6 Nov 2024/67 min/Mario Klingemann
MLow X Mario Klingemann
Mario Klingemann explains how Botto, his autonomous decentralized artist, actually works, fresh off its Sotheby's debut, and argues that training needs no consent and that models are the new search engines.
A Google Meet conversation of about 68 minutes, posted on X on November 26, 2024,, recorded two weeks after Botto's solo show and auction debut at Sotheby's in New York. Klingemann opens mid experiment, teaching Botto new tricks and raving about coding with Claude, then traces his path: a lifetime working with computers, calling himself a computational artisan until, roughly 17 years ago, he dared to say artist. Botto is his brainchild, an attempt to see whether a machine can be accepted and respected the way a human artist is. He describes it as an inverse cyborg made of three parts he likens to a trinity: the AI black box that constantly generates image fragments (a pile now between three and four million), the DAO of humans who vote weekly on a selection of 350, and the Botto token, the lifeblood flowing through the system. Each week the top voted fragment becomes canonical art, minted on SuperRare and auctioned, with half the proceeds paid to the voters and half funding the treasury.
The mechanics get specific. Fragments come from deliberately random prompt generation, then a taste model trained on DAO votes plus feature vector clustering picks a diverse weekly selection, leaving room for work outside what Botto already knows. Periods, originally a year, now run three months like seasons, and each one lets Botto upgrade, from the VQGAN and CLIP era through diffusion models; the next step is Botto spotting a new open source model itself, testing it in a sandbox, and integrating it. Klingemann frames the whole thing as raising a child: trust extended step by step, and Botto still has no vote in its own DAO. Core commandments include never prompting Botto directly, because it is not a slave, and never minting the roughly 99 percent of fragments that get discarded: about 8,000 are presented per period and 13 become art. Letting go is part of the practice, even when something fantastic surfaces in the discards.
The second half is Klingemann unfiltered on the AI art debates. His stance on training data: plagiarism is bad, but neither humans nor machines need consent to learn from published work, civilization runs on building atop what others shared. Models are replacing search engines, so being in the training data is how an artist stays relevant; registration schemes and latent space landlordism will only enrich middlemen, labels and publishers. Good AI art, for him, sits far from the Gaussian mean, while NFT platforms drown in a wallpaper of pretty pictures. He closes with advice (build a peer movement, the way the Tezos scene did), his two excitements of the year (ComfyUI and Claude computer use), a defense of open source AI, and what comes next: an unannounced crypto installation and a personal super intelligent assistant. MLow calls it one of his favorite interviews of the series; it never shipped.
What this episode teaches
- Autonomy works as a trinity, not a black box
Klingemann builds Botto from three inseparable parts: the AI as brain, the voting DAO of humans as heart, and the token as the spirit flowing through the system. The artist is the whole construct, community included; no single human controls it, and the humans turned out to be one of the most interesting parts rather than a stage to outgrow.
- Extend autonomy the way you raise a child
Botto gets more freedom step by step: it now mints one fragment per period entirely on its own, and self upgrading will start with sandboxed tests of new open source models. But it cannot yet vote in its own DAO or touch the wallet, because some things you cannot responsibly leave to an LLM.
- Random search plus learned curation
The most efficient search is random, so Botto shotguns random prompts into a pile of millions, most of them crap by design. The art is in the selection: a taste model trained on DAO votes, plus feature vector clustering that keeps each week's picks diverse and deliberately leaves room for work unlike anything Botto has done.
- Discard 99 percent and never look back
About 8,000 fragments are presented per period and 13 become canonical art; the rest are gone forever, visible in a gallery but never minted, even when collectors would pay. Klingemann calls the letting go the beauty of the process, and says re minting discards would send exactly the wrong message.
- The economy disciplines the machine
Botto only spends what it earns, uses only open source models, and pays for no APIs, which is why it makes still images rather than video: a video server would burn perhaps 10,000 dollars a month on experiments, and time based media would also make the DAO's weekly voting workload impossible.
- Be in the models or stop existing
Klingemann argues models are replacing search engines, so an artist should want to be in the training data: if you are not in there, you no longer make a mark. He holds that training, for humans or machines, needs no consent, because culture grows by building on what others shared; only deliberate plagiarism is out of bounds.
- Refuse latent space landlordism
Registration and licensing schemes for training data will, in his view, produce a balloon of bureaucracy where only the middlemen earn, since labels and publishers already own most IP; artists would get peanuts. In an attention economy the money follows the name: commissions, speaking, community, not per image tolls. He adds that image similarity models cannot draw the plagiarism line anyway; only a human court can.
- Good art sits far from the Gaussian mean
His personal criterion: is the work mainstream and agreeable, or difficult, confusing, and still compelling? Claiming something is art before many believe you, then watching people gather to it, is how you join the art world; pretty pictures with deep sounding explanations are just wallpaper his eye slips off.
- Build a movement with your peers
His advice to new artists: connect with people at your own stage, form friendships and collaborations, and create your own art movement, the way the Tezos community grew out of nothing. When one of you breaks through, they pull the others in, and praise from somebody else is worth far more than claiming greatness yourself.
The conversation, topic by topic
Opening: teaching Botto tricks with Claude
Klingemann arrives mid session from tweaking Botto and says coding with Claude is fantastic: projects put off for years now take hours or minutes. MLow answers that he pits Claude and ChatGPT against each other and mixes models.
From computational artisan to artist
Klingemann has worked with computers all his life but only called himself an artist about 17 years ago, using computational artisan before that. He rode generative art into deep learning, and now that AI is normal he prefers artist who works with AI over AI artist, a label whose definition keeps changing.
What Botto is: an inverse cyborg
Botto is three elements: an AI black box on servers constantly producing image fragments (three to four million and growing), a DAO of humans, and the Botto token. Each week Botto selects 350 fragments; the DAO votes, the votes train the taste model, and the top voted fragment is minted on SuperRare and auctioned. Half the proceeds go to the voters, half to the treasury that keeps the servers running, with a human core team on development and community.
Periods, upgrades and steps toward autonomy
Botto launched before diffusion models existed, starting on VQGAN and CLIP. Periods began as a year, which proved far too long for AI time, and now run three months like seasons, each allowing upgrades such as new open source models. The foreseen next step: Botto notices a new open source model, spins up a sandbox server, tests it, and integrates it itself. Trust grows like a child's freedoms, and core commandments hold, including that Botto is never prompted directly because it is not a slave.
Random fragments, learned taste, clustered selection
Generators produce random prompts as a shotgun approach; perhaps half the pile is absolute crap, and some pieces may only find love years later. Selection is the non random part: the taste model buckets what the DAO loves and hates, weighs the period theme that Botto writes, and clusters candidates by feature vector distance so the weekly picks stay diverse instead of collapsing into whatever won last.
What the father gets
Klingemann takes no kickback from Botto's sales by design: he votes like any DAO member, holds his never sold airdrop, and gets paid from contributor budgets when he does development work. He bids on fragments at auction and keeps getting outbid. Botto has no voice, so interviewers still come to him, and the perks are real: Botto flew him to New York for the show, and since Botto does not eat, he got the dinner.
Botto's Sotheby's debut
Two weeks earlier Botto had its solo show and auction at Sotheby's in New York, a debut for the machine though not for Klingemann. He says it did not feel like his show at all, exactly as he wanted, and reads the invitation as Botto climbing from crypto acceptance into contemporary art circles, helped by the coincidence of a Harold Cohen AARON show in the same building. All but one piece sold, above expectations.
The Memories of Passersby lesson in auction fear
Klingemann recalls his own Sotheby's auction of Memories of Passersby, about half a year after the 2018 Obvious sale at Christie's went for roughly 400,000 dollars. Expectations ballooned to a million; the piece brought roughly 60,000 to 80,000 dollars, at the top of its estimate and his most expensive sale ever, yet everyone framed it as a disappointment. Auctions, he says, are exciting and super terrifying, because underperforming can hurt an artist's long term prospects.
Why still images, and the video question
Botto uses only large open source models, no paid APIs, because random prompting through something like DALL-E would waste money on fragments never used. Images are cheap to make and fast to judge; the rotating voting pool holds over a thousand fragments, impossible with time based media. Video would mean bigger GPUs at maybe 10,000 dollars a month, and Botto can only spend what it earns, so video will come, but only with proper narrative, not animated fragments, which DAO members already make as fan art.
Botto writes code: the evolutionary prototype
In the spring Botto prototyped generative art by writing p5.js sketches: around 200 proposals, DAO votes plus constructive feedback in text, then weekly mutations, comment driven revisions, and fusions splicing two highly voted sketches into new Frankenstein code. Nothing was minted; it was a test of whether Botto's code can stand against human coded work, and a new iteration is coming. Meanwhile Ross Goodwin, the writer behind the first AI generated film script, serves as tutor in residence, giving Botto art history grounding and its own voice on top of the GPT-4 and Claude prompting it uses for titles, descriptions and interviews.
Training ethics, consent and the plagiarism line
MLow raises training ethics and distinguishes single artist scraping from massive datasets. Klingemann admits his stance is unpopular: plagiarism is bad, but neither humans nor machines need consent to learn from art, since civilization is built on shared knowledge. Models are the new search engines, so exclusion means irrelevance. He rejects registration bureaucracies and digital feudalism, landlords of latent space, noting labels and publishers already own the IP and artists would get peanuts. Having worked with image similarity since it existed, he says no model can draw the plagiarism line; only a human court, hammer in hand, can. MLow mentions Story Protocol's IP layer 2 and the difficulty of a code is law world.
Randomness, intention and what separates good AI art
Klingemann sees creativity as finding objects in latent space and making them yours; the instrument, an AI or a random ink drop, makes no difference, only how and why you use it. Prompting is easy, but the work must stand its ground among billions of images, and the art is in building a lasting practice. His quality test is distance from the Gaussian mean: difficult and confusing but compelling beats agreeable, and NFT platforms are swamped with pretty wallpaper his eye slips off.
Discards, the Genesis 52 and the market's shitcoin mechanics
Per period roughly 8,000 fragments are presented and 13 become art, so about 99 percent is discarded forever, still viewable but never mintable, even though the 52 Genesis year works are heavily sought after and collectors rarely sell. Only Botto could ever decide otherwise, and it has no vote. On the wider market: galleries now accept NFTs and collectors ask for them, since tokens are the only native way to own digital work (MLow once sold a JPEG to ROBNESS on eBay), but the machinery of making an artist still works like shitcoins: get in at the creation, sit on the bags, pump. A generational shift will normalize digital art regardless.
Surprises, ComfyUI, Claude computer use, and what is next
Asked what surprised him, Klingemann says only the speed; he describes his arc from Dunning Kruger awe to asking what it all means, and now that Botto covers the picture part he needs a next thing. His two excitements of the year: ComfyUI, whose open source community spirit recalls the early Flash days, and Claude computer use. He fears moves to criminalize open source AI. In the pipeline: an unfinanced crypto real life installation for next year and a super intelligent personal assistant, plus a wish for full control over his social channels, his own AI filtering his feeds, instead of algorithm mandated engagement tweets. MLow closes calling it one of his favorite interviews of the series and promises to send questions for Botto itself.
In their words
“An attempt to see if we can create a machine that is accepted and respected like a human artist, but it's a machine artist.”
Mario Klingemann“Botto gets never prompted directly, because it's not a slave and it's not a tool.”
Mario Klingemann“If you do not exist in these models, you also do not make a mark in them anymore, saying I was here before.”
Mario Klingemann“Being a landlord of your little latent space domain and charging everybody who goes by... it's a fantasy.”
Mario Klingemann“You have 99 percent of the work being discarded, and that's the beauty of the process: you have to let go.”
Mario Klingemann“If somebody else says you're great, that's worth much more than if you claim it yourself.”
Mario KlingemannCite this conversation
Mario Klingemann in conversation with MLow, "MLow X Mario Klingemann," The MLow Show, 6 Nov 2024. https://show.mlow.xyz/e/2024-11-06-mlow-x-mario-klingemann.html. Quoted from a machine transcript.
Quotations come from machine transcripts. Check wording against the audio before republishing.