LM
Uploaded byLoic Merckel
285 views

Introduction to Generative Artificial Intelligence

The buzz around Generative AI (GenAI) is louder than ever—but are we seeing the whole picture? This presentation was designed for a broad audience, avoiding technical jargon while addressing the real opportunities and challenges AI brings. 🔍 A few key insights from the presentation: 💡 Innovation under constraint: DeepSeek achieved remarkable results at a fraction of competitors' costs. 🔄 The productivity paradox: AI may boost output for writers and coders but may reduce job satisfaction for scientists. 🤖 The shifting definition: "AI is what computers can't do; once they can, it's just software." — Mustafa Suleyman ⚡ The GPU Bottleneck: Big Tech is turning to custom chips and nuclear energy to meet AI's soaring computational demands. ♻️ Energy-Hungry AI: AI's energy needs are driving investments in dedicated nuclear power for datacenters. Slide decks are meant to support live talks, so they might not capture the full story on their own. If you are curious to dive deeper, feel free to reach out—I would be happy to discuss these ideas in a more interactive setting. Feel free to check out the slides here and my LinkedIn post for further discussion: https://www.linkedin.com/posts/merckel_intro-to-genai-activity-7300095492862930946-rnL9

Embed presentation

Intro to GenAI Loic Merckel Frankfurt (DE), 02/2025
1966: ELIZA “While ELIZA was capable of engaging in discourse, it could not converse with true understanding. However, many early users were convinced of ELIZA's intelligence and understanding, despite Weizenbaum's insistence to the contrary.” en.wikipedia.org/wiki/ELIZA.
nytimes.com/2011/10/26/science/26mccarthy.html AI is "making a machine behave in ways that would be called intelligent if a human were so behaving." — McCarthy et al. (1955) McCarthy, J., Minsky, M. L., Rochester, N., & Shannon, C. E. (1955). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence
AI are "all those unfinished computer developments brimming with fantasies and capable of generating buzz." — Attributed to Cédric Villani (Fields Medalist 2010) "l'IA, ce sont tous les développements informatiques non encore aboutis qui sont porteurs de fantasmes et permettent de faire le buzz" linkedin.com/posts/merckel_lintelligence-artificielle-va-t-elle-sauver-activity-7269447993265733632-pTPx Cédric Villani at his office 2015, ©Marie-Lan Nguyen / CC BY 3.0
AI is "a system’s ability to interpret external data correctly, to learn from such data,and to use those learnings to achieve specific goals and tasks through flexible adaptation." — Haenlein & Kaplan (2019) Haenlein, M., & Kaplan, A. (2019). A Brief History of Artificial Intelligence: On the Past, Present, and Future of Artificial Intelligence. California Management Review, 61(4), 5-14.
AI are "software (and possibly also hardware) systems designed by humans that, given a complex goal, act in the physical or digital dimension by perceiving their environment through data acquisition, interpreting the collected structured or unstructured data, reasoning on the knowledge, or processing the information, derived from this data and deciding the best action(s) to take to achieve the given goal. AI systems can either use symbolic rules or learn a numeric model, and they can also adapt their behaviour by analysing how the environment is affected by their previous actions."
"AI is as those of us building it like to joke—'what computers can’t do.' Once they can,it’s just software." — Mustafa Suleyman (2023) Suleyman, M. (2023). The coming wave: technology, power, and the twenty-first century's greatest dilemma. Audible. ASIN B0C78KDTPM
AI is "the frontier of computational advancements that references human intelligence in addressing ever more complex decision- making problems." — Berente et al. (2021) Berente, N., Gu, B., Recker, J., & Santhanam, R. (2021). Managing artificial intelligence. MIS quarterly, 45(3).
Generative Models: Statistical Modeling vs. Deep Learning A class of statistical models that learn the underlying probability distribution of data. – Can generate new data samples similar to the training data (e.g., creating realistic images, synthesizing text). – Contrast with discriminative models, which learn decision boundaries (e.g., classifying emails as spam or not). – Traditional Statistical Modeling (SM) Explicit Distributional Assumption: Starts by assuming a specific probability distribution (e.g., Normal, Poisson, etc.). – Parameter Estimation: Uses data to estimate the parameters of the chosen distribution (e.g., mean, standard deviation). – Generative Capability: Once parameters are estimated, the model can generate new data by sampling from the assumed distribution. – Example: Gaussian Mixture Model (GMM). – Deep Learning (DL) Generative Models Implicit Distribution Learning: Learns the data distribution implicitly through the network architecture and training process. – Representation Learning: Focuses on learning complex, high-dimensional representations of the data. – Generative Capability: Uses the learned representation to generate new data samples. – Example: Transformer-based Large LLMs, GAN. –
Transformers: ML, but Statistically Grounded Google introduced the Transformer, which rapidly became the state-of-the-art approach to solve most NLP problems. Vaswani (2017), Attention Is All You Need (doi.org/10.48550/arXiv.1706.03762) Transformers (e.g., The 'T' of ChatGPT) are machine learning techniques: They do not make explicit distributional assumptions like GMMs. – But: They implicitly learn a statistical model of the data through attention and their layered architecture (i.e., an approximation of the probability of the data). – Generative models can be built using both SM (explicit probability distributions) and DL (learned representations). –
The Euphoria for "GenAI" “ChatGPT, the popular chatbot from OpenAI, is estimated to have reached 100 million monthly active users in January, just two months after launch, making it the fastest-growing consumer application in history” — Reuters, Feb 1, 2023 (reut.rs/3yQNlGo) ©ESA/CNES/ARIANESPACE/Activité Photo Optique Video CSG - 2008
One Common Categorization of AI en.wikipedia.org/wiki/Artificial_general_intelligence Weak AI ("narrow AI") is focused on one narrow task: self-driving cars, vocal assistants (e.g., Siri or Alexa), AlphaGo, ... – Strong AI ("artificial general intelligence" or AGI) is at par with human cognitive capabilities across a wide range of cognitive tasks. – Created by Winston Duke as part of the Visualising AI project launched by Google DeepMind. unsplash.com
"It is LeCun who said AlphaGo was impossible just days before it made its first big breakthrough." — Mustafa Suleyman (2023, ch. 8) Suleyman, M. (2023). The coming wave: technology, power, and the twenty-first century's greatest dilemma. Audible. ASIN B0C78KDTPM
Big Names Forecasts McKinsey (2023): "AI could increase corporate profits by $4.4 trillion a year, according to new research." – Goldman Sachs (2023): "Generative AI could raise global GDP by 7%." – mckinsey.com/mgi/overview/in-the-news/ai-could-increase-corporate-profits-by-4- trillion-a-year-according-to-new-research – goldmansachs.com/intelligence/pages/generative-ai-could-raise-global-gdp-by-7- percent.html –
Forecasts by Big Names Are Not Always Reliable AT&T started to investigate “Mobile Telephony” in 1980. – McKinsey projected then that the size of mobile phone market in 2000 should be < 1 Million subscribers. – It turned out that the size was > 120 Million, and several billion today... – Cutting the cord, economist.com/special-report/1999/10/07/cutting-the-cord – Statistics, itu.int/en/ITU-D/Statistics/Pages/stat/default.aspx – Danny Ralph & Marc Jansen, Lecture Slides 2015, Management Science, Judge Business School, The University of Cambridge, UK –
Agência Brasi | CC-BY-3.0-BR | commons.wikimedia.org/wiki/File:Oil_platform_P-51_(Brazil).jpg Remember When Data Was Called the New Oil?
The major bottleneck in the further development of AI is not necessarily in the availability of data, but in the computational power of AI infrastructure and in companies' ability to attract top-level talents. — Cant et al. (2024, ch. 3)1 Cant et al. (2024). Feeding the Machine: The Hidden Human Labor Powering AI. Audible. ASIN B0D9YSH6WT. 1.
GPUs Demand Skyrockets How to Build the Ultimate GPU Cloud to Power AI | Odd Lots (Jul 20, 2023). youtube.com/watch? v=9OOn6u6GIqk&t=1308s Before LLMs, GPUs were primarily needed for training, and CPUs were used for inference. However, with the emergence of LLMs, GPUs have become almost essential for both tasks. – Paraphrasing Brannin McBee, co-founder of CoreWeave, in Bloomberg Podcast: While you may train the model using 10,000 GPUs, the real challenge arises when you need 1 million GPUs to meet the entire inference demand. This surge in demand is expected during the initial one to two years after the launch, and it's likely to keep growing thereafter. –
Big Tech’s Rush on GPUs Estimated Purchases of Nvidia Hopper GPUs (2024) Shift Toward Custom AI Chips FT (2024). Microsoft acquires twice as many Nvidia AI chips as tech rivals. on.ft.com/3XwDTVJ Microsoft: 485,000 – Meta: 224,000 – Amazon: 196,000 – Google: 169,000 – Google: 1.5m TPUs – Meta: 1.5m Meta Training & Inference Accelerator chips – Amazon: 1.3m Trainium & Inferentia chips – Microsoft: 200,000 Maia chips – NVIDIA H100 - computerhistory.org/collections/catalog/102801460
Energy-Hungry Datacenters Are Getting Their Own Nuclear Plants Big Tech’s Nuclear Shift (Datacenters) BLOOM, trained on Jean Zay, is considered one of the cleanest LLMs Companies like Google and Microsoft/OpenAI are turning to dedicated nuclear power for their AI-driven datacenters. – AI workloads demand enormous energy, requiring a stable, clean, and high-output power source. – Nuclear energy provides: – Reliability (unlike wind/solar, which depend on weather). – Low carbon emissions (cleaner than fossil fuels like coal/gas). – High energy density (small footprint, large power output). – Jean Zay is a supercomputer, not a standard datacenter. – Jean Zay benefits from France’s nuclear-heavy grid (~70%), but does not have a dedicated nuclear plant. –
European Commission (11/02/2025): EU launches InvestAI initiative to mobilise €200 billion of investment in artificial intelligence – WSJ (21/01/2025): Tech Leaders Pledge Up to $500 Billion in AI Investment in U.S. – Barron's (09/02/2025): Macron Says France To Receive 109 Bn Euros Of AI Investment In 'Coming Years' –
Oops! Did someone turn off the music?
Under $6 Million? Covers only the final pre-training, context extension, and post-training runs (for V3, not R1, which training cost is unknown). – Excludes data acquisition, hyperparameter tuning, and R&D costs. – Anyone replicating this model from scratch would likely pay much more. – Yet, an impressive achievement when considering the other similar models reported costs (e.g., possibly an order of magnitude lower than Llama 3). – DeepSeek-AI et al. (2024). DeepSeek-V3 Technical Report. arxiv.org/abs/2412.19437. – Grattafiori et al. (2024). The llama 3 herd of models. arxiv.org/abs/2407.21783. – Gibney, E. (2025). China’s cheap, open AI model DeepSeek thrills scientists. Nature 638, 13-14. doi.org/10.1038/d41586-025-00229-6 –
Innovation Under Constraint? Dave Snowden (2006) argued that innovation thrives on "starvation of familiar resources," – while Eric Schmidt (2015, loc. 3179) observed that "when you want to spur innovation, the worst thing you can do is overfund it." – Schmidt cites Frank Lloyd Wright's observation that "the human race built most nobly when limitations were greatest." – The release of DeepSeek V3/R1—labeled AI's "Sputnik moment" (FT, 2025)—challenges assumptions about the need for billion-dollar investments in AI infrastructure. – Could this be a case that validates these long-standing observations about innovation under constraint? – Snowden, D. (2006). Culture and Innovation. thecynefin.co/culture-and-innovation. – Schmidt , E. (2015). How Google Works. Hodder & Stoughton. Kindle Edition. – FT (2025). Tech stocks slump as China’s DeepSeek stokes fears over AI spending. The Financial Times. on.ft.com/4gQtxGK – on.ft.com/3X5sdJ4
on.ft.com/4b8AD8f $6M? Not Without Help? DeepSeek-R1 was trained using reinforcement learning (RL), but RLHF is not explicitly confirmed. – RLHF is expensive—DeepSeek may have used OpenAI's API instead of human feedback. – DeepSeek denies wrongdoing, but OpenAI is investigating potential misuse. – (A growing share of the web is now AI-generated —so may training on the web leave some flavor of model distillation?) – DeepSeek-AI et al. (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arxiv.org/abs/2501.12948. – Bloomberg (2025). Microsoft Probes if DeepSeek-Linked Group Obtained OpenAI Data. bloomberg.com/news/articles/2025-01-29/microsoft-probing-if-deepseek-linked-group-improperly-obtained- openai-data. – Thompson et al. (2024). A Shocking Amount of the Web is Machine Translated: Insights from Multi-Way Parallelism. arxiv.org/abs/2401.05749. –
Will OpenAI & Microsoft Actually Do Anything? OpenAI & Microsoft benefited from outsourced cheap RLHF labor under alleged dubious conditions, raising ethical questions—can they call out DeepSeek without backlash? – Legal action would raise questions about OpenAI's own data practices. – Microsoft has business interests in China—would they really risk them over this? – Admitting model distillation works could encourage more competitors to do the same. – (Microsoft's Bing was accused of copying Google search results—calling out DeepSeek too fast for model distillation would be ironic.) –
Now What?
Demis Hassabis - britannica.com/biography/Demis-Hassabis AGI at the Corner? Are we Doomed? "You can find people on both sides of the argument, very eminent people, just take Jeff Hinton versus Yann LeCun. Both are Turing Award winners. I know them both very well, Yoshua Bengio, these are some of the top people who originally were in the field. The fact that they can be completely in opposite camps, to me suggests that actually we don't know. With this transformative technology, so transformative, it's unknown. So I don't think anyone can precisely, I think it's kind of a nonsense to precisely put a probability on it. What I do know is it's non-zero, that risk, right?" — Demis Hassabis (2024) Hard Fork: Google DeepMind C.E.O. Demis Hassabis on the Path From Chatbots to A.G.I., Feb 23, 2024. podcasts.apple.com/de/podcast/hard-fork/id1528594034?i=1000646586260&r=2389
Ray Kurzweil: Toward an Abundance Era Cade Metz (2024). Ray Kurzweil Still Says He Will Merge With AI. The New York Times. nytimes.com/2024/07/04/technology/ray-kurzweil-singularity.html Predicts the arrival of the Singularity by 2045: A merging of humans and AI that will enhance human intelligence. – Advocates for Universal Basic Income (UBI), as automation could eliminate most jobs—though he sees this outcome as positive for society. – Believes technology will create an abundance era by eliminating scarcity through exponential advancements in AI, nanotechnology, and biotechnology. – nrkbeta - Ray Kurzweil @ SXSW 2017 - CC BY-SA 2.0
"Can we forge a narrow path between catastrophe and dystopia?" (Mustafa Suleyman, 2023) Catastrophe: Dystopia: Uncontrolled AI development could lead to unintended consequences, destabilizing societies. – Risks include economic collapse, job displacement, and existential threats from autonomous systems. – Authoritarian regimes may exploit AI for mass surveillance and control. – Advanced technologies could entrench inequality, limiting freedom and privacy. –
Global Job Projections: Wishful Thinking? "In the next five years,170 million jobs are projected to be created and 92 million displaced ...a net [gain of] 78 million jobs." — Future of Jobs Report (2025, p. 18)
The Future of Jobs Report relies heavily on a survey of large companies' executives. This presents several critical limitations. Uncertainty and the "Unknown Unknowns" of AI Biases and the "Ackoff Problem" Rapid Technological Change: The pace of AI development makes accurate mid- to long-term prediction extremely challenging. – Limited Expertise: Survey respondents are not necessarily AI experts; they are making projections based on current (and potentially flawed) understanding. – "Unknown Unknowns": The most disruptive impacts may be those we cannot currently foresee. – Executive Bias: The survey captures the perspective of large-company executives, not workers or smaller businesses. – Self-Reported Data: Responses may be subject to biases, organizational pressures, and the desire to appear forward-thinking. – Ritualistic Planning: The report may reflect aspirations rather than realistic assessments, echoing Ackoff's critique of corporate planning: "...like a ritual rain dance..." –
If you ask which jobs are safe, my best bet about a job that's safe is Plumbing— Geoffrey Hinton (2024)1 Financial Wise (2024). Godfather of AI Geoffrey Hinton Believes Only Plumbing Jobs Will Survive in the Short-Term. youtu.be/uqXUDFi-BsU. 1.
Impact for Writers and Coders Increase productivity and quality for tasks such as writing and software development. Generative AI tools benefit professionals at all levels but provide the greatest boost for less experienced workers. These tools not only improve output but also enhance job satisfaction and confidence in the workplace. For writers, tools like ChatGPT increase productivity by reducing time spent on tasks by 37% and improving output quality by 0.4 standard deviations, as rated by professional evaluators based on writing, content, and originality (Noy & Zhang, 2023). – For coders, AI coding assistants increase completed tasks by 26%, with the largest gains observed among less experienced developers (Cui et al., 2025). – Cui et al. (2025). The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. dx.doi.org/10.2139/ssrn.4945566 – Noy, S. and Zhang, W. (2023). Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. dx.doi.org/10.2139/ssrn.4375283 –
Scientific Research: New Materials Discovery Toner-Rodgers, A. (2024). Artificial intelligence, scientific discovery, and product innovation. arxiv.org/abs/2412.17866 "AI-assisted researchers discover 44% more materials, resulting in a 39% increase in patent filings and a 17% rise in downstream product innovation." – These gains are concentrated among top scientists, whose output nearly doubles due to their ability to efficiently leverage AI suggestions. – However, 82% of researchers report reduced job satisfaction due to decreased creativity and underutilization of skills. –
AI in Business: Efficiency vs.Profitability AI investments boost efficiency—but profits remain elusive. Businesses are shifting focus from efficiency gains to revenue generation as the primary measure of AI’s return on investment (ROI). – Despite an estimated $38.8 billion in AI spending for 2024, many companies have yet to see substantial financial returns. – While AI tools improve productivity (e.g., a 20% boost for software developers reported by Goldman Sachs), translating these gains into direct profits remains a challenge. – Tech companies integrating AI into their products see faster returns, but for other industries, measurable revenue impacts may take years to materialize. – Key challenge: High costs and complex implementation processes often outweigh short-term financial benefits. – WSJ (2025). It’s Time for AI to Start Making Money for Businesses. Can It?. wsj.com/articles/its-time-for-ai-to-start-making-money-for-businesses-can-it-b476c754 – WSJ (2024). Goldman Sachs Deploys Its First Generative AI Tool Across the Firm. wsj.com/articles/goldman-sachs-deploys-its-first-generative-ai-tool-across-the-firm-cd94369b –
Potential Downsides Cyber Warfare & Autonomous Weapons – Weapons with Systemic Risk (Bioweapons & Beyond) – Market Instability & Economic Risks – Job Displacement – Bias and Discrimination – Privacy Concerns –
Potential Upsides Scientific Discovery (e.g., AlphaFold) – Healthcare Improvements – Automation of Dangerous and Repetitive Tasks – Enhanced Accessibility – Environmental Sustainability – Economic Growth –
It will require an awesome effort to fundamentally change our societies, human instincts, and the patterns of history. It’s far from certain. It looks impossible. But meeting the great dilemma of the twenty-first century must be possible.—Mustafa Suleyman (2023)1 Suleyman, M. (2023). The coming wave: technology, power, and the twenty-first century's greatest dilemma. Audible. ASIN B0C78KDTPM 1.
So long, and thanks for all the fish... Or any questions?

More Related Content

PPTX
Generative AI Deep Dive: Architectures, Mechanics, and Future Applications
bygenz9514
 
PPTX
Artificial Intelligence
byPrakhyath Rai
 
PDF
Artificial Intelligence in Business Management
byDharun Krishna K B
 
PDF
Introduction to Artificial Intelligence.pdf
byfizzamscs1105
 
PPTX
unleshing the the Power Azure Open AI - MCT Summit middle east 2024 Riyhad.pptx
byUsama Wahab Khan Cloud, Data and AI
 
PDF
AI Lect 2 Identifying AI systems, branches of AI, etc.pdf
bymswindow00
 
PPTX
Module 3_AI.pptxaaaaaaaaaaaaaaaaaaaaaaaa
bypokah34509
 
PDF
Exploring AI as tools in your career.pdf
byvideongamesrfun
 
Generative AI Deep Dive: Architectures, Mechanics, and Future Applications
bygenz9514
 
Artificial Intelligence
byPrakhyath Rai
 
Artificial Intelligence in Business Management
byDharun Krishna K B
 
Introduction to Artificial Intelligence.pdf
byfizzamscs1105
 
unleshing the the Power Azure Open AI - MCT Summit middle east 2024 Riyhad.pptx
byUsama Wahab Khan Cloud, Data and AI
 
AI Lect 2 Identifying AI systems, branches of AI, etc.pdf
bymswindow00
 
Module 3_AI.pptxaaaaaaaaaaaaaaaaaaaaaaaa
bypokah34509
 
Exploring AI as tools in your career.pdf
byvideongamesrfun
 

Similar to Introduction to Generative Artificial Intelligence

PDF
Technologies Demystified: Artificial Intelligence
byPioneers.io
 
PPTX
Artificial Intelligence
bydikshendrasarpate2
 
PPTX
Lesson 1 intro to ai
byankit_ppt
 
PDF
SparkLabs Primer on Artificial Intelligence 2025
bySparkLabs Group
 
PDF
AI and Healthcare: An Overview (January 2024)
byKimberley Barker
 
PDF
SparkLabs Primer on Artificial Intelligence 2025
bySparkLabs Group
 
PPTX
Artificial Intelligence in Emerging Technology
bySitanathBiswas
 
PPTX
Basics and History of Machine Learning.pptx
byAnuragBarthwal4
 
PDF
AI Evolution Beyond Humans _The Age of Machine Superiority.pdf
byImpaakt Magazine
 
PDF
AI Evolution Beyond Humans _The Age of Machine Superiority.pdf
byImpaakt Magazine
 
PPTX
Machine learning help in engineering .pptx
byAMITKUMAR86671
 
PPTX
Artificial intelligence ( AI ) | Guide
byTo Sum It Up
 
PPTX
ai and smart assistant using machine learning and deep learning
byanujapawar1950
 
PPTX
artificial intelligenc AI types of AI ppt
byButterfly education
 
PPTX
AI_Introduction.pptx for gaining knowledge about ai
byhasishhasish
 
DOCX
one hour presentation content for introduction to....docx
bykannakalamonline
 
PPTX
chapter-1-1.pptxmsdsdvsdvnsdvndsvdvds i.
byAkhileshPansare
 
PPTX
AI 21-22 chpt 1 .pptx
byDrChhayaPawar1
 
PDF
Artificial Intelligence.pdf nhbhbuhuuhjuj
byarpitdhagate08
 
PPTX
AI_DGV.pptx
byMANIGANDANKRISHNAN1
 
Technologies Demystified: Artificial Intelligence
byPioneers.io
 
Artificial Intelligence
bydikshendrasarpate2
 
Lesson 1 intro to ai
byankit_ppt
 
SparkLabs Primer on Artificial Intelligence 2025
bySparkLabs Group
 
AI and Healthcare: An Overview (January 2024)
byKimberley Barker
 
SparkLabs Primer on Artificial Intelligence 2025
bySparkLabs Group
 
Artificial Intelligence in Emerging Technology
bySitanathBiswas
 
Basics and History of Machine Learning.pptx
byAnuragBarthwal4
 
AI Evolution Beyond Humans _The Age of Machine Superiority.pdf
byImpaakt Magazine
 
AI Evolution Beyond Humans _The Age of Machine Superiority.pdf
byImpaakt Magazine
 
Machine learning help in engineering .pptx
byAMITKUMAR86671
 
Artificial intelligence ( AI ) | Guide
byTo Sum It Up
 
ai and smart assistant using machine learning and deep learning
byanujapawar1950
 
artificial intelligenc AI types of AI ppt
byButterfly education
 
AI_Introduction.pptx for gaining knowledge about ai
byhasishhasish
 
one hour presentation content for introduction to....docx
bykannakalamonline
 
chapter-1-1.pptxmsdsdvsdvnsdvndsvdvds i.
byAkhileshPansare
 
AI 21-22 chpt 1 .pptx
byDrChhayaPawar1
 
Artificial Intelligence.pdf nhbhbuhuuhjuj
byarpitdhagate08
 
AI_DGV.pptx
byMANIGANDANKRISHNAN1
 

More from Loic Merckel

PDF
Intro to LLMs
byLoic Merckel
 
PDF
Introduction to LLMs
byLoic Merckel
 
PDF
Generative Models and ChatGPT
byLoic Merckel
 
PDF
An Introduction to AI (Formerly Data Science)
byLoic Merckel
 
PDF
Agile Infinity: When the Customer Is an Abstract Concept
byLoic Merckel
 
PDF
iQHorse—Towards Beating the Market
byLoic Merckel
 
PDF
Ideation: First Step Towards Innovation
byLoic Merckel
 
PDF
Are Decisions From a Single Point Wise?
byLoic Merckel
 
Intro to LLMs
byLoic Merckel
 
Introduction to LLMs
byLoic Merckel
 
Generative Models and ChatGPT
byLoic Merckel
 
An Introduction to AI (Formerly Data Science)
byLoic Merckel
 
Agile Infinity: When the Customer Is an Abstract Concept
byLoic Merckel
 
iQHorse—Towards Beating the Market
byLoic Merckel
 
Ideation: First Step Towards Innovation
byLoic Merckel
 
Are Decisions From a Single Point Wise?
byLoic Merckel
 

Recently uploaded

PDF
20251202 Frank_Melis Vittal_Prabhu Platform_Talk v3b.pdf
byhome
 
PPTX
[DSC Europe 25] Andy Cotgreave - Nothing is new in analytics.pptx
byDataScienceConferenc1
 
PPTX
[DSC Europe 25] Dusan Jovicic - AI Story: From on-prem to cloud and back agai...
byDataScienceConferenc1
 
PDF
[DSC Europe 25] Marija Vlajkovic & Andrea Radonjanin - Integration of AI tool...
byDataScienceConferenc1
 
PDF
The Best Data Pipeline Tools in 2025: Automating Your Data Stack
bySkyvia
 
PDF
[DSC Europe 25] Nikola Rajovic - Hardware Technologies Under the Hood: RISC-V...
byDataScienceConferenc1
 
PDF
Data+Security+Masterclass Presentation.pdf
byckurt9676
 
PDF
AI Business Analyst — Transforming Real Estate Data Into Insight
byLeni Analyst
 
PPTX
[DSC Europe 25] Selena Milanovic - Educational AI in Sales.pptx
byDataScienceConferenc1
 
PPTX
L34 Applications and Limitations of Unsupervised Learning.pptx
byadityasaxena378335
 
PPTX
[DSC Europe 25] Jim Sterne - Adopting Generative AI Capabilities Into the Ent...
byDataScienceConferenc1
 
PPTX
[DSC Europe 25] Bogdan Daniel Maruneac - AI - It starts with you.pptx
byDataScienceConferenc1
 
PPTX
CC-105-Semi-Final-Topicdsgsdvsdgsdb.pptx
bykhajezjhozhua131
 
PPT
Database Concept with SQL Commands for grade 12
byVinod Srivastava
 
PPTX
284761009-Statistics-Quiz-Questions.pptx
bymarckevinsbarrida
 
PPTX
Presentation 1-PUM Singapore 2015- 2D.pptx
byWilliamChong71
 
PDF
OPPOTUS - Malaysians on Malaysia (MOM) Q3 2025
byOppotus
 
PDF
7-IntroToRecursionprgramminginc++course.pdf
byFantaneshTegegn1
 
PPTX
CHAPTER-2hdueidijxjxixidididis%20HTML.pptx
byssuser273bea
 
PDF
District cooling graphs 2024 Finland.pdf
byEnergiateollisuus ry - Finnish Energy Industries
 
20251202 Frank_Melis Vittal_Prabhu Platform_Talk v3b.pdf
byhome
 
[DSC Europe 25] Andy Cotgreave - Nothing is new in analytics.pptx
byDataScienceConferenc1
 
[DSC Europe 25] Dusan Jovicic - AI Story: From on-prem to cloud and back agai...
byDataScienceConferenc1
 
[DSC Europe 25] Marija Vlajkovic & Andrea Radonjanin - Integration of AI tool...
byDataScienceConferenc1
 
The Best Data Pipeline Tools in 2025: Automating Your Data Stack
bySkyvia
 
[DSC Europe 25] Nikola Rajovic - Hardware Technologies Under the Hood: RISC-V...
byDataScienceConferenc1
 
Data+Security+Masterclass Presentation.pdf
byckurt9676
 
AI Business Analyst — Transforming Real Estate Data Into Insight
byLeni Analyst
 
[DSC Europe 25] Selena Milanovic - Educational AI in Sales.pptx
byDataScienceConferenc1
 
L34 Applications and Limitations of Unsupervised Learning.pptx
byadityasaxena378335
 
[DSC Europe 25] Jim Sterne - Adopting Generative AI Capabilities Into the Ent...
byDataScienceConferenc1
 
[DSC Europe 25] Bogdan Daniel Maruneac - AI - It starts with you.pptx
byDataScienceConferenc1
 
CC-105-Semi-Final-Topicdsgsdvsdgsdb.pptx
bykhajezjhozhua131
 
Database Concept with SQL Commands for grade 12
byVinod Srivastava
 
284761009-Statistics-Quiz-Questions.pptx
bymarckevinsbarrida
 
Presentation 1-PUM Singapore 2015- 2D.pptx
byWilliamChong71
 
OPPOTUS - Malaysians on Malaysia (MOM) Q3 2025
byOppotus
 
7-IntroToRecursionprgramminginc++course.pdf
byFantaneshTegegn1
 
CHAPTER-2hdueidijxjxixidididis%20HTML.pptx
byssuser273bea
 
District cooling graphs 2024 Finland.pdf
byEnergiateollisuus ry - Finnish Energy Industries
 

Introduction to Generative Artificial Intelligence

  • 1.
    Intro to GenAI LoicMerckel Frankfurt (DE), 02/2025
  • 2.
    1966: ELIZA “While ELIZAwas capable of engaging in discourse, it could not converse with true understanding. However, many early users were convinced of ELIZA's intelligence and understanding, despite Weizenbaum's insistence to the contrary.” en.wikipedia.org/wiki/ELIZA.
  • 3.
    nytimes.com/2011/10/26/science/26mccarthy.html AI is "makinga machine behave in ways that would be called intelligent if a human were so behaving." — McCarthy et al. (1955) McCarthy, J., Minsky, M. L., Rochester, N., & Shannon, C. E. (1955). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence
  • 4.
    AI are "allthose unfinished computer developments brimming with fantasies and capable of generating buzz." — Attributed to Cédric Villani (Fields Medalist 2010) "l'IA, ce sont tous les développements informatiques non encore aboutis qui sont porteurs de fantasmes et permettent de faire le buzz" linkedin.com/posts/merckel_lintelligence-artificielle-va-t-elle-sauver-activity-7269447993265733632-pTPx Cédric Villani at his office 2015, ©Marie-Lan Nguyen / CC BY 3.0
  • 5.
    AI is "asystem’s ability to interpret external data correctly, to learn from such data,and to use those learnings to achieve specific goals and tasks through flexible adaptation." — Haenlein & Kaplan (2019) Haenlein, M., & Kaplan, A. (2019). A Brief History of Artificial Intelligence: On the Past, Present, and Future of Artificial Intelligence. California Management Review, 61(4), 5-14.
  • 6.
    AI are "software(and possibly also hardware) systems designed by humans that, given a complex goal, act in the physical or digital dimension by perceiving their environment through data acquisition, interpreting the collected structured or unstructured data, reasoning on the knowledge, or processing the information, derived from this data and deciding the best action(s) to take to achieve the given goal. AI systems can either use symbolic rules or learn a numeric model, and they can also adapt their behaviour by analysing how the environment is affected by their previous actions."
  • 7.
    "AI is asthose of us building it like to joke—'what computers can’t do.' Once they can,it’s just software." — Mustafa Suleyman (2023) Suleyman, M. (2023). The coming wave: technology, power, and the twenty-first century's greatest dilemma. Audible. ASIN B0C78KDTPM
  • 8.
    AI is "thefrontier of computational advancements that references human intelligence in addressing ever more complex decision- making problems." — Berente et al. (2021) Berente, N., Gu, B., Recker, J., & Santhanam, R. (2021). Managing artificial intelligence. MIS quarterly, 45(3).
  • 9.
    Generative Models: Statistical Modelingvs. Deep Learning A class of statistical models that learn the underlying probability distribution of data. – Can generate new data samples similar to the training data (e.g., creating realistic images, synthesizing text). – Contrast with discriminative models, which learn decision boundaries (e.g., classifying emails as spam or not). – Traditional Statistical Modeling (SM) Explicit Distributional Assumption: Starts by assuming a specific probability distribution (e.g., Normal, Poisson, etc.). – Parameter Estimation: Uses data to estimate the parameters of the chosen distribution (e.g., mean, standard deviation). – Generative Capability: Once parameters are estimated, the model can generate new data by sampling from the assumed distribution. – Example: Gaussian Mixture Model (GMM). – Deep Learning (DL) Generative Models Implicit Distribution Learning: Learns the data distribution implicitly through the network architecture and training process. – Representation Learning: Focuses on learning complex, high-dimensional representations of the data. – Generative Capability: Uses the learned representation to generate new data samples. – Example: Transformer-based Large LLMs, GAN. –
  • 10.
    Transformers: ML, but StatisticallyGrounded Google introduced the Transformer, which rapidly became the state-of-the-art approach to solve most NLP problems. Vaswani (2017), Attention Is All You Need (doi.org/10.48550/arXiv.1706.03762) Transformers (e.g., The 'T' of ChatGPT) are machine learning techniques: They do not make explicit distributional assumptions like GMMs. – But: They implicitly learn a statistical model of the data through attention and their layered architecture (i.e., an approximation of the probability of the data). – Generative models can be built using both SM (explicit probability distributions) and DL (learned representations). –
  • 11.
    The Euphoria for "GenAI" “ChatGPT,the popular chatbot from OpenAI, is estimated to have reached 100 million monthly active users in January, just two months after launch, making it the fastest-growing consumer application in history” — Reuters, Feb 1, 2023 (reut.rs/3yQNlGo) ©ESA/CNES/ARIANESPACE/Activité Photo Optique Video CSG - 2008
  • 12.
    One Common Categorization of AI en.wikipedia.org/wiki/Artificial_general_intelligence WeakAI ("narrow AI") is focused on one narrow task: self-driving cars, vocal assistants (e.g., Siri or Alexa), AlphaGo, ... – Strong AI ("artificial general intelligence" or AGI) is at par with human cognitive capabilities across a wide range of cognitive tasks. – Created by Winston Duke as part of the Visualising AI project launched by Google DeepMind. unsplash.com
  • 15.
    "It is LeCunwho said AlphaGo was impossible just days before it made its first big breakthrough." — Mustafa Suleyman (2023, ch. 8) Suleyman, M. (2023). The coming wave: technology, power, and the twenty-first century's greatest dilemma. Audible. ASIN B0C78KDTPM
  • 16.
    Big Names Forecasts McKinsey (2023):"AI could increase corporate profits by $4.4 trillion a year, according to new research." – Goldman Sachs (2023): "Generative AI could raise global GDP by 7%." – mckinsey.com/mgi/overview/in-the-news/ai-could-increase-corporate-profits-by-4- trillion-a-year-according-to-new-research – goldmansachs.com/intelligence/pages/generative-ai-could-raise-global-gdp-by-7- percent.html –
  • 17.
    Forecasts by Big NamesAre Not Always Reliable AT&T started to investigate “Mobile Telephony” in 1980. – McKinsey projected then that the size of mobile phone market in 2000 should be < 1 Million subscribers. – It turned out that the size was > 120 Million, and several billion today... – Cutting the cord, economist.com/special-report/1999/10/07/cutting-the-cord – Statistics, itu.int/en/ITU-D/Statistics/Pages/stat/default.aspx – Danny Ralph & Marc Jansen, Lecture Slides 2015, Management Science, Judge Business School, The University of Cambridge, UK –
  • 18.
    Agência Brasi |CC-BY-3.0-BR | commons.wikimedia.org/wiki/File:Oil_platform_P-51_(Brazil).jpg Remember When Data Was Called the New Oil?
  • 19.
    The major bottleneckin the further development of AI is not necessarily in the availability of data, but in the computational power of AI infrastructure and in companies' ability to attract top-level talents. — Cant et al. (2024, ch. 3)1 Cant et al. (2024). Feeding the Machine: The Hidden Human Labor Powering AI. Audible. ASIN B0D9YSH6WT. 1.
  • 20.
    GPUs Demand Skyrockets How toBuild the Ultimate GPU Cloud to Power AI | Odd Lots (Jul 20, 2023). youtube.com/watch? v=9OOn6u6GIqk&t=1308s Before LLMs, GPUs were primarily needed for training, and CPUs were used for inference. However, with the emergence of LLMs, GPUs have become almost essential for both tasks. – Paraphrasing Brannin McBee, co-founder of CoreWeave, in Bloomberg Podcast: While you may train the model using 10,000 GPUs, the real challenge arises when you need 1 million GPUs to meet the entire inference demand. This surge in demand is expected during the initial one to two years after the launch, and it's likely to keep growing thereafter. –
  • 21.
    Big Tech’s Rushon GPUs Estimated Purchases of Nvidia Hopper GPUs (2024) Shift Toward Custom AI Chips FT (2024). Microsoft acquires twice as many Nvidia AI chips as tech rivals. on.ft.com/3XwDTVJ Microsoft: 485,000 – Meta: 224,000 – Amazon: 196,000 – Google: 169,000 – Google: 1.5m TPUs – Meta: 1.5m Meta Training & Inference Accelerator chips – Amazon: 1.3m Trainium & Inferentia chips – Microsoft: 200,000 Maia chips – NVIDIA H100 - computerhistory.org/collections/catalog/102801460
  • 22.
    Energy-Hungry Datacenters Are Getting TheirOwn Nuclear Plants Big Tech’s Nuclear Shift (Datacenters) BLOOM, trained on Jean Zay, is considered one of the cleanest LLMs Companies like Google and Microsoft/OpenAI are turning to dedicated nuclear power for their AI-driven datacenters. – AI workloads demand enormous energy, requiring a stable, clean, and high-output power source. – Nuclear energy provides: – Reliability (unlike wind/solar, which depend on weather). – Low carbon emissions (cleaner than fossil fuels like coal/gas). – High energy density (small footprint, large power output). – Jean Zay is a supercomputer, not a standard datacenter. – Jean Zay benefits from France’s nuclear-heavy grid (~70%), but does not have a dedicated nuclear plant. –
  • 23.
    European Commission (11/02/2025): EUlaunches InvestAI initiative to mobilise €200 billion of investment in artificial intelligence – WSJ (21/01/2025): Tech Leaders Pledge Up to $500 Billion in AI Investment in U.S. – Barron's (09/02/2025): Macron Says France To Receive 109 Bn Euros Of AI Investment In 'Coming Years' –
  • 24.
  • 25.
    Under $6 Million? Coversonly the final pre-training, context extension, and post-training runs (for V3, not R1, which training cost is unknown). – Excludes data acquisition, hyperparameter tuning, and R&D costs. – Anyone replicating this model from scratch would likely pay much more. – Yet, an impressive achievement when considering the other similar models reported costs (e.g., possibly an order of magnitude lower than Llama 3). – DeepSeek-AI et al. (2024). DeepSeek-V3 Technical Report. arxiv.org/abs/2412.19437. – Grattafiori et al. (2024). The llama 3 herd of models. arxiv.org/abs/2407.21783. – Gibney, E. (2025). China’s cheap, open AI model DeepSeek thrills scientists. Nature 638, 13-14. doi.org/10.1038/d41586-025-00229-6 –
  • 26.
    Innovation Under Constraint? Dave Snowden(2006) argued that innovation thrives on "starvation of familiar resources," – while Eric Schmidt (2015, loc. 3179) observed that "when you want to spur innovation, the worst thing you can do is overfund it." – Schmidt cites Frank Lloyd Wright's observation that "the human race built most nobly when limitations were greatest." – The release of DeepSeek V3/R1—labeled AI's "Sputnik moment" (FT, 2025)—challenges assumptions about the need for billion-dollar investments in AI infrastructure. – Could this be a case that validates these long-standing observations about innovation under constraint? – Snowden, D. (2006). Culture and Innovation. thecynefin.co/culture-and-innovation. – Schmidt , E. (2015). How Google Works. Hodder & Stoughton. Kindle Edition. – FT (2025). Tech stocks slump as China’s DeepSeek stokes fears over AI spending. The Financial Times. on.ft.com/4gQtxGK – on.ft.com/3X5sdJ4
  • 27.
    on.ft.com/4b8AD8f $6M? Not Without Help? DeepSeek-R1was trained using reinforcement learning (RL), but RLHF is not explicitly confirmed. – RLHF is expensive—DeepSeek may have used OpenAI's API instead of human feedback. – DeepSeek denies wrongdoing, but OpenAI is investigating potential misuse. – (A growing share of the web is now AI-generated —so may training on the web leave some flavor of model distillation?) – DeepSeek-AI et al. (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arxiv.org/abs/2501.12948. – Bloomberg (2025). Microsoft Probes if DeepSeek-Linked Group Obtained OpenAI Data. bloomberg.com/news/articles/2025-01-29/microsoft-probing-if-deepseek-linked-group-improperly-obtained- openai-data. – Thompson et al. (2024). A Shocking Amount of the Web is Machine Translated: Insights from Multi-Way Parallelism. arxiv.org/abs/2401.05749. –
  • 28.
    Will OpenAI & MicrosoftActually Do Anything? OpenAI & Microsoft benefited from outsourced cheap RLHF labor under alleged dubious conditions, raising ethical questions—can they call out DeepSeek without backlash? – Legal action would raise questions about OpenAI's own data practices. – Microsoft has business interests in China—would they really risk them over this? – Admitting model distillation works could encourage more competitors to do the same. – (Microsoft's Bing was accused of copying Google search results—calling out DeepSeek too fast for model distillation would be ironic.) –
  • 29.
  • 30.
    Demis Hassabis -britannica.com/biography/Demis-Hassabis AGI at the Corner? Are we Doomed? "You can find people on both sides of the argument, very eminent people, just take Jeff Hinton versus Yann LeCun. Both are Turing Award winners. I know them both very well, Yoshua Bengio, these are some of the top people who originally were in the field. The fact that they can be completely in opposite camps, to me suggests that actually we don't know. With this transformative technology, so transformative, it's unknown. So I don't think anyone can precisely, I think it's kind of a nonsense to precisely put a probability on it. What I do know is it's non-zero, that risk, right?" — Demis Hassabis (2024) Hard Fork: Google DeepMind C.E.O. Demis Hassabis on the Path From Chatbots to A.G.I., Feb 23, 2024. podcasts.apple.com/de/podcast/hard-fork/id1528594034?i=1000646586260&r=2389
  • 31.
    Ray Kurzweil: Toward an AbundanceEra Cade Metz (2024). Ray Kurzweil Still Says He Will Merge With AI. The New York Times. nytimes.com/2024/07/04/technology/ray-kurzweil-singularity.html Predicts the arrival of the Singularity by 2045: A merging of humans and AI that will enhance human intelligence. – Advocates for Universal Basic Income (UBI), as automation could eliminate most jobs—though he sees this outcome as positive for society. – Believes technology will create an abundance era by eliminating scarcity through exponential advancements in AI, nanotechnology, and biotechnology. – nrkbeta - Ray Kurzweil @ SXSW 2017 - CC BY-SA 2.0
  • 32.
    "Can we forgea narrow path between catastrophe and dystopia?" (Mustafa Suleyman, 2023) Catastrophe: Dystopia: Uncontrolled AI development could lead to unintended consequences, destabilizing societies. – Risks include economic collapse, job displacement, and existential threats from autonomous systems. – Authoritarian regimes may exploit AI for mass surveillance and control. – Advanced technologies could entrench inequality, limiting freedom and privacy. –
  • 33.
    Global Job Projections:Wishful Thinking? "In the next five years,170 million jobs are projected to be created and 92 million displaced ...a net [gain of] 78 million jobs." — Future of Jobs Report (2025, p. 18)
  • 34.
    The Future ofJobs Report relies heavily on a survey of large companies' executives. This presents several critical limitations. Uncertainty and the "Unknown Unknowns" of AI Biases and the "Ackoff Problem" Rapid Technological Change: The pace of AI development makes accurate mid- to long-term prediction extremely challenging. – Limited Expertise: Survey respondents are not necessarily AI experts; they are making projections based on current (and potentially flawed) understanding. – "Unknown Unknowns": The most disruptive impacts may be those we cannot currently foresee. – Executive Bias: The survey captures the perspective of large-company executives, not workers or smaller businesses. – Self-Reported Data: Responses may be subject to biases, organizational pressures, and the desire to appear forward-thinking. – Ritualistic Planning: The report may reflect aspirations rather than realistic assessments, echoing Ackoff's critique of corporate planning: "...like a ritual rain dance..." –
  • 35.
    If you askwhich jobs are safe, my best bet about a job that's safe is Plumbing— Geoffrey Hinton (2024)1 Financial Wise (2024). Godfather of AI Geoffrey Hinton Believes Only Plumbing Jobs Will Survive in the Short-Term. youtu.be/uqXUDFi-BsU. 1.
  • 36.
    Impact for Writersand Coders Increase productivity and quality for tasks such as writing and software development. Generative AI tools benefit professionals at all levels but provide the greatest boost for less experienced workers. These tools not only improve output but also enhance job satisfaction and confidence in the workplace. For writers, tools like ChatGPT increase productivity by reducing time spent on tasks by 37% and improving output quality by 0.4 standard deviations, as rated by professional evaluators based on writing, content, and originality (Noy & Zhang, 2023). – For coders, AI coding assistants increase completed tasks by 26%, with the largest gains observed among less experienced developers (Cui et al., 2025). – Cui et al. (2025). The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. dx.doi.org/10.2139/ssrn.4945566 – Noy, S. and Zhang, W. (2023). Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. dx.doi.org/10.2139/ssrn.4375283 –
  • 37.
    Scientific Research: New Materials Discovery Toner-Rodgers,A. (2024). Artificial intelligence, scientific discovery, and product innovation. arxiv.org/abs/2412.17866 "AI-assisted researchers discover 44% more materials, resulting in a 39% increase in patent filings and a 17% rise in downstream product innovation." – These gains are concentrated among top scientists, whose output nearly doubles due to their ability to efficiently leverage AI suggestions. – However, 82% of researchers report reduced job satisfaction due to decreased creativity and underutilization of skills. –
  • 38.
    AI in Business:Efficiency vs.Profitability AI investments boost efficiency—but profits remain elusive. Businesses are shifting focus from efficiency gains to revenue generation as the primary measure of AI’s return on investment (ROI). – Despite an estimated $38.8 billion in AI spending for 2024, many companies have yet to see substantial financial returns. – While AI tools improve productivity (e.g., a 20% boost for software developers reported by Goldman Sachs), translating these gains into direct profits remains a challenge. – Tech companies integrating AI into their products see faster returns, but for other industries, measurable revenue impacts may take years to materialize. – Key challenge: High costs and complex implementation processes often outweigh short-term financial benefits. – WSJ (2025). It’s Time for AI to Start Making Money for Businesses. Can It?. wsj.com/articles/its-time-for-ai-to-start-making-money-for-businesses-can-it-b476c754 – WSJ (2024). Goldman Sachs Deploys Its First Generative AI Tool Across the Firm. wsj.com/articles/goldman-sachs-deploys-its-first-generative-ai-tool-across-the-firm-cd94369b –
  • 39.
    Potential Downsides Cyber Warfare &Autonomous Weapons – Weapons with Systemic Risk (Bioweapons & Beyond) – Market Instability & Economic Risks – Job Displacement – Bias and Discrimination – Privacy Concerns –
  • 40.
    Potential Upsides Scientific Discovery (e.g., AlphaFold) – HealthcareImprovements – Automation of Dangerous and Repetitive Tasks – Enhanced Accessibility – Environmental Sustainability – Economic Growth –
  • 41.
    It will requirean awesome effort to fundamentally change our societies, human instincts, and the patterns of history. It’s far from certain. It looks impossible. But meeting the great dilemma of the twenty-first century must be possible.—Mustafa Suleyman (2023)1 Suleyman, M. (2023). The coming wave: technology, power, and the twenty-first century's greatest dilemma. Audible. ASIN B0C78KDTPM 1.
  • 42.
    So long, andthanks for all the fish... Or any questions?