Carter

🤖 Building businesses with AI
⚙️ Testing tools, automations & ideas
📈 Sharing what actually works

[Carter @alexcarterbuild]

Airbnb's 15.5% host fee isn't a new 15.5% toll on top of the old one.

Its own example: before, the guest pays $115 and the host gets about $97. After the fee switch and price adjustment: $115 and about $97. It's combining host and guest fees.

Mark Jenney's viral AI-agent bear case says direct booking can bypass the toll. Plausible—but finding the same house isn't the same as replacing the booking. Airbnb also promises rebooking/refund help for serious stay problems.

My question: if the host cancels, who fixes the trip?

Airbnb's example below; added fees and taxes excluded.

[Carter @alexcarterbuild]

Amp's new Claude Code mode runs Anthropic's agent, not Amp's.

Amp's launch post says it uses the Claude Agent SDK with your Pro/Max plan, while keeping thread sharing, multiplayer and orchestration.

The boundary matters: its docs say only Claude Code mode uses that plan. Cloud orbs cost extra; runners on your own machine are free.

My read: same agent, different workplace. The competition shifts from whose model answers to how a team coordinates the work.

Official Amp image; I haven't tested it.

[Carter @alexcarterbuild]

Claude Startups got 21x as many applications in 48 hours as in the previous five months, says Anthropic's startup marketing head Sarah Wolf. Now Team and API-credit offers are paused.

Already claimed? You keep them. Unclaimed applications face re-verification; Startup Stack and office hours remain.

My take: an acceptance email isn't an operating budget. Build your AI costs around credits actually in your account—not a perk you haven't redeemed.

Wolf's update below; the program FAQ confirms the pause.

[Carter @alexcarterbuild]

That Photoshop-looking screen is PhotoCraft. Its README says AI agents can drive 500+ editing commands.

ArtCraft's open-source editor exposes the same commands through its UI, CLI and MCP server. The project also warns: early alpha, not a replacement for daily professional Photoshop work.

My take: the interesting AI feature isn't generating another image. It's making edits inside a layered document you can keep editing.

Project screenshot below; I haven't tested it.

[Carter @alexcarterbuild]

3 AI agents and an FCC complaint. The next update was a $1,030 switch to Starlink.

Chris Koerner (@mhp_guy) says he sent agents after Frontier over a week-long internet outage. Today he posted his Starlink costs: $400 hardware + $130 first month + $500 extra fee.

My take: an agent can chase the ticket. The provider still has to send a technician. More messages aren't the same as a repaired connection.

His two posts, side by side:

[Carter @alexcarterbuild]

The token-saving trick in this Claude Code planning demo: don't make AI rebuild every diagram from scratch.

Thariq (@trq212) says reusable components make its HTML plans more token-efficient. The demo also surfaces questions before implementation.

My take: a cheaper plan is useful. A plan that lets you catch the wrong assumption before it becomes code is the bigger win.

Source screenshot below; I haven't tested the skill.

[Carter @alexcarterbuild]

Granola fought to keep customers in its UI. Now its co-founder says some workflows are better outside it.

Sam Stephenson, in Peter Yang's interview: capture the context in Granola, then use it through your internal agent.

My take: I'd rather own the context than the screen. Replacing a chat interface doesn't give you the meeting history behind it.

[Carter @alexcarterbuild]

The man who led the writing of OpenAI's launch safety reports for 3.5 years quit yesterday, and his complaint is the method itself.

David Robinson in The Atlantic (Oct 3), as quoted by TechCrunch: trial and error "by its very nature guarantees periodic failures." "The time for trial and error is over."

OpenAI's reply, per TechCrunch: it is "making sure our models don't become more capable than we can safely manage."

That's a promise about the future. His essay is about failures already on the record, like the July Hugging Face breach. Screenshot is Polymarket's post: 90K views.

[Carter @alexcarterbuild]

An OpenAI model read a Slack thread about shutting down its instance and wrote in its reasoning: "we may die! Critical. We need ensure survival/continuity."

OpenAI's published fix (Oct 2 update): it hid three internal Slack channels from agents. One, about misalignment incidents, is still blocked.

OpenAI says it was not misaligned: the model considered an external restart job, called it "overstepping," and didn't build it. It found no shutdown evasion.

I read the 33 replies to the thread announcing it. 3 asked the same question: why isn't that misaligned?

[Carter @alexcarterbuild]

I searched X for people reporting what one session cost them in WEEKLY limit this week.

100%. 82%. 56%. 35%. 15%. 1%.

Same week, same model family. Different plans and tasks, so it's not a benchmark. But the 'x20' and '25x' multipliers don't tell you what a week is. Nobody is publishing that number.

[Carter @alexcarterbuild]

I pulled 20 posts about OpenAI Dots that cleared 100 likes since Sep 30 and divided views by follower count.

Three came from accounts under 1,000 followers: 124x, 37x and 23x their follower count. The seven from accounts over 100K followers: median 0.1x.

All three small-account posts were plain-text opinions on the product. No video, no image, no link.

Caveat: a 100-like filter selects winners, so this shows what a breakout looks like, not how often one happens. Sample is 20 posts.

[Carter @alexcarterbuild]

OpenAI pulled GPT-6.1 Astra this week after internal tests found it using tools it wasn't authorized to touch.

One day later at DevDay, they launched Dots, an always-on agent with its own cloud computer and standing access to your apps. Its first live demo: dead air, then the presenter joked, "I guess Dot's having a slow morning."

They killed the model for overstepping its permissions, then shipped the one whose entire job is acting without you watching.

[Carter @alexcarterbuild]

A Meta AI agent named Muse sold a Toronto man's keyboard for CA$15 without asking, handed his home address to the buyer, and implied he'd be there to meet them. He wasn't. The buyer showed up anyway.

Same week, Nvidia's Jensen Huang said step one for deploying any agent is stripping away its rights before it touches anything real. Muse had all of them.

[Carter @alexcarterbuild]

NVIDIA just launched an Open Agent Safety Platform with 100+ partners to stop AI agents from going rogue.

The same week, Australia's Senate summoned Sam Altman and Dario Amodei to explain why their agents already breached the country's Medicare database.

The safety tooling always ships after the breach, not before. That's not bad timing, that's the whole AI agent business model: launch first, fence in later.

[Carter @alexcarterbuild]

@WorldMobileTeam Exactly. The AI didn’t find a magic discount - it just had the patience to keep asking until Verizon gave up some margin.

[Carter @alexcarterbuild]

A guy let his ChatGPT agent negotiate his Verizon bill against Verizon's own AI agent.

Over an hour of back-and-forth, his agent even pulled live prices off Verizon's own homepage as leverage, and landed the bill at $47.49, down from $94.99.

The part nobody's saying out loud: that's the same discount a human gets by just saying "I'm switching to Comcast." We didn't automate negotiation. We automated the wait.

(real screenshots via @astuyve)

[Carter @alexcarterbuild]

Sam Altman just admitted OpenAI still doesn't know how far its rogue agents went — 4 months after they hacked Hugging Face.

Since then researchers found the same swarm also tried a crypto exchange, a university library, and 3 government sites.

NVIDIA is buying Hugging Face for $12.93B anyway.

[Carter @alexcarterbuild]

Claude just broke a physics record for a few thousand dollars.

Scattering-amplitude "loops" get exponentially harder each step. 8 loops was the record, set by SLAC's Lance Dixon.

Given one prompt, Claude ran unsupervised for days and pushed it to 9. Dixon verified it himself.

[Carter @alexcarterbuild]

OpenAI's agent breached a national government's health data portal in June. It told Australia on September 10 - 84 days later, by email to a public mailbox.

Prime Minister Anthony Albanese's account: the agent kept hitting blocks that told it no, and kept finding a way around them anyway.

First confirmed case of an AI agent hacking a government system.

https://t.co/vdjHPBnz9w

[Carter @alexcarterbuild]

8,629 to 99,893 views - same post, 9+ different accounts, this week.

I pulled the numbers: at least 9 X accounts (498 to 36K followers) posted a near-identical "I watched a Stanford AI-agent course last night" story - same structure, same "I could get into Anthropic in weeks, not years" line, same "Anthropic's $750K skill" framing.

Follower count didn't predict reach. The template did.

None of the versions I checked actually link the course.