes down.

And Mercury 2 isn't alone in challenging the autoregressive paradigm. Logical Intelligence, founded by Eve Bodnia with Yann LeCun as founding chair of their technical research board, is building energy-based reasoning models that attack a different weakness — accuracy under constraints.

Where diffusion models generate and refine text holistically for speed, energy-based models score partial reasoning traces to catch errors mid-thought, acting like a built-in fact-checker. Their model Kona scored 96% on Sudoku benchmarks where LLMs scored 2%, and their orchestration system Aleph solved 99.4% of PutnamBench, a formal math reasoning benchmark.

The through-line: both approaches reject the one-token-at-a-time bottleneck that every major lab is currently built around. Diffusion gives you speed and efficiency. Energy-based models give you verifiable reasoning. Diffusion makes generation 10x faster. EBMs make verification nearly instant and pinpoint errors instead of just flagging them. Combined, you could get something like (napkin math here) ~30-50x faster reasoning loops at a fraction of current costs — and with mathematical guarantees that the answer is correct (for verifiable domains, of course). Read more here.

Try Mercury's chat playground yourself — make sure you click the "diffusion" button and watch how it generates. Once you see it, you'll understand why we couldn't stop talking about it.

Watch and/or Listen now: YouTube | Spotify | Apple Podcasts

Read more: 

• Our Mercury 2 deep dive 

• Mercury 2 vs. GPT 5 Mini vs. Claude Haiku cost comparison 

• Eve Bodnia: Why AI Needs Energy-Based Models

• Watch Stefano’s segment on TBPN!

Keep scrolling for details on this week’s livestream on AI for therapy, and three more episodes we think you’ll love.

Real quick: Want to see your AI-adjacent product or service show up right here, below these podcast promos? Click the button below to advertise to our 650K readers!

LATER THIS WEEK: 🔴 Can AI actually do therapy — or is it dangerous to even try?

OpenAI got sued. Character.AI got sued. Clinical reports keep raising red flags. And there’s no denying that real harm has happened in many tragic incidents.

Luckily, Daniel Reid Cahn and his team at Slingshot AI raised $93M from a16z to build Ash — an AI purpose-built for therapeutic support from the ground up, trained on structured therapeutic conversations across CBT, DBT, and psychodynamic therapy.

We’re going LIVE this Thursday, Feb 26 at 10AM PT / 1PM ET (YouTube |
LinkedIn | X) to talk to Daniel and learn all about Ash and the different path they are blazing in this very sensitive area.

We plan to ask him: What makes this different from ChatGPT playing therapist? What does the data say about using AI for therapy? And can AI actually, safely expand access to mental health care — or should we keep our counseling on a couch with an actual human?

We’re so excited to talk to Daniel. This will be one of the most important AI conversations we have all year. And seriously, if you are already using AI for therapy and support use-cases today, you’ll absolutely want to hear what Daniel has to say. Don’t miss this one. Click here to get notified when we go live on YouTube on Thursday.

IN CASE YOU MISSED IT…

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

• Google's Secret Coding Tool Just Went Free (Gemini CLI Deep Dive) — The guy who built GitHub Copilot at Microsoft is now at Google, and his team ships 100-150 features a week using AI. (Spotify | Apple)

• Can AI Improve Customer Service Without Killing Jobs? — Matt Price spent 13 years at Zendesk. Now he's building an AI-native CX platform that automates 90% of tickets with 99.8% accuracy. (Spotify | Apple)

• This New AI Model Thinks Without Language — Eve Bodnia of Logical Intelligence explains energy-based models, a completely different approach to AI reasoning. Pairs perfectly with today's episode. (Spotify | Apple)

• We Added Our Own Brain to This Robot (LIVE Demo) — Flexion Robotics showed us what embodied AI actually looks like when it has to walk, grasp, and not fall over.

Last week, we went hands-on with Nikita Rudin of Flexion Robotics who demoed humanoid robotics intelligence live. While yes, we did demo a REAL robot, we also had a great chat that covered the following:

• (2:07) How reinforcement learning trains robots — it takes "tens of years of virtual experience" to learn to walk, but since it's simulated, that's only a few hours of actual compute.

• (18:12) Why most robot demos look painfully slow — they're trained on human teleoperator data, which is inherently clunky. RL-trained robots have the opposite problem: they move too fast and have to be told to slow down.

• (48:11) The surprise: today's vision-language models have basically solved high-level planning for robots. The hard part is still motor control — bending, grabbing, walking on stairs.

• (53:48) Nikita's timeline: robots deployed in factories by end of this year, but the "ChatGPT moment" for robotics won't be one moment like the iPhone reveal — it'll look more like the iPhone-to-Android ecosystem expansion.

Watch the full convo here

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 50K subscribers by the end of the year (if not 100K), and we’re closing in on 20K! If you like learning about AI, and already watch some of our videos, do us a favor and click here to subscribe today.

Stay curious,

The Neuron Team

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.