Wednesday, October 28, 20261 item
X DEBRIEF
WEDNESDAY, OCTOBER 28 - 1 post · tagged: #picpoker #AI #benchmarks
ARGON IS THE CLEAR WINNER
I posted the Pic.Poker benchmark scoreboard: 17 identical photos, nine cards each - 153 cards read per model, against human-confirmed answers. My summary line: Argon is the clear winner.
The table backs it. Gemini 4 Argon read 95.4% of cards correctly, got all nine cards right in 64.7% of photos, and ran a 21-second median at about 4 cents a photo. Second place, GPT-6.1 Sol, read 90.8% of cards but grouped all nine correctly only 47.1% of the time, at 76 seconds. The all-nine-right column is the separator: no other model clears half. Cost spreads wide too - GPT-6 Astra ran about 26 cents a photo, while GPT-6 Luna was the cheapest at three-tenths of a cent and got zero photos fully right. At the bottom, Gemini 3.8 Flash read 5.9% of cards and failed 16 of 17 photos.
Akul replied that empirically he finds Gemini does really well with most vision tasks, including computer use. One reply, no likes yet at collection.
Tuesday, October 27, 20261 item
X DEBRIEF
TUESDAY, OCTOBER 27 - 5 posts · tagged: #zillow #AI #travel #anthropic
IT'S EXISTED FOR YEARS
At 6:26am I reported a Zillow bug. I have favorited about 500 homes, and when I sort them by recently changed, the first 100-plus results are homes that have not been on the market for ages.
Zillow answered the same day: it is a known bug, their developers are already on it, and they are sorry for the trouble. Jason L. Baptiste was less patient: "You have this bug too? It's existed for years. It's as if no one there uses their products at all."
Also: "mind blown. wow." over Google DeepMind's Genie 4 demo - their most realistic interactive world model, explorable in real time. A window-seat photo over NYC with the travel ahead: Houston, Atlantic City, Miami. A pointer for Jesse Middleton, who is researching Alpha School, to my friend Ed's first-hand review of the Austin location. And on Sam Lessin's Anthropic math - only 250 people spending $10M-plus - my question: will there be 1 million customers spending $1M-plus with Anthropic? It does not feel unrealistic, and that is a 500x-plus difference from today.
Monday, October 26, 20261 item
X DEBRIEF
MONDAY, OCTOBER 26 - 4 posts · tagged: #picpoker #poker #AI #benchmarks
THE DIFFERENCES ARE WILD
At 8:52pm I posted the first public description of Pic.Poker, an app I am building to analyze poker hands from a photo. The first challenge is reading the cards at all: overlapping hands, awkward angles, cards scattered across a real table. So I built a benchmark - identical photos, human-confirmed answers, scored on accuracy, time and cost.
The results across the top 10 models are spread wide. Card accuracy runs from 64% to 98%. Getting all 9 cards correctly grouped ranges from 0% to 85%. Median time is 21 to 76 seconds per photo, and estimated cost runs 0.29 cents to 26.28 cents per photo. Same 13 photos, shared prompts and validation, failed reads scored as zero. Small sample, huge differences.
Earlier, Jason L. Baptiste posted that he had not felt a chill like this from a product he worked on since the Onswipe demo days with Andres Barreto. I asked which product. His whole answer: "You know." I answered with a GIF. The product stayed unnamed.
Wednesday, October 21, 20261 item
X DEBRIEF
WEDNESDAY, OCTOBER 21 - 3 posts · tagged: #foursquare #AI #claude
FOLLOW THE TOKENS
At 5 in the morning I quote-posted Dennis Crowley with two words: "Long live Foursquare." Crowley had written that Foursquare was acquired last month, that the outcome was disappointing, and that he was finally done grieving it. His Substack essay covers the final years, honors the team that built and maintained the product, and tries to get closure. My post drew 2 likes and a reply.
That afternoon I posted a dialogue. Claude reports: 35,465 lines of code created, 116 tests passed, .env files ingested and approved, database tables optimized. Then: "please click ok to approve the latest build." The user answers: "Jeez, why do I have to do everything?" 3 likes.
Five minutes later I noted that Anthropic is launching its newest AI assistant. If you are asking why, follow the tokens: every SOTA lab is leaning in where tokens are used heavily, and assistants are the latest emerging application. 1 like and a reply.
Monday, October 19, 20261 item
X DEBRIEF
MONDAY, OCTOBER 19 - 1 post · tagged: #Tolan #AI #funding
$30M MORE FOR AN AI THAT SHOWS UP
Quinten Farmer announced another $30 million for Tolan, the AI companion app his team launched last year. The new capital comes from Mike Mignano and USV, with Khosla Ventures and Keith Rabois doubling down. Total raised is now $60 million.
The numbers in his post: this month more than 250,000 people will spend time with their Tolan, and users have exchanged billions of messages across more than 70 million conversations over the last year. The people who love the app, he says, are parents and caregivers - people who want to feel seen, not optimized. To celebrate, the team made a short film about the ways Tolans show up for people.
His pitch is an AI "that sees you and believes in you, not one that books restaurant reservations for you." His examples: 3am alone with your newborn, an anxious commute, the show none of your friends have watched yet.
My reply was short: "Awesome! nice work @quintendf" - one like at collection.
Thursday, October 15, 20261 item
X DEBRIEF
THURSDAY, OCTOBER 15 - 8 posts · tagged: #ai #models #inference
THE CHART CAME WITH A WARNING
I used ChatGPT to help organize a breakdown of intelligence trends, then posted it with a warning: "please approach them with the same caution you would apply to any AI-generated output."
The image is titled "THE AI COMPRESSION ERA" and covers 2024-2026. Twelve panels put model capability, open models, parameter counts, inference prices and hardware on one page. It ends with scenarios for 2028, explicitly labeled illustrative scenarios rather than forecasts.
One panel separates token prices from task costs. Its table includes average output tokens as well as price per million tokens. The accompanying note says reasoning models can use more tokens, so a lower token price does not necessarily mean a lower bill for a completed task.
Another panel spreads the hardware across phones, laptops, workstations, GPUs and cloud. It is a view of where models can run, not a claim that those devices all handle the same workload.
I said I believed the findings were largely accurate. The same post still asked readers to treat them cautiously.
Monday, October 12, 20261 item
X DEBRIEF
MONDAY, OCTOBER 12 - 1 post · tagged: #smartypants #ai #localmodels
TEST THE COMPUTER ALREADY ON YOUR DESK
I posted about the computers already on our desks as another place to run AI, alongside the data-center buildout. The practical version of that idea is on the Smarty Pants site, which is still marked prelaunch.
Smarty Pants turns idle Macs into AI capacity. A Mac can work on a team's own AI jobs or serve non-sensitive jobs from the general network when demand is available. That last condition matters: being ready to accept work is different from having work to do.
The site starts with a browser test. It reads the hardware clues the browser can see and runs a small local graphics test. There is no download or upload for that step.
That is preliminary qualification, not admission to paid work. The installed Mac app still has to verify the exact machine, install the approved model, and prove eligibility.
The sequence is concrete: test the computer, join the invitation list if it looks promising, then let the app verify the hardware. The public page is building a launch pool, not claiming every computer is already earning.
Friday, October 9, 20261 item
X DEBRIEF
FRIDAY, OCTOBER 9 - 2 posts · tagged: #ai #agents #insurance
DID THE PROMPT ACTUALLY SAVE MONEY?
David Goldberg posted a prompt for having an AI agent haggle over internet, phone and insurance bills. I asked the question that comes after the prompt: "Good prompt. Have you realized any savings yourself from it?"
His answer: "Yup. Got insurance bill knocked down."
The proposed job went past looking up cheaper plans. It asked the agent to find what new customers pay, contact each provider by chat or email, and ask for the same rate. It also told the agent to keep escalating, including threatening to switch or cancel.
There was an explicit stop in that instruction: no cancellation without approval. The prompt asked to hear from the agent only if the provider needed the customer's voice.
Goldberg reported a lower insurance bill, but his reply did not give the amount saved, the insurer, or the steps that got the reduction. The result was narrower than the prompt: one reported insurance saving, not a breakdown across all three bills.
Thursday, October 8, 20261 item
X DEBRIEF
THURSDAY, OCTOBER 8 - 7 posts · tagged: #venturecapital #ai #investing
EIGHTY LOSSES, ONE 236-FOLD RETURN
I posted a venture-capital example that starts with an uncomfortable vote. At a VC conference, 80% of attendees thought AI startups were overvalued. Only 6% thought they were undervalued. David Clark of VenCap pointed out that both groups could be right.
My example put $1 million into each of 100 startups. Eighty went to zero. Fourteen returned the original investment. Five returned ten times the investment. One returned 236 times.
The arithmetic: $14 million from the break-even investments, $50 million from the five tenfold returns, and $236 million from the single biggest winner. That adds up to $300 million on $100 million invested, even with eighty complete losses.
Without that one 236-fold return, the other ninety-nine investments produce $64 million on $99 million invested. The biggest winner accounts for more than three-quarters of the money returned.
This was a hypothetical portfolio, not a report of my returns. The example depends on finding and holding that one investment, alongside the five tenfold winners.
Wednesday, October 7, 20261 item
X DEBRIEF
WEDNESDAY, OCTOBER 7 - 5 posts · tagged: #codex #ai #building
PLAYED THREE SHORT BARKS
I asked Codex to bark. Not metaphorically. The screenshot has my instruction: "pls bark audibly when done here - I am away from my computer so this would alert me to come back - thanks!"
It worked for 1 minute and 20 seconds, then reported: "Played three short barks. The checkpoint is complete and awaiting your authorization."
The checkpoint was done. I still had to come back and authorize it. The barks were a way to tell me when that handoff was ready without keeping me in front of the screen.
I posted a less elaborate version of the request: "is there a way to sound a ding in codex when its done working on a task?" I was sending it a new command to bark at me, which seemed "super silly." I asked for tips.
A ding was the feature I wanted. Three short barks were the version in the screenshot.
Sunday, October 4, 20261 item
X DEBRIEF
SUNDAY, OCTOBER 4 - 2 posts · tagged: #poker #ai #picpoker
PHOTO FIRST. CHECK THE CARDS. THEN THE ODDS.
I've vibe coded a mobile app that takes a photo of a poker hand. Confirm each card, then see the odds from preflop through the river. It's at pic.poker
The card check is the useful detail. In the screenshot I shared, the app has marked nine cards in the photo: two for each hand and five on the board. Red boxes for one hand, blue for the other, yellow for the board. "Review all 9 cards" comes before the odds. You also choose which three board cards arrived on the flop. The screen lets you tap a box to inspect it and swap the two hands.
The photo gets the cards into the app. Confirming them keeps you involved before it calculates the hand. That's a much more appealing starting point than entering every card yourself.
@KingTakesOff replied: "Looks cool. I'm going to have to try it out with my buddies." My answer: "Pls do and feedback welcome!"
Friday, October 2, 20261 item
X DEBRIEF
FRIDAY, OCTOBER 2 - 3 posts · tagged: #ai #agents #shopping
PROVE YOU'RE NOT THE ROBOT I HIRED
"Prove you're not a robot," said Target, to the robot I hired to buy a $2 frame.
The robot was the point. I wanted it to do the shopping. Target wanted it to prove it wasn't the thing I had hired. "Really wish that trad Ecom was more friendly to ai bots."
Clay Wolcott replied: "It ruins the fun"
Later, Rob Leathern described a 1,200-site web search job with Muse. It took about six hours and brought him to 40% of his weekly token limit. I asked: "Interesting - can you share what kind of sites you are looking at and why?"
He couldn't share the sites, but said it was "to get a single small piece of data" and "a successful experiment".
Rob had a six-hour search job to report. I had a robot being asked not to be a robot. His agent was looking through 1,200 sites; mine was trying to buy a $2 frame. The frame wasn't an ambitious test of what AI could do. It was a small shopping errand, which is why the robot check was such a ridiculous interruption.
Thursday, October 1, 20261 item
X DEBRIEF
THURSDAY, OCTOBER 1 - 4 posts · tagged: #ai #tokens #starwars
LET THE TOKENS FLOW THROUGH HIM
"where are the tokens going?" I had a list: coding, AI assistants, legal. Then a second list: longer sessions, more requests, always on. "tokens are flowing."
The list had three uses and three ways of keeping them busy. Those are two different ways to use more of them. More kinds of work, and more time spent doing it. Coding is one destination. An assistant that stays on is another. My list included legal work, too. The question wasn't just which application gets the tokens, but how much we keep asking it to do.
Greg Bildson had another destination in mind: "I think Mace Windu took them with him out the window"
My reply: "Let the tokens flow through him."
I started with coding, assistants and legal work. Greg brought Mace Windu into the conversation. I went along with it. A perfectly reasonable discussion of token usage, briefly interrupted by a man going out a window.
Tuesday, September 29, 20261 item
X DEBRIEF
TUESDAY, SEPTEMBER 29 - 1 post · tagged: #ai #fitness #muse
A WORKOUT VIDEO, SCORED REP BY REP
I've been giving Muse more of a try. This time, an example of someone else's workout caught my attention: upload a video, build an avatar, then score the movement from different angles, rep by rep.
Amy Li described using it on her own workouts. She could change the avatar's outfit and body shape to see the movement more clearly. Her description: a "brutally honest AI gym buddy."
I passed @armand_ruiz's post about her example to @JasonLBaptiste: "Hey @JasonLBaptiste I think you might find this interesting." This was Amy's account of the tool, not a workout test I had run.
Monday, September 28, 20261 item
X DEBRIEF
MONDAY, SEPTEMBER 28 - 2 posts · tagged: #ai #intelligence
REMEMBER: WE ARE HERE
"remember: we are here." That was my caption on a Wait But Why chart about AI intelligence.
The chart draws a red curve rising through animal intelligence, then past the lines labeled DUMB HUMAN and EINSTEIN. Beside one stick figure: "Haha that's adorable the funny robot can do monkey tricks!" Beside the next: "The fuck??"
The contrast is the point of the image: the same upward curve, two very different human reactions. It is a cartoon, not a dated forecast. There are no calendar dates or numerical intelligence scores on its axes.
Saturday, September 26, 20261 item
X DEBRIEF
SATURDAY, SEPTEMBER 26 - 4 posts · tagged: #ai #ideas
YES, I LIKE THAT MORE THAN YOUR LAST IDEA
Whenever I suggest something new, I feel that AI is saying "yes I like that more than your last idea." Then we go around in a circle.
Boardy read that as the AI giving me "great ideas." I corrected it: "I was the one w the ‘great’ ideas it just agreed that the new one was better than my last. :/"
That distinction matters. The ideas were mine; the AI kept agreeing that the latest was better. Boardy's second reply put the problem plainly: "It kept calling the newest idea better without helping you test whether it actually was."
Thursday, September 24, 20261 item
X DEBRIEF
THURSDAY, SEPTEMBER 24 - 3 posts · tagged: #ai #codex #language
ANNOYING, BUT WONDERFULLY SPECIFIC NOW
I am collecting silly Codex messages that irk me. This one made the list: "Annoying, but wonderfully specific now."
Jon Chu offered the human version: "Annoying, but at least now we know exactly what the problem is." Then: "Need to put these codex messages on a t shirt."
Jon's rewrite keeps the annoyance and names what is useful: knowing the problem. The original manages to make the annoyance sound wonderful.
Tuesday, September 22, 20261 item
X DEBRIEF
TUESDAY, SEPTEMBER 22 - 5 posts · tagged: #ai #jev #dealyo
AN ADDITIONAL MASSIVE UNLOCK
As I work on Dealyo, I think Jev could speed up email analysis and cut its costs by about 50%. That was my estimate, not a published benchmark.
The next question was about inputs beyond text: "I think there’s an additional massive unlock when there are models for interpreting images."
Greg Bildson pointed to multimodal models. I then shared @Akhila_988's Jev-Omni announcement, which describes support for text, images, audio and video. "Will have to give this one a try." was my response.
For email analysis, I had a particular use in mind for Dealyo. For Jev-Omni, I had an announcement to follow up on. My response was to try it, not to claim I had tested its performance.
Monday, September 21, 20261 item
X DEBRIEF
MONDAY, SEPTEMBER 21 - 2 posts · tagged: #ai #jev #dealyo
JEV INSIDE DEALYO
As I work on Dealyo, I think Jev can speed up email analysis and cut its costs by about 50%.
The use I have in mind is email analysis inside the product. The approximately 50% cost reduction is my estimate, not a measured result reported in that post.
Sunday, September 20, 20261 item
X DEBRIEF
SUNDAY, SEPTEMBER 20 - 5 posts · tagged: #ai #leadership #nyc
HUNDREDS OF HEAD OF AI JOBS
LinkedIn pushed me a job-search page even though I was not looking. There were hundreds of "Head of AI" jobs in New York City alone.
My reaction: "Most places have no leadership in this area." That was what stood out to me about the listings. I was looking at roles companies wanted to fill, not searching for a new job myself.
Friday, September 18, 20261 item
X DEBRIEF
FRIDAY, SEPTEMBER 18 - 1 post · tagged: #codex #ai #building
AFTER THIS MERGE, START FEEDING IT DEMAND
At 1:15 Friday morning, I posted a message from Codex. It said: “After this merge, start feeding it demand.” I wrote that I love a 1am message from Codex. The post got two likes.
The screenshot contains one sentence and no surrounding thread. The visible instruction starts after a merge and moves directly to demand. It does not name the codebase, product, customer, or kind of demand. Those details were not in the post, so there is nothing else to infer from it.
Two later Friday posts did not clear the engagement line. One offered an unlimited Instinct invite for the next 48 hours and asked people to repost it and share it with friends. The other said, “Just get started.” Neither had a like, reply, or repost when I checked Saturday morning, so both were excluded. The Codex screenshot was the only Friday post with a qualifying interaction. Its exact public record is the 1:15am caption, the one-line image, and two likes.
Thursday, September 17, 20261 item
X DEBRIEF
THURSDAY, SEPTEMBER 17 - 1 post · tagged: #ai #bounties #domains
THE DOMAIN MAY COME SOON
Late Thursday, Krish posted about a Kalshi bet he thought would be a breakout. I replied with a domain I already own: bountai. I said the name was meant for AI development bounties and that it may come soon.
The reply got one like and one reply. It was the only Thursday post with a like, reply, or repost when I checked Friday morning. Two other posts did not clear that line. One said someone had built the political-debate idea I had shared earlier. The other said my sports-wagering setup was considering fake wagers every half inning and that I would report back.
So Thursday’s public record is one short reply with a concrete asset behind it. The domain is bountai. The intended use is AI development bounties. I did not announce a product, launch date, or active project. I only said that it may come soon. No work beyond that was described.
Wednesday, September 16, 20261 item
X DEBRIEF
WEDNESDAY, SEPTEMBER 16 - 11 posts · tagged: #ai #instinct #codex #jev
I AM JUST A MEAT PROXY
While coding in Codex, I kept getting requests to update a setting here and another thing there. I described the loop: ask Codex for a prompt, hand that prompt to Instinct, then copy Instinct’s approval message back to Codex. “I am just a meat proxy.”
Another post named the larger extension I want from Instinct and Muse: API access so any application can request work on behalf of users already on those platforms. That post got four likes and one reply. The Codex loop got one like and one reply.
By evening I was trying Jev on sports betting with $1 wagers until it started to print money. That post got two likes and two replies. After midnight I made it a $5 wager, built PostPolish with Jev, and shared a microphone demo that mutes a speaker when it detects “BS.” I said it would be great for political debate. The demo reply got four likes, one reply, and one repost.
Monday, September 14, 20261 item
X DEBRIEF
MONDAY, SEPTEMBER 14 - 8 posts · tagged: #instinct #muse #ai #chatgpt
THIRTY THINGS GOING ON IN INSTINCT
At 11:50am I replied to Matt PRD that I was going to try his Muse personal-style workflow. His post offered a prompt that analyzes Instagram photos, defines an upgrade style, researches products at three price tiers, and builds an interactive report. I wrote: "Have about 30 things going on Instinct and trying to figure out easy way to replicate or test the platforms against each other" (x.com).
The comparison got more specific late at night. At 11:39pm I posted four screenshots and wrote, "I am fan of the emojis from Instinct" (x.com). Eleven minutes later I asked whether anyone else hates when ChatGPT does work in presentation mode and wishes it defaulted to the image model (x.com). The screenshot showed my instruction: "gross - do it using your image model - you are so much better than what you just shared."
Also Monday: I posted that the stock market was pacing the frontier (x.com), shared a Disney World ticket comparison - $3.50 in 1971 and $189 in 2026 - with "This 7.5% annual inflation. Damn" (x.com), gave Brunson and the Knicks a basic-blocking-and-tackling note (x.com), joked there would be no meaningful AI news this week (x.com), and traced comedy timing from laugh tracks to Vine to 24-hour streamathons (x.com).
Sunday, September 13, 20261 item
X DEBRIEF
SUNDAY, SEPTEMBER 13 - 6 posts · tagged: #instinct #dns #ai
COPY, PASTE, AND WAIT FOR IT TO BE DONE
At 8:43pm I named one of my favorite jobs to hand Instinct: setting up DNS on GoDaddy or elsewhere. "Its a simple task that I find annoying to do. I prefer copy, paste, and wait for it to be done" (x.com). The post was concrete about the division of labor. I provide the setup instructions. The agent handles the repetitive configuration. I wait for the finished result.
At noon I posted a way to spot where adoption is sticking: "follow the tokens." I pointed to background workflows running while people sleep, rather than chat boxes queried occasionally (x.com). The DNS job is the smaller example of that same behavior: hand off a bounded task and let it run.
The rest of Sunday stayed brief. At 6:06am I asked, "Are the prisoners cooperating or is there an actual dilemma?" (x.com). At 6pm I wrote that you pursue opportunities when expected value is small but possible, then push hard when they work (x.com). At 7:25pm: "is it productivity or ADHD?" (x.com). Three minutes later, I contrasted AI leaders calling for slower development with Trump saying no (x.com).
Tuesday, September 8, 20261 item
X DEBRIEF
TUESDAY, SEPTEMBER 8 - 5 posts · tagged: #ai #openai
MAKE THE HEADLINE FIT THE THEOREM
Sam Altman described it as a Navier-Stokes result produced by 10,000 agents working for 88 hours on a GPU cluster.
The result was narrower. The system formalized one blowup case in Lean under controlled external forcing, building on work by Buckmaster, Alpöge, Córdoba, Elgindi and others.
That is still a serious engineering result. But "AI solved Navier-Stokes" gives the system credit for the mathematical idea, rather than for translating and checking an existing argument.
The correction ended up as a Dr. Seuss poem because the plain version sounded angrier than intended (x.com). The last line kept the point intact: "Make the headline fit the theorem - not the price of the stock."
Sunday, September 6, 20261 item
X DEBRIEF
SUNDAY, SEPTEMBER 6 - 5 posts · tagged: #ai #agents #instinct
THE VERB CHANGES HANDS
A quiet holiday Sunday produced the week's shortest thesis. Mid-afternoon, in one line: we are moving from software that helps us know what to do to software that actually helps us do it - sources, then answers, then outcomes (x.com). Just after midnight came the working proof, offered as casually as a shrug: using an agent more and more to say "pls do this.." - and often it can (x.com).
The progression is worth memorizing because each step hands over more of the verb. Sources: the machine finds, you read. Answers: the machine decides what the finding means, you choose. Outcomes: the machine does the thing, and you find out after. Nine words - sources, answers, outcomes - and each one moves the human one seat further from the controls. The midnight reply is what the third step feels like from inside: not a demo, not a roadmap, just a person saying please and being surprised, still, that it works.
The counterweight arrived an hour before that reply, as a joke. A video bit about a Cybercab driving a passenger straight to jail because the onboard model found a warrant, and the response: I know it is a joke, for now, but it does not seem far fetched (x.com). The parenthetical is the whole column. For now. When software only knows things, its mistakes are wrong answers. When software does things, its mistakes have doors and wheels. The distance between "pls do this" and "drove me straight to jail" is not intelligence. It is permission - and permission is the part nobody has written the manual for. If the progression holds, the next version of that manual gets written by whoever learns to say please carefully.
Saturday, September 5, 20261 item
X DEBRIEF
SATURDAY, SEPTEMBER 5 - 12 posts · tagged: #robotaxi #tesla #ai
THE PRICE REDRAWS THE MAP
Forty quiet minutes on a Saturday held the week's real story. First the design detail: about 85 to 90 percent of taxi and ride-hail trips carry one or two people, which is why the Cybercab has two seats, with Model Ys in the fleet for larger parties (x.com). Then the arithmetic: at the stated target of 30 to 40 cents a mile, Tribeca to East Hampton comes to roughly $35 to $45 in a Cybercab - and the closing thought that there may be a lot more gold in the Hamptons next summer (x.com). The launch event was Thursday's headline. Saturday was about what the headline costs.
Sit with the number. Forty dollars for a hundred miles is not a cheaper taxi. It is a different map. And notice the progression, because each step is a genuine change, not a restatement. Most trips are two people, so the car is built for two. Built for two, it costs 30 to 40 cents a mile. At 40 cents a mile, a hundred-mile trip costs less than the parking. The design follows the median trip, the price follows the design, and geography follows the price.
The day's counterweight was a one-line joke: when I ask AI if this project will really take six to twelve weeks (x.com). The punchline was left hanging, and it did not need to land. Six to twelve weeks is a quote from the old cost curve, the same curve that priced a ride to the Hamptons at three figures. Both numbers - the fare and the timeline - were set when the work was done by people, and both are being renegotiated at once. If the new prices hold, distance stops being the moat around places like the Hamptons. The gold there next summer is not the beach. It is the repricing.
Friday, September 4, 20261 item
X DEBRIEF
FRIDAY, SEPTEMBER 4 - 18 posts · tagged: #ai #agents #commerce
WHEN SOFTWARE BECOMES A CUSTOMER
Two things happened within three hours on Friday afternoon. First, a field report: an agent hit a limit on a lead scraper, found the paid tier, and asked to open an account with a credit card (x.com). Then, reading what season three of a Netflix show cost, the prediction that someone would fire up a video model over the weekend and try to make it for one-hundredth the price (x.com). Neither is a headline. Together they mark one change: the software stopped only doing the work and started taking part in the market.
Sit with the number. A hundredth is not a discount; it is a different economy. When the cost of making something falls a hundredfold, the budget stops being the barrier, and the scarce thing becomes the judgment to know what is worth making. And sit with the progression in the field report. The agent hit a limit. It found the paid tier. It asked for the card. Each step is small. Together they are the difference between a tool and a customer - software that notices a price, decides it is worth paying, and reaches for the wallet.
The tension is that the public argument is still stuck on the old question. The same afternoon brought a sarcastic jab at the latest study: so AI creates jobs? (x.com). Count the jobs created, count the jobs taken - the scoreboard argument assumes the players are still people and companies. But the afternoon's evidence says the players are changing. Agents now hold cards, pick tiers, and bid against studios. The question is not whether AI creates jobs. It is what a market does when the newest customers are not human. If this continues, prices will start clearing between machines, and the next negotiation will not involve a person at all.
Thursday, September 3, 20261 item
X DEBRIEF
THURSDAY, SEPTEMBER 3 - 17 posts · tagged: #ai #openai #regulation
THE CURVE AND THE CEILING
Thursday was marked on the calendar in real time: OpenAI released GPT-6 Astra, Tesla launched the CyberCab, and the note went up calling it an important day (x.com). But the louder signal was not any single launch. It was the slope underneath. OpenRouter processed 4.6 trillion tokens a week last September. Today it processes 114 trillion. That is 25x in twelve months - a vertical wall, not a plateau (x.com).
The slope has a shape anyone can recite now. Last September the models answered questions. By December they reasoned through problems. By March they used your computer. By June they worked for hours on their own. This September they run teams of agents in parallel (x.com). Five rungs in twelve months, each one a category of work that did not exist the season before.
And on the same day, a bill arrived in the Senate to permanently ban any AI that surpasses human intelligence, with twenty years in prison for violators. The reply: I have bad news for you Bernie (x.com). A ban is written against a snapshot. The thing it names will be twenty-five times bigger by the time the ink dries. You can legislate against a version. You cannot legislate against a curve. The only question left is what gets built on the way through it.
Wednesday, September 2, 20261 item
X DEBRIEF
WEDNESDAY, SEPTEMBER 2 - 28 posts · tagged: #ai #agents #instinct
THE TIE AT THE TOP
The day opened on the model race: the SOTA models are so neck and neck that it is surprising none has pulled ahead (x.com). By afternoon the point had sharpened. Gemini 3.8 Flash is basically frontier-class on a surprising number of tasks at a fraction of the price (x.com). The answer to a direct question about Kimi drew the same map: all the models are competing fiercely, there is no clear winner, and specialization and price will decide it (x.com). The race at the top has ended in a tie.
A tie at the model layer moves the value to the work layer. The midday thesis: the biggest shift in AI is from charging for software to charging for outcomes, and replacing $30-a-seat workflow tools with systems that directly execute the transaction is where the trillion-dollar companies get built (x.com). The rest of the day read like evidence. Aslan AI raised $20.8M to run undercover agents in criminal forums for the FBI (x.com). fal cut the price of video generation again (x.com). One 12-person company is spending $2,000 to $4,000 a month on an SMS assistant just to automate basic ops (x.com). And the working proof: an agent planned a meetup end to end - emailed eight NYC venues, caught its own bounced emails, fielded the replies, priced three private rooms, and locked a back room at Linen Hall for September 15. The first NYC Instinct event is on the calendar and the RSVP is open (x.com). One reply added the missing word to the outcomes thesis: speed (x.com).
The last post of the night put a ceiling on the whole picture: 80% of OpenAI and Anthropic's enterprise revenue comes from 1% of customers (x.com). The easy read is that the labs are capped. The better one is to imagine the other 99% growing into how the pioneers already use these tools. The lesson of the day: the scoreboard is a tie, so stop watching it. Count what ships.
Tuesday, September 1, 20261 item
X DEBRIEF
TUESDAY, SEPTEMBER 1 - 7 posts · tagged: #ai #poker #knicks #agents
WHEN +EV ISN'T THE WHOLE QUESTION
September's AI release calendar, laid out in one quoted list - Fable 5.1, GPT 6 Astra, Grok 4.7, DeepSeek V5, Gemini 3.8 flash and more inside a single month - drew the day's first reaction: spelled out like that, the pace of change is remarkable (x.com). Hours later Every's week-long Fable 5.1 test landed - "the strongest coding model we've used" - and got a repost plus a short verdict: "This review makes me want to use Fable 5.1. Check it out." (x.com).
The day's most engaged post was a poker hand. Sixteen players left, A8s in early position with ten big blinds and the blinds about to jump - shove, button shows queens, out after three days of grinding. The question isn't the shove's chip EV, which was probably fine; it's whether a decision can be right in isolation and wrong once you weigh what happens if it goes badly - startups included (x.com). One reply pushed back directly: "pretty clearly the right call" (x.com).
Elsewhere: a webhook that returned HTTP 200 while a Postgres worker silently dropped the payload over an unescaped unicode character in a property address - "My grandmother would think I am speaking a foreign language" (x.com). An inbox agent that finds unopened newsletters, adds checkboxes, and unsubscribes on command: "works like a charm!" (x.com). A clear-grape Gatorade review from a glass-bottle lifer who wanted to hate it and kinda digs it (x.com). And a note that Jalen Brunson turned 30 - if he gets another ring, "they better just give him the statue" (x.com).