About
The modern enterprise is a data-driven business. Generative AI is responsible for analyzing and interpreting the data, but the journey goes much further – to autonomous units where artificial intelligence increasingly takes control and decision-making.
Discussion
One year ago, during the CIOMove conference in Zurich, AI was not as relevant of a discussion nor talked about as commonly as it is today. In this year’s conference in Porto, AI and especially Chat GPT dominate the discussion. Things have changed quite a lot.
What has happened in the last 12 months with AI?
– There has been continual and steady progress in the AI space, however this year, public perception of AI has heightened.
– In enterprise IT, disruptions tend to happen when certain technologies hit the public.
- This is an important factor to be considered.
Where do we currently stand with Generative AI?
– A functioning and ready to use Generative AI model is years away, as is complete autonomous driving Level 5
- This type of technology tries to predict and assume the next thing to come. This may not always be a good thing, as seen in the case of autonomous driving.
– Generative AI is a tool such as Excel and we will have to find solutions and ways to apply it in the best manner once we start utilizing it in our everyday work.
- We will need to learn the limits of Generative AI so and define processes on how to use it.
- The most important thing is we need to get it right and ensure the regulation of AI. The CEO of Chat GPT agreed to this as well.
– Why is it a race against complexity?
- The complexity of traffic is so large with autonomous cars that we are going back to Level 3 from Level 5
- Controlled complexity is a drive on a straight road from point A to point B
- However, driving in New York City for example would be too complex and therefore presents the issue we are facing with complexity.
– There also exists narrow domains and specific use cases for these AI solutions to advance a business, rather than reliance on Generative AI
- For example: At Wendy’s fast-food restaurant, the individual ordering will speak to an AI model, and it then will process your order.
- Wendy’s will train the model with their specific business language to make this happen.
- We want to start with IP and data on a domain model and not wait for a complete generative model because that is years out.
– Pan is the model working behind the scenes that allow AI technology to become more specialized.
- there will be a medical or finance version of Pan. A company will take the finance model and enrich it with its own IP and company-specific data.
What should regulations on AI look like?
– Regulating AI becomes both necessary and potentially risky.
– Large-scale regulation does not refer to the regulation of “criminals” because those individuals will find ways around regulations regardless.
- We need to determine how to regulate and standardize AI for everyone.
– When it comes to political or spiritual beliefs, it is difficult to know how to regulate them.
- Who is able to delete certain content? Who censors what is wrong or right?
What should AI Authenticity look like? How do we have proof of authenticity?
– To “clarify” the rules of the game, AI-generated content needs to be flagged and shown that it is an AI image.
- We are currently working on creating a watermark to understand to do just this. This is something that should be coming very soon.
- The goal is for the watermark to become so commonly used and expected to be used, that providers of AI have to adhere, Make it a global standard. The watermark will not remove “fake” or false information from the internet, rather making the user aware of the content they are consuming.
- For example, football records are already generated by AI, individuals are okay with this as long as they have the knowledge and therefore a choice to consume.
– Technicalities of an AI watermark:
- Any document with a certain set of metadata or picture can be digitally signed to indicate this is generative AI created content.
- The developer conference is available on YouTube for additional information
– Is it always possible from a tech perspective, to determine if it is AI generated or not?
- Usually yes, and today there is software on the market that can tell you which is AI, and which is not.
- However, it is not 100% possible, but if usage becomes broad, then the mere absence of proven authenticity speaks for itself.
– Currently, Google and Microsoft are companies producing those models to show AI transparency, but what happens when the everyday person starts generating AI?
- This is already commonplace, such as adhering to HTML.
- Technologies will also be developed that can sift through and find the AI watermark.
- This currently occurs within YouTube for example.
– New AI models will eventually be using the content of the internet that has already been generated by other AI models. Are we in jeopardy of everything becoming AI and cannot trust it?
- Watermarks will help a little bit but not solve the general problem.
- There will inherently be bias in the AI space since it input base dataset is largely biased to a degree.
AI Education
– As a society, we will need to find ways to not lose basic skills and qualities such as geography that can be lost when depending heavily on AI
– We need to adjust our education expectations and understand how to utilize AI in the best way.
- This would be the job of not just schools and technology companies, but your organization and even taught within families.
– There is no return from where we are, things are rapidly developing. So, it is our responsibility to get things done the right way and stay on top of AI education.
Job Security and AI
– Instead of worrying about AI making an individual obsolescent in their job, understand that it will transform the way you do your job.
– You can leverage the building of the neuro network for a model and have an algorithm that processes a lot of information, for the tasks required in your job.
– The benefit of AI is it never gets tired nor makes mistakes on repetitive tasks and lacks fatigue and increased error rate.
- The jobs in which these human error rates could occur due to repetition over time are the ones that will benefit the most from AI usage.
- For example, a data processing role will benefit greatly from the incorporation of AI tools.
- It is predicted that ~80% of code can be created from AI but will still require human revision.
– Examples of AI usage in different professions:
- Medicine: AI can be used to make a suggestion or catch an error; however medical decisions still are done by the physician. AI is used as a support tool.
- Lawyers: only lawyers can give legal advice. So, AI cannot legally give legal advice, rather again can be used as a support tool.
– Overall, it is very difficult to predict how AI will affect jobs until its implementation.
- Then we learn the benefits and dangers and evolve from there, it is all a learning experience.
How can CIOs implement AI?
– How can a CIO use these technologies to automate and synthesize and accelerate tasks?
– For example: tracking the efficiency of a car is an intense simulation and is very expensive.
- There is an AI model that will determine the efficiency of the model with 95% accuracy.
– It would be beneficial to use AI to create content that you know is so appealing to a specific target audience to enhance marketing to drive up revenue. Do not think we are at a place where we can use technology to do this?
- For example, AI has been given input data to create something unique and successful.
- AI was given 30 pictures of shoes and then trained to build proposals for a shoe as though it was from the 70s. It created shoes that fit the mold exactly.
- A marketing test was done with one ad created by Chat GPT and one created by a marketing expert. The chat GPT one had a 50% more click rate.
– But how do you know how much you will sell?
- With AI, you can tell how often a shoe is being sold, sold for, how much is in stock in the supply chain, etc. This is currently being utilized.
– Who is responsible for the outcome of AI usage? Would the CIO be held responsible?
- We must first ensure that individuals are aware of the content generated by algorithms and AI.
- if an autonomous vehicle knocks someone off of the sidewalk, this is very bad press, and you question who is liable in this scenario.
- These are unsolved questions.
- To avoid issues with the outcome of AI, we need to create sets of rules similar to that of regulated industries such as banking.
- Ideally, if you feed the same input, you want to see the same output every time.
- This needs to be the expected norm to avoid issues with the outcome of AI.
- The pharmaceutical industry has been doing this for a long time now.
Who own’s AI generated data?
– AI models that are trained using the company’s data should have no issues with copywriting.
- When you fine-tune the model, it becomes the customer’s intellectual property.
- After that, you feed it with data from your own infrastructure.
– For example, Med-Palm in the medical domain, you train with your company IP data, which is then never fed back into the central library for security purposes.
– Med-Pam has 500 billion parameters; however, every parameter has a compute cost.
- If you choose to trim down the neuro network and complexity of the AI, you need less computing power and bring therefore down costs.
– Information that hits cyber networks can see a lot of patterns and correlations within the network. Are companies allowed to use this information?
- Even today, reinforced AI algorithms will give you a hint which is gathered from identifying patterns and correlations within the data it has access to.
AI Advantages Expected in 5-10 years.
– In individuals’ private lives, AI will help a lot too, such as GPS systems.
– AI can act as a “stand-in” for meeting for a meeting or virtual event and return with a summary, enhancing the efficiency of time.
– Ai will be able to read an official document or sift through the transcript of a video conference and then present a thorough summary.
– The importance of data quality and collection is increasing.
- Improving data quality is difficult.
- Knowledge about external and traffic data is also important.
– Skip L and emberscript
- Emberscript: transcribe conversation from one language to another and can do so in ~ 5-10 minutes.
– In many years AI may be better than we are. Reminds me of when machines started making clothes and individuals did not want clothes made by machines.
Three Interesting Theses
- Companies will not benefit from AI if they cannot manage or control their data. Those who do not have a data strategy today will be left behind by competitors.
Agree: 29
Disagree: 2 - A centralized multiparty generalize AI model will lead to data loss. It will be important to manage isolated instances of AI to prevent IP data leakage.
Agree: 27
Disagree: 2 - The union of OT And IT is the starting point for an autonomous manufacturing future where generative AI can control processes without human intervention.
Agree: 19
Disagree: 4.5