TOGETHER WITH

Good morning, human brains, and welcome back to your daily munch of AI news.

Here’s what’s on the menu today:

• AlphaFold folds proteins like I do under pressure 🧬🚀

  Google DeepMind released Alphafold 3.

• Massive prompts outperform fine-tuning 💵 👨‍💼

  A study shows that thousands of examples improve AI performance.

Sponsor Bot Eat Brain (Now 50% off) 🤯

New here? Subscribe! 😎

MAIN COURSE

Fold me tender… Fold me sweet. 🧬🚀

Yesterday, Google DeepMind and Isomorphic Labs released AlphaFold 3. 

It’s the follow-up to AlphaFold 2, which advanced the field of protein folding and drug discovery.

AlphaFold’s favorite movie…? Amino Royale 😂 🧬

What does it do?

It predicts 3D protein structures from amino acid sequences. It’s predicts antibodies twice as accurately as AlphaFold 2. It also outperforms specialized physics-based systems in predicting small molecule interactions with proteins.

So this one is better?

Yes. Alphafold 2 only predicted the 3D structures of proteins. AlphaFold 3 predicts 3D structures of proteins, ligands, ions, DNA, and RNA, and more.

For more on AlphaFold 2, click here.

So, what?

Alphafold 2 led to the discovery of new proteins which advanced heart mapping, antimicrobial resistance modeling, extinct bird egg identification, and more.

Alphafold 3 promises to predict the structures of not only proteins, but many biological elements. It has the potential to revolutionize pharmaceutical research and expand the boundaries of biological understanding.

What does AlphaFold drive…? An El Amino. 🛻

Whoa.

Plus, Google DeepMind created a free, fully hosted web app.

Click here to take it for a spin.

SPONSORED BY TAPLIO

Save hours every week and expand your reach.

Taplio writes and schedules personalized LinkedIn content for you, so you can focus on growing your business.

• Suggests posts based on proven viral content

• Creates carousels using any URL or topic

• Expands your network with access to quality leads

• Monitors your performance with advanced analytics

The reviews are stellar:

“Taplio brought me a million views in 6 months. Up from 15k views in the prior 6 months.” – Emil Sterndorff, Founder of The Capital.

“I create content for LinkedIn all the time, and Taplio has greatly helped me analyze my past posts and understand what type of content is working and what type is not.” - Awais Kahn, Co-Founder of Superhuman.

Join 7,000+ founders, marketers, and professionals and try Taplio for free!

SIDE SALAD

Oversharing is an asset? 💵 👨‍💼

Last week, Carnegie Mellon and Tel Aviv Researchers published a paper on ICL (In-Context Learning).

It shows that massive prompts can outperform fine-tuning in AI models. Using hundreds or thousands of examples drastically improves their results.

Verbal diarrhea: the new secret to AI success! 💩🔥

What does that mean?

The study shows that as text prompts grow longer, the performance impact of individual examples or their order diminishes, making prompts more robust.

So, it’s better than fine-tuning?

Fine-tuning requires more data and has a higher one-time cost, but ICL requires more computing power due to extensive prompts.

Does it work for every AI model?

Nope. ICL only works with models that can handle long context windows. Fine-tuning is the better strategy with smaller models.

In AI land, the incessant rambler reigns supreme… 💪

WHAT WOULD YOU DO?

“Congrats, you win a free vacation.”

Where are you going? 👇

Good choice, are we there yet? ✈️

Ideas? Comments? Complaints?

Respond to this email or hit me up on 𝕏.

Now on to the Daily Munch. 🤗

YOUR DAILY MUNCH

Think Pieces 🧠

The U.S. government invested $285 million to develop “digital twins.“ They simulate the behavior of physical AI chips.

An open-source code library for robotics? It’s called LeRobot and Hugging Face just launched it.

Startup News 💰

OpenAI and Stack Overflow announced a partnership. They will integrate vetted tech content into OpenAI’s models to increase their accuracy and reliability.

OpenAI is reportedly launching a search engine. A web developer unearthed details on “SearchGPT,” that offers text search, image search, widgets, and more.

Google launched Google Threat Intelligence. It’s a cybersecurity tool that combines Mandiant and VirusTotal in its Gemini AI model.

Research 👨‍🔬

Clover — enhances parallel decoding with sequential knowledge, surpassing Medusa in efficiency and accuracy.

Extend Llama-3’s Context — lengthen Llama-3-8B-Instruct's context length from 8K to 80K with QLoRA fine-tuning, training samples, GPT-4, and more.

SAGS — (Structure-Aware 3D Gaussian Splatting) encodes scene geometry using graph neural networks to improve quality and efficiency.

MEMES FOR DESSERT

TWEET OF THE DAY

OpenAI started as a nonprofit to “benefit humanity” and turned into a for-profit Microsoft subsidiary. It’s faces more than 8 lawsuits from newspapers that claim it unlawfully scraped their content to train AI models. Yesterday, it published a blog to “deepen the public conversation” about how AI models should behave.

Thoughts?

Tag us on Twitter @BotEatBrain for a chance to be featured here tomorrow.

AI ART SHOW

Until next time 🤖😋🧠