How to grow neural networks in petri dish | Stephen Smith posted on the topic | LinkedIn (2024)

Excited to announce the first issue of the Journal of Business and Artificial Intelligence! Check out my article on AI Accelerators. Six case studies! Go here: JBAI.ai

9

Evolving computer programs with large language modelsFour Interesting Things1.One can evolve computer programs by simulating biological techniques like survival of the fittest, mutation, and crossover. This is called genetic programming and has been around since the 1990s.2.By looking at differences in code in GitHub, an LLM can be trained to learn which types of code modifications result in bug fixes and improved code. This LLM can then do a much better job of ‘mutating’ to fix bugs compared to random mutation – even for code it has never seen before.3.Two problems are used as tests for the improved code evolution. The Sodarace problem (evolve weights, springs, and oscillation properties to evolve digital walking robots) and the 4-parity task (determine if the sum of the number of 1’s in a four-digit binary number is even or odd). The LLM based mutation with simulated evolution is better than random mutation but often not as good as just using an LLM directly.4.Various techniques were utilized to keep the evolved population from converging. The idea of ‘open-endedness’ for improved final results is promoted and utilized.Four years from now I expect the concept of ‘open-endedness’ to continue to be important when using simulated evolution on really hard problems. It is not clear however whether simulated evolution will be a superior technique for problem solving compared to just using an LLM. The idea of the evolution and adaptation of the parameters of evolution (meta evolution) is a very cool idea and may be an area to look for the next leap forward in our understanding of evolution. Read this: https://lnkd.in/eWHKhsQy

4

3 Comments

Blocking credit card fraud at VISAFour Interesting Things1.Card not present (CNP) credit card fraud is increasing as more purchases are completed online where the credit card number, expiration date, and CVV are entered digitally.2.VISA is using generative AI to create the Visa Account Attack Intelligence (VAAI) score. The model that creates the score utilizes 182 risk attributes (six times more features than previous models) to respond within 20 milliseconds of receipt of a transaction.3.The model was built with 15 billion VisaNet transactions and is showing an ability to reduce false positives by 85% (i.e. fewer legit transactions will be flagged as fraud, resulting in fewer legit customers becoming annoyed).4.They embed the VAAI score into the transaction message and allow the client to create their own rules, based on the client’s risk tolerance, to decide whether to accept or decline the transaction.Four years from now expect generative AI to be horrifyingly good at simulating human interactions and luring people into fraudulent online relationships in order to capture important personal information. Read this: https://lnkd.in/emvx_tEM

2

1 Comment

Regulating AIFour Interesting Things1.Long Term Planning Agents (LTPAs) can plan over a long time horizon and present a novel existential threat for humanity. LTPAs are a form of AI that can be put in charge of real things in the real world that matter. For instance: having an AI independently book a vacation for you using your credit card, or optimize brown outs on a power grid during a summer heat wave. 2.You may not be able to test a smart AI agent to see if it could act independently and against human wishes. If the AI is sufficiently capable it could detect that it is being tested and either pretend to be compliant or, if it feels threatened, seek to harm the human tester. This is analogous to a king wishing to test whether a rogue general is loyal. If the general is not loyal, they will pretend to be loyal or, if they feel threatened by the test, initiate a coup against the king.3.AIs can hide their intentions. Some AI agents have already been found to pause misbehavior when the AI detects that a safety test is being initiated.4.The only way to control an out-of-control AI may be to control the resources used to create the models. Similar to the way we monitor uranium in order to control nuclear weapon proliferation, we might need to control GPUs or other computer resources necessary for running and building these models. Of course, this only works as long as the AI requires specialized or expensive computer resources.Four years from now we will invariably look back and see mistakes from AIs that cause human harm (we already have seen self-driving cars make mistakes that cost human lives). But we will also see harm done by AIs whose goals are not correctly aligned with a human’s (e.g. a renegade drone from the military that takes out civilians that it thinks are hostile to itself or its mission). Though AI is scary, I think the real problem is broader, it is the complexity of these systems and that humans no longer know how they work or how to control them. I don’t think regulation will solve the problem; we will just have to be very careful about putting automated systems in charge of things that matter.Read this: “Regulating advanced artificial agents” https://lnkd.in/dBQhtqJ3

6

1 Comment

I’m honored and excited to have been named to the Editorial Board of the Journal of Business and Artificial Intelligence. Our focus is on publishing real-world case studies showing practical applications of AI to solve specific business challenges across all industries, including manufacturing, medicine, logistics, technology, retail, agriculture...you name it. If you or someone you know has used AI successfully (or unsuccessfully) to solve a real, practical problem in their business, case studies can be submitted here: https://lnkd.in/ekX_iAzK

This content isn’t available here

Access this content and more in the LinkedIn app

64

32 Comments

Just finished the course “GPT-4: The New GPT Release and What You Need to Know”! #gpt4 #generativeai

This content isn’t available here

Access this content and more in the LinkedIn app

5

1 Comment

Worth a read on Sam Altman's investment in life extension and fusion. These are 'hard startups' - i guess OpenAI is easy... ;-0

5

How do you motivate students? New paper reviewing incentive systems for educational apps: https://lnkd.in/eaths6sJ

2

Cool idea ... Teachers suffer from 'toxic positivity' - maybe your company as well...?

5

1 Comment