Share this page!

Enter URL or ID to Unroll

You can paste full URL like: https://x.com/threadreaderapp/status/1644127596119195649
or just the ID like: 1644127596119195649

How to get URL link on X (Twitter) App

  1. On the Twitter thread, click on or icon on the bottom
  2. Click again on or Share Via icon
  3. Click on Copy Link to Tweet
  4. Paste it above and click "Unroll Thread"!
  5. More info at Twitter Help
Javi Lopez ⛩️ Profile picture
@javilopen
Aug 5, 2023 • 15 tweets • 5 min read • Read on X
Scrolly
🔴 PERFUSION: a generative AI model from NVIDIA that fits on a floppy disk 💾

It takes up just 100KB. Yes, you heard it right, much less than any picture you take with your mobile phone! Why is this revolutionary and can change everything?

I'll tell you 🧵👇 Image
Perfusion is a really lightweight "text-to-image" model (100KB) that also trains in just 4 minutes.

🔗 Link: https://t.co/502nEzWL2eresearch.nvidia.com/labs/par/Perfu…
Image
It allows creatively portraying objects and characters, maintaining their identity, using a novel mechanism they have called "Key-Locking." Image
Perfusion can also combine individually learned concepts into a single generated image. Image
Moreover, it allows controlling the balance between visual alignment and the text prompt at the time of inference, covering the entire Pareto front with just a single trained model. Image
And why is this revolutionary?

For several reasons. Image
1️⃣ Such great optimization means that we will soon have truly powerful AI models integrated into our mobile phones, computers, etc. Much lighter, faster to train, and consuming less computing power. Image
2️⃣ The costs of training models will be drastically reduced in the future with optimizations like this and new techniques that allow everything to be streamlined.
3️⃣ If, in just 100KB, a new technique (key-locking) has achieved such a large increase in the coherence of objects/characters between generations, as in this example, it means that we have only SCRATCHED THE SURFACE of what the future Generative AI will be able to do. Image
In short, a massive piece of news that I don't understand why it's going so unnoticed. Don't be fooled by "the low quality" of the images. The potential it has is truly MASSIVE. Image
If you liked this and would like me to continue writing similar threads, an RT on the first tweet of the thread will encourage me to keep doing so. Thanks! 😉👇

More info here: decrypt.co/150861/nvidia-…
Watch out for this important detail 👇

More info: fagenwasanni.com/news/nvidia-in…
Mentally, I'm already calling it: 'the miniLoras'

• • •

Missing some Tweet in this thread? You can try to force a refresh
　

Keep Current with Javi Lopez ⛩️

Javi Lopez ⛩️ Profile picture

Stay in touch and get notified when new unrolls are available from this author!

Read all threads

This Thread may be Removed Anytime!

PDF

Twitter may remove this content at anytime! Save it as PDF for later use!

Try unrolling a thread yourself!

how to unroll video
  1. Follow @ThreadReaderApp to mention us!

  2. From a Twitter thread mention us with a keyword "unroll"
@threadreaderapp unroll

Practice here first or read more on our help page!

More from @javilopen

Javi Lopez ⛩️ Profile picture
Apr 29
⚡ IT'S FINALLY HERE!

F-Lite: our first foundational model for image generation. A collaboration between Freepik ♥️ Fal.

• Open Source
• Fully commercially usable
• 10B parameter DiT trained on 80M images
• Trained with 100% licensed data

Link + info 🧵👇 Image
We’ve been secretly working on this for months! It feels good to finally share it!

LINKS:

• Regular version: more predictable and prompt-faithful, but less artistic: fal.ai/models/fal-ai/…

• Texture version: is more chaotic and error-prone, but delivers better textures and creative compositions: fal.ai/models/fal-ai/…

• Paper: github.com/fal-ai/f-lite/…

Enjoy!Image
Image
Image
Image
Congrats to @cloneofsimo, @ivanprado, @kuer5ord, @info_libertas and the rest of the team that built this model from scratch!

I did nothing except to test it 😍 Image
Image
Read 7 tweets
Javi Lopez ⛩️ Profile picture
Apr 15
🦾 Skin in the AI game

This is a side of me I don’t usually share publicly: my investment thesis based on my vision of the future. Because investing is exactly that: a bet that we’ll be able to guess the future.

Go grab a coffee, ‘cause this one’s gonna be long. It’s been a while since I put this much effort into a thread:

1. 🔮 My predictions.

• AI and everything that supports it (GPUs, datacenters, etc) will keep growing exponentially and steadily over the coming years, impacting every field of human knowledge.

• In the near future, every work process that happens in front of a computer will be affected by AI (if not completely swept away). And soon after that, every process that happens away from a computer too, thanks to robotics. And when AI and robotics converge, we’re in for some very interesting times (hopefully not terrifying).

• Pay close attention to what I’m about to say, it might blow your mind: I believe software (and a big chunk of audiovisual entertainment) will become a commodity, like electricity. Which means all the digital tech value will be concentrated in just a few companies: those who win today’s multimodal LLM race and those who provide the infrastructure they run on. You might understand this better if you imagine a world where you can just say: “I want a SaaS like this site” or “make me a movie in this style with my dog as the main character” and an LLM creates it on the spot, with a quality far beyond today’s best productions. Basically, I believe all logic and visual layers will be run on advanced LLMs we can barely imagine today. So, building apps/webs/entertainment the way we do now will stop making sense, and the ability to do so will be concentrated in companies with the best LLMs and the compute power to run them at scale. We’ll choose between “AI providers” based purely on price, and not so much on features/capabilities (just like we do today with electricity companies; or like PS vs Xbox if they get some exclusive IPs that make a difference).

2. 💰 My general investment thesis.

• There will be investment opportunities in everything that drives this paradigm shift (AI itself), but also in things that will still exist with or without AI (like food, real estate, or tourism (though I won’t cover these here, even if they’re still interesting and I might invest in them outside the stock market).

• As for AI, I’ll invest in both the “gold hunters” 🥇 (the companies in the race to build the foundation models) and the ones selling picks and shovels ⛏️🪏 (the companies building the hardware and infrastructure that make AI possible).

• Trying to “time the market” to find the perfect entry point is impossible. But there are some strong signs that the market is currently overvalued (see attached screenshot, data from CurrentMarketValuation).

• Concentrating your investment increases potential return, but also the risk. And vice versa.

3. 💸 My specific investment thesis.

• I want very high concentration in AI companies and everything that supports it, both in pre-IPO and in public markets.

• I think not only the US, but also China, will play a huge role in AI’s future. I have less faith in my dear Europe, because of its obsessive regulatory spiral and its ink-stained bureaucrats. Yes, I believe the US and China will devour the AI pie. But with China I sadly assume regulatory risks, so I won’t go above 10%-20% exposure in my portfolio.

• I don’t want to go all in at once in case the market is, in fact, overvalued: so I’ll be investing through monthly/quarterly contributions (TBD) over the next 5-6 years. In other words, I’ll avoid Lump Sum and follow a DCA (Dollar-Cost Averaging) strategy. This also lets me easily tweak the strategy later through future contributions if my portfolio drifts off course. Detail: historically, Lump Sum performs better... except when you hit the market at its peak. And since all signs point to us being maybe too high right now, I don’t want to risk it.

• But I don’t do trading. I actually DON’T believe in trading. Over 90% of active traders underperform the market in the long run. Even professional fund managers can’t consistently beat a simple index like the S&P 500 or MSCI World. So my plan is to build the portfolio over time, according to the weights in the screenshot, and never sell (unless I ever really need the cash). If anything, if I see the market drop hard, I’ll “buy the dip” and invest 2x or 3x the regular amount to take advantage of the discounts.

• Related to the above: author funds and picking individual stocks usually perform worse on average than simply indexing. So I want at least 70% of my portfolio to be indexed. But I’ll trust my own judgment and pick a few individual ones (30% of the portfolio). Again, I’m not planning to buy and sell often, just enter regularly over time.

• TER (fees) of funds and ETFs are super important and should be studied carefully. If not, they’ll eat you alive long-term. I’ve looked for the best products that match my thesis, but also the cheapest ones.

• I prefer accumulation over distribution for tax efficiency (I want at least 75% of my portfolio in accumulation stocks/ETFs). Long live compound interest!

• In Spain, moving between funds doesn’t trigger taxes (until you sell). The only downside is that fees are several points higher. But I want to keep at least a portion in funds so I can move things around easily and tax-free if needed.

• I think some of the best opportunities aren’t in public markets, but in pre-IPOs. I’ve managed to get into OpenAI, xAI, SpaceX, Freepik and Canva. I’d love to get into Anthropic, Inflection AI, Cohere, Hugging Face, Cerebras and Midjourney if I ever get the chance. If the stock market is already risky, the barrier to entry and risk for pre-IPOs or startups is way higher.

4. 🤯 Key risks to keep in mind.

• If you run this investment thesis through Gemini, Grok or ChatGPT’s deep research mode, their heads will explode 😂 (yep, I’ve tried them all, of course, I actually built this AI-focused portfolio partly using AI). Any LLM will lose its mind over the extreme AI concentration in this portfolio. If you concentrate, you increase risk but also potential return. If you diversify, you reduce risk but also reduce returns. I chose the former and I’m okay with the risks.

• “IE00BLRPRL42 (similar to TQQQ but accumulation)”: not for the faint of heart. It’s leveraged 3x, can go up fast... but also vanish at the speed of light.

• Cathie Wood’s ARKs are risky by nature. “Author ETFs” tend to underperform index funds, so they’re a risky bet on extreme concentration.

• KSTR is a Chinese AI companies ETF. Many are opaque, government-dependent, and vulnerable to sanctions or bans.

• The fact that I chose to enter gradually (DCA) means I’ll need to stay alert and rebalance in the future, sell duds before they crash and keep an eye especially on author ETFs and individual stocks. No one wants a 3dfx or a BlackBerry in their future portfolio... but it’s sooo easy to end up with one!

5. 🦄 Disclaimer: this is *definitely* not investment advice.

These are just my personal predictions about the future (which I might totally get wrong, because predicting the future is nearly impossible) and my investment thesis based on those predictions, which I decided to share. You’d be nuts to take this as investment advice. Everyone should make their own decisions.

So... how’s your brain doing after all that? Can’t wait to hear your thoughts!Image
Image
Link to my portfolio:

docs.google.com/spreadsheets/d…
Follow me at @javilopen if you found this interesting!

And a repost to the first post is always welcome 😜 Thanks! 👇
Read 4 tweets
Javi Lopez ⛩️ Profile picture
Apr 3
⚡ Let's play a game!

Just reply with your own image of the next frame you imagine.

I’ll be selecting the images and adding them to the thread so you’ll know what’s “canonical story”.

Finally, I’ll interpolate all the frames into a full video. Let’s see where this goes! Image
Style: "Retro tech-noir anime, like Akira, Ghost in the Shell, or Cyber City Oedo 808: cool tones and neon lights, strong shadows, intense expressions, and a futuristic, dark, and dramatic atmosphere."
Frame 2: (by @billywoodward)

Aspect ratio should be 9:16 👌 Image
Read 8 tweets
Javi Lopez ⛩️ Profile picture
Feb 27
IT'S FINALLY HERE!

🔥 Mystic Structure Reference! 🔥

Generate any image controlling structural integrity ✨ Infinite use cases! Films, 3D, video games, art, interiors, architecture... From cartoon to real, the opposite, or ANYTHING in between!

Details & 12 tutorials 🧵👇
Available NOW at Magnific 🪄 for all users! 👇

magnific.ai
Super easy to find:

👉 Magnific > Mystic > Structure reference

Don't forget a good prompt to guide your generation!

ℹ️ Currently Style Ref and Structure Ref can't be used at the same time. We are working on that! Image
Read 24 tweets
Javi Lopez ⛩️ Profile picture
Feb 18
⚡ Magnific on the big screen!

I CAN FINALLY TALK ABOUT THIS!

The VFX team of Here (directed by Robert Zemeckis and starring Robin Wright & Tom Hanks) used Magnific for their FX 🤯

To break it all down (+more), I interviewed VFX supervisor Kevin Baillie! 🧵👇
An incredibly exciting conversation where @kbvfx shares how he got started in the world of VFX, his career journey, what it’s been like working with directors like George Lucas and Robert Zemeckis, and the impact of generative AI in Hollywood plus much more! Image
So happy to finally be able to share the details!

I've been biting my nails for months but we weren’t allowed to make it public until NOW. Huge thanks to Kevin, finally!

As a co-founder of Magnific, seeing our creation used in a film directed by Robert Zemeckis (Back to the Future, Forrest Gump, etc.) is a dream come true.

We've always known that Magnific is a tool used by professionals (Dior, MrBeast, Adobe, Beeple, etc.), but seeing something this incredible with our own eyes makes it all feel so much more real.

You have no idea how happy this has made me. I don’t think I’ve felt this proud since that one time, by some miracle, I beat Emilio at Age of Empires.

Anyway, I won’t ramble on. Here are the questions Kevin was kind enough to answer, you won’t want to miss them!Image
Read 12 tweets
Javi Lopez ⛩️ Profile picture
Feb 14
🪄 Magnific.

From $0 to $10M ARR in just one year with only two people.

This was supposed to be a secret, but I think even my grandma knows by now.

Yep, small teams are the future. Image
I wasn’t sure whether to post it or not, but since X’s algorithm has me in the shadows, no one’s gonna see it anyway. 😆
Ah! And PURE bootstrapping.

ZERO investment and ZERO debt.

Thanks to our loyal customers 😘😘😘
Read 5 tweets

Did Thread Reader help you today?

Support us! We are indie developers!

This site is made by just two indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3/month or $30/year) and get exclusive features!

Become Premium

Don't want to be a Premium member but still want to support us?

Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal

Or Donate anonymously using crypto!

Ethereum

0xfe58350B80634f60Fa6Dc149a72b4DFbc17D341E copy

Bitcoin

3ATGMxNzCUFzxpMCHL5sWSt4DVtS8UqXpi copy

Thank you for your support!

Follow Us!

Send Email!

:(