Build Fast with AI

We ship AI apps and show you exactly how.
20K+ builders following along.

[Build Fast with AI @BuildFastWithAI]

PewDiePie just launched Odysseus on May 31st.

44,000+ GitHub stars in under 5 days.

The project: a self-hosted AI workspace that runs
100% on your machine. Free. Open source. MIT license.

No subscriptions. No cloud. No tracking.
"It's yours and yours forever."

What it actually does:

→ AI chat + autonomous agents
→ Deep research tools
→ Email management
→ Document editing
→ Persistent memory
→ Local model serving (Ollama, llama.cpp, vLLM)
→ OpenAI/OpenRouter if you want cloud too
→ Windows, macOS, Linux

He spent 12 months building it.
Built half of it using AI.
Called the launch video "MY trillion $Dollar Project is finally OUT!"

The line he opened with:
"The war on big tech has just begun."

Most AI startups backed by millions in VC never
see 44,000 GitHub stars in a full year.

A YouTuber with 110M subscribers just out-shipped them
in 5 days.

The reaction says something bigger than the hype though:
People want AI tools. They just don't want to rent them
from someone else's servers forever.

Have you tried it?

[Build Fast with AI @BuildFastWithAI]

NVIDIA just shipped Nemotron 3 Ultra.

550B MoE. 55B active parameters per token. 5× faster inference.
30% lower cost than competing open models.

But the number that tells the real story: 300–400+ tokens/sec
in real-world deployment. That's not a benchmark lab result.
That's what your agent actually experiences in production.

What makes this architecture different:

Hybrid Mamba-Transformer + LatentMoE means the full 550B never
activates on a single token. Smart routing pays compute only for
what the task needs. Efficiency at scale, not just in demos.

The specs that change production decisions:
→ 1M token context window
→ 20 trillion training tokens
→ Full open weights + training recipes + reward models + datasets

That last line is unusual. Most labs open the weights.
NVIDIA opened everything - including exactly how they trained it.

Highest-ranking U.S. open-weight model right now.
Beats Gemma 4 31B on major open benchmarks.

The bigger strategic move: NVIDIA now controls GPU → inference stack
→ foundation model
→ agent frameworks.

They're not competing in the model race.
They're making their hardware the only rational choice by also
shipping the best software running on it.

Running it through our 5-task agent evaluation suite this week.
Full results dropping soon.

What workload are you stress-testing this on first?

[Build Fast with AI @BuildFastWithAI]

Google: "People love AI Mode."

Also Google users: *traffic to DuckDuckGo's
"No AI" search page triples*

The gap between what tech CEOs say on earnings calls
and what users actually do has never been wider.

DuckDuckGo isn't even anti-AI.
They just let you choose.

Turns out that's a competitive advantage now.

[Build Fast with AI @BuildFastWithAI]

@RamirezTims i wish it was cheaper tho

[Build Fast with AI @BuildFastWithAI]

I made a 30-second Visa ad using Seedance 2.0.

No production crew. No location shoot. No budget.

Just references, prompts, and a clear storyboard.

The result is below.

A year ago this would have cost alot of money and weeks of production time.

Today it cost an afternoon.

AI video isn't coming for advertising.
It's already here.

Is this the future of brand advertising especially with the launch of such amazing models like Gemini Omni and Seedance 2.0?

[Build Fast with AI @BuildFastWithAI]

I did 10,000 AI video generations in 10 months.
Here's what actually matters.

The mindset shift that changed everything:

You're not a creative. You're a system builder.

The creators making money aren't the most artistic.
They're the most systematic.

The 6-part prompt structure that works across everything:

[SHOT TYPE] + [SUBJECT] + [ACTION] +
[STYLE] + [CAMERA MOVEMENT] + [AUDIO CUES]

Most people stop at action. Audio cues alone
doubled my engagement.

Instead of: "Person walking through forest"
Try: "Person walking through forest, Audio: leaves
crunching underfoot, distant bird calls, gentle wind"

The difference is not small.

What nobody tells you about cost:

Google's direct pricing = $0.50/second = $30/minute.
Factor in failed generations = $100+ per usable video.

Volume testing at those rates is impossible.
Find cheaper resellers. Makes the whole model viable.

The camera movements that always work:

→ Slow push/pull - most reliable, professional feel
→ Orbit around subject - great for product reveals
→ Static with subject movement - often highest quality

Avoid complex combinations. One movement per generation.

The biggest mistake I see:

Fighting the AI aesthetic instead of embracing it.

Beautiful impossibility performs better than
uncanny valley realism every time.

Stop trying to make AI look real.
Make it look like nothing else can.

The workflow that generates profit:

Mon: Analyze performance, plan 10-15 concepts
Tue-Wed: Batch generate 3-5 variations each
Thu: Select best, create platform versions
Fri: Schedule for optimal posting times

10,000 generations later:
Volume + selection > perfection + luck.

What's been your biggest AI video breakthrough?

[Build Fast with AI @BuildFastWithAI]

Gemma 4 12B just dropped.

The move nobody's talking about:
No multimodal encoders at all.

Vision and audio go directly into the LLM backbone.
Less memory. Less latency. Same output quality.

Runs locally on 16GB VRAM.
Benchmarks near their 26B model.
Apache 2.0.

Google is making "run a capable multimodal model
on your laptop" genuinely possible.

That changes what's viable to build without cloud costs.

[Build Fast with AI @BuildFastWithAI]

The text-in-image problem has been "unsolved" since 2022.

Ideogram 4.0 just solved it.

9.3B open-weight parameters. Elo 1285 on Design Arena.
115 points ahead of FLUX.2. #1 open image model on the planet.

But the number doesn't tell the story. The video does.

Watch what it does with typography - logos, stickers, posters,
mixed fonts - all rendered correctly. No hallucinated letters.
No broken words. Just clean output.

What makes this different from every model before it:
→ Trained on structured JSON captions, not freeform prompts
→ Native 2K generation
→ Up to 16 controllable colors per image
→ Precise object + text positioning
→ Open weights - fine-tune on your own data, run on your infra

For AI builders: this changes the design layer of every product
you're building. Marketing assets, UI mockups, branded outputs -
all now doable without a designer or a closed API.

We're running it on 5 real client design briefs this week.
Full unfiltered results dropping soon.

What's the first thing you're generating with this?

[Build Fast with AI @BuildFastWithAI]

Qwen 3.7 Max one shots a 5 slide deck with AI generated images and video :

[Build Fast with AI @BuildFastWithAI]

@zavxai gemini because i like to live life on the edge

[Build Fast with AI @BuildFastWithAI]

@VazeKshitij ai slop in the post too lol

[Build Fast with AI @BuildFastWithAI]

@aryaan_smm yeah

[Build Fast with AI @BuildFastWithAI]

OpenAI just launched Sites for Codex.
Natural language → deployed web app → shareable URL.

No code. No hosting setup. No deployment pipeline.

Here's what the announcement doesn't say - and what it means for the tools we actually recommend to builders:

Sites connects to Slack, Google Calendar, Google Drive, and your internal data. Then it deploys with Cloudflare infrastructure, auth controls, and access management built in.

That's not a "no-code tool." That's the full software creation stack - planning, coding, hosting, auth, maintenance - collapsed into a single prompt.

The number that tells the real story: 5 million weekly Codex users. And 20% of them aren't engineers. They're knowledge workers - analysts, ops leads, PMs - who are now shipping production web apps from a text description.

This is what disruption actually looks like when it arrives. It doesn't announce itself loudly. It just makes the question "do I need a developer for this?" answerable with "no" for 20% more use cases overnight.

The platforms that should be paying close attention: Lovable, Replit, Bubble, Glide - every no-code tool whose value proposition was "build apps without code." OpenAI just made that their value proposition too. With 5M weekly users already in the funnel.

We're going to build 3 internal tools using Sites this week and compare directly against what we'd build with our current stack.

What's the first internal tool you'd build if you had this access today?

[Build Fast with AI @BuildFastWithAI]

@m13v_ hmm , interesting

[Build Fast with AI @BuildFastWithAI]

@i_nehasingh_ cursor

[Build Fast with AI @BuildFastWithAI]

@Pallavi_345 hey claude, make an app that makes 10k a month, make no mistakes . the make no mistakes part is most important btw

[Build Fast with AI @BuildFastWithAI]

Gemini Omni is impressive. Seedance 2.0 is still better.

Here's why the gap hasn't closed:

1. Motion quality
Action scenes, fight choreography, dancing - Seedance still
produces more convincing body mechanics. Fewer floaty movements,
better spatial consistency between frames.

2. Character consistency
Same character, multiple shots - Seedance was built around
reference-driven workflows. It holds. Gemini drifts.

3. Native audio-video generation
Seedance generates audio and video together, not audio
bolt-on after the fact. Better lip-sync, better sound timing,
better scene coherence.

4. Director control
The phrase that keeps appearing in creator reviews:

"Gemini looks good. Seedance feels directed."

Gemini creates attractive clips.
Seedance lets you control exactly how a scene unfolds.

5. Benchmarks
Seedance 2.0 leads the most recent academic benchmark
for audio-video physical consistency.

Gemini Omni is closing the gap on aesthetics.
The gap on craft is still real.

Which one are you actually using for production work?

[Build Fast with AI @BuildFastWithAI]

@Surendar__05 J*b

[Build Fast with AI @BuildFastWithAI]

@shiri_shh hope they dont add ai slop in windows , its already very bloated

[Build Fast with AI @BuildFastWithAI]

@MiniMax_AI another day another frontier model , but not one can provide sufficient rate limits