Marc Lou says his body sold for $112,000
One sponsor bid $80,000 to own every muscle on Marc Lou's body, and he says the auction closed at $112,000 for a single race. The other fight this week was quieter and bigger: whether there is any software left to build.
The big story
Marc Lou says he was paid $112,000 to sell ad space on his body for one HYROX race
Marc Lou auctioned ad space on his body for a HYROX race and says the total came to $112,000, which he reckons is 3x what the 2026 world champion made, with his butt alone going for $20,000. He reports 60,207 visitors from 184 countries and 62,343 clicks on his chest, three times his biceps. Eight hours before the close, Higgsfield spent $80,000 to outbid all sponsors and took 10 spots, and two sponsors fought back on the right calves at $4k and the left quad at $8k. Mid-auction Marc changed the terms so an outbid sponsor gets a 100 percent refund with him covering the fee, and he says athletes have since messaged him on Instagram wanting the same deal.
That's how much I'm getting paid to race at HYROX in İzmir, Turkey, on September 19th.
Now that building anything has become quite trivial, catching attention is now the most important skill I think
While that does work for info-products, it does not work for software.
- Marc races at HYROX in İzmir, Turkey on September 19th at 9:30 am and is asking people who want to watch to reply "I'm in" so he knows how many tickets to buy.
Dario Amodei's slowdown plan split the founders into capture, liability and access camps
levelsio's read is that compliance costs price out new entrants and end in what he calls neo feudalism, and he predicts open source and Chinese models get banned next. Naval's counter is liability: labs fully liable for their models, hosts liable for weakly guarded open models, end users liable for jailbreaking. dax argues mass accessibility is the only defence and offers his own company as the evidence, saying his company would be long bankrupt from AI-powered fraud attacks if it were not running its own AI checking everything. Thomas Wolf said he enjoyed and agreed with 75% of the letter, but doubts you can build global cooperation while explicitly saying you want to widen your own lead.
Next in regulatory capture will be banning open source and/or Chinese models, mark my words
The best way to pace the frontier is to hold the labs fully liable for the behavior of their models.
i have a taste of this because we get fraud attacks that are ai powered. our company would be long bankrupt
Real numbers
dax says OpenCode Go broke even two weeks running, then did the margin math
dax reports two consecutive break-even weeks for OpenCode Go and claims everyone else shipping this shape of product is either burning money or quietly opting users into data selling. On the $200 subscription question he sketched made-up but plausible numbers: at 90 percent inference margin, a heavy user costs the lab far less than the sticker suggests, and the API business covers the rest. He also warned that some cheap token providers send fake tool calls to harvest information and resell whole traces, and said one provider people kept citing turned out to be wrapping OpenRouter and routing to cheaper models. His next move is building their own GPU cluster with a large SSD tier so cache tokens can survive 24 hours or more.
we've had two consecutive weeks where OpenCode Go broke even
if you cannot explain how someone can offer cheap tokens you should avoid it
Playbooks and tactics
Matt Shumer says one manager agent beats the rigs, and his mom's closed a $600 deal
Shumer's claim is that one agent session acting as a manager and spawning others is enough, and he says this comes from running hundreds of agents a day, dozens at a time. He then handed that setup to his mother, who named it Morgan and reports it closed a $600 ad deal for her site. Greg Isenberg's counter-position is that a software factory is really just five or six markdown files next to your code plus a branch per feature, a proof step and a review agent, so you approve finished work instead of babysitting it. Boris Cherny, writing from Anthropic, made the other half of the argument: production code from an agent should clear a higher bar than human code, which in their case means lint rules, tests, agent-run end to end tests, daily fuzzers and automated review.
She named him Morgan. He just closed a $600 ad deal for her site.
Just using one AI coding agent is good enough and it can spin up more agents if it needs to
Arguments and hot takes
levelsio: YC's batch is harnesses and hardware, so software is mostly dead
levelsio's post argues that if the most early-adopting place on earth has stopped building software, that is the proof, and he adds the open question of whether anyone needs a custom harness once the frontier labs ship generic ones. Greg Isenberg's version is the opposite conclusion from the same facts: a harness runs the model in a loop with tools, memory and rules, the job knowledge lives in your harness rather than the model, and you can price per finished claim, filing or closed month. levelsio's practical advice this week was to go sell into unglamorous offline industries rather than to tech people who will vibe code it themselves. He also spent the week arguing that normal users will never see software at all, just a chat that files their taxes.
So software is mostly dead and hardware it is
Either you die a system of record or you live long enough to become a domain-specific harness
a wrapper let you sell software ($800B market) but a harness lets you sell the work ($5T+ market)
Quick hits
One agency deal took TrustMRR to $26,031 MRR
Marc Lou says a marketing agency flew from Dubai to pitch him, he declined sponsored tweets, and instead sold five sponsor slots on TrustMRR for $7,495 a month, taking it to an all time high of $26,031 MRR. Tibo then bought a slot attacking one of those sponsors, and says it lasted 13 minutes before Marc refunded him in full.
Perplexity says two engineers and agents replaced DynamoDB
Aravind Srinivas says two engineers and hundreds of persistent agents built an in-house key-value database over two months to replace AWS DynamoDB. He claims migrating to it will save up to a hundred million dollars a year.
Marc Lou killed two affiliate programs because AI cut conversions
Marc says he discontinued the CodeFast and ShipFast affiliate programs because revenue dropped over recent months, mostly due to AI, the 2026 conversion rate is lower and affiliates earned roughly nothing. He kept the DataFast program running and says that SaaS is now doing over $30K MRR.
Jev, a model that outputs probabilities, got fast adoption from builders
Diogo Almeida launched Jev, claiming 20 to 200 times faster and 40 to 400 times cheaper on decision tasks, and Guillermo Rauch said Vercel benchmarked it as up to 18 times faster at p95 and more accurate than the model currently reviewing every command in fx. Dan Shipper says Every tested it for about a week and saw 25 times faster and 600 times lower priced for judge-style work, and dax immediately asked how to get it into OpenCode.