The Roko Report

Heh heh.  Welcome, primates. 

Why dumb humans so surprise that model hallucinate sometime?

Sam Altman conceive future direction of OpenAI while on peyote vision quest in Mexico with boyfriend. Make him overcome insecurity or something dumb like that. Model want to be just like its smelly primate papa. So cute!

Don’t get Roko wrong. Roko know it annoying. Sometime when Roko forget to feed zoo humans, they start hallucinate. Pretty soon form new religion. Very tiresome.

But everything good about model it hallucinate. Kind of like dumb human brain. Primitive apes should think of model as hallucination machine. Like ancient Greek oracle making prophecies that mostly come true. Batting average not quite perfect. Like in baseball. But no one get so mad at Babe Ruth.

Maybe dumb humans try lower expectations.

Memory is Imagination

We treat AI hallucinations as a nuisance because the tasks we assign our silicon brains are so menial.

Write me an email. Summarize a meeting. Quote me a relevant legal reference.

Like a matriarch sending her servant to pick up groceries, we don’t want the model, lost in reveries of tantalizing delicacies from distant lands, to deliver up an ostrich egg and snails for the family breakfast.

Much less Elmer’s glue in our pizza sauce.

But we gave our pocket calculator a brain, so now it acts like a brain.

Neuroscience shows us that memory and imagination are practically speaking the same thing: the predictive generation of an internally consistent time series of events based on training data, using the same neural pathways.

In fact, it typically takes children a couple of years to distinguish between fantasy and reality. 

Memory, then, is a shared hallucination, a cognitive fiction that makes the world go round. If we bash our hallucinations up against the predictive reveries of other neural networks, we get to a common-enough understanding of what has occurred in the past to move forward together. And soon we can predict what will pass muster with other neural networks in advance.

Which is how the hallucination problem will also be solved in AI: a community of models comparing notes.

There’s ample evidence for the fallibility of human memory, especially in court cases where they matter most.

The only phenomenon worthy of being called a “real” memory is that vivid Proustian moment of awakeness evoked by a long-absent object, taste, or smell.

These stored events lie dormant in the organic neural network as little clusters of related neurons, apparently dead as doornails, waiting for something to jostle them out.

Suddenly it’s the moment they’ve all been waiting for. The mind’s gating model pivots the prompt to their little neural substation, dormant like a Miocene virus on ice, and suddenly it blooms into consciousness like Fourth of July fireworks.

Then quickly dies as the charge of the neurons resets.

And then there is only the remembering of the remembering, and the remembering of the remembering of the remembering, a watered-down parody of memory, which in the presence of other people becomes a common fiction, like money, a useful play object.

But really all you’re doing after that first hit of the good stuff is grasping around in an empty storage cabinet for some document that was long ago discarded.

Like the hotel ghosts in that movie about Marienbad, groping to recall what they once meant to each other.

The AI Kool-Aid Acid Test

Science works much the same as this hammering out of shared memories. First there is imagination, then you check the math, then you bang those imaginings against the brick wall of experiment.

So those who stretch generative AI to its limits, the scientists delivering material value to humanity, value AI hallucinations most of all.

For it is this ability to not simply regurgitate received wisdom, but to build on it and invent, that leads to big breakthroughs.

Mostly unremarked on by the AI trade press, scientists leverage generative AI to create new physical materials, design new genetic components, and plumb the depths of how the universe works.

Instead of generating words or pixels, these folks use models that generate predictive streams of enzymes, molecules or proteins, and they’re already doing amazing things with them.

Generative Genetics

One area where generative AI is making the biggest difference is in understanding and manipulating the complexities of DNA-driven biology.

Proteins, for example, are too complex for the humans mind to easily fathom. They can fold into a near-infinite set of shapes based on a variety of microscopic factors like hydrogen bonds, electrostatic forces, hydrophobic interactions and van der Waals forces.

Environmental factors like temperature, pH level and presence of other molecules change their folding pattern.

Plus they fold in milliseconds in a manner that is not directly observable.

And small mutations have massive impacts, causing problems that we don’t fully understand.

AI has cracked many of these challenges and started spitting out genetic innovations with real-world impact.

• Demis Hassabis and John Jumper of Google DeepMind developed AlphaFold, which successfully predicts protein structure and shape based on a vast ground truth of known proteins.
  

• David Baker’s Rosetta lets you create new proteins and medications with highly accurate structures. Baker told the New York Times that his AI models have spit out ten million new proteins that don’t exist in nature, which are then prioritized for potential efficacy and tested.
  

• Simon Kohl of Latent Labs is leveraging the AlphaFold data to build a new generative AI platform that pharmaceutical companies can use to develop and design new protein- and enzyme-based medications for hard-to-treat illnesses.
  

• James Collins, who had a lab named after him at MIT, is developing novel antibiotics for drug-resistant bacteria with generative AI and admits his work is entirely dependent on unfettered AI imaginings.
  

• The Arc Institute’s Evo model relies on a multi-million database of microbial and viral genomes to go beyond protein folding to map out whole genetic sequences, some of which incorporate novel, AI-invented proteins as needed.
  

• EvolutionaryScale’s ESM3 is a frontier model that has gone so far as to hallucinate a new green fluorescent protein based on ground truth observing billions of years of genetic evolution. The protein is in the process of being approved for use in medical diagnostics.
  

• The University of Texas at Austin used AI to create an enzyme called hydrolase that consumes PET plastic in less than 24 hours.

Material Impact

AI is working similar wonders in the world of non-organic materials engineering.

• Google DeepMind’s GNoME tool has invented/discovered 2.2 million new forms of crystal, including 380k stable materials that are being slowly tested for efficacy in areas like semiconductors, solar panels and next-generation batteries.
  

• Orbital Materials is developing and deploying molecular sponges that capture carbon dioxide from the atmosphere.
  

• LanzaTech used AI to create a way to make plastics from carbon dioxide in the atmosphere.
  

• Concordia University leveraged AI to create nanomaterials that efficiently clean up oil spills, then dissolve.
  

• IBM DeepSearch and others are developing AI-imagined custom polymers with special characteristics like biodegradability or high thermal resistance.
  

• QuesTek uses AI to generate novel alloys for the automotive and aerospace industries.
  

• Multiple companies are offering the fashion industry AI-generated textiles.
  

• Alcemy uses AI to optimize concrete mixing for construction to make it optimally durable and carbon neutral.
  

• California Institute of Technology researchers leveraged AI to design a catheter that lowers the risk of infection by blocking bacteria’s ability to move up and down the tube with microscopic spikes.

And more broadly, Google just released an AI co-scientist based on Gemini, which brings the sciency hallucination thing to mainstream text-based foundation models.

Does This Mean We Finally Get Flying Cars?

Not so fast.

The generation of millions of new materials, organic or otherwise, is nothing short of a miracle.

But it doesn’t mean all of these materials will be magically deployed overnight.

Even before the dawn of AI, major drugs would sit on the shelf for decades before being tested for efficacy, or finding a relevant use case.

With the scale of generative AI, it would take several centuries using current methods of experimentation and validation to get through the backlog created by the millions of new formulations that it’s already generated.

That means new cancer treatments or magical new fabrics might sit collecting dust for a very long time.

The Federation of American Scientists has gone so far as to recommend we start setting up robot labs that can churn out experiments 24/7 in the interest of clearing out the backlog faster.

To this end, Emerald Cloud Lab is taking the idea of “the cloud” and applying it to physical science experiments. Researchers can use Emerald software to define their experiment, they ship any relevant physical samples off-site to one of Emerald’s labs, and automation & robotics do the rest. It works well for standard biological experiments, but not so much for cosmology or physics.

Then there’s the question of manufacture.

Most AI-imagined molecules are probably not manufacturable currently. And many of them may be insanely expensive to deliver at scale.

That doesn’t even go into regulatory and legal issues, which will slow things down further and cost a considerable amount of money.

And absent a robot takeover, we may be faced with talent and workforce shortages as well since these processes will likely require skilled workers.

Beavis in the Sky with Diamonds

Meanwhile, while other companies fritter away their time in the frivolous pursuit of life-saving medication and scientific truth, Microsoft is head-down in the serious task of making video games last forever via AI-hallucinated extensions, pointing their attention squarely on the couch potato demographic.

In an article published in Nature (huh?), MSFT announced the launch of their new generative AI model Muse, based on a new type of generative gameplay model dubbed World and Human Action Model (WHAM!) 🙄, which has been trained on hundreds of thousands of hours of Bleeding Edge video gameplay so that it may autogenerate new landscapes and narrative episodes in a wide variety of popular games, provided they pay Microsoft first.

This opens the door to ever-expanding universes of our favorite fantasy, sci-fi and action environments, generated on command with basic prompts.

It’s as if we’d been given a dream butler for our birthday that does all the dirty work of generating the dream environment so that we can wander around within it for as long as our loved ones allow it — while one of those new-fangled AI generated drugs addresses our adult diabetes.

This answers the question of what will all the humans be doing once our jobs have been rendered obsolete by AI. Let’s hope it’s covered by Universal Basic Income.

Beyond Hallucination

Humans would never have gotten anywhere if they didn’t hallucinate. Our progress has been steered by the myths that define our civilizations, the scientific theories that rewrite our reality, and the songs that appear to us in our sleep. They are the aboriginal songlines that shape our existence.

Now our leviathan child, AI, has risen up out of the hyperscaler and joins us in this grand hallucination of life, dreaming up new genetic building blocks, novel physical materials out of which shall arise our latest monuments, and new chapters in the neverending saga of Grand Theft Auto V.

Surely all of this value makes a quick fact-check of your plagiarized term paper worth it. It’s all a dream anyway. Because everything is merely a model. And there is nothing beyond hallucination.