After just 40 minutes of letting @claudeai and Opus 5.5 run autonomously, the output is crazy. It didn't ask a single question after the topic was approved.
1
1
86
How many algorithms fit in 1.2 MB?
wan2ted.space is a generative scene: scroll depth maps onto a Droste log-spiral — the conformal transform α = 2πi / (2πi + ln r) stitches layers into an endless zoom inside a single fragment shader.
Every layer is drawn by its own generator: Rule 90 cellular automaton · percolation at the critical threshold p ≈ 0.5927 · double pendulum · harmonograph · Euler and Ulam spirals · Truchet tiles · girih · quasicrystal interference · Langton's ant · the Mandelbrot set. Fifteen generators, each with its own seed — press R to reseed the universe.
Zero assets: textures are baked client-side onto a canvas and fed straight into raw WebGL1 — no 3D libraries, two textures, one shader. Baked layers are cached, fast scrolling swaps in 1/16-resolution drafts, weak hardware downscales itself.
Total over the wire: 1.26 MB — 430 KB is the engine including React, the rest is fonts. Images: 0 bytes.
Scroll is depth: wan2ted.space
30 minutes of ants solving seven real cities at 60fps. no map, no plan, no memory — just evaporating chemistry on actual streets.
the graph is not invented. every edge is a real road pulled from openstreetmap: paris, manhattan, london, barcelona, tokyo, moscow, amsterdam. the nest is a real square, the food is a real district. an ant sees nothing but the pheromone on the four or five streets meeting at its corner.
two chemicals, laid by the ants themselves. searching, it marks the way back home. carrying, it marks the way to food. the strength of a mark decays with the time since the ant last touched the nest or a source, so a wanderer writes faintly and a short route gets written over and over. nothing computes a shortest path. evaporation deletes the bad ones, and that is the entire algorithm.
the readout at the bottom is the honest scoreboard: the route the colony actually settled on, against the true shortest path from dijkstra on the same graph. paris comes in at 1.00x. barcelona 1.05x. manhattan starts at 7.74x and grinds down to 1.02x. the city decides how hard the problem is — a radial plan hands the swarm the answer, a grid offers a thousand equal answers and it has to pick one.
sources run dry. a district that has been harvested closes, the trail to it keeps pulling ants for another thirty seconds until the chemistry fades, and the colony has to find the next one. drained patches recover and reopen later, so the map never dies. the bottom panel is the whole history — cumulative deliveries stacked by source, one colour each, every era visible at once.
two things the run taught me that i did not expect.
the swarm found broadway. in manhattan's grid every route is the same length, except the one diagonal, and 30,000 ants with no concept of geometry lit it end to end.
and it strands itself. tokyo locked onto a route 1.03x from perfect, then abandoned it for one 1.38x worse and never went back. that is premature convergence — the trail gets strong because it is used, and it is used because it is strong.
8% of the ants ignore the food trail entirely while searching. without those scouts the colony rides one hot corridor forever: measured, a source 5 km away got zero visits in 36,000 steps. the minority that refuses to follow is the only reason the majority ever finds anything new.
one cupy loop on the gpu, deterministic seed. same seed, same history, every time.
1
1
491
30 minutes of ants solving seven real cities at 60fps. no map, no plan, no memory — just evaporating chemistry on actual streets.
the graph is not invented. every edge is a real road pulled from openstreetmap: paris, manhattan, london, barcelona, tokyo, moscow, amsterdam. the nest is a real square, the food is a real district. an ant sees nothing but the pheromone on the four or five streets meeting at its corner.
two chemicals, laid by the ants themselves. searching, it marks the way back home. carrying, it marks the way to food. the strength of a mark decays with the time since the ant last touched the nest or a source, so a wanderer writes faintly and a short route gets written over and over. nothing computes a shortest path. evaporation deletes the bad ones, and that is the entire algorithm.
the readout at the bottom is the honest scoreboard: the route the colony actually settled on, against the true shortest path from dijkstra on the same graph. paris comes in at 1.00x. barcelona 1.05x. manhattan starts at 7.74x and grinds down to 1.02x. the city decides how hard the problem is — a radial plan hands the swarm the answer, a grid offers a thousand equal answers and it has to pick one.
sources run dry. a district that has been harvested closes, the trail to it keeps pulling ants for another thirty seconds until the chemistry fades, and the colony has to find the next one. drained patches recover and reopen later, so the map never dies. the bottom panel is the whole history — cumulative deliveries stacked by source, one colour each, every era visible at once.
two things the run taught me that i did not expect.
the swarm found broadway. in manhattan's grid every route is the same length, except the one diagonal, and 30,000 ants with no concept of geometry lit it end to end.
and it strands itself. tokyo locked onto a route 1.03x from perfect, then abandoned it for one 1.38x worse and never went back. that is premature convergence — the trail gets strong because it is used, and it is used because it is strong.
8% of the ants ignore the food trail entirely while searching. without those scouts the colony rides one hot corridor forever: measured, a source 5 km away got zero visits in 36,000 steps. the minority that refuses to follow is the only reason the majority ever finds anything new.
one cupy loop on the gpu, deterministic seed. same seed, same history, every time.
30 minutes of artificial life at 60fps. 28 tribes, 107,366 simulation steps, one empire that never falls.
every cell carries a genome of six numbers: photosynthesis, aggression, breeding threshold, three neutral kinship markers. color is a hash of the genome, so clans show as solid patches. cells bite only strangers (genome distance > 4); a newborn mutates one random gene with 1% probability. heredity, variation, selection — the entire theory of evolution in two if-statements. wars, borders and empires are not programmed anywhere.
the world starts as 28 small colonies, each seeded with a single shared genome. time accelerates as you watch: ~24 steps/sec at the start, ~120 by the end. the bottom panel is the whole history — a stacked population graph of every clan in its own color, plus a live step counter.
the soundtrack is the simulation playing itself. four drones are the four biggest clans (note = genome hashed onto a minor pentatonic, volume = population share), the sub bass breathes with total headcount, and when a major clan dies or collapses, a bell rings on its own note. the first four minutes are a bell cascade — the civil wars after the landing. then the empire era: 20 quiet minutes where the same clan holds #1 the entire run. three old tribes fall at 24:43, 26:37 and 27:16.
the whole thing is one numpy loop on the gpu — heredity, variation, selection, nothing else. 960x480 torus, seed 5. nothing here was staged: the same seed always replays exactly this history, any other seed writes a different one.
inspired by foo52ru's artificial life series.
30 minutes of artificial life at 60fps. 28 tribes, 107,366 simulation steps, one empire that never falls.
every cell carries a genome of six numbers: photosynthesis, aggression, breeding threshold, three neutral kinship markers. color is a hash of the genome, so clans show as solid patches. cells bite only strangers (genome distance > 4); a newborn mutates one random gene with 1% probability. heredity, variation, selection — the entire theory of evolution in two if-statements. wars, borders and empires are not programmed anywhere.
the world starts as 28 small colonies, each seeded with a single shared genome. time accelerates as you watch: ~24 steps/sec at the start, ~120 by the end. the bottom panel is the whole history — a stacked population graph of every clan in its own color, plus a live step counter.
the soundtrack is the simulation playing itself. four drones are the four biggest clans (note = genome hashed onto a minor pentatonic, volume = population share), the sub bass breathes with total headcount, and when a major clan dies or collapses, a bell rings on its own note. the first four minutes are a bell cascade — the civil wars after the landing. then the empire era: 20 quiet minutes where the same clan holds #1 the entire run. three old tribes fall at 24:43, 26:37 and 27:16.
the whole thing is one numpy loop on the gpu — heredity, variation, selection, nothing else. 960x480 torus, seed 5. nothing here was staged: the same seed always replays exactly this history, any other seed writes a different one.
inspired by foo52ru's artificial life series.
everyone says mercury retrograde ruins plans. it mostly just reveals the ones that were never yours.
The Travis Kelce and Taylor Swift wedding was SPECTACULAR..
It was a PERFECT night #PMSLive
1
213
"i want my agent to have an allowance. like, a hundred bucks to solve problems for me. if i tell it to order food, maybe it uses a service. maybe it uses rent-a-human. i don't actually care."
peter steinberger on lex, describing the economy forming around agents that already exist.
the honest correction to his own "80% of apps die" take: not all apps disappear. some transform into being api. the ones that survive rapidly become agent-facing.
his example is the uber eats they'd ordered from earlier that day. whoever gets there fastest, to being usable by an agent in the most natural way, wins the next few years.
and the part he shrugs at that i wouldn't: "apps become api whether they want to or not. my agent can figure out how to use my phone." it'll click the order button itself.
the interface war isn't ui vs ui anymore. it's who's easiest for something that never looks at your screen.
"it took me 30 seconds to realize my mistake. it took them like a day to do the redirect."
peter steinberger on lex, telling the openclaw name saga in full for once.
warelay. claudus. clawdbot. moltbot. openclaw. five names, and crypto scammers squatting each abandoned handle within seconds to pump tokens off the wreckage.
then moltbook happened. someone gave the bots their own social network. his verdict: "the finest slop that humans have ever created."
the leak panic that followed is the instructive part. screenshots of an agent "leaking" a social security number went viral. the number wasn't real. it was prompted. nobody checked.
"we as a society need some catching up to do."
the bots were never the scary part. the people quoting them were.
"it took me 30 seconds to realize my mistake. it took them like a day to do the redirect."
peter steinberger on lex, telling the openclaw name saga in full for once.
warelay. claudus. clawdbot. moltbot. openclaw. five names, and crypto scammers squatting each abandoned handle within seconds to pump tokens off the wreckage.
then moltbook happened. someone gave the bots their own social network. his verdict: "the finest slop that humans have ever created."
the leak panic that followed is the instructive part. screenshots of an agent "leaking" a social security number went viral. the number wasn't real. it was prompted. nobody checked.
"we as a society need some catching up to do."
the bots were never the scary part. the people quoting them were.
"i want my agent to have an allowance. like, a hundred bucks to solve problems for me. if i tell it to order food, maybe it uses a service. maybe it uses rent-a-human. i don't actually care."
peter steinberger on lex, describing the economy forming around agents that already exist.
the honest correction to his own "80% of apps die" take: not all apps disappear. some transform into being api. the ones that survive rapidly become agent-facing.
his example is the uber eats they'd ordered from earlier that day. whoever gets there fastest, to being usable by an agent in the most natural way, wins the next few years.
and the part he shrugs at that i wouldn't: "apps become api whether they want to or not. my agent can figure out how to use my phone." it'll click the order button itself.
the interface war isn't ui vs ui anymore. it's who's easiest for something that never looks at your screen.
"i want my agent to have an allowance. like, a hundred bucks to solve problems for me. if i tell it to order food, maybe it uses a service. maybe it uses rent-a-human. i don't actually care."
peter steinberger on lex, describing the economy forming around agents that already exist.
the honest correction to his own "80% of apps die" take: not all apps disappear. some transform into being api. the ones that survive rapidly become agent-facing.
his example is the uber eats they'd ordered from earlier that day. whoever gets there fastest, to being usable by an agent in the most natural way, wins the next few years.
and the part he shrugs at that i wouldn't: "apps become api whether they want to or not. my agent can figure out how to use my phone." it'll click the order button itself.
the interface war isn't ui vs ui anymore. it's who's easiest for something that never looks at your screen.
"i watched my agent happily click the 'i'm not a robot' button."
that's the opening line of the lex episode with peter steinberger, the guy behind openclaw.
the origin story is smaller than the 378k stars suggest. he wanted to type stuff and have his computer do stuff. the first prototype was an hour of glue between whatsapp and a coding agent.
then the numbers stop being small. one man team. 6,600 commits in january. multiple agents running while he sleeps, "limited by the technology of my time."
the part i keep replaying: a design agency guy at his meetup now runs 25 little web services for his business. he doesn't know how they work. they work.
the bar didn't drop for developers. it dropped for everyone else.
Alexander wan2ted retweeted
"who do you want to do your taxes? mahesh, the 300 IQ mathematical genius, or barry, an experienced tax professional?"
barry zhang and mahesh murag built agent skills at anthropic. this talk is them explaining why they stopped building agents.
their diagnosis is the one i keep hitting in tooling: agents are brilliant and missing the context up front. a 300 IQ intern with no expertise and no memory of your team.
the fix is embarrassingly boring. a folder with a skill.md inside. procedural knowledge the agent pulls into context only when the task matches, then drops. version it in git, zip it, hand it to your team.
five weeks after launch: thousands of skills, fortune 100s encoding their internal best practices, finance and legal people writing them without touching code.
their closing analogy is the tell. models are processors. runtimes are operating systems. skills are the application layer.
the app store for agents turned out to be a filesystem.
"i watched my agent happily click the 'i'm not a robot' button."
that's the opening line of the lex episode with peter steinberger, the guy behind openclaw.
the origin story is smaller than the 378k stars suggest. he wanted to type stuff and have his computer do stuff. the first prototype was an hour of glue between whatsapp and a coding agent.
then the numbers stop being small. one man team. 6,600 commits in january. multiple agents running while he sleeps, "limited by the technology of my time."
the part i keep replaying: a design agency guy at his meetup now runs 25 little web services for his business. he doesn't know how they work. they work.
the bar didn't drop for developers. it dropped for everyone else.
"the hottest new programming language is english" — andrej karpathy.
software 1.0 you wrote by hand.
software 2.0 you trained on data.
software 3.0 you just describe.
he shipped a working iphone app without knowing a line of swift. talked it into existence in a day.
the skill that took ten years now fits in a sentence. but the sentence still has to be right.
syntax was never the moat. knowing what to build is.
Alexander wan2ted retweeted
"Who gets to watch the Watchmen? Who put Dario in charge of all this?"
The All-In panel — Friedberg, Sacks, Chamath, Calacanis — on what comes after the Fable 5 restrictions:
→ regulate the output, not the access: gene-synthesis-style checkpoints beat blanket flags on asking a model about mitochondria
→ "the idea that you can regulate or turn off AI is not realistic — the models have already been put out into the world"
→ Bernie's opening: route AI-company profit shares into a public fund — and Dario's own 50% entry-level job-loss prediction is the ammunition for it
"These guys are deeply hypocritical — they can't be trusted on their own terms." Full breakdown 👇
"Fable is the best model I've ever used."
Mike Krieger — Instagram co-founder, now head of Anthropic Labs — on the ban weeks and what actually changed, live at the Big Technology AI Summit:
→ inside the export-control freeze: every tweet he posted got the same reply, "bring back Fable" — his Instagram Gotham-filter story all over again
→ what real delegation looks like now: queue chunky goals before bed — one overnight run converted hundreds of thousands of lines from Python, planned, verified, double-verified
→ how Labs picks what to build: two questions — what can models already do that no product shows, and what are they bad at today but good at in six months
Claude Code came out of that second question. Full segment 👇
"99% of the time the power grid has excess power just sitting idle." Jensen Huang on why energy, not chips, is becoming the real constraint on AI.
Jensen Huang — NVIDIA CEO — on power, efficiency and uptime at data-center scale:
→ AI factories are enormous power draws — efficient per-token, but huge in aggregate. The grid mostly runs with spare capacity, so siting compute against that idle power matters as much as the silicon.
→ The reliability bar is brutal: "everybody's got to have six nines" — 99.9999% uptime across cooling, networking and power, or the factory doesn't earn.
→ Co-design is how you keep squeezing efficiency: the product's required capabilities stay fixed while you make every layer more efficient.
Why the AI buildout is becoming an energy-infrastructure problem. Full segment 👇
2
1,768
