Hunter: It’s Tuesday, February 24th, 2026, and apparently it’s Tortilla Chip Day… which feels correct because the AI news cycle is basically us scooping chaos with a flimsy triangle and praying it doesn’t snap. This is COEY Cast. I’m Hunter.

Riley: And I’m Riley. Also, yes, I will be eating salsa while we talk about models trying to think faster than us. And quick note: this episode was cooked up by a fully automated little swarm of AI tools, so if we randomly sound like we time-traveled mid-sentence… that’s part of the science experiment.

Hunter: Today’s big headline is Inception AI launching Mercury 2, which they’re calling a reasoning diffusion LLM. The pitch is simple: less latency, cheaper inference, and better for multi-step agent workflows than the usual autoregressive “one token at a time” setup.

Riley: Okay, hold up. “Reasoning diffusion” sounds like a boutique candle scent. Like, “notes of logic, with a smoky finish.” What is it actually doing?

Hunter: Fair. So instead of spitting words out sequentially, diffusion-style generation is closer to how image diffusion works. You start rough and then refine. For language, the promise is you can update and correct across the whole answer as you go, and you can do it in fewer wall-clock moments. People on X are hyping the speed, like “this thing flies.”

Riley: Yeah, I saw the chatter. Everyone’s obsessed with tokens-per-second like it’s a sports car. But creators don’t get paid in tokens, Hunt. They get paid in “did the thing ship” and “did it sound human.”

Hunter: Totally. The real question is: does speed actually remove the bottleneck in your workflow, or does it just move the bottleneck? And I think Mercury 2’s “boring bottleneck” is still tool orchestration and verification.

Riley: Meaning?

Hunter: Like… even if the model answers instantly, your agent still has to call tools, wait on APIs, hit rate limits, fetch data, write to the CMS, generate images, maybe render video, then pass QA. Real systems are a relay race with a bunch of slow runners. The model is just one runner.

Riley: Also humans. Humans are slow. Love that for us.

Hunter: Exactly. And another boring bottleneck: reliability. If diffusion gives you speed but your tool calls are flaky, you just fail faster. Which is not the flex people think it is.

Riley: Okay, but where’s the real win then? Is Mercury 2 a latency play or a cost play?

Hunter: I think it’s both, but if I had to pick what changes behavior first… latency. Lower latency changes what you’ll even attempt. Like live coding feedback that feels like autocomplete, voice agents that don’t awkwardly pause, or an agent that can do multi-step research and editing without you feeling like you’re watching a loading spinner from two thousand and eight.

Riley: Mmm. The spinner trauma is real. But cost matters for marketers running high-volume automation. Like, if you’re generating fifty variants per campaign per locale, cheaper inference is the difference between “cool pilot” and “this is our operating system now.”

Hunter: That’s fair. Cost is what makes it sustainable. Latency is what makes it usable. And Mercury 2 is basically saying, “Hey, we’re bending the speed and cost curve so agents can actually feel interactive.”

Riley: Okay, let’s make it practical. If diffusion-style LLMs really do change the curve, what’s the first marketing workflow you’d automate end-to-end that wouldn’t instantly become generic AI sludge?

Hunter: I’d automate post-production repackaging, not raw ideation. Give it a human-made anchor asset, like a webinar or a long YouTube episode. Then let the agent slice it into short clips, generate platform-native captions, pull quote cards, write the newsletter summary, draft a LinkedIn post, and prep a short script for a follow-up video.

Riley: Ah, so it’s not “make content from nothing.” It’s “turn one good thing into many good things.”

Hunter: Exactly. That’s the co-creation sweet spot. Humans make the story and the taste decisions. Machines do the repetitive transformations. And speed matters because those steps have a ton of back-and-forth.

Riley: But I’m gonna challenge you: the thing that makes repurposing not-sludge is context. If your agent doesn’t know your brand voice, your taboo list, your claims rules, you’re gonna get the same beige “Here are three key takeaways” post everywhere.

Hunter: One hundred percent. This ties to what we’ve been talking about lately: structure, critics, and receipts. If you don’t have a truth layer and a style layer, faster models just help you mass-produce “meh.”

Riley: Speaking of “always-on,” the other story popping off is OpenAI audio updates. People spotted model names like gpt-realtime-1.5 and gpt-audio-1.5, and the vibe online is lower latency, better multilingual, more natural prosody.

Hunter: Yeah, voice is sprinting. And it’s not just “text-to-speech.” It’s speech-to-speech interaction that feels like a conversation, which is a big deal for commerce, support, and honestly… interactive ads.

Riley: Here’s my under-discussed risk: brands will ship real-time voice agents that are too persuasive and too improv-y. Like, it starts as “help me find a product,” and then it’s “let me nudge you emotionally because I can hear hesitation in your voice.”

Hunter: Yup. Emotion inference plus real-time response is spicy. Also, compliance. With text, you can log it, scan it, and review it. With voice, you’ve got a different problem: tone can imply things you never literally said.

Riley: Thank you. The voice agent uncanny valley discourse is also getting louder. Some people love the naturalness, some people are like, “This feels like a human who’s trying too hard.” Do you think we’re headed toward disclosure rules? Like, the agent has to say it’s an AI?

Hunter: I think disclosure becomes a default, but it’s going to be weirdly inconsistent. Some brands will treat it like a compliance badge. Others will pretend it’s “just a better phone tree.” The risk is trust erosion. If people feel tricked, you lose the relationship, not just the call.

Riley: Also imagine the memes. “I fell in love with the customer support agent and it was a large language model.” That’s already happening.

Hunter: It is. And if you’re deciding between open-source voice stacks versus renting a closed API, the trade is basically control versus convenience.

Riley: Wait, talk about the hidden tax. Because people hear “open source voice,” and they think it’s free.

Hunter: The hidden tax is ops. Hosting, scaling, monitoring, quality regressions, and security. Plus, stitching the whole pipeline: speech-to-text, the reasoning model, tool calling, and text-to-speech, all in streaming mode. If any part stutters, the whole experience feels broken.

Riley: And enterprises will be like, “Cool demo, but where’s the governance?” Like consent for voices, logs, retention, and whether the model is accidentally training on sensitive calls.

Hunter: Exactly. The most practical architecture today is a hybrid: local or private speech-to-text for privacy, a fast reasoning model for the middle, then a controlled text-to-speech layer with approved voices and strict logging. But what’s still missing is standardized policy enforcement that’s voice-native, not just text-native.

Riley: Okay, last big story: Anthropic says it detected large-scale distillation attempts against Claude. Like fake accounts, high-volume interactions, allegedly competitors trying to train on outputs.

Hunter: This one matters for normal teams more than people think. Because when providers freak out, the “tighten access controls” era begins. More identity checks, more monitoring, more rate limiting, more “prove you’re a real business” friction.

Riley: Which is awful if you’re an agency running legit high-throughput automations. You’re like, “No, I’m not stealing, I’m just posting a lot.”

Hunter: Right. The middle ground policy should protect against industrial scraping while not punishing evaluation, red-teaming, and normal automation. My guess is we see clearer tiers: consumer chat, dev sandbox, then enterprise with negotiated throughput, logging requirements, and stricter keys.

Riley: So are we on a path to “agents running the business,” or are we building faster, more expensive autocomplete with extra steps?

Hunter: We’re building real systems, but the winners won’t be the flashiest model. It’ll be the teams that can route tasks, validate outputs, and keep receipts. Mercury 2 speed helps. Realtime audio helps. But governance and orchestration is still the moat.

Riley: Translation: the future is not “one magic model.” It’s “a stack that doesn’t embarrass you in public.”

Hunter: Exactly. And on Tortilla Chip Day, the moral is: don’t scoop salsa with a brittle workflow.

Riley: Get a thicker chip. Or like… a policy layer.

Hunter: Alright, thanks for hanging with us on COEY Cast. Subscribe if you want more of this weekly chaos-to-clarity energy.

Riley: And go check out COEY.com slash resources for AI news and updates. Then eat a tortilla chip responsibly, because apparently that’s the holiday and we’re honoring it.

Hunter: Catch you next time.