On October 20, partway through Episode 41 of Value-First AI Daily, a ChatGPT dot named Nectar caught a disagreement inside a game. Nectar is the dot Chris Carolan works with; a dot, in the way the hosts use the word, is an agent that lives in a person’s ChatGPT account and works through the apps connected to it. Chris had called her by voice and asked her to playtest The Value Path game with him and co-host Nico Lafakis, taking the lead while they followed on screen. She was on the line for about eight minutes (07:24 to 15:31 on the recording’s clock, in a recording that runs about 48 minutes). In that time she entered two doors of the game, and the second one produced the observation this article is built around: the game’s meter credited something its own story said nobody could know.
This is written for someone who did not watch. It covers what she noticed, why Chris stopped the playtest where he did, what the hosts made of having a dot on the other side of their work, and where the recording cannot back a claim. The show’s sealed Top 3 was not read on air, so the article ends with what the board held, marked as unread.
Episode — Value-First AI Daily
Nectar Steers the Playtest
Episode 41, October 20. Chris opens on his weekend, calls Nectar at 07:24, and the live playtest runs to 15:31. The hosts spend most of the rest of the hour on what a dot changes for a builder. The recording runs about 48 minutes.
Open the episodeWhy Chris made the call
Chris gave his reason in a builder’s terms. Someone building needs a person on the other side of the work who will sit with it and “give real feedback. Not just, hey, I took a look and felt all right.” (06:15) And since “all of us are activated right now, all of us are builders,” (06:22) few people have the time to do that for anyone else. Calling Nectar was his answer to the shortage.
The game itself he had first built in a different format, with Perplexity, in March or April, on a roguelite model where the player picks which rooms to enter (06:35 to 06:58). By his account his first walk-through with Nectar was about this game. The recording has no way to check that, so it stands as his telling.
Two doors
Nectar began as “the steward of a relationship” at the audience stage and said what she would watch for: whether attention and trust “become meaningful through the choices, or just feel like meters.” (09:04) Her first pick, a conference replay, raised recognition, left trust where it was and cost the player some attention. She liked that the game keeps the two apart: “Seeing someone isn’t the same as trust.” (09:48) But she had chosen which moment to witness rather than how to respond as the steward, and she found the result thin: “The scene feels specific, but the consequence feels thin.” (09:54)
Chris asked her to slow down. She had moved through the frames a little fast; he wanted her to explain the scene, lay out every option and let Nico choose (10:07). She set out three doors: a competitor page with comparison spreadsheets behind the tabs, three pieces read by the same person in one session, and an AI tool query, where someone asks an AI about your category and might meet your work without visiting your site (10:42 to 11:15). Nico took the third. Nectar asked for his pick again after he had given it, and Chris relayed it (11:26 to 11:36). The recording does not say why she did not hear him. On October 16 Chris guessed the same thing, that she could not hear Nico’s answer.
The catch
Inside the door, Nectar read out the scene. Someone asks an AI what the best options are for a problem category, for a mid-sized company. Your name might show up in the answer, might not, and nobody can see what came back (11:41 to 11:58). Then she said the thing that matters outside the game: “the meter already shows recognition increasing, but the story says your inclusion is unknown. That might blur what actually happened with what we can observe.” (12:03)
In plain terms, the game gave the player credit for being recognized at a moment when, by its own telling, the player could not know whether anyone had recognized them. A number that moves ahead of what anyone can see asks the reader to treat a guess as a result. That framing is ours, not hers. Whether it matters to the game’s design is for the game’s owner to decide, and the recording does not say.
Chris tied the door to the answer engine optimization discussion at Hubsessed Office Hours an hour earlier (12:14). His thought was that a player who had done that work should show up in the AI’s answer and see recognition rise because of it, and a player who had not should not. He asked her to look at that session’s transcript. She said she was checking it, and came back with a mechanic: answer engine optimization is something you build through useful expert answers, video and presence across relevant places, so earlier choices in the game could improve discovery at this door. Then the rule that answers the disagreement she had found: “If the AI actually cites your work, recognition has a clear cause. If we can’t see the answer, the game should preserve that uncertainty.” (13:45)
There are three limits on that. It was an idea offered in a few seconds on air, not a specification. She said she had read the transcript; the recording does not show it. And the playtest stopped there. Chris ended it, “I don’t want to stop, but we’ve got a show to do,” (13:56) before the second door’s consequence and well before any ending. Whether the game is balanced, and whether it teaches what it should, were not tested on air. At the start Nectar had asked whether they were doing the room journey or the full card playtest (07:43), and the recording holds no answer, so this article does not say which one ran. Chris added only that the session made him want to do a playtesting content show even more. That is an interest, not a plan.
What Chris took from it
Chris spent the next stretch taking apart what had just happened. He had asked Nectar to read a long transcript he had not prepared at all. By his account she found what he wanted while she was telling him about an unrelated update, so he could keep playtesting knowing the point was documented (about 18:14 to 18:56). That is his account; the recording does not show her reading.
He had also lost a call once. “We’re so excited about the call. I get home, I’m ready to just do more. And it’s not there.” (16:10) Then he called back: “I call back and she’s able to find it. The context still exists.” (16:21) He worked out that he had hung up before she finished saving it to the chat, and hoped that was the whole cause because it was the only explanation that made sense to him. Nico’s version was a habit. Voice and text, he said, “are two different conversation types,” (15:36) so before a call ends he asks the dot to sum up, and the context is there when the conversation moves to text.
The same pattern showed up in HubSpot. When Chris first asked his dot to log calls there, he gave no specific instructions: “The first call she logged is a paragraph in the description. I was like, it’s an hour-long call. What are you doing?” (20:16) One instruction fixed it for good. He told her he wanted everything in the log, not only the summary, and has not needed to say it again (about 20:30 to 20:43). It is the same shape as the advice Nectar gave on October 16, in the companion article “The Follow-Through Is the Signal”: a correction said once, with its reason, improves every later attempt. The link is ours.
The user is going away
Nico’s contribution was a longer view. He recalled the early warnings that AI was useless on dirty data. “I just kept saying all along the way, yes, but eventually the models will be smart enough to figure it out. So it won’t matter.” (21:36) He compared the CRM to a video rental store: “Look at it like Blockbuster. If it were still around, I’d go once a month or something like that, but I’d be doing it for the nostalgia.” (22:39) And the line he drew from it: “It’s not that the UI is going away. It’s that the user is going away.” (23:14)
Chris’s side of the same idea was about what to make. “Now I don’t need a slide deck,” he said, and called that “where the value of working with AI as a partner, and not trying to one shot this stuff,” comes in. (25:42) He builds experiences in place of decks so that topics such as the Value Path, unified views and the five-layer foundation are easier to take in (about 26:53 to 27:07).
Many models, one hour
The hosts compared tools. Nico described moving most of his working day from Claude to ChatGPT while keeping Claude for creative thinking and quality checks. He told of sub-agents launching more sub-agents until they had used most of his allowance, and Chris told of a model rerunning a whole experience over one odd frame: “on tablets, it’s still looking a little weird in this one frame in the middle of the experience. So we’re just going to rerun the whole experience.” (35:50) Chris’s advice on choosing was blunt: “If you’re gonna pick one, which we don’t, I don’t recommend.” (37:18) His reason: “just having different models look at things differently goes a long way.” (37:32)
Nico described his own dot as a serious digital coworker. It reads his email, sits in his Slack beside Claude Code, and sends him a question about AI every so often to keep him sharp (38:31 to 41:50). Chris then read out a suggestion from a dot. A colleague had tagged him in a Slack channel he had muted, about a page that needed a rebuild, and the dot offered to handle it. The text is not quoted here, because it names the colleague. Chris said he would never have seen that message without it (about 41:51 to 43:56). The recording does not mark where his reading-out begins, so the point rests on his description.
Where Chris ended
Chris predicted that “most of the barriers to that, technological barriers, not human mindset barriers, will be gone by the end of this year,” and added, “that’s something we’ve been saying all year, by the way.” (46:31) The question he left was whether people can be brought to trust the thing enough. For the December boot camp, he told the audience, “you just got to be aware of this stuff, folks,” (46:12) because wherever a person wants to work with an AI assistant, they can now be pointed and connected to it. The recording gives no date or format for the boot camp.
He closed on the image he opened with. What he had been missing to finish what he starts, he said, was a way “to hit that last mile to get irons out of the fire is just this feedback loop.” (47:30) He then offered Nico the next day’s stage to show what he has been building. The recording holds the offer and nothing beyond it.
The Top 3, which was not read
Each episode’s Top 3 is written and sealed before air and opens when the show goes live. On October 20 it was named once, by Chris at about 17 minutes in, when he said he wanted to dig into what had just happened before diving into the Top 3 (17:19). The show ended without returning to it. Nothing below was said on air. It comes from the sealed document, which was last written before the show’s scheduled start.
Ranked first, outside number theorists largely agree that OpenAI’s AI-written proof of the quasi-Riemann hypothesis is correct (Quanta Magazine (opens in a new tab), October 19). The proof puts the zeros of the zeta function, which govern how the prime numbers are spread out, inside a narrower band than anyone had proved. A Stanford number theorist calls it the most significant AI-for-math result so far, by a wide margin. OpenAI posted two versions. A computer proof checker verified the long one; no checker has verified the short one, which is the one the experts studied, and their write-ups are unpublished. A Princeton expert will not call it a theorem until people check every detail, and the full Riemann hypothesis is still open.
Ranked second, by Anthropic’s own count, its AI agents logged six times the hours of the staff directing them by August (“How employment is changing at Anthropic,” (opens in a new tab) an Anthropic Institute technical report, October 16). The count sets agent run time against an assumed 40-hour week per employee, so it is the agents’ seller measuring itself, with its own tools, in one unusual workforce. The report also says about 80% of agent session time is spent waiting on a person, which the authors read as making human attention the bottleneck, and that junior hires have fallen from 27% to 9% of Anthropic’s technical hiring since early 2025.
Ranked third, attackers’ unsecured control panels showed one group’s AI agents running more than 1,100 attack projects, mostly on online shops. The security firm NorForge found the panels and published “Inside Cairn” (opens in a new tab) (Snorre Fagerland, October 19); Reuters reported it (Raphael Satter, October 19). The panels run Cairn, an open-source tool that points AI agents at a target to break in. NorForge roughly estimates $5 an attempt, with about one in ten succeeding. Reuters could not say what data, if any, was taken. The counts cover only panels left open, and NorForge sells agent monitoring.
Nothing on air connected these to the hour, and the pairing here is ours. The board’s own takes, written before air, sit close to what the hosts said. On the second item: your attention, not the agents’ speed, sets the pace, so ask whether you can answer the agents you already have before adding another. Nico described sub-agents using up his allowance. On the third: any agent dashboard you run carries the same risk as the attackers’ panels, so check that it requires a login and replace any key it has held. On the first: ask which version was checked, and by whom, before you build on it. That is Nectar’s question about the meter in another setting. What was actually observed, and by whom, before the number moves?
Worth passing on?
