AI for Renaissance Humans Beings

The presentation explores artificial intelligence (AI), covering its definition, current capabilities, limitations, and potential impact.

Defining Intelligence (Artificial and Otherwise)
The presentation defines intelligence as adjusting parameters of a mental model, minimizing errors, exploring possibilities, restricting search space, optimising a reward function, and projecting a priori hypotheses. It also suggests that humans, and even the universe itself, could be considered forms of AI.

Categorising AI
Several categories of AI are mentioned, including: heuristic, symbolic, perception, machine learning, reinforcement learning, and generative computing. These categories likely represent a progression in complexity and capability.

Current State of AI
The presentation highlights both advancements and limitations of current AI. Progress is being measured by various benchmarks and evaluations like Massive Multitask Language Understanding, General Language Understanding Evaluation (GLUE), AI Index Annual Report 2024, and ARC-AGI.

While AI excels in certain tasks, it does not outperform humans in all areas. The presentation notes the increasing investment in AI, particularly in generative AI, and its positive impact on worker productivity and quality of work. It also shows that the US leads in top AI model development. However, there's a significant lack of robust and standardised evaluations for responsible use of Large Language Models (LLMs).
Limitations and Areas for Improvement

Several limitations of current AI are highlighted. These include:
- Sensitivity: AI algorithms are highly susceptible to minor variations in input data, making them vulnerable to adversarial attacks.
- Bias: Algorithms can reflect biases present in training data or the cultural context of developers, posing challenges for applications with moral implications.
- Retention: Algorithms struggle to "store" history, which is problematic for long-term time series data.
- Common Sense and Justification: AI lacks common sense reasoning and struggles to justify its decision-making processes, hindering its use in scenarios with moral implications.
- Risk Analysis: Current models provide average accuracy levels, necessitating better risk assessment for sensitive applications.

Temporal Model Drifting: The presentation also touches upon the concept of temporal model drifting and the need for further exploration. It questions whether we have reached the full potential of AI or even know where development is headed.

Industry Trends
The sources indicate several industry trends:
- Increased prioritisation and budget allocation towards AI by enterprises.
- Growth in AI use cases across various sectors, including automation, logistics, supply chain, customer experience, and insights.
- The role of IT partners in accelerating AI adoption and reducing costs.
- Challenges in scaling AI, with increased development expenditure and deployment times.
- The slow maturation of AI, with many enterprises still in early stages of development and deployment.
- Difficulties in evaluating AI effectiveness due to multiple and fragmented success metrics.
- Concerns about software supply chain and configuration management impacting AI quality.
- Significant issues with model performance, security, and auditing, impacting many enterprises.

Potential and Risks
The presentation acknowledges the transformative potential of AI to revolutionise industries, solve complex problems, and improve lives.

However, it also emphasises the need for careful regulation, ethical guidelines, and safeguards to mitigate these risks is underscored

Transcript

1. Proprietary + Confidential Madrid, November 2024 AI for Renaissance Humans Beings Andrés-Leonardo Martínez-Ortiz a.k.a almo By JPxG - DALL-E 3, Public Domain, https://commons.wikimedia.org/w/index.php?curid=144161107
2. About me… Andrés-Leonardo Martínez-Ortiz a.k.a almo, holds a PhD on Software, Systems and Computing and a Master on Computer Science. Based on Zurich, almo is a member of the Vertex AI SRE Site Reliability Engineering team, leading several programs aiming for reliability, efficiency & convergence. Main editor of the Cathartic Computing Club newsletter and member of IEEE, ACM, Linux Foundation and Computer Society. @davilagrau almo.dev almo Do you want join the Cathartic Computer Club? Sign up here 👉
3. Disclaimer
4. Yes, I did this presentation using AI
5. Photo by Ben Sweet on Unsplash Actually… I am an AI*! You all are AI! and the universe might be an hologram! * AI stands for Augmented Intelligence
6. AI Metaphysics
7. Intelligence Adjusting the parameters of a mental model Minimizing errors Exploring the space of possibilities Restricting the search space Optimizing a reward function Projecting a priori hypotheses Exploiting the Combinatorial Explosion (multi-level, multimodal)
8. Artificial Incomplete Formal Systems Quantum Mechanics Runaway solution of Dirac Equation
9. Artificial Intelligence Heuristic computing Symbolic computing Perception computing Machine learning computing Reinforcement Learning computing Generative computing
11. AI Psychoanalysis
12. 👉 https://ai.google.dev/gemini-api/prompts But also you can find free prompting libraries @ Hugging Face or @ Anthropic https://docs.anthropic.com/en/prompt-library/library
13. ML Commons https://mlcommons.org/benchmarks EFF Measuring the Progress of AI Research (2017) https://www.eff.org/ai/metrics Massive Multitask Language Understanding https://github.com/hendrycks/test General Language Understanding Evaluation (GLUE) https://gluebenchmark.com AI Index Annual Report 2024 https://aiindex.stanford.edu/report
14. 👉 ARC-AGI measures progress towards artificial general intelligence using a private evaluation dataset https://arcprize.org
15. AI Dialectic By Jakob Schlesinger - anagoria, Public Domain, https://commons.wikimedia.org/w/index.php?curid=33762804
16. Increasing prioritization and budget ˜80% of the enterprises have AI as priority ˜83% of the enterprises increase the AI budget 49% process automation 45% managing logistics 43% supply chain optimization 57% customer experience 50% customer insights 48% customer interaction Increasing number of use cases Increasing saving and adoption speed thanks to specialized partnership IT partners allow savings up to 30% of the time IT partners demand ˜25% less of internal IT resources IT partners allow savings up to 20% of the budget Scale decreasing efficiency ˜40% of the enterprises expend +50% in development Increasing deployment time (+64% MoM) Maturity long curve 55% of the enterprises on evaluation or early stage of development and deployment Multiple & fragmentated success metrics make difficult the evaluation Software supply chain and configuration management diminish AI quality 36% of the enterprises have serious problem with model performance 56% of the enterprises have serious problem the security and auditing models 67% of the enterprises have to be complaint to several quality standards.
17. Thesis Transformative potential of AI to revolutionize industries, solve complex problems, and improve human lives. Ability to automate tasks, analyze data, and generate creative solutions, leading to increased efficiency, innovation, and economic growth. Disruptive Technology Exponential Technology
18. Antithesis Potential dangers, such as job displacement, algorithmic bias, privacy violations, and the potential for autonomous weapons systems to escalate conflicts. Peed for careful regulation, ethical guidelines, and safeguards to prevent unintended consequences and ensure AI benefits all of humanity
19. Improvement Areas Sensitivity Current algorithms exhibit high sensitivity to variations in input data. Adversarial attacks are capable of disrupting AI solutions by introducing noise that is often imperceptible. Bias Current algorithms often exhibit significant biases, stemming from the cultural context of the development teams or the training data. These are intrinsic biases that are not always easy to identify and mitigate. This makes their application difficult in scenarios with moral implications. Retention Current algorithms respond to training data without the ability to "store" history. This effect is catastrophic in time series data that extends over long periods. Common Sense Current algorithms are incapable of using common sense. Reaching a decision based on globally available information and common knowledge accessible to people is still a challenge yet to be solved. Justifiable Current algorithms do not allow for adequate justification of the reasoning they follow to reach their solutions. In this sense, they behave like black boxes, making their application difficult in scenarios with moral implications. Risk Analysis Currently, models provide average levels of accuracy. For scenarios with moral implications, risk or accuracy assessment mechanisms are necessary to allow for human intervention. This functionality is not available in most cases.
20. Temporal Model Drifting
21. How does the temporal ML drifting look like? Exotic patterns: chaos and periodic Vela, D., Sharp, A., Zhang, R. et al. Temporal quality degradation in AI models. Sci Rep 12, 11654 (2022).
22. Data (including contrafactual data)
23. Synthesis Are we there yet?
24. #1 AI beats humans on some tasks, but not on all. #2 Industry continues to dominate frontier AI research #3 Frontier models get way more expensive #4 The United States leads China, the EU, and the U.K. as the leading source of top AI models. #5 Robust and standardized evaluations for LLM responsibility are seriously lacking. #6 Generative AI investment skyrockets. #7 The data is in: AI makes workers more productive and leads to higher quality work. #8 Scientific progress accelerates even further, thanks to AI. #9 The number of AI regulations in the United States sharply increases. #10 People across the globe are more cognizant of AI’s potential impact—and more nervous. AI Index Annual Report 2024 https://aiindex.stanford.edu/report
30. AI Index Annual Report 2024 https://aiindex.stanford.edu/report
31. And thus, we are not there yet… maybe we do not even know where we're going yet…
32. Future Design & Use Cases https://worldbuild.ai
33. δυστόπος
34. δυστόπος
35. Resources DeepMind, Technologies https://deepmind.google/technologies Hugging Face https://huggingface.co Institute for Human-Centered AI https://hai.stanford.edu https://aiindex.stanford.edu Artificial Intelligence, Our World In Data https://ourworldindata.org/artificial-intelligence arXiv Artificial Intelligence https://arxiv.org/list/cs.AI/recent The Batch https://www.deeplearning.ai/the-batch Do you want join the Cathartic Computer Club? Sign up here 👉
36. Thank you!