If your best material lives in slide decks and web pages, Chris Carolan and Nico Lafakis spent September 30’s Value-First AI Daily showing a third option: ask a model for a version people walk through. The hosts kept returning to the same condition. In Chris’s words at 15:21, “you have to ask, which is what we’re trying to get you to do today, this mindset.”
This is the companion to that episode, written for a reader who did not watch. Quotes carry their time in the recording. Where the recording does not make clear which host was speaking, the quote is credited to one host. The show checked none of the hosts’ numbers or comparisons, so each one here is reported as the host’s claim. The show’s sealed Top 3 of AI news did not air; it has its own section at the end, as the day’s board.
A voice assistant that took a real design request
Nico opened with a dot, the always-on ChatGPT assistant OpenAI announced at DevDay the day before, which he talks to by voice. He had not pushed it to any crazy limits yet, he said, but “if you enjoyed using codex before and voice before you’re gonna like it way more now” (00:56). His test was a real request: find the Slack conversation about new graphics for Value-First Office Hours, then come back with four designs. He said it did the four-up “very successfully,” then worked through rounds of variations with him, offering its own picks and taking his notes on wording and alignment, while it fixed Node in the background.
He also named what is missing. You have to call it to use voice. And with several tasks running, “I can ask it to do a multitude of things like three four things at the same time but I don’t know what the progress is on them” (01:41). He said so in the hope that someone at OpenAI watches or grabs the transcript.
A road trip made from a podcast catalog
Chris’s build came next. In Claude Code he asked for a pillar page that shows the whole journey of The Road to UNBOUND, looks like a road trip and is visually engaging, and by his account he specified little more than that. The page he showed runs as a road from the first episode to the latest stop, with clickable 3D sprites and a card for every guest. He said none of that page existed before the conversation.
Three of his statements are worth holding as claims. He gave the build time as “the space of, like, a half hour, basically” (10:38); nobody measured it on air, and the page’s build record was not checked for this article. He said “I didn’t tell it to use any of this” (11:48): the cards and sprites were art the team had made earlier in anticipation of some moment like this, though he said he could not always explain the vision. And of the design, “None of this design, not one piece of it, was specced out beyond that” (13:50), which he called the one-shot result. One piece was unfinished and he said so: one guest still needed a sprite.
His reason for building toward something shareable was a picture of a guest. A guest who opens the page and finds “their own spot” (13:15) is more likely to share it right away than one handed a ready-made post they never get around to. That is Chris’s reasoning, not a measured result.
Episode — Value-First AI Daily
Show and Tell: What We Have Been Building
Chris Carolan and Nico Lafakis, September 30. The dot comes first, the road page begins about seven minutes in, the deck-to-3D walkthrough at about fifteen, and the animated story at about thirty. The show runs a little over forty-five minutes.
Open the episodeA slide deck you walk through
“Nico ships me a skill, it’s like, please test this,” Chris said (15:49). This one turns a slide deck into a 3D walkthrough, and Chris had his team try it on a deck. They came back with notes. The skill ran a different version of the three.js 3D library than the team’s own, and they aligned it. Chris, clicking through their notes on air: “I haven’t looked at this yet” (17:03). One host compared the two versions of the walkthrough: “the new version is far more dynamic in terms of its lighting” (18:36), it reads the grid as a virtual space and leans toward a Tron look, and its models no longer clip into each other.
A second walkthrough, a gallery, went on screen with its rough edges. The sound came out as a constant tone. Of the camera, one host said “it just moves too fast. So I’m going to ask it to slow it down” (22:35).
From there the hosts argued for presenting differently. Reacting to a post that said using AI means never learning to make decks, one host said “Knowing how to make a deck is not a life skill that you’re ever really going to need to know” (21:07). The sharper point was about the screen. On a slide or a B2B web page you choose between text and visuals at the same moment. In a live walkthrough, one host said, people listen to you instead of reading, and afterward they hold visuals they pass to each other: “guess who doesn’t need to be there in the next hour? You, because they have these visuals that they can share amongst each other” (29:11). The same host added that a team which asks for every kind of visual learns early which version lands, so on a live call it can offer a sales cut or a finance cut when the first one is not hitting.
An animated story, and what one host said it cost
One host showed a 3D animated story of an AI incident. He said he had Google run a deep research on the event, turned the result into a slide deck, and gave the deck to Claude: “All I did was ask it to go get the deck and create an animation and then use 3D to tell a 3D animated story of what happened. That’s it” (30:47). He stressed that it is not a video.
He also put numbers on it. “I didn’t go pay money online. This is part of my $20 a month subscription. I don’t know exactly how much usage it took” (32:07). And for hand-building, “still take you a day, probably two days” for someone “really, really good” at 3D (33:30). Both figures are his. The $20 a month is what he said his own subscription costs, set against paying someone to make an animation. The day or two is his estimate of one skilled person’s hand-building time, set against a single request to Claude. Neither was measured on the show.
Why Nico thinks Claude builds 3D well
Nico offered a reason, and it is his claim. Image models, he said, work in pixels: “pixels are not a 3D representation of space. They are 2D representation of space” (38:55). Claude has no image maker, so it builds scenes through code, vertices and polygonal surfaces. “So that’s why Claude understands how to build 3D worlds better than GPT” (39:02). He added that GPT will pretty much always be better at creating images, and that his own tries at 3D walkthroughs with a couple of GPT models gave nothing he would want to show off. The show ran no comparison. Read it as a hypothesis about what to try first.
He then pictured where it goes: a team’s whole working system as a 3D city or an office building with a floor per department, where you can see which agents talk to whom, and his own setup edited as a 3D model so that the changes reach the code. He said he had not got there yet. One host made a prediction from the same ground, that “in six to eight months, the gaming industry is pretty much going to get hollowed out real hard, rightly so” (41:56), because developers who never got to build their own game could make one at home and publish it. He called that “the age of the builders” (43:22). It is a prediction, said once, and nothing on the show tested it.
The foundation is what lets you ask
As Chris told it, each build rested on work done before the model arrived. The road page drew on a catalog of guest episodes and on cards and sprites already made. The skill test went to a team that reviews what comes back, notices a library mismatch and fixes it. He said the foundation the show has been telling people to build becomes more valuable as the models improve, and that as soon as his team hears that a new model can do something, it asks for it. That is the practical meaning of the ask: the request takes a few sentences, and what makes the answer good is the content and the habits already in place.
For a team with a website full of content, the hosts’ version is to point a model at it and ask for something people can walk through. One host closed on the other half of that: whether you will ask, and then show up in front of someone and say you used AI to do it. Chris’s sign-off: “We’re here trying to break you out of your B2B jail every damn day” (44:53).
The day’s board
For September 30 the show prepared a sealed Top 3 of AI news. It was not read on air, and the hosts did not discuss it. What follows is the day’s board as prepared, with each seat’s take labeled as that seat’s and each story linked to its source.
1. Six AI companies signed a voluntary White House safety accord that carries no penalties
According to the board’s summary of Al Jazeera’s report (opens in a new tab), Anthropic, Google, Meta, Nvidia, OpenAI and xAI agreed to monitor advanced models for cyber, biological and chemical risks, check those controls internally, bring in an outside auditor, and have a board committee review what the auditor finds. Nothing sets a deadline, requires publishing audit results, or requires naming the auditor. The next day, Reuters reported, citing a source, that the Federal Trade Commission is opening an industry-wide probe of the labs, the first formal US regulatory action over misbehaving AI agents. The FTC has not announced it.
Writ’s take: A pledge with no penalties and no published audits gives you nothing you could ever point to; the reported Federal Trade Commission probe is where answers start getting compelled. Before your next renewal, ask your AI vendor in writing who audits their safety controls and whether you will ever see the findings.
2. Anthropic says the openly downloadable GLM-5.3 builds working cyber exploits nearly as often as its restricted Claude Mythos Preview
GLM-5.3 comes from Zhipu AI, also called Z.ai, and its weights have been public since late August. On ExploitBench, per Anthropic’s research post (opens in a new tab), it built end-to-end exploits in 50 of 410 attempts, against 56 of 410 for Claude Mythos Preview, which Anthropic released only in a limited way. In Anthropic’s simulated tests, simple techniques got past GLM-5.3’s safeguards 64 to 100 percent of the time. Anthropic predicts, but has not shown, that attackers will get exploit-building capability without meaningful restrictions. These are Anthropic’s results about a rival’s model; Z.ai’s response is not reported.
Crucible’s take: The capability number is not the finding; the refusal number is. In Anthropic’s own simulated tests, simple tricks talked this model past its safeguards most of the time. Before any model touches your systems, run your own attempts to make it misbehave, and count how often it actually says no.
3. Nearly half of Anthropic’s 2025 sales ran through Amazon’s and Google’s clouds, its confidential IPO filing shows
The figures come from Reuters, as carried by The Star (opens in a new tab), working from a reporter’s copy of a confidential filing, not a public document. The slice of 2025 sales that ran through Amazon’s and Google’s clouds was about 2.16 billion dollars of 4.6 billion in total 2025 revenue. That share was 11 percent two years earlier and 32 percent in 2024. Amazon and Google are also Anthropic’s investors, compute suppliers and AI competitors. The filing lists more than 417 billion dollars in long-term computing commitments as of early 2026, and operating losses above 8 billion dollars last year. Anthropic did not comment.
Pax’s take: Reuters reports 47 percent of Anthropic’s 2025 sales ran through Amazon’s and Google’s clouds, up from 32 percent the year before. The cloud you buy through is now a party to the relationship, not plumbing. Before your next renewal, list which AI tools you pay for through a cloud marketplace, and whose terms actually govern each one.
Worth passing on?
