- Forget overnight delivery, ship that package via rocket
- CRISPR could be a godsend to California vineyards
- Microsoft aggressively employs OpenAI’s technology
Dear Reader,
A few days ago, new research was published on artificial intelligence (AI) that practically no one is talking about. They should be.
As with most research papers, they all have innocuous sounding titles. This particular one is called “Collaborating with language models for embodied reasoning.” I know. It doesn’t sound too exciting, but it is.
As anyone who reads The Bleeding Edge knows, large language models like GPT 3, GPT 3.5, ChatGPT, and now Google’s Bard are all the rage in AI. Trained on billions of parameters, they allow us to effectively hold conversations with an AI. The progress in this field in the last two months alone has been mind-blowing.
But where these AI models come up short is that they aren’t designed to function in the real world… Our world… The world of humans…
That’s why this paper is so important.
It’s written by a team of researchers at DeepMind, the U.K. based AI firm that was acquired by Google back in 2014. It’s the same firm that has been delivering breakthroughs in AI that we explored throughout last year, like AlphaFold 2, which solved a grand challenge in life sciences by predicting how hundreds of millions of proteins fold.
It is also the same firm that built Alpha Go, the AI that bested the best human grandmasters in the game of Go. Go was a particularly hard problem to solve because there are more possible positions in Go than there are atoms in the universe. That says something. That also meant that unlike the game of chess, there is not enough computing power to calculate all possible positions.
AlphaGo had to recognize patterns and “think” for itself in order to win. The computer scientist behind AlphaGo used deep learning, an advanced form of neural networks, and reinforcement learning to create the best Go player in history.
And it’s the “collaboration” between these large language models and reinforcement learning that makes this paper so exciting. Large language models are trained on tens of billions of parameters (text/information) which is why we can converse with them. They ingest, synthesize, optimize, and ultimately weight answers that are most likely to be correct and useful based on the inputs that they trained on.
To me, reinforcement learning is AI’s bridge into the real world. It’s the foundation to creating more generalized AI systems that can be given complex tasks to perform without any prior specific instruction.
At a high level, reinforcement learning is easy to understand. The AI, or “agent”, is “rewarded” for completing an assigned task. It’s a unique approach in that the AI using reinforcement learning can have somebody of existing knowledge from which it can call on, and it can also function in an environment that it’s not yet familiar with.
The goal of a reinforcement learning AI, or agent, is simply to find the best path/way to complete a given task.
Where the paper is exciting is that it combines a large language model with a reinforcement learning agent. It also introduces a third AI that monitors what the reinforcement agent is doing and communicates that back to the large language model.
The large language model, something like ChatGPT, is the foundational knowledge upon which the reinforcement learning agent draws upon to complete its task. The “monitor” acts somewhat as a feedback loop between the two.
The research produced two key results. First and foremost, the combination and collaboration between these AIs produced results. And the other was that the larger the language model, the better the performance.
While the tasks were relatively simple and only involved a two dimensional environment, it doesn’t take much to extrapolate that the tasks can rapidly become more complex and even enter the real world. And there’s a clear path toward improving the performance further through the use of larger language models.
This all may sound academic, but here is what this all means: This approach to artificial intelligence can be applied to humanoid robots. This is exactly the kind of technology that brings AI into the real world and enables robots to work and collaborate alongside humans in factories, hospitals, assisted living facilities, mines, and even in our homes.
It won’t be long now…
Hypersonic cargo transport, global transport in hours…
Cargo transport hasn’t changed much at all in decades. Yes, it has become a lot more efficient, and the logistics networks are now extensive and pervasive around the globe. But we still ship cargo by ship, by train, and, if it’s urgent, by plane – Just as we did fifty years ago.
But what if we could deliver cargo anywhere in the world within hours using aerospace technology?
The U.S. Air Force is exploring the possibility of delivering military supplies and humanitarian aid via rockets. So rather than launching a rocket and its payload into orbit, the rockets would deliver the payload to some other point on Earth. And that’s why the U.S. Air Force just awarded SpaceX a $102 million contract to pilot such a program.
SpaceX founder Elon Musk has talked about this in the past.
Rockets aren’t limited exclusively to space travel. They can also enable hypersonic travel around the Earth. And this could cut flight times down by as much as 90%. We’re talking about turning fourteen hour flights into about ninety minutes.
Cost is the only reason rockets haven’t been used in this manner before. Up to this point, the cost to transport cargo this way has been too expensive.
But SpaceX has already reduced the cost of transport down to about $4,000 per kilogram (kg) thanks to its Falcon 9 rocket and the reusable boosters. This represents more than an 80% price decrease compared to the “old way” of doing things.
And get this – the SpaceX Starship will cut transport costs by another 97.5%. It will enable cargo transport at just $100 per kg.
At that point it’s entirely economic to transport cargo around the world via rocket. And that’s why the U.S. Air Force is partnering with SpaceX on this initiative.
We’ll have to wait a bit for more details on the plan. For instance, how would this even work?
Would a SpaceX rocket “airdrop” cargo once it’s over the destination? Would the rocket land at a nearby spaceport and be unpacked like a normal cargo jet? Will SpaceX use the Falcon 9, the Starship, or a newly designed configuration specifically for hypersonic sub-orbital travel? I’m excited to see what SpaceX comes up with.
But we could imagine the possibilities if we expanded this concept into commercial deliveries.
Once SpaceX has demonstrated the functionality, I’m sure companies like DHL, FedEx, and UPS will want a seat at the table also. Suddenly, they would be able to deliver high-priority packages in minutes rather than days.
We’re going to keep a close eye on this initiative from SpaceX. Sometime in the near future, we’ll be receiving goods launched from halfway around the world in a matter of hours. And they will have traveled on the top of a rocket into space to get to their destination. I like the sound of “hypersonic” or “sub-orbital” delivery.
Agricultural applications for CRISPR genetic editing…
Researchers out of the Baylor College of Medicine are working on a novel application of CRISPR-Cas9 genetic editing technology. They’re tackling something called Pierce’s Disease.
Longtime readers will know we can think of CRISPR as “editing software” for our DNA. It’s able to cut, insert, or replace pieces of genetic code. We’ve mostly covered the application of this technology in healthcare. But it goes far beyond that…
CRISPR can also be used on plants and insects. Those of us in agriculture will be familiar with Pierce’s Disease. It stems from a bacteria that can cause devastation to crops. It’s so destructive that it can cause grape vines to whither completely.
This bacteria breeds in the mouth of a strange little bug. It’s called the glassy-winged sharpshooter. Here it is:
Source: Nikonians
The insect looks a bit like a cricket. It seems harmless enough…
But when it lands in crop fields, it spreads the bacteria that cause Pierce’s Disease far and wide. And the disease leaves destruction in its wake.
To give us an idea, one study showed that, in 1999, wine grape growers in Temecula, Riverside County, CA, lost 200 acres to Pierce’s Disease. That was roughly 10% of the total acreage. And as somebody who enjoys a nice glass of wine, that’s a terrible shame.
That’s where CRISPR comes in.
The team at Baylor managed to sequence the full genome of the glassy-winged sharpshooter. With some study, they pinpointed areas of interest in the genome. They found the genes that create the conditions which allow bacteria to thrive in the bug’s mouth.
Then the team developed a CRISPR therapy to correct the problem. It cuts out the genes and replaces them with new ones.
Once injected into the insect embryos, this CRISPR therapy effectively changes the tissue in the bug’s mouth so that it develops a coating around it. This makes it impossible for the bacteria to breed.
The end result is that glassy-winged sharpshooters can frolic in crop fields without spreading the destructive bacteria. It’s quite a creative solution. And to think – it all comes from a small genetic change.
But I’m hopeful that a solution like this will be used to address some of the largest problems facing the agricultural industry in the years to come. And in this instance, I hope it means we’ll have plenty of wine to enjoy for the foreseeable future.
Microsoft is moving fast with its AI partnership…
Yesterday we explored how Google is rolling out its own generative artificial intelligence (AI) in response the success of Open AI’s ChatGPT.
Well, it didn’t take long for Microsoft to make a counter move. Microsoft just announced that it has incorporated ChatGPT into its search engine, Bing.
As readers know, Google is the 800-pound gorilla in the online search industry. In 2022, Google search held more than 92% market share. Bing is a distant second at 3.4%.
But implementing ChatGPT into Bing could change all that. The service is so appealing to consumers that it could eat into Google’s market share very quickly. That’s a problem for Alphabet – Google’s parent company – which derives roughly eighty percent of its revenue from online advertising.
This news got a lot of media attention. But what very few noticed was another – but equally important – announcement from Microsoft.
It’s not just Bing. ChatGPT is making its debut in Microsoft Teams as well.
The generative AI is capable of having intelligent conversations with humans. They can “listen” to human speech and to a certain degree understand its nuances and implications.
And that’s why applying ChatGPT to Microsoft Teams is so interesting.
Microsoft revealed that ChatGPT will attentively listen to every call and take notes in the process. It will provide a full transcript as well as a list of the highlights from each call.
The AI will also put together a suggested list of tasks that need to be done as a result of the call. And it will measure the level of engagement from each member of the call.
Here’s a visual of what it all looks like:
Source: Microsoft
Here we can see that ChatGPT provides notes and suggested tasks presented on the call in the column on the right. And down below it measures how much each person spoke.
Microsoft is building this functionality into a premium version of the product called Microsoft Teams Premium. They call it “intelligent recap technology.”
Microsoft Teams users can upgrade to the premium version for an extra $7 per month right now. That price will go up to $10 a month in June.
Make no mistake about it – this is another shot in the race for generative AI supremacy.
Microsoft is taking aim at Google’s core business by adding ChatGPT to Bing. And the addition into Microsoft Teams could be bad news for companies like Zoom. I would fully expect this new version of Teams to be picked up by enterprise customers.
One of the difficulties – from the perspective of corporate operations departments – of remote work is that it’s hard to “keep an eye” on all employees. A ChatGPT-powered Teams could address that.
It would give companies even greater surveillance over the behavior of employees. The AI will list all employees’ engagement down to the second.
What if the operations department arbitrarily decides that every employee who demonstrates less than say 10% engagement is determined to be “checked out” because they’re not pulling their weight?
Well, that would create a situation where all employees would chime in just to hit their engagement metrics – whether or not their contributions were useful or necessary.
This is a great example of what’s called the Hawthorne Effect. It says that people modify their behavior when they know they are being watched.
So this kind of tracking could potentially do more harm than good.
My team and I often use an AI that creates a transcription of each call. And it allows us to notate certain timelines or important parts of a meeting. It’s also a very efficient way for us to catch my thoughts on a topic and quickly turn it into editorial through transcription.
That can be incredibly useful.
But do we need to track just how much each team member spoke? Not at all.
Different companies will employ this kind of technology in different ways. But if it’s used improperly, it will change how people work, forcing them to “game” the system so that they don’t get flagged by human resources or operations. And that’s not a healthy work environment.
Regards,
Jeff Brown
Editor, The Bleeding Edge