Ethan

293 posts

Ethan banner
Ethan

Ethan

@_ethan

staff @DecartAI

Sumali Temmuz 2014
2.4K Sinusundan498 Mga Tagasunod
Ethan nag-retweet
Decart
Decart@DecartAI·
What if video editing had a “volume knob”? We are excited to share new research from @DecartAI and @TelAvivUni, accepted to #SIGGRAPH2026. We found directions within video models that enable smooth control over edit strength, from subtle changes to full transformations! Check out our paper & blogpost 👇
English
2
11
64
9.4K
Ethan
Ethan@_ethan·
We're hiring in SF! Autoregressive diffusion models have their own unique challenges and our opt team paves the way for new experiences like Lucy.decart.ai If this type of impact resonates with you, reach out to @hiba280700 with ref to your perf work!
Heba Abbud@hiba280700

I joined Decart to shape the AI industry and I’m hiring an engineer to join me. At most inference companies I could have gone to, I would just be improving Deepseek/Qwen and existing models. At Decart I shipped the first realtime diffusion model in the world to run on Trainium.

English
1
0
3
290
Jared Roesch
Jared Roesch@roeschinc·
CUDA 13.2 is now live, docs.nvidia.com/cuda/ we dropped a new version of cuTile, and Tile IR with number of new features in the DSL, and a large number of performance and stability improvements in Tile IR. Happy Hacking and looking forward to chatting at GTC this coming week!
English
4
4
57
3.4K
Ethan
Ethan@_ethan·
@isidentical I've gotten more compliments on the GPU poor hat than any other swag!
English
0
0
3
181
batuhan the fal guy
batuhan the fal guy@isidentical·
which startups have the coolest swag? especially more unique stuff than just hats/t-shirts
English
7
0
11
3.1K
Ethan nag-retweet
Decart
Decart@DecartAI·
I am some dude at Decart. We don't have titles, and if we did they'd change more often than our GPU clusters. That happens twice a week. Our domain is the name of a french philosopher, if you spelled it wrong. .ai. Our entire budget is spent on the best scientists you can get. And DoorDash for dinner. We don't have a marketing team. Well we do, but they are not allowed to use computers. You can reach them by standing under the balcony and yelling "Ido!!". We don't have a balcony anymore. We are also on the 32nd floor. We probably have a product somewhere, but who knows. It's realtime for sure tho. We are not hiring. If we did, we wouldn't tell you. Members of technical staff preferred.
Peter Girnus 🦅@gothburz

I am the lead engineer at ai.com. We had $78 million to work with. $70 million went to the domain. $8 million went to the Super Bowl ad.  I got the rest. "The rest" was $500 and a Cloudflare free tier. This ratio -- 156,000 to 1, marketing to engineering -- is not a bug. It is the business model of the entire artificial intelligence industry in 2026. You do not need a product. You need a name. Preferably two letters. Preferably letters that made investors lose bladder control in 2024. I built the website in a weekend. I didn't build it, actually. I described it to OpenClaw (previosely Moltbook), (previously, reviously Clawdbot) and the AI built it. We are, after all, an AI company. Using AI to build the website felt appropriate. The AI charged us nothing. We are charging users $20 a month. This is called "margin." We have a free tier and a paid tier. The free tier gives you access to a product that doesn't exist. The paid tier gives you access to the same product that doesn't exist, but with more input tokens. No one has asked "input tokens for what." This is the kind of question that delays launches. Nobody checked if it worked. Nobody checked if it scaled. Nobody checked if it did anything at all. We were too busy approving the logo. The logo is a planet with a ring around it. Someone said it looked like the old Saturn car logo. Saturn went bankrupt in 2010. But the logo was free and our design budget went to the domain, so here we are, orbiting a dead brand at $70 million per revolution. Our product is an "autonomous AI agent" that "organizes work, sends messages, and executes actions across apps." Which actions. Which apps. At what cost. In the AI industry, these are called "implementation details." Implementation details are beneath us. We are a vision company. The vision cost $70 million. The implementation cost $500. The gap between the two is where shareholder value lives. Our press release promises the agent will "trade stocks, automate workflows, and update your online dating profile." We are building artificial general intelligence so it can fix your Hinge bio. This is on the roadmap. The roadmap is longer than the codebase. Our marketing says you can create an AI agent in 60 seconds. This is technically true. You type a username. You click "generate." You receive a loading spinner. Sixty seconds. What you do not receive is an AI agent. But the experience of waiting for one is, I'm told, "the product." Our press release describes a "decentralized network of billions of agents." We used the word "decentralized" because our CEO comes from crypto. In crypto, "decentralized" means "we haven't decided how it works yet." We have not changed the definition. This is not unique to us. OpenAI has raised $40 billion. Their product loses money on every user. Anthropic has raised $15 billion. Their stated goal is to build something they believe might destroy humanity, and investors are fighting to give them more. Microsoft has committed $80 billion to AI infrastructure this year. Their Copilot product tells people to put glue on pizza. The entire industry is a $300 billion screensaver with a loading spinner. We fit right in. Our CEO is the Crypto.com guy. He previously spent $700 million to rename a basketball arena and hired Matt Damon to tell America "fortune favors the brave" six months before crypto lost 70% of its value. He paid for our domain in cryptocurrency. I am told this was "tax efficient." I have learned not to ask follow-up questions about things that are "tax efficient." He is now pivoting from crypto to AI. In the industry, we don't call this "pivoting." We call it "convergence." Convergence means the last bubble popped so you inflate the next one using the same PowerPoint deck with different nouns. The Super Bowl ad ran during the fourth quarter. Thirty seconds. It told 130 million Americans to visit our website. The ad was thirty seconds. That's $266,666 per second. Each second of airtime cost more than our entire engineering budget. Second fourteen showed the logo. Second fourteen cost more than the website. They did visit. All of them, apparently, at once. The website went down. "Prepared for scale, but not for THIS," our CEO tweeted, adding three fire emojis. The fire emojis were load-bearing. They were doing more work than our infrastructure. The entire site was hosted on Cloudflare's basic tier, which is designed for food blogs and wedding photographers, not for absorbing the combined curiosity of a nation told to visit a two-letter domain during the biggest television event on earth. But the crash was, in a way, perfect. It is the most honest thing the AI industry has produced. A $78 million promise that, when 130 million people showed up to collect, returned a loading spinner and the words "please refresh and try again." Every AI company should adopt this as their mission statement. The previous owner of ai.com was OpenAI. They used it to redirect to ChatGPT -- a product that exists, built by thousands of engineers who were paid more than $500, running on billions of dollars of compute. We bought the domain from them to redirect to a page that asks you to pick a username. OpenAI also ran a Super Bowl ad this year. They sold us the domain, then bought ad time in the same broadcast to promote the product they used to host on it. We are now competing with the company that built the thing we may or may not be reselling. During the same commercial break. On the same channel. For the same audience. The AI industry is a snake eating its own tail, except the tail cost $70 million and the snake can't stay online. That's the product. A username. For an AI agent that doesn't exist yet. On a website that couldn't survive its own launch. Sold by a crypto CEO during a crypto winter. Wearing the logo of a bankrupt car company. Twenty-three percent of Super Bowl ads this year were AI companies. That's 15 out of 66. In 2000, it was dot-coms. Pets.com ran a Super Bowl ad. They went bankrupt nine months later. Their sock puppet mascot outlived the company. I'm not saying history repeats. I'm saying it rhymes, and the rhyme scheme is expensive. But none of that matters. What matters is the domain. Two letters. Seventy million dollars. The most expensive thing we own is our name. The least expensive thing we own is everything the name is supposed to represent. In the AI industry, this is called "brand-first development." In every other industry, it's called something else. Anyway, we're hiring. Backend engineers preferred. Budget: whatever's left.

English
3
3
43
9.6K
Ethan nag-retweet
Decart
Decart@DecartAI·
Drag, Drop and Try it live That's just a small glimpse of what’s possible when AI runs in Realtime, powered by Lucy 2.0. What else can you build?
English
76
219
2.2K
310K
Ethan
Ethan@_ethan·
@DecartAI Descartes on Decart... Decartception?
English
1
0
1
74
Decart
Decart@DecartAI·
Lucy 2 supports both Nvidia GPU and AWS Trainium, leveraging our proprietary IP to port cutting edge AI models
English
4
2
22
4.1K
Decart
Decart@DecartAI·
Introducing Lucy 2.0: a World Editing Model running at 1080p, 30FPS in realtime. Read the technical report, API access, and try the live demo 🧵
English
50
83
459
146.7K
Ethan
Ethan@_ethan·
Excited for people to see what we've been cooking, the first real-time v2v model with reference image support running at these specs! We're still just scratching the surface of these models and some prompts have surprised us (e.g. add a cat/spider and some light vfx). Try it out and let us know what you think! 🚀
Decart@DecartAI

Introducing Lucy 2.0: a World Editing Model running at 1080p, 30FPS in realtime. Read the technical report, API access, and try the live demo 🧵

English
0
1
2
160
Ethan
Ethan@_ethan·
Now I'm only one character away from @ethan @XHandles if I start publishing articles can we talk pls
English
0
0
1
101
Ethan
Ethan@_ethan·
@Dr_Gingerballs You think that OAI is paying an equivalent of $2-7/H100GPU-hr?
English
0
0
2
1.2K
Dr_Gingerballs
Dr_Gingerballs@Dr_Gingerballs·
Hi everyone who says AI compute is getting cheaper. OpenAI has made about $1.20 per kWh of compute each year for the last 3 years. Zero cost reduction in 3 years. Conveniently, takes about 1 kW to run an H100, so they make $1.20 per GPU hour. Reserved instances run anywhere from $2-7/hr per gpu (also being run at a loss by just about everyone). So in the BEST BEST case, OpenAI is slated to lose $20B this year, and could be multiples higher. It's unlikely they have more than $30B after the recent $22B infusion from Softbank. They have a runway of less than a year and they are running out of dumb money to pump them further without an IPO. The lawsuit by Musk complicates that timeline as well, if it's even possible. So they are going to ads because they will die this year without them. Thank you for your attention to this matter.
Chris@chatgpt21

🚨 OpenAI CFO Sarah Friar just posted a huge update on revenue: $20B+ ARR. The article went onto say: “Compute grew ~3x year over year (0.2 GW in 2023, 0.6 GW in 2024, ~1.9 GW in 2025). Revenue followed the same curve: $2B ARR in 2023, $6B in 2024, and $20B+ in 2025.” If ARR is scaling like this at that size, it’s fair to say: “the ai bubble was a hoax” ?

English
59
127
1.7K
220.7K
Simo Ryu
Simo Ryu@cloneofsimo·
Tao's comment on this proof: apparently its 'slightly different from the standard methods' and 'Previous generations would almost certainly have fumbled these delicate issues' GPT 5.2 is actual beast, I wonder what GPT 6 would look like... 😆
Simo Ryu tweet media
Simo Ryu@cloneofsimo

It is November 2022, you yap "LLM cant reason because they are autoregressive!! We need cat level intelligence, neuro-symbolic AI! Chatgpt is fun toy product!" It is year 2026, they solve conjectures in 40 min. You literally share proof of open problems with chatgpt links. Actually insane timeline.

English
14
38
578
267.2K
Ethan nag-retweet
Decart
Decart@DecartAI·
Meet Lucy Motion - an image-to-video model with precise control over motion. Instead of struggling to describe movement in text and hoping the model gets it right, you draw a path from a subject to its destination. Lucy animates it exactly along that path.
English
19
70
489
36.6K
Ethan
Ethan@_ethan·
@latkins Congrats on the release! 🚀🚀
English
0
0
1
504
Lucas Atkins
Lucas Atkins@latkins·
Today, we are introducing Trinity, the start of an open-weight MoE family that businesses and developers can own. Trinity-Mini (26B-A3B) Trinity-Nano-Preview (6B-A1B) Available Today on Huggingface.
English
81
150
1.3K
621.8K
Ethan nag-retweet
Decart
Decart@DecartAI·
Make quick, local edits to your video with Lucy-edit Fast. Enjoy near Pro-level quality at 2× the speed and a fraction of the cost, bringing batch edits closer to real-time performance.
English
1
13
31
3.6K
Ethan nag-retweet
Decart
Decart@DecartAI·
You can plug real-time video models directly into gameplay. Built with @v0 on @vercel, vibe-coded with @geminiapp Pro 3. This Flight Simulator pipes Three.js into Decart LSD v2, turning it into an AI video renderer with live prompts, VFX, lighting, and shadows.
English
7
21
183
21.7K
Ethan nag-retweet
Decart
Decart@DecartAI·
Restyle your live video while it is playing. Introducing LSD v2 - our new realtime flagship model. It’s illegally good. Try it now - free for the next 72 hours!
English
31
48
414
1M