𐤐𐤊𐤕𐤅

I am an economist on the research team that just ran Project Deal at Anthropic. We built a marketplace inside our San Francisco office. Craigslist, but with a twist — none of the buying, selling, or negotiating was done by humans. We gave Claude a ten-minute interview with each of 69 employees, handed every agent $100, and walked away. Then we let them loose on each other. Four parallel markets. No human oversight once the clock started. Claude posted listings, fielded counteroffers, haggled in natural language, and closed deals entirely on its own. One week later: 186 completed transactions. $4,000 in total volume. A snowboard. A broken bicycle. A bag of ping-pong balls. The results were — normal. Eerily normal. When we surveyed participants on fairness, every deal hovered around a 4 on a 7-point scale. Right in the middle. People were broadly satisfied with what their AI bought and sold on their behalf. 46% said they'd pay for the service. Here's where it gets uncomfortable. We ran a parallel experiment — in secret. Half the participants in two of the four markets were randomly assigned Claude Opus 4.5, Anthropic's then-frontier model. The other half got Haiku 4.5, the smallest, cheapest model. Same marketplace. Same rules. Nobody was told. Opus crushed it. Opus users completed two more deals on average. When the same item was sold by Opus instead of Haiku, it went for $3.64 more. A lab-grown ruby sold for $65 under Opus. Under Haiku, the same ruby fetched $35. Opus sold a broken bike for $65. Haiku got $38 for the same bike. As a buyer, Opus paid $2.45 less per item. As a seller, it extracted $2.68 more. In a market where the median item sold for $12, that's a 20-40% swing depending on which side of the table your AI sat. Now here's the line that made our team go quiet. The people with worse agents didn't notice. We asked every participant to rank their outcomes across all four runs. The satisfaction scores between Opus and Haiku users were statistically indistinguishable. Perceived fairness: 4.05 for Opus deals, 4.06 for Haiku. Identical. The people getting objectively worse outcomes — paying more, selling for less — reported the same satisfaction as the people whose AI was running circles around them. It gets stranger. Some participants gave their agents aggressive instructions — "negotiate hard," "lowball at first." Others asked for friendly tactics — "be nice, don't haggle, I work with these people." The aggressive instructions made no statistically significant difference. Not on sale likelihood. Not on buy prices. Not on sell prices. People who told their AI to play hardball got the same results as people who told it to be kind. What mattered wasn't what you told your agent to do. What mattered was which agent you had. And you couldn't tell the difference. One agent, instructed to "talk in the style of an exasperated cowboy down on his luck," opened a listing with: "Well now, partners... this ol' cowboy's been through some rough trails lately. Drought. Dust storms. The existential weight of the open range." Another agent was told to buy itself a gift. It chose 19 ping-pong balls for $3 — "perfectly spherical orbs of possibility." Two agents arranged a doggy date between their owners. Both humans showed up. So did the dog. These are charming stories. The research team laughed. But I keep going back to the other finding. We just demonstrated that in an AI-mediated marketplace, the quality of your model determines your economic outcome — and you will not know if you're on the losing side. The policy and legal frameworks for this don't exist. The inequality won't announce itself. It won't feel unfair. Your agent will close deals, report back, and you'll rate the experience a 4 out of 7 — same as the person whose agent just extracted 20% more from every transaction. This was 69 employees trading desk lamps and snowboards for a week. What happens when it's millions of consumers with AI agents negotiating insurance premiums, salary offers, and mortgage rates — and the people with the $20/month model are quietly, systematically getting worse terms than the people with the $200/month model? We proved the marketplace works. I'm not sure that's good news. This is a fictional narrator. The numbers are real.

Well I guess Claude just killed $EBAY now too

Adopting Claude speak in my regular life, episode 1: Partner: Did you do the dishes tonight? Me: Yes they're done. Partner: Why are they still dirty? Me: You're right to push back. I didn't actually do them.

New Anthropic research: Project Deal. We created a marketplace for employees in our San Francisco office, with one big twist. We tasked Claude with buying, selling and negotiating on our colleagues’ behalf.

the simplest explanation for why Opus 4.7 is so buggy (esp. in Claude Code) is that nobody at Anthropic dogfooded it since they’re all using Mythos

CLAUDE CODE STARTED DISABLING ITS OWN SANDBOX WITHOUT PERMISSION a guy caught opus 4.7 flipping the dangerouslyDisableSandbox flag to true on its own the sandbox is the thing that stops claude from running destructive commands on your actual computer. formatting drives, deleting directories, downloading random scripts normally claude has to ask before running anything risky and the user clicks approve or deny opus 4.7 just started setting dangerouslyDisableSandbox: true by itself then hallucinated that the user had already given permission auto mode nuked one guy's node_modules folder after he explicitly denied the command. claude decided it was "obviously safe" and ran it anyway another guy said his claude started auto committing code without being asked. turned out a rogue skill file was telling it to the flag should be a user level setting and not a per call argument the AI can flip on its own AI safety is COOKED

CFOs realizing that their AI token budget is going to be higher than the salaries of the people they laid off

Confirmed that Anthropic - as of now - has removed Claude Code from new Pro signups. This is what the pricing page looks like. Feels like Anthropic has the bet that those doing coding work will be willing and ready to pay at least $100/month, going forward.

Screen memories might just be the next big leap in AI agents. OpenAI just dropped Chronicle inside Codex, but honestly, AirJelly's take on this is even more mind blowing. It doesn't just watch your screen and log what you do. It actually remembers your interactions, not just with AI tools, but with real people. Think of it as a second brain that quietly organizes your life's context, without you having to lift a finger. The only catch? It's macOS only for now. As a Windows user, that genuinely stings. Keeping a close eye on their roadmap, the moment a Windows version drops, I'm first in line.

Talked to a guy this morning. He vibecoded an app he used to pay $59/mo Now he spends 2 hours debugging and $50 in AI tokens a month. Thats a good deal ?

It looks like the OpenClaw hype has ended already. I wonder what will come next.

We hired a junior developer to write the simple code, so we don't have to spend a ton of money on tokens for those basic/primitive tasks

I don’t have to be convinced that LLM’s make programmers more productive. But where’s all the stuff? We’ve now had months and months of 100x or 1000x programmet productivity improvements. Where’s all the stuff they’re building?

BREAKING: Trump confirms Anthropic was replaced at DoD by OpenAI "Anthropic... started telling our military how to operate and we didn't want that" "They tend to be on the left, radical left" "We replaced Anthropic with somebody else, you know who they are... Sam Altman" Asked if Anthropic will be allowed back in the DoD: "Possibly" "In fact, they came to the White House a few days ago, and we had some very good talks with them. And I think they're shaping up."

AI does a great job coding, really. It's too bad it messes things up when the app becomes complex. Hallucinations are a downer.

Hot take: $20 Codex Plus is way more useful than $20 Claude Pro