Two Nobel Prizes: AI is Still resting on Giant Shoulders
John Hopfield and Geoffrey Hinton got the Nobel Prize of Physics, Demis Hassabis and John Jumper the nobel Prize of Chemistry. It is obvious that the first Nobel Prize was not given merely for their contributions to physics, but mostly for their profound and foundational contributions to what is today modern AI.
Let's talk about the second Nobel prize.
AlphaFold was put on map by beating other methods on a competition (CASP14/CASP15) that has been running for year on a well established dataset. As such, AlphaFold winning is more like an ImageNet moment (when the team of Geof Hinton demonstrated the superiority of Convolutional Networks on Image Classification), than a triumph of multi-disciplinary AI research.
The dataset of Alphafold rests on many years of slow and arduous research to compile a dataset in a format that could be understood not by machines, but by computer scientists. This massive problem of finding the protein structure was, through that humongous work, reduced to a simple question of minimizing distances. A problem that could now be tackled with little to no knowledge of chemistry, biology or proteomics.
This in no way reduces the profond impact of AlphaFold. However it does highlight a major issue in applied AI: computer scientists, not AI, are still reliant on other disciplines to drastically simplify complex problems for them. The contributions and hard work required to do so gets unfortunately forgotten everything has been reduced to a dataset and a competition.
What to do when we do not have problems that computer scientists can easily understand? This is true for all fields that require a very high level of domain knowledge. Through experience, I came to consider the pairing of AI specialists with specialists of other disciplines, a sub-optimal strategy at best. The Billions of dollars invested in such enterprises have failed to produce any significant return on investment.
The number one blind spot of these endeavours is the supply chain, it usually takes years and looks like this:
1- Domain specialists identify a question
2- Years are spent to develop methods to measure and tackle it
3- The methods are made cheaper
4- The missing links: Computational chemists, Bioinformaticians, ... start the work on what will become the dataset
5- AI can finally enter the scene
Point number (1) is the foundation. You can measure and ask an infinite number of questions about anything. Finding the most important one is not as obvious as it seems. For example, it is not at all obvious that a protein structure is an important feature a priory. Another example, is debugging code. A successful debugging session involves asking and answering a succession of relevant questions. Imagine giving a code to someone with no programming experience and asking them to debug it. The probabilities of them asking the right questions is very close to 0.
Identifying what is important is called inserting inductive Biases. In theory LLMs could integrate the inductive biases of a field and generate interesting questions, even format datasets from open-source data. However until this ability has been fully demonstrated, the only cost efficient way to accelerate AI driven scientific discoveries is to build the disciplinarily into the people: AI Researchers that know enough about the field to be able to identify the relevant questions of the future.
Partages
Two Nobel Prizes: AI is Still resting on Giant Shoulders
copier:
https://bluwr.com/p/40004649
In the wake of two Nobel Prizes AI is Still resting on Giant Shoulders
John Hopfield and Geoffrey Hinton got the Nobel Prize of Physics, Demis Hassabis and John Jumper the nobel Prize of Chemistry. It is obvious that the first Nobel Prize was not given merely for their contributions to physics, but mostly for their profound and foundational contributions to what is today modern AI.
Let's talk about the second Nobel prize.
AlphaFold was put on map by beating other methods on a competition (CASP14/CASP15) that has been running for year on a well established dataset. As such, AlphaFold winning is more like an ImageNet moment (when the team of Geof Hinton demonstrated the superiority of Convolutional Networks on Image Classification), than a triumph of multi-disciplinary AI research.
The dataset of Alphafold rests on many years of slow and arduous research to compile a dataset in a format that could be understood not by machines, but by computer scientists. This massive problem of finding the protein structure was, through that humongous work, reduced to a simple question of minimizing distances. A problem that could now be tackled with little to no knowledge of chemistry, biology or proteomics.
This in no way reduces the profond impact of AlphaFold. However it does highlight a major issue in applied AI: computer scientists, not AI, are still reliant on other disciplines to drastically simplify complex problems for them. The contributions and hard work required to do so gets unfortunately forgotten everything has been reduced to a dataset and a competition.
What to do when we do not have problems that computer scientists can easily understand? This is true for all fields that require a very high level of domain knowledge. Through experience, I came to consider the pairing of AI specialists with specialists of other disciplines, a sub-optimal strategy at best. The Billions of dollars invested in such enterprises have failed to produce any significant return on investment.
The number one blind spot of these endeavours is the supply chain, it usually takes years and looks like this:
1- Domain specialists identify a question
2- Years are spent to develop methods to measure and tackle it
3- The methods are made cheaper
4- The missing links: Computational chemists, Bioinformaticians, ... start the work on what will become the dataset
5- AI can finally enter the scene
Point number (1) is the foundation. You can measure and ask an infinite number of questions about anything. Finding the most important one is not as obvious as it seems. For example, it is not at all obvious that a protein structure is an important feature a priory. Another example, is debugging code. A successful debugging session involves asking and answering a succession of relevant questions. Imagine giving a code to someone with no programming experience and asking them to debug it. The probabilities of them asking the right questions is very close to 0.
Identifying what is important is called inserting inductive Biases. In theory LLMs could integrate the inductive biases of a field and generate interesting questions, even format datasets from open-source data. However until this ability has been fully demonstrated, the only cost efficient way to accelerate AI driven scientific discoveries is to build the disciplinarily into the people: AI Researchers that know enough about the field to be able to identify the relevant questions of the future.
Partages
In the wake of two Nobel Prizes AI is Still resting on Giant Shoulders
copier:
https://bluwr.com/p/39066143
In the wake of two Nobel Prizes AI is Still resting on Giant Shoulders
John Hopfield and Geoffrey Hinton got the Nobel Prize of Physics, Demis Hassabis and John Jumper the nobel Prize of Chemistry. It is obvious that the first Nobel Prize was not given merely for their contributions to physics, but mostly for their profound and foundational contributions to what is today modern AI.
Let's talk about the second the second Nobel prize.
AlphaFold was put on map by beating other methods on a competition (CASP14/CASP15) that has been running for year on a well established dataset. As such, AlphaFold winning is more like an ImageNet moment (when the team of Geof Hinton demonstrated the superiority of Convolutional Networks on Image Classification), than a triumph of multi-disciplinary AI research.
The dataset of Alphafold rests on many years of slow and arduous research to compile a dataset in a format that could be understood not by machines, but by computer scientists. This massive problem of finding the protein structure was, through that humongous work, reduced to a simple question of minimizing distances. A problem that could now be tackled with little to no knowledge of chemistry, biology or proteomics.
This in no way reduces the profond impact of AlphaFold. However it does highlight a major issue in applied AI: computer scientists, not AI, are still reliant on other disciplines to drastically simplify complex problems for them. The contributions and hard work required to do so gets unfortunately forgotten everything has been reduced to a dataset and a competition.
What to do when we do not have problems that computer scientists can easily understand? This is true for all fields that require a very high level of domain knowledge. Through experience, I came to consider the pairing of AI specialists with specialists of other disciplines, a sub-optimal strategy at best. The Billions of dollars invested in such enterprises have failed to produce any significant return on investment.
The number one blind spot of these endeavours is the supply chain, it usually takes years and looks like this:
1- Domain specialists identify a question
2- Years are spent to develop methods to measure and tackle it
3- The methods are made cheaper
4- The missing links: Computational chemists, Bioinformaticians, ... start the work on what will become the dataset
5- AI can finally enter the scene
Point number (1) is the foundation. You can measure and ask an infinite number of questions about anything. Finding the most important one is not as obvious as it seems. For example, it is not at all obvious that a protein structure is an important feature a priory. Another example, is debugging code. A successful debugging session involves asking and answering a succession of relevant questions. Imagine giving a code to someone with no programming experience and asking them to debug it. The probabilities of them asking the right questions is very close to 0.
Identifying what is important is called inserting inductive Biases. In theory LLMs could integrate the inductive biases of a field and generate interesting questions, even format datasets from open-source data. However until this ability has been fully demonstrated, the only cost efficient way to accelerate AI driven scientific discoveries is to build the disciplinarily into the people: AI Researchers that know enough about the field to be able to identify the relevant questions of the future.
Partages
In the wake of two Nobel Prizes AI is Still resting on Giant Shoulders
copier:
https://bluwr.com/p/39050240
In the wake of two Nobel Prizes AI is Still resting on Giant Shoulders
John Hopfield and Geoffrey Hinton got the Nobel Prize of Physis, Demis Hassabis and John Jumper the nobel Prize of Chemistry. It is obvious that the first Nobel Prize was not given merely for their contributions to physics, but mostly for their profound and foundational contributions to what is today modern AI.
Let's talk about the second the second Nobel prize.
AlphaFold was put on map by beating other methods on a competition (CASP14/CASP15) that has been running for year on a well established dataset. As such, AlphaFold winning is more like an ImageNet moment (when the team of Geof Hinton demonstrated the superiority of Convolutional Networks on Image Classification), than a triumph of multi-disciplinary AI research.
The dataset of Alphafold rests on many years of slow and arduous research to compile a dataset in a format that could be understood not by machines, but by computer scientists. This massive problem of finding the protein structure was, through that humongous work, reduced to a simple question of minimizing distances. A problem that could now be tackled with little to no knowledge of chemistry, biology or proteomics.
This in no way reduces the profond impact of AlphaFold. However it does highlight a major issue in applied AI: computer scientists, not AI, are still reliant on other disciplines to drastically simplify complex problems for them. The contributions and hard work required to do so gets unfortunately forgotten everything has been reduced to a dataset and a competition.
What to do when we do not have problems that computer scientists can easily understand? This is true for all fields that require a very high level of domain knowledge. Through experience, I came to consider the pairing of AI specialists with specialists of other disciplines, a sub-optimal strategy at best. The Billions of dollars invested in such enterprises have failed to produce any significant return on investment.
The number one blind spot of these endeavours is the supply chain, it usually takes years and looks like this:
1- Domain specialists identify a question
2- Years are spent to develop methods to measure and tackle it
3- The methods are made cheaper
4- The missing links: Computational chemists, Bioinformaticians, ... start the work on what will become the dataset
5- AI can finally enter the scene
Point number (1) is the foundation. You can measure and ask an infinite number of questions about anything. Finding the most important one is not as obvious as it seems. For example, it is not at all obvious that a protein structure is an important feature a priory. Another example, is debugging code. A successful debugging session involves asking and answering a succession of relevant questions. Imagine giving a code to someone with no programming experience and asking them to debug it. The probabilities of them asking the right questions is very close to 0.
Identifying what is important is called inserting inductive Biases. In theory LLMs could integrate the inductive biases of a field and generate interesting questions, even format datasets from open-source data. However until this ability has been fully demonstrated, the only cost efficient way to accelerate AI driven scientific discoveries is to build the disciplinarily into the people: AI Researchers that know enough about the field to be able to identify the relevant questions of the future.
Partages
In the wake of two Nobel Prizes AI is Still resting on Giant Shoulders
copier:
https://bluwr.com/p/39050236
