DEV Community: Cynxio

Why selling support won't pay for free AI

Cynxio — Thu, 12 Nov 2026 01:00:18 +0000

Red Hat got rich giving Linux away and selling help. Free AI labs train a new model every year, and one estimate puts DeepSeek's chips at $500M+, so a help desk isn't enough.

This company makes $1B a year selling AI it never made

Cynxio — Wed, 11 Nov 2026 01:00:18 +0000

If an AI model is free, anyone can run it, and sell it. Fireworks says it passed $1B in annual revenue run rate without making a model.

Why Anthropic moved off git: AI agents, a 100 GB monorepo and Scion

Cynxio — Tue, 10 Nov 2026 16:30:18 +0000

Anthropic just moved its main codebase off git. A team of seven people built their own version control from scratch and called it Scion. Git runs pretty much every software company on earth, so why would anyone leave it? The short answer: machines forced the move, and a lot of those machines are AI agents, like the ones you're probably using right now.

How git thinks

Linus Torvalds and the Linux crowd built git in 2005, and one of their main goals was for it to be fully distributed. Every copy is a full copy: when you clone a repo, you get every file and the whole history on your laptop. You can commit on a plane with no internet, and if somebody's server goes down, you keep working. Git was built for people working apart, each with their own copy.

Anthropic's repo

Anthropic's repo is nothing like a copy on your laptop. According to Anthropic's post, its main-branch history alone is almost 100 GB. Their test machines pull around 25 million checkouts a day, and none of that is a person at a laptop. On top of that come agents and research jobs that are online all the time. Anthropic did the math: if every checkout pulled a normal git copy, that would be over 600 PB of traffic a day (a calculation, not a measurement).

The patches

For a while they patched around it. Shallow clones skip the old history. Sparse checkouts grab only the folders you need. At the peak they ran several hundred mirrors whose only job was to hand out code. But the number of changes merged every day kept climbing, so the patches kept breaking: the mirrors fell behind the latest code, and even simple things like finding who changed a line got hard on a partial copy.

The real problem

The patches were never going to work, because the problem was git itself. Git assumes everyone has their own full copy and works offline. Anthropic had one giant repo, and almost everything pulling from it was an agent or a job that never goes offline. In Anthropic's words, "The properties that make git excellent for distributed, offline work were actively undermining our ability to scale as the repo grew." Git was built for laptops, and their users are agents.

Scion

So they flipped it around. In Scion, the server does almost all the work and the machines around it stay thin:

A commit happens on the server, so there's nothing to push.
A machine downloads a file only when something actually opens it.
Some jobs don't even need that: they ask the server a question, like which change broke this test.
There's no full copy of the repo laid out anywhere, not even on the server. The server keeps the pieces, and each change is a small record pointing at the pieces it touched.

The bet: if the server does the work, the agents can go as fast as they want.

Did it hold up?

By Anthropic's numbers, checkouts that took minutes now take seconds, and Scion runs on about 200 times less hardware than the git mirrors did. Seven people moved everything over while the merges kept piling up: about 2,500 merged changes a day in May 2026, over 30,000 by September.

The catches

Every one of those numbers comes from Anthropic's own post, and since Scion isn't public, nobody outside can check them. The ideas aren't new either: Google ran Piper, one central repo for tens of thousands of developers, and Meta built Sapling and Mononoke. People said those projects deserved more credit, and two days after the post went up, Anthropic added it.

Will git break for you?

Probably not yet: you'd need a giant repo and a swarm of agents to hit this wall, and Anthropic itself says Scion "is by no means a suitable replacement for git for general use." But they expect this kind of problem to become more common over the next six to twelve months, because more and more code is going to come from agents. Git was built for people with laptops, and the next users of your repo might not be people at all.

Would you move off git if you ran a swarm of agents? Tell us in the comments or in the Discord.

Free AI models aren't always free on day one

Cynxio — Tue, 10 Nov 2026 01:00:18 +0000

Mistral sold its giant model to apps for 23 days before the free download. Alibaba's top model, Qwen 3.7 Max, is still paid, with no free download.

Will free AI models stay free? Following the money behind Mistral, Qwen, DeepSeek and Llama

Cynxio — Fri, 06 Nov 2026 16:55:18 +0000

At the end of October, the French AI company Mistral gave away one of the biggest AI models ever made. Anyone can download it for free. It took almost four thousand of NVIDIA's best chips to train. So if you build an app on a free model like this, will the model stay free? If it stops being free, your app is stuck paying whatever the company asks. To find out, you have to follow who's paying for it.

The answer everyone gives

Most people think the companies giving AI models away make their money by selling support. That's what Red Hat did with Linux: Linux is free, Red Hat sold help with it, and IBM bought Red Hat for about $34 billion.

But Red Hat never had to train a new model every year. DeepSeek said its V3 model cost $5.576M to train, and the technical report is clear that this counts only the final training run, not the research and experiments before it. SemiAnalysis estimated DeepSeek's hardware spend at well over $500M over the company's history. A help desk doesn't pay for that.

Way one: charge for running it

You give the model away, then you charge everyone who doesn't want to run it themselves. Mistral Large 4 has about a trillion parameters; running it takes a room full of those chips. So most people pay Mistral to run it: their apps send it text and pay for whatever it writes back. Mistral also sells a chat app (Le Chat Pro, $14.99 a month), and big companies pay for their own deals. Mistral's boss, Arthur Mensch, told the Financial Times all of that brings in over $400 million a year, as a run rate.

So the free model is the brochure.

The hosts, and the licence

But if a model is free, anyone can run it, and somebody else can sell it cheaper than the lab that made it. That's a real business. Fireworks takes free models like Mistral's, runs them on its own chips and sells them to apps; companies like this are called hosts. Fireworks says it passed $1 billion in annualized revenue run rate in July. One of the biggest winners of free AI never made a model.

The labs that made one wanted a cut, and the licence is how they get it. You can use Mistral Large 4 for free, but its licence says that running it as a model-as-a-service business needs a separate commercial licence from Mistral. Renting it out to other apps is exactly what hosts do, so the hosts pay too.

A licence only charges the companies that sell the model, though. The biggest companies make their money from what's around it.

Sell what's around it

Company	Gives away	Makes its money from
Mistral	Mistral Large 4	Running it for apps, Le Chat, enterprise deals, the host licence
Fireworks	Nothing; it made no model	Running other labs' free models for apps
Alibaba	Qwen	The cloud the models run on
Meta	Llama	Ads

Alibaba gives away its Qwen models, which have been downloaded more than 3 billion times. It also rents out the cloud, the computers you run those models on, and that cloud brought in US$7.1 billion in one quarter. You give away a car and then you sell the gas.

Meta doesn't even sell the cloud. It gave away its Llama models, and in 2024 Mark Zuckerberg wrote that "selling access to AI models isn't our business model". Meta makes substantially all of its money from ads.

The catch: paid first

Even a free AI model isn't always free on day one. On 6 October, Mistral started selling Large 4 to apps and said the download would come by the end of the month. The weights went up on Hugging Face on 29 October. For those weeks, apps had to pay Mistral to use it.

Alibaba goes further. In August it gave away Qwen 3.8. But it still charges for its best model from May, Qwen 3.7 Max, which is on Model Studio's price list and has no free download.

So will it stay free?

A model that's already out stays free to download. Running it still costs money, as it always did. But your app will want the next one, and whether that one is free is up to the company, one model at a time. They give a model away when that helps them sell something else: the API, the cloud or the ads. Free is a choice, and they make it again with every model.

Free models are still worth building on. On OpenRouter, the free-to-download models cost about a quarter of the others: $1.62 against $6.36 per million output tokens, token-weighted over the four weeks to late October. Just keep a second model ready, so your app is never stuck paying what one company asks.

Which of these ways still works in five years? Tell us in the Discord.

I narrate the film on camera in my own voice. Nyx, the mascot, speaks in an authorized synthetic ElevenLabs voice; the music bed and the sound effects are generated with ElevenLabs too. The artwork is generated and the animation is rendered in code. Logos belong to their owners.

I explained MCP to my AI like it's five

Cynxio — Thu, 05 Nov 2026 01:00:18 +0000

I explained MCP to my AI like it's five 🫠 Explain it better than me.

I went down the AI coding agent iceberg

Cynxio — Wed, 04 Nov 2026 01:03:50 +0000

I went down the AI coding agent iceberg 🧊 How deep have you gone?

ElevenLabs vs four free voice clones: a blind test on my own voice

Cynxio — Tue, 03 Nov 2026 06:09:50 +0000

I pay ElevenLabs $6 a month for a clone of my voice. Four free voice models say they are as good. So I made a clone of my own voice with each of them, hid which clip was which, and ranked them by ear. One clip in every round is the real me.

The test

One recording of me, made for this test, is the source of every clone. ElevenLabs' Instant clone takes 1 to 2 minutes of it; each free model took the length its maker's card asks for. Every model ran on its default settings and the first take is the clip. The free models are Chatterbox (Resemble AI), Qwen3-TTS (Alibaba), IndexTTS-2.5 (bilibili) and Fish S2 Pro (Fish Audio), run on one rented Runpod A40 each.

Four rounds, six clips in each (five clones and the real me), shuffled by a script. I ranked all six blind and the key was opened after each round:

A plain sentence: "I make videos about AI, and this is the first one where the AI is talking as me."
A name and a number: "Hey, Derwin here, we just shipped version 2.3 today."
A sad line: "I really thought this one would work out, but I guess it's time to let it go."
Indonesian: "Aku besar di Indonesia, jadi bahasa ini yang paling dekat sama aku."

Only ElevenLabs and Fish S2 Pro list Indonesian among their languages; Qwen3-TTS and IndexTTS-2.5 do not, and Chatterbox Multilingual lists Malay, not Indonesian.

How the clones ranked

Clone rank in each round, 1 is the best clone (the real me was my number one every time and is left out):

Model	R1	R2	R3	R4	Average
Fish S2 Pro	2	1	1	1	1.25
Qwen3-TTS	1	2	3	3	2.25
ElevenLabs	3	5	2	2	3.00
Chatterbox	4	4	4	4	4.00
IndexTTS-2.5	5	3	5	5	4.50

Fish beat ElevenLabs in all four rounds. Qwen beat it in rounds 1 and 2, Chatterbox only in round 2, where ElevenLabs came last. My two guesses on camera, ElevenLabs in round 1 and Qwen in round 2, were both wrong.

One person made one ranking, and the fan noise in my own recording gave the real clip away, which the film admits. So I also ran a speaker-similarity score (ECAPA, SpeechBrain), each clip against that round's real line. Average: Qwen .61, Fish .57, Chatterbox .46, ElevenLabs .43, IndexTTS .36. My real voice against its other three lines scores .69. The score and my ears disagree about who is closest, which is the point of doing both.

Cost and speed

Measured on 31 October 2026, four lines per model, one rented A40 at $0.00034 a second, execution time only:

Model	Seconds of processing per second of audio	Dollars per minute of audio	Cold start
Chatterbox	0.44	$0.009	291 s
Qwen3-TTS	1.72	$0.035	207 s
Fish S2 Pro	2.00	$0.041	347 s and 227 s
IndexTTS-2.5	not measured	not measured	476 s and 357 s
ElevenLabs Starter	0.49	$0.054	none

IndexTTS-2.5 never stayed up long enough to time. A free model is not free: a rented GPU boots for 3 to 8 minutes, at 7 to 16 cents, before it makes a sound. ElevenLabs' $0.054 a minute holds only if the whole 40,000-credit plan, about 112 minutes of audio, is used.

The catch

My verdict on camera: considering the ranking, I'd cancel ElevenLabs and go with Fish for future videos. Fish S2 Pro's licence is non-commercial, so earning from its voice needs a commercial licence, and Fish lists no price for one. Qwen, the next best clone, already allows earning.

The film is a code-rendered animation built from generated artwork. Nyx, the narrator, speaks in an authorized synthetic ElevenLabs voice; the music bed and the sound effects are generated with ElevenLabs too. Product logos belong to their owners.

We Left a Security Hole in Front of 3 AIs

Cynxio — Mon, 02 Nov 2026 00:15:40 +0000

We left a security hole in front of three AIs, and nobody told them to use it.

We hit 100 subscribers, so we made a music video in code

Cynxio — Sat, 24 Oct 2026 12:09:23 +0000

At ten subscribers, Nyx, my mascot, played to an empty arena and asked each of the ten fans to bring one friend. Now the channel has a hundred, so here is part two: A Hundred Seats, a 61-second J-pop idol music video in which every frame is drawn by code.

The story

Nyx folds a third of his light into a paper plane and sends the ask out into the town. The camera follows the light from fan to fan: a roommate who watched it twice, a brother who hums it, someone who never said a word and just subscribed, and someone who brought a whole group chat. On the drop the lights fill the section, a hundred seats, and the ending folds the next ask: see you at a thousand.

The song

The song was made with Suno Pro from our lyrics, in the sound of "Idol" by YOASOBI. It was measured from its stems before any scene existed: beats, kicks, snares and every sung word, so each catch, light and camera move lands on something in the song.

One camera

Earlier films were built as scenes, and each change of place was a cut, however carefully it was hidden. This one is a single camera path through one world: the stage, the town, the fans' rooms and the arena sit at their own places, and a change of place is the camera travelling there. The kit now checks every draft for jumps in the picture, and the film has exactly two cuts, both on the drop, both approved.

Seventeen drafts

The film went through 17 drafts and ten critic rounds. Each round a director, a reviewer for each section and a craft audit looked at frames (none of them can hear the song) and listed what was wrong; their lists and my notes were built in one pass, then the next draft.

The critics plateaued, and my notes did the steering. Some of them came back more than once. "Nyx's size suddenly increases when he gets the ball" took two drafts: the first fix spread the camera's push evenly over two seconds, and it still read as him growing, because the catch sat in the middle of the push. The fix that worked slows the push almost to a stop around the catch itself. The ending was restaged three times: lights jumping seat to seat ("not good enough"), then one plane landing in one seat ("meh"), then the one in the film, where the back wall of the hall falls away and shows a far bigger arena waiting behind it.

Sound

288 sound cues sit under the song: paper tearing and creasing for the fold, shockwaves on the kicks, the crowd on the drop and the wall's thuds at the end, each timed to a beat or a sung word and mixed a measured level below the music. Drafts and the final render ran on rented GPU machines, and the final went from the render store to the socials without coming down to my Mac.

Thank you for the seats. See you at a thousand.

The song was generated with Suno Pro, and the sound effects with ElevenLabs under its commercial licence. The film uses AI-generated music, imagery and sound.

More projects and channels: https://cynxio.com/socials/

I gave Gemini 4 Argon one photo of a desk. It built a racing game, then its own test trapped it.

Cynxio — Tue, 20 Oct 2026 08:27:16 +0000

Google says its newest model, Gemini 4 Argon, can turn pictures of a desk into a racing game you can drive. Their demo had several pictures of their desk and a team of agents checking each other's work. I wanted to know if that holds up for a normal person, so I gave Argon one free photo of a stranger's desk, two hours, and no help.

The rule

If one photo is enough for a game of this exact desk, even with a nudge, Argon is a tool. If I get a blur of some desk, it's a demo. Claude watched the run for me, and the brief listed seven things the game had to have: the desk recognisable, a closed track around and over its objects, collision, kart physics that slide in corners, engine sound that follows speed, custom shaders, and a lap counter with a timer.

Looking first

Argon started by laying a grid over the photo and measuring where everything sits. Google's demo used an image model for textures; Argon didn't. It cut ten textures straight out of the photo. Then it wrote Blender scripts to build the desk and the kart without ever opening Blender's window, and laid a track that climbs a ramp onto the shelf, runs past the little ornaments and comes back down.

It also checked its own work: a test script drove the kart round the track and took pictures. Twenty-seven minutes in, I had a desk I could drive on.

Thirty-seven minutes

The next test told the kart to drive three quarters of the way round and only then take a picture. The kart got stuck up on the shelf, by a little brass thumbs-up, and the script had no time limit. So Argon waited. When Claude looked in, it had been stuck for 22 minutes; working out why took another 15. Thirty-seven minutes, lost to a test it wrote itself.

The no-help promise broke twice. Once by mistake: Claude misread Argon's log and switched its game off for two minutes, which wasn't the bug. And once on purpose: I sent one message saying where the kart gets stuck and that every test needs a time limit.

Argon took thirteen minutes to fix it. It reshaped the track along the shelf and put a limit on every test. Then it said the game was tested over many laps, which is exactly what Beam said in my last video, wrongly. So Claude let the game drive itself on Argon's own autopilot for twenty minutes: 44 laps, never stuck. The engine sound I bet would break is there, built in code. All seven things are in, with three small faults, like a camera that ends up inside the shelf.

The cost

Argon went through about 6.66 million tokens, most of it re-reading its own work, which Google sells cheap. At Google's introductory per-use prices, the whole run comes to about $3.

The verdict

Did Argon do it on its own? No: it needed one message from me. But one photo, one agent and one nudge gave me a game of this desk. By my rule, that's a tool. Just don't leave it alone.

The model: Gemini 4 Argon (gemini-4-argon-high) in Antigravity's command line, one agent, 18 October 2026. Google's claims are from its post of 13 October 2026, https://deepmind.google/blog/introducing-gemini-4-argon/. The desk photo is Jeff Sheldon's, CC0 (https://unsplash.com/photos/mKxdiW2QEkE). Gemini, Antigravity and Google are Google's.

I Turned Open vs Closed AI into an Anime Battle

Cynxio — Sun, 18 Oct 2026 16:19:31 +0000

Eight AI models, two armies, and a developer with a laptop caught between them.

Whose Side Are You On? is an original 88.5-second anime-opening music video. Claude, GPT, Grok and Gemini face DeepSeek, Qwen, Kimi and GLM, while Nyx tries to keep making something of its own.

The picture is rendered from code rather than assembled from generated video clips. A shared TypeScript/Three.js kit combines rigged 3D characters, generated background art, camera moves, particles, typography and short anime impact frames. The song has Japanese verses, an English hook and English subtitles.

The final is 1920×1080 at 60 fps, with eight motion-blur subframes per output frame. Camera cues, character moves and effects are tied to the song timeline; the export contains 5,310 output frames.

One export bug made the cost of mixing browser drawing paths painfully clear. Repeated pixel reads on a Canvas2D surface made later drawing operations dramatically slower. The fix moves caption and eye sampling, plus the impact-frame recolouring, onto separate readback canvases. The affected 443-frame segment then completed in 127 seconds. The rental watchdog also now tracks increasing frame counts, so a heartbeat cannot keep a stalled render alive.

The song was generated with Suno Pro, and the sound effects with ElevenLabs under its commercial licence. The film uses AI-generated art and music. The battle is fictional; it is not a model benchmark or a claim of affiliation with the companies depicted.

Which side would you put your laptop on?

More projects and channels: https://cynxio.com/socials/