Using Beethoven to Discover New Materials

THEWHITEBOX
TLDR;

• ✋🏼 OpenAI Sora Leaked

• 🥸 The Scam That Wasn’t A Scam After All?

• 🤖 Olmo2, The Best Fully Open LLM to Date

• 🫡 Anthropic Launches ‘Claude Styles’

• 🤨 Anthropic Models Engage in Uncanny Conversation

• ✒️ AI Poetry is Indistinguishable to Humans

• 📹 Runway Launches ‘Frames’ Model

• [TREND OF THE WEEK] Using Beethoven to Discover New Materials

Writer RAG tool: build production-ready RAG apps in minutes

RAG in just a few lines of code? We’ve launched a predefined RAG tool on our developer platform, making it easy to bring your data into a Knowledge Graph and interact with it with AI. With a single API call, writer LLMs will intelligently call the RAG tool to chat with your data.

Integrated into Writer’s full-stack platform, it eliminates the need for complex vendor RAG setups, making it quick to build scalable, highly accurate AI workflows just by passing a graph ID of your data as a parameter to your RAG tool.

Learn more about our production ready RAG tooling here.

NEWSREEL
OpenAI Sora Leaked

OpenAI’s Sora model was leaked for a short time. A group of artists serving as beta testers for the unreleased text-to-video AI model, Sora, have leaked access to the tool in protest.

They allege that OpenAI exploited their unpaid labor for research and public relations purposes, using them as free bug testers and for “art washing,” which lends artistic credibility to the product.

The leak enabled the widespread generation of AI-created videos resembling OpenAI’s demos before the company revoked access. OpenAI responded by stating that participation in the research preview was voluntary and emphasized its support for artists through grants and other programs. The company is working to balance creativity with robust safety measures before Sora’s broader release.

TheWhiteBox’s takeaway:

Without weighing in the dispute, which I do not know who is at fault here, the only thing I can say is that video generation is the biggest “fun toy you’ll play with for 10 minutes and never use it again” technology ever if it doesn’t improve its understanding of the world.

I mean, what on Earth is this thing below?

It can generate moving dogs, yes, but it doesn’t understand normal dog behavior and interactions, let alone one of the most basic concepts in physics: solid objects can’t go through each other. And even if it seems it can understand shadows, these behave in supernatural ways (the tails make shadows only at some points).

The videos are impressive, but you can instantly observe a tremendous amount of inaccuracies. These make it clear that the model simply imitates things it has seen many times without fully understanding them. While LLMs can hide their poor understanding by writing text, the gaps in knowledge and proper understanding from AI are clear with video models.

The only thing that makes me interested in video generation is that video carries much more information about the world than text, which means that, at scale, these models seem to be more suited to becoming world models (models that can predict events in the real world, much needed for things like embodied AI) than models like GPT-4. In other words, I believe the path to AGI is closer to video models than to LLMs.

HARDWARE
The Scam That Wasn’t A Scam After All?

A while back, we discussed the Rabbit R1, visionary hardware that allowed users to interact via voice with a Generative AI model that executed actions on their behalf.

However, when the product was first released, it received fierce hatred because it was basically non-functional. Coffeezilla’s research even accused it of not having an actual AI behind it.

Six months later, it seems that some unfulfilled promises are starting to come true. As this video shows, they have released ‘Teach mode’ to users, who can now teach the model using a very intuitive interface and the actions it has to execute in different software, such as Spotify or YouTube.

TheWhiteBox’s takeaway:

The AI running under the hood is particularly interesting. It’s not your usual LLM agent but a LAM (Large Action Model). The critical difference is that it doesn’t interact with other software through APIs but with the software’s interface, just like a human would using a cloud service.

However, just as we saw how primitive Anthropic’s results were, we must be very cautious about this feature. Without having tested it but knowing the dubitable reputation of the company behind the R1, it’s most likely still very buggy (nonetheless, it’s in beta testing).

That said, even if the R1 eventually underdelivers, it’s nonetheless a door to the future—a future where hardware products will work seamlessly with your voice in a declarative approach, a world where you declare what you want, and AIs do your bidding. It’s just unclear whether we are ready for that era.

OPEN-SOURCE
Olmo2, The Best Fully Open LLM to Date

The Allen Institute of AI (AI2) has released a second version of Olmo, their fully open-source model. According to their claims, it’s the best fully open model in the world, and it must be said that, at its size, it’s scarily good.

It’s one of the so-called ‘Pareto LLMs’ LLMs that aim to maximize performance as a unit of running costs. In other words, while they might not be the best model in raw performance, they offer the best performance per unit of compute, as seen in the above graph.

TheWhiteBox’s takeaway:

I know what AI2 did with that title. If you watch carefully, Olmo2 13B isn’t even close to being the best open-source model. Llama 3.1 70B, 405B, DeepSeek v2.5 Coder, or Qwen 2.5 32B are among the many “open-source” models with better performance.

But here’s the thing: those aren’t open-source; they are open-weight. True open-source comes when the lab releases the model’s weights for download and the training dataset, which neither Meta, Alibaba, nor DeepSeek does.

Thus, Olmo2 is the best truly open model.

In more detail, the model was trained for 6e23 FLOPs (600 YottaFlops, or 60 trillion trillion operations). If that sounds like a lot, it’s because it is.

To put this into perspective, if the Allen Institute wanted to train this model with just one state-of-the-art GPU, like an NVIDIA H100, which has a peak performance of 1,979 TeraFLOPs (2,000 trillion operations per second, which is a huge number), it would still take 3,509 days (or 9.6 years). This gives quite the insight into why AI labs need so many GPUs.

PRODUCT
Anthropic Launches Claude Styles

Anthropic has launched a new Claude feature that allows users to customize the AI assistant’s responses to align with their specific communication styles and workflows.

Users can select from preset styles—Formal, Concise, and Explanatory—or create custom styles by uploading sample content that reflects their preferred communication approach. This functionality enables Claude to adapt its tone and structure to various contexts, such as technical documentation, marketing materials, or project planning.

TheWhiteBox’s takeaway:

While I’ll die on the hill arguing that you should never use LLMs to write content for the public (because it’s generally pretty mid-level and lacks any meaningful trail of emotion), this tool can help rewrite emails, create internal documentation about your products and processes, and enabling—mid—content creation at scale.

However, to me, the news here is that frontier AI labs have clearly parked their ‘AGI aspirations’ (despite what their CEOs will say) to focus on what matters: product.

I’m sure investors aren’t buying the AGI hype anymore and are asking for revenues. And fast.

FOR THE LAUGHS
Anthropic Models Engage in Uncanny Conversation

Anthropic models act as if they were conscious when engaging in dialogue between them, as shown in the following tweet. I’m showing you this not because I am afraid of this but to expose how delusional some folks in the AI space are.

LLMs simply imitate human patterns, which can very well engage in these types of conversations by themselves. It’s not like they have developed consciousness or agency; they are literally parroting past human discussions.

Please do not give these people the benefit of the doubt. They will make you believe that AI is something it isn’t.

DETECTING AI
AI Poetry is Indistinguishable to Humans’

AI poetry is preferred over human poetry. And it’s also indistinguishable from human-generated poetry.

This is the result obtained by a study published in Nature, frankly quite eye-opening, as humans perform below chance (46% accuracy, meaning that you would be better off by choosing among the two options by pure luck) when detecting whether a poem is generated by a human or by an AI.

TheWhiteBox’s takeaway:

The implications of this are tremendous. If AI can fool us in poetry, there’s no way to tell whether the content is AI-generated or not. But, to me, this is nuanced.

For starters, they use non-experts who can’t claim to understand poetry, let alone choose between two options. Researchers also point out the secret behind this finding is the participants' faulty heuristics.

Specifically, most participants resorted to assuming that if something was ‘very simple to understand,’ it was human.

The inductive bias was to assume that human poems would be more straightforward to understand when the opposite was, in fact, true: AI generated much simpler poems that were easier to understand, which, for non-experts, made them more tempting to flag as human.

Conversely, as non-experts evaluated this, human poems were usually flagged as AI-generated because they were much more complex and nuanced (as real poetry is), deceiving the participants because they simply couldn’t understand it.

Long story short, the key heuristic for detecting AI-generated poems was literally the opposite of the one most broadly used, suggesting that these results are more due to a lack of experience from the average participant.

This points to a common theme in society's use of AI: We evaluate their intelligence and capabilities in the wrong way. While we should prompt them on topics we are experts on, we desire to be impressed and, thus, ask them questions on topics we do not fully comprehend. In that scenario, the AI appears much more competent than it really is.

VIDEO GENERATION
Runway Launches ‘Frames’

Talking about indistinguishable AI-generated content, we have Runway’s latest image foundation model, called ‘Frames.’ The model is so accurate with its representations that you cannot truly know whether it’s AI-generated.

As the examples show, the model provides what appears to be unmatched personalization capabilities, allowing you to create almost anything you wish.

TheWhiteBox’s takeaway:

It’s clear that soon, most content will be not only AI-generated but also indistinguishable from human content. Now more than ever, AI companies need to implement watermarking somehow. As much as I appreciate the industry's progress, I would like to know whether the person I’m talking to is a human.

I predict that the open Internet will consolidate in vetted communities where proof-of-humanity credentials are provided to ensure you interact with humans, not human-like bots.

I also expect many humans to return to the real world. Over the last decade, we have become increasingly immersed in the digital world and simply forget about the real one. I hope this flood of AI-generated content will make the Internet less ‘human’ and attractive, forcing us to return to the real world.

Maybe it’s too late, though.

TREND OF THE WEEK
Using Beethoven to Discover New Materials

Did you know that an AI has found patterns between Beethoven’s Ninth Symphony and biological materials, which could lead to discovering never-before-seen ideas, concepts, and designs related to these biological materials just by using music?

Or that we can propose new materials based on the structure of a Kandinsky painting?

Yes, may I present to you what is the pinnacle of the trend we have hyping in this newsletter for months: using AIs to discover the world.

Now, we can use music and art to discover new materials and properties. Markus Buehler, an MIT star professor and researcher, has proposed this fascinating discovery that will blow your mind.

This method could also elevate a Large Language Model (LLM)’s reasoning capabilities to new heights, allowing them to think more clearly about the context and potentially find hidden patterns invisible to the human eye.

Today’s piece will introduce you to the world of graph-based reasoning, a concept that, according to some, holds the secret to unlocking AI’s true powers.

The Power of Graphs

If you’ve been a subscriber for a while, you will know what in-context learning is. In simple terms, one of LLM’s best features is that they can take in a new context they have never seen before, learn about it, and perform accurate predictions on it.

For example, we can connect an LLM to a browser API (as OpenAI does) and let it use the Internet. It can then scrape relevant information related to the user query and respond to recent events it couldn’t have possibly known about.

This is a powerful tool, but it’s highly dependent on context quality. Thus, what if we could enrich the provided context so that LLMs can traverse it and discover new latent knowledge that is not visible initially?  

From Text to Graphs

Professor Buehler argues that while standard text can give you the ‘what,’ ‘when,’ ‘where,’ and ‘which,’ the real value comes from the ‘how;’ that is, context can be improved if expressed as a form of knowledge (about the ‘how’) instead of raw information, because it then allows LLMs to reason better over it.

But how do we do this? Well, graphs.

Let’s take the following example: “Gene A, through the production and interaction of proteins B and P and the application of silk fibers C as scaffolds, can indirectly contribute to wound healing technology” (D).

Next, we can turn this sentence into a graph {Gene A → produces → Protein B; Protein B → interacts with → Protein P; Protein P → helps form → Silk fibers C; Silk fibers C → are applied as → Scaffolds; Silk fibers C → can be utilized as → Scaffolds in wound healing (D)}.

Therefore, we can conclude that Gene A → indirectly contributes to → Wound healing technology (D).

See what we’ve done? We’ve abstracted the key patterns in the data, leaving out the noise and focusing on the fundamental relationship hidden beneath it.

We can observe this visually below. Using concepts as ‘nodes’ and relationships as ‘edges,’ we can connect seemingly unrelated concepts, like Gene A and wound healing, by traversing a chain of connected concepts and relationships.

Through this method, the LLM can find connections in data much faster and more abstractly (eliminating irrelevant words and only focusing on the essential structures).

Another powerful feature of graphs is that they can connect information from different sources. For instance, using the example above, one piece of content may connect the three first concepts (from Gene AI to Protein P), while the other may discuss the connection between ‘Protein P’ and ‘Wound healing.’

Although the previous example is simple and doesn’t need an LLM, consider a graph based on 1,000 research papers with thousands of nodes and connections between nodes that can be 100-hopped relationships long, like the one below:

That is where the true power of AI comes in.

When sending these graphs to LLMs like GPT-4, Professor Buehler noted that the model generated deep insights and profound reasoning about never-before-established relationships between different concepts.

This clearly illustrates how LLMs, thanks to their in-context learning capabilities, could be used to find new paths of exploration to discover new materials and their properties, as well as relationships between apparently disconnected concepts.

It doesn’t take much to picture a world where this method is used to discover new ways to fight wounds or illnesses, new materials, or being applied to other areas beyond biology.

But Professor Buehler didn’t stop there. Actually, things are only getting wilder.

Finding the Common Structures in the Universe

While the method we have described is excellent for allowing LLMs to ‘close the gap’ between concepts that, while previously unnoted, share common connections through direct relationships, Buehler then focused on isomorphisms.

An isomorphism is a common structure across two or more seemingly disconnected data distributions. In layman’s terms, while two topics might seem completely unrelated, how their data is structured may be common.

But what does that even mean?

The Kaleidoscope Theory

The Kaleidoscope Theory (TKT) claims that while the universe appears infinitely novel and changing, its underlying structure is surprisingly common in nature and physics. In layman’s terms, some general laws rule the world, even if they express themselves differently depending on the medium.

Knowing this, Buehler posited: Could it be that if I find isomorphisms between two unrelated topics, I can use my knowledge of one subject to discover things about the other?

Using graphs again, it found that the answer was yes. For instance, the underlying structure (graph) of Beethoven’s Ninth Symphony was identical to that of a dataset of biological data.

Knowing this, assuming the data structures are identical, if one of the graphs is bigger than the other, we can extend the latter to discover new patterns in data.

For instance, the subgraph identified from the bioinspired corpus connects to many other nodes, whereas the Beethoven-based graph is much smaller and limited. Hence, we can use the first graph's known structural extensions to estimate how and in what specific manner the second graph may be extended. This can lead to an extension of knowledge in a less well-studied field, aka amplifying our knowledge in that given area.

Crucially, these isomorphisms, as present everywhere in the universe, are modality agnostic, meaning we can also use the underlying structures in images.

For example, Buehler found an uncanny resemblance between Kandinsky’s “Composition VII” and sustainable mycelium-based materials, which could lead to the discovery of new materials simply by understanding the patterns in the painting.

Whenever people described AI as magic, it was never about ChatGPT; it was always about stuff like this.

But the biggest contribution of this paper is that it could finally answer the secret to unlocking AI ‘generalization.’

Current AI models cannot reason over data they haven’t fully analyzed before. But if the model can apply the isomorphisms of known data to new data, it will be able to reason over that unknown data, tearing down what seems to be the ultimate blocker to real AI progress: the capacity of AIs to reason over “unknown unknowns.”

And with that, we can really start asking ourselves whether we are truly creating machine intelligence.

TheWhiteBox’s takeaway

Magic, that’s how I see this research.

It provides a refreshing view of AI’s current limits, and offers a novel approach to dealing with them: the use of graphs and the underlying patterns in data as a tool to discover and extrapolate, one thing that AIs have continued to fail at.

It also illustrates the universal nature of certain principles across different fields of study, going beyond the conventional boundaries of disciplines and exploring how ideas from one area can enrich understanding in another.

But if there’s one thing I want you to take away from all this, it is that, beyond the ChatGPTs of the world (which will still play a role as we’ve seen today), AI’s most significant contribution to society will be the acceleration of scientific discovery.

And the weapon isn’t superintelligence as people like Sam Altman frame it; it’s a much boring but equally powerful yet true reality about AI; they can pick up patterns we simply can’t. And by using graphs, we are simply making their greatest feature even more powerful.

And that, AGI and superintelligence aside, is the real AI revolution.

THEWHITEBOX
Premium

If you like this content, by joining Premium, you will receive four times as much content weekly without saturating your inbox. You will even be able to ask your questions you need answers to.

Until next time!