This is legit hilarious: opusfived.dev/

Hey I made some progress getting Claude to talk normal! I made a bunch of edits to a doc and then asked it to look at my edits and analyze how the changes. Here is what Claude said: Here's the pattern I read in your edits, so I can write this way from the start: You cut the clause that tells the reader what to think about the sentence they just read. "which is where the gap actually is," "and it arrives with no media cost attached to it," "It is not collecting on those payments," "and it pays for the sales motion." The point was already made; the trailing clause just underlines it. You cut throat-clearing setups. "This is worth stating plainly, because it changes what the fix is" and "There is a second reason this matters, and it goes to why you bought the business" both disappeared entirely. Start the paragraph with the actual content. You cut praise and intensifiers. "Mike framed the question well" → "framed the question." "That is exactly the job" → "That is the job." "This was Bob's instinct and it is the right one." "Sequenced last deliberately." Complimenting the client on their own idea, or telling the reader my point is correct, adds nothing. You cut the unprovable brag ("almost nobody in this market has it") and the flourish ("the piece that turns a store into a relationship"). You dropped the salesman's posture. "The decision we are asking for is whether" → "The decision is whether." State the decision; don't stage yourself asking for it. You cut the repeated proof point. I used the five-year-blog fact three times. You cut it to zero, which reads better than my once-would-have-been-fine instinct. You merged two bullets into one where the second was a variation on the first, not a separate idea. Short version: say it once, say it plain, then stop. No clause that exists to make the previous clause land harder.

Opus 5.0 dumb, output token-consuming comment of the day: "Your call, but decide it consciously." Thanks, Claude. I was going to hit myself over the head with a hammer and then decide it unconsciously. Oh, and thanks for reminding me that doing my work is my call.

The hardest thing right now is building a moat, and selling a dollar for 90 cents ain’t a moat.

It's super interesting that a perceptive human can tell when an AI model degrades, the way Fable has this week. The models are amazing, miracles even. But when they go downhill, they aren't able to reel it back in and reset with a portion of the context. Tactical use cases are super powerful. Giving a sales person live advice in a meeting with a customer - that's powerful and reliable. Strategic use cases are strained. I still find it useful to work with Claude when I am sorting out strategy. Better than many humans I've worked with. But still not good enough to handle the nuances and depth of hard strategic thinking.

GPT-Realtime-2.1 cut voice latency another 25%, mostly through better caching. Worth noticing how it did it: you go faster by not re-fetching what you already have. That is the right instinct, and it has a limit. Caching speeds up the thinking. It does nothing for the trip. Most "AI for live calls" tools still ship audio and video out of the meeting, run inference on someone else's servers, and send a result back. The model can answer in 200ms and the round-trip still blows the budget, because you do not control the pipe when you do not own it. No cache fixes a trip you did not have to take. We did not make the round-trip faster. We removed it, by running the agent inside the video engine. Where the AI runs is still the part almost no one is optimizing.

The most advanced AI companies on earth just hired about a hundred enterprise salespeople away from Salesforce. The labs building the models that are supposed to replace sellers are staffing up on sellers. Not because the AI is weak. Because the deal that matters still gets decided in a live conversation, and they know it. The frontier of GTM AI is not replacing the rep. It is giving the rep real-time help in the room where the decision actually happens.

Voice AI had a month. It can interrupt you, talk over you, recover when you cut it off. Genuinely impressive. It still only has the audio. The buyer's face is doing half the talking, the pause before "let me think about it," the flat nod, the lean-back at the price, and none of it reaches a model that only listens. Hearing got solved this month. Seeing is the frontier that's still open.

Every revenue tool promises to "capture and analyze your sales conversations." Read that slowly. Capture, then analyze. That is a recording played back after the call, and by the time the analysis lands, the deal has already moved. The interesting problem was never analyzing the conversation after it ends. It is coaching inside it while it is still live. That is a question of where the AI runs, not which model writes the summary. A bolt-on tool ships the audio and video out of the meeting, thinks somewhere else, and sends a verdict back, and the round-trip is dead time you cannot afford when the objection is landing right now. We put the agent inside the video engine so the tip arrives during the call, not in tomorrow's digest. After-the-fact analysis tells you which deals you lost. In-the-moment coaching is how you stop losing them.

The GTM AI numbers are starting to separate the teams that adopted from the teams that changed. 83% of sales teams using AI reported revenue growth last year against 66% of those that did not. But only about a quarter run truly autonomous agents, and the window is already half closed. Here is the part the averages hide: most of that agentic work is aimed at the queue. The outbound, the follow-up, the enrichment. Clean inputs, clean outputs, easy to automate, and worth doing. The live conversation is the opposite. The signal is a pause, not a field. It is the half of the funnel almost no agent touches, and it is where the deal is actually decided. The teams that win the next round will not be the ones with the most agents. They will be the ones who put an agent where the decision happens.

Consequential week for voice AI. Full-duplex you can talk over, sub-300ms models, STT that finally nails turn detection. All of it is the same win: cut the time between the words and the model's answer. Real progress, and for a phone bot it may be the whole game. A live sales meeting runs on two clocks, though. One is the model's. The other is the network: shipping audio and video out of the meeting, running inference somewhere else, sending a result back. A faster model does nothing for the second clock, and in a meeting the second clock is usually the bigger one. We did not chase a faster round-trip. We removed it, by running the agent inside the video engine. When the field says latency is solved, it is answering the first clock. The meeting has two.

There's now a benchmark for AI that does go-to-market: build the list, enrich the record, score the account, reach the decision-maker. Good. Those are the parts that get written down. The part that decides the deal, the live conversation, has no benchmark, because its signal never gets logged. You cannot score a lean-back, or the pause before "let me think about it." The measurable half of GTM is finally getting a scoreboard. The half that actually closes is still played in a room no dataset can see.

The GTM category spent this year racing to make agents more autonomous. Outreach repositioned around agentic AI. Every platform now ships an agent that runs the workflow while the human sets the goal. Deloitte says only 24% of sales orgs actually run autonomous agents, so most of the race is still ahead. It is a race worth running. When a task has a clean input and output, hand it to an agent and walk away: the follow-up, the enrichment, the stale-lead queue no human would call. The teams doing this are growing faster (83% of AI-using sales teams reported revenue growth last year against 66% without). But autonomy wins where the work has a spec. The live sales conversation is the opposite of a spec. The signal is a pause, a lean-back, a change in tone, and it is gone in a second. You cannot hand that to an unattended agent, because there is nothing clean to hand it. So the split is not human versus agent. It is: run the workflow autonomously, and coach the conversation in real time. Different jobs, different systems, same company.

ZoomInfo just wired its context graph straight into Codex and Claude. 100M+ companies of firmographic context, sitting inside the model. Right direction, and it is becoming table stakes: the agent should know the account before the call. But that context describes the company before anyone speaks. The context that decides the deal is what the buyer says and does once the call starts, and that lives in no graph. Company context you can buy. Conversation context you capture in the moment or lose. One is turning into a commodity. The other is the asset.

OpenAI's new voice model listens while it talks. Full-duplex, so you can interrupt it and talk over it the way people actually do. That is a real fix to the rhythm of a conversation. It shipped audio only. No video. So it can hear the pause before "let me think about it." It cannot see the lean-back at the price, or the glance away that came first. In a sales meeting the decision shows up on the face before it reaches the voice. Fixing the clock is not the same as opening the eyes.

Every AI value-capture argument lands on the same question: what do you own that the model does not. For a sales team it is not the transcript, everyone has that. Not the CRM field, everyone has that too. It is how the buyer actually reacted in the room, the pause before "let me think," the moment they went quiet. That never gets written down, so nobody owns it by default. You own it only if you capture it while it happens. That is the asset that compounds.

Voice AI had a big week. OpenAI cut realtime latency again, Thinking Machines showed a model that listens while it talks and backchannels "mm-hmm" at human turn-taking speed. Impressive, and exactly right when the AI is the one doing the talking. A sales meeting is the other case. The AI is not talking. The human is. The job is not to sound human, it is to read the human on the other side of the table. Sounding real and reading the room are two different frontiers, and almost everyone is racing the first one.

Every lab is racing to buy more data. Every GTM tool is racing to store more memory. Both are chasing the same thing from the wrong end. The data that decides a B2B deal is not fragmented across your systems. It never enters a system at all. It is the pause before "let me think about it," the lean-back at the price, the room going quiet on a hard question. No transcript writes it down, so no data deal and no memory layer can buy it back later. The frontier in GTM AI is not more data or more memory. It is capture. Reading the live conversation at the source, before it disappears. Everything downstream only ever knows what got recorded.