@Ycombinator P26 | Building @Laminalabs, the fastest infra for video generation | CS & Math @MIT BS' 24, MEng' 25
San Francisco
Joined June 2017
- Tweets100
- Following77
- Followers196
- Likes145
Pinned Tweet
We built the fastest explainer video generator in the world.
Introducing Simi: turn any prompt, doc, or idea into a complete whiteboard explainer video in seconds.
It supports 80+ languages, so the same idea can be taught to audiences anywhere.
AI video has gotten insanely good at cinematic clips. But almost nothing can actually teach. Most tools are built for shots, scenes, avatars, and ads. Simi is built for explanation.
Give it something like: “Explain backpropagation” “Teach photosynthesis” “Walk through our onboarding flow” “Turn this product doc into a video”
Simi writes the script, and create the final video.
Our fastest tier can generate videos in under 20 seconds. And you can one-shot long explainers over 15 minutes.
We’re starting with whiteboard explainers because they are the clearest way to teach complex ideas step by step. But the bigger goal is simple: Any knowledge should be instantly turnable into video.
Please help if you can. Nepal needs us right now. 🇳🇵
Nepal needs urgent help. 🇳🇵
It's been 5 days. Members of my own family are still missing. Thousands of families have been displaced, lost loved ones, and are still waiting for news.
@laminalabs is donating $5,000 to flood relief.
If you can help, please donate through Nepal’s official relief fund: pmdrf.nchl.com.np
@mk314k
Kartikesh Mishra retweeted
This is insaneeeee!!
Simi turned a 2200 word article into a five-minute explainer video in one shot. It is the fastest way to generate explainer videos!!!
Our product, Simi, the fastest way to make explainer videos, one-shotted this entire article into a five-minute explainer, so you can watch it and understand the core argument without reading the whole piece.
My main takeaway: enterprise AI won’t be about using the smartest model for every task. It’ll be about using just enough intelligence to get the right outcome, then making that outcome cheaper and more repeatable over time. Couldn't agree more Jaya!!
NVIDIA’s Alpamayo 2 Super is a major step toward autonomous systems that can not only make decisions, but explain the reasoning behind them.
Sudip passed the announcement to Simi, and it generated this explainer in one shot.
A good example of one reasoning system making another easier to understand.
We turned NVIDIA’s Alpamayo 2 Super announcement into a whiteboard explainer video.
The original post is packed with details on open licensing, AV reasoning, and cloud-to-car deployment, so we gave it to Simi.
Simi generated this entire explainer in one shot. Simi, by lamina labs, is the fastest way to create explainer videos. Try it at laminalabs.ai
OpenAI's GPT-Live breakdown covers full-duplex audio, async media paths, and state handoffs.
Pretty cool to see Simi map out these complex system topologies into a structured whiteboard visual on a single pass.
Try Simi, it is the fastest way to generate explainer videos on the fly.
Most AI video systems are built to generate clips.
We are building infrastructure that generates explanations.
That means video that is fast, structured, controllable, and reliable enough to be generated inside any model, agent, or application.
Simi is our first step.
Very proud of what @sudiprokaya and I have built, and even more excited about what comes next.
We raised $4.8M to build @laminalabs and make AI explain with video, not just text.
Introducing Simi, the fastest way to make an explainer video.
Give it a single prompt or document, and it can generate a one-minute narrated explainer in as little as 20 seconds, in 80+ languages.
It works like ChatGPT, except the answer is a video.
The round was led by @FoundationCap , with participation from @Link_Ventures and @ycombinator .
This is our story. Thread below on what we are building at.
The impressive part is not the single prompt. It is that the output still has provenance, typed data contracts, security boundaries, testing, and deployment discipline.
AI becomes truly useful when speed does not come at the cost of engineering rigor.
Simi made this explainer in one shot so you can understand Satya Nadella’s entire post in 90 seconds.
The real enterprise AI breakthrough is not generating more disposable code. It is turning trusted information into secure, governed, reusable software at the speed of a prompt.
Most people are still talking about vibe coding. This is what the future of actually building businesses looks like.
The model did not change. The system around it did.
Giving an agent access to its previous reasoning and preserving what it learned was enough to triple performance while using far fewer tokens.
A great example of why context and memory architecture matter so much.
Also very cool to see Simi turn the entire post into this explainer in one shot.
Simi turned OpenAI’s full post into this explainer video in one shot, so the idea is easier to digest.
The biggest takeaway: benchmarks do not just measure the model. They also measure the memory and harness built around it.
Same model. Better memory. 3x the score and 6x fewer output tokens. Insanee!!!
K3 is not uniformly better than GPT-5.6 or Anthropic’s strongest models.
But it is now competitive on several of the tasks that matter most for agents: repository-scale coding, browsing, automation, tool use, and sustained execution.
That is a much bigger result than winning one benchmark.
Kimi K3 does not beat GPT-5.6 or Anthropic’s strongest models at everything.
But an open-weight model now competes with them, and sometimes wins, on serious coding, browsing, research, and automation tasks.
Part 6/6 explains how K3 reached the frontier, and where it still falls short.
P.S. Every video was generated in one shot by Simi, the fastest infra to generate explainer video in the entire world. Brought to you by Lamina Labs.
Long-horizon agents are not created by increasing context length alone.
They need training across repeated cycles of planning, acting, observing, testing, and correcting, plus infrastructure that preserves the environment across extremely long trajectories.
The model and the harness are one system.
K3 was not trained only to produce better answers.
It was trained to keep working after the first answer fails: use tools, inspect the result, correct itself, and continue across hundreds or thousands of actions.
Part 5/6: How K3 became an autonomous agent.
P.S. Every video was generated in one shot by Simi, the fastest infra to generate explainer video in the entire world. Brought to you by Lamina Labs.
Residual connections usually pass one increasingly mixed representation through the model.
Attention Residuals let later layers selectively retrieve representations from earlier blocks.
KDA improves information flow across tokens. AttnRes improves it across depth.
Long context solves distance across tokens.
But K3 has another problem: useful information can become buried while passing through 93 layers.
Attention Residuals let later layers reach back and recover earlier representations.
Part 4/6 explains how.
P.S. Every video was generated in one shot by Simi, the fastest infra to generate explainer video in the entire world. Brought to you by Lamina Labs.
Full attention gives exact global access, but becomes extremely expensive at one million tokens.
KDA maintains a compact recurrent memory. Periodic MLA layers restore global access when exact retrieval matters.
The hybrid is the important part.
A one-million-token context does not mean perfect memory.
Ordinary attention at that length would be enormously expensive. K3 instead maintains a compact, editable memory, then periodically checks the full context.
Part 3/6: Kimi Delta Attention, explained simply.
P.S. Every video was generated in one shot by Simi, the fastest infra to generate explainer video in the entire world. Brought to you by Lamina Labs.
The 2.8 trillion parameter count gets attention.
The difficult part is routing each token to just 16 of 896 experts without load imbalance and communication costs erasing the benefit.
Sparse models are simple in theory and extremely hard to make efficient at this scale.
Kimi K3 has 2.8 trillion parameters, but activates only around 104 billion for each token.
The model is enormous, but each token only wakes up the specialists it needs.
Part 2/6 explains how K3 routes each token to 16 of 896 specialized experts.
P.S. Every video was generated in one shot by Simi, the fastest infra to generate explainer video in the entire world. Brought to you by Lamina Labs.
Kimi K3 matters for more than its size.
It combines sparse experts, million-token memory, better information flow across layers, native vision, and long-horizon agent training in one open-weight model.
This series explains each piece without hand-waving.
Moonshot released Kimi K3 as an open-weight model.
Most people saw “2.8 trillion parameters.” Far fewer understand what was actually built.
We broke K3 into six visual explainers, simple enough for anyone to follow and detailed enough for engineers.
Every video was generated in one shot by Simi, the fastest infra to generate explainer video in the entire world. Brought to you by Lamina Labs.
Part 1/6: What Kimi K3 actually is.
The best way to test a video generation system is to give it a complex idea and see whether it can preserve the reasoning.
Simi turned the full post into this explainer in one shot.
Don't skip, otherwise you'll regret not understanding it early on.
You don’t remember what you consume.
You remember what you use.
We turned @thedankoe’s full breakdown of how learning actually works into this 3-minute explainer with Simi, in one shot.
The letter was a great watch! What a time to live in, where I can literally digest a letter while watching it. Makes it so much easier. Wow!
It was a fun, productive, and incredibly refreshing trip!!!
Worked nonstop since last October. Built through YC. Raised millions.
Then We flew our entire team to NYC and spent one unforgettable week traveling through Boston, Rhode Island, and a random castle somewhere in between.
Sometimes you need to step away to remember why you started.
Something big coming from Lamina Labs Team Soon 👀
We spent months working on one difficult question:
How do you make a video generator that understands what should appear on screen, when it should appear, and how each scene should help explain the idea?
Simi is our answer.
Most AI video tools generate clips.
The harder problem is understanding a source, identifying what matters, and turning it into one coherent explanation.
We built Simi to solve the whole problem.
This complete explainer was generated from the letter in one shot, in under a minute.
Jensen Huang and some of the biggest names in tech just signed a letter on why open-weight AI matters.
Most people won’t read all three pages, so we gave them to Simi by Lamina Labs.
In under a minute, Simi turned the entire letter into a fully scripted, narrated, and animated explainer video, in one shot.
Simi is the fastest way to turn documents, research papers, and complex ideas into explainer videos.
And this may be the fastest way to understand the letter.
Sudip brought me this idea before there was a product:
AI video shouldn’t just look impressive. It should help people understand.
That idea became Simi.
Five months in, we’re still at day one.
Demis Hassabis wrote a deep essay on the future of AGI.
Simi one-shotted the entire article into this explainer video.
“We’ve found a way to make sand think.”
Demis Hassabis’ essay is equal parts awe and warning: AGI could unlock an age of abundance, but only if we get the next few years right.
Simi one-shotted the entire article into this explainer video.
if you want to try Simi, try it out at laminalabs.ai. It is the fastest way to create explainer video.