Question 1

How do biases in narrow AI systems create real-world harm?

Accepted Answer

_AI systems enforce human prejudices at massive scale._

Narrow AI amplifies existing biases at scale. In parole decisions, computers were meant to remove human bias, but when fed biased data, they kept more minorities in jail. Loan application systems with few women in training data couldn't correctly handle female applicants. The stop and frisk program in New York used a computer to select precincts, but it was racially unjust. When you remove skin color from the data, the system became even more skewed. There's a logical data science explanation for this.

— [Bias in, bias out: Addressing the impact of bias in AI](https://www.lutzfinger.com/transcript/2023-06-07-bias-in-bias-out-addressing-the-impact-of-bias-in-ai/) · _The Edge_

Question 2

How do you spot social media bots?

Accepted Answer

_Detection heuristics._

> Figure 6-4. Subset of a social network, created using LinkedIn Labs Since bot networks will first create their own group by linking to one another before they start to infiltrate the rest of the network, especially in the beginning, you will see a highly dense network with very low cohesion to the rest of the network. However, those measures using network topology only work in the beginning as the bot network is formed. Once the bots…

— [Finger Dutta Ask Measure Learn (1)](https://www.lutzfinger.com/book/) · _Book — Ask, Measure, Learn_

Question 3

How does Catalog prevent large language models from hallucinating when retrieving enterprise information?

Accepted Answer

_Stopping hallucinations by feeding LLMs the right information first._

The way to avoid hallucination is to stop the model from reaching into the model layers for information. If you give it structured or unstructured information in raw format and then ask it to summarize or list, it gives you exactly what you provided. Without that context, it makes things up because it's doing probabilistic retrieval from its training. This is retrieval augmented generation. You get the information first, augment the prompt with it, then generate the response based on that specific context.

— [Large Language Models: one interface to rule them all?](https://www.lutzfinger.com/transcript/2023-08-15-large-language-models-one-interface-to-rule-them-all/) · _The Edge_

Question 4

Is AGI near?

Accepted Answer

_Lutz's stance on the AGI timeline._

> Both have been hot topics in AI, sparking both fear and excitement. Sam Altman predicted AGI would arrive in 2025, and Elon Musk 2026. Both are more about hype than reality. In 2025, there will be no AGI but large language models will find their "killer app." What are AGI and the Singularity?

— [No AGI But A 'Killer App' — 2025 AI Prediction](https://www.forbes.com/sites/lutzfinger/2025/01/03/no-agi-but-a-killer-app2025-ai-prediction-110/) · _Forbes_

Question 5

Is OpenAI going bankrupt?

Accepted Answer

_Lutz's take on OpenAI economics and moats._

> OpenAI — "Experts predict bankruptcy" — was attention-grabbing, a deeper look reveals that the company has simply seen a drop in its users. Should you fear ChatGPT will disappear? No, quite the opposite.

— [OpenAI Isn't Going Bankrupt, But It Has A Business Model Problem](https://www.forbes.com/sites/lutzfinger/2023/08/18/is-openai-going-bankrupt-no-but-ai-models-dont-create-moats/) · _Forbes_

Question 6

Should you trust OpenAI blindly?

Accepted Answer

_Vendor risk and architectural diversification._

> Stargate's $500B investment — message to the competitors — meant to intimidate potential challengers with overwhelming financial backing. AI is extremely dangerous; only a select few should be allowed to develop it — OpenAI's storyline to use safety concerns as justification for a regulatory setup that creates barriers to entry and gives them more monopolistic power. From now on, models should respect IP rights — OpenAI ignored IP…

— [Is OpenAI A Victim — DeepSeek And IP Rights — A Playbook For Leaders](https://www.forbes.com/sites/lutzfinger/2025/01/29/is-openai-a-victim---deepseek-and-ip-rights---a-playbook-for-leaders/) · _Forbes_

Question 7

What about deepfakes?

Accepted Answer

_How society will adapt to synthetic media._

> As AI lowers the barriers to content creation, abuse is also becoming more feasible. Deepfakes — AI-generated images, videos and audio designed to mimic real people — pose serious risks. Already in 2022, I demonstrated how easily they can be created. This increases the risk of misinformation, identity theft and political manipulation.

— [Goku, A Chinese Open-Source Model, Challenges OpenAI And The U.S.](https://www.forbes.com/sites/lutzfinger/2025/02/11/goku---a-chinese-open-source-model-challenges-openai-and-the-us/) · _Forbes_

Question 8

What are the fundamental limits of large language models that people should understand?

Accepted Answer

_Language models face limits in reasoning, planning, and true understanding._

First, language is only part of human knowledge. We know more than we can describe. A four year old has had way more sensory impressions than any large language model, despite not being able to read. Second, LLMs can't truly reason or explain their decisions because they're built on neural networks, which work like gut feel rather than rule based systems. They predict the next word based on probability distributions, not logical reasoning. Third, there's the data quality problem. And finally, LLMs are fundamentally probabilistic, so they can't reliably predict specific individual choices or plan multi step processes independently.

— [A reality check on AI](https://www.lutzfinger.com/transcript/2024-07-10-a-reality-check-on-ai/) · _The Edge_

Question 9

What are the main limitations of large language models that knowledge graphs help address?

Accepted Answer

_LLMs are powerful but flawed. Knowledge graphs provide the missing structure._

Large language models have several undesired behaviors. They hallucinate, generating false information. They need guardrails to prevent answering certain prompts inappropriately. They have gaps in their encoded knowledge and sometimes fail to retrieve information even when it exists, often due to how prompts are formulated. The prompt itself can cause the model to deprioritize relevant information. Knowledge graphs help with all these issues by adding structure and control. They make models more reliable and accurate at every step, from training to retrieval.

— [The synergy between LLMs and knowledge graphs](https://www.lutzfinger.com/transcript/2023-08-24-the-synergy-between-llms-and-knowledge-graphs/) · _The Edge_

Question 10

What framework helps assess whether imperfect AI decisions are acceptable?

Accepted Answer

_Match your AI quality threshold to the stakes of the decision._

The framework asks: How good is your decision quality, and how important is the decision? How bad is a false decision? We discussed two opposite examples at Station F. One company did predictive maintenance for infrastructure where errors are costly, requiring high quality predictions. Another did social media outreach for hiring where low quality matters less since you already have low close rates. You can mitigate both. For low-stakes, automate with feedback loops. For high-stakes infrastructure, use augmentation first, guiding engineers and creating Jira tickets.

— [Building AI products: 5 lessons from our founders' workshop](https://www.lutzfinger.com/transcript/2024-08-01-building-ai-products-5-lessons-from-our-founders-workshop/) · _The Edge_

Question 11

What is the hallucination problem with large language models and why does it matter for businesses?

Accepted Answer

_LLMs invent answers when they don't know, risking business credibility_

Large language models will make things up when they don't know the answer. I demonstrated this by asking ChatGPT to summarize a completely fake Cornell website that doesn't exist. It generated a summary saying my course was the most amazing and made very good logical arguments, but it was completely fake. For businesses, this is a huge problem. If a customer asks a shop for something it doesn't have, the model might just make it up and promise red shoes with blinking laces. That's why proprietary data matters. You need to ground the model on your local data.

— [AI as an interface](https://www.lutzfinger.com/transcript/2023-04-24-ai-as-an-interface/) · _The Edge_

Question 12

What is your perspective on AGI and when it might happen?

Accepted Answer

_AGI remains limited by data and lacks human agency for decisions._

AGI means artificial general intelligence, models that learn and develop capabilities by themselves, becoming vastly more intelligent than humans. But I disagree with this focus. AGI still has the data problem and only one type of prediction modeling, while humans have emotional prediction and other ways of understanding. Crucially, we humans have agency. We decide what data to clean, what to feed models, when to correct them. Models can't make those decisions themselves. I'm not against smarter models, but why not focus them on expert tasks like predicting climate change? I'm focused on using data and machine learning to improve things, not worried about some ominous AGI taking over.

— [A reality check on AI](https://www.lutzfinger.com/transcript/2024-07-10-a-reality-check-on-ai/) · _The Edge_

Question 13

What lessons from social media regulation are shaping today's AI discussions?

Accepted Answer

_Social media's harms taught us to regulate AI before it's too late._

Julie Owono notes that the conversation we're having today about AI is probably the lessons we've learned from social media. When Facebook launched, there was this famous phrase, move fast and break things. At the time, we didn't think about the negative externalities. Innovation was central. Now we are all in this room to discuss regulation, which is historically a quite interesting and big step in the right direction. We may not all agree on the details, whether lawmakers understand technology or technologists understand the importance of protecting democracy, but there is agreement virtually around the world.

— [SZ 139: AI in the EU Ft. Yaad, Gerard, Florence, Julie, & Lutz](https://www.lutzfinger.com/transcript/2024-12-09-sz-139-ai-in-the-eu-ft-yaad-gerard-florence-julie-lutz/) · _Silicon Zombies_

Question 14

What problems with RAGs still need to be solved before widespread adoption?

Accepted Answer

_The observability and control challenges that remain in RAG systems._

RAGs don't completely stop hallucination yet. The model uses both the trained corpus and new data, but we don't have a very good solution to manage that difference. Who decides when new information should override training data? Is the LLM taking that decision or the RAG? We lack observability and control. For example, I created a document saying I'm Captain America, but ChatGPT rejected it as fake. What if that was actually true breaking news? Amazing startups will help create this observability, solve access control, and manage the difference between what the corpus was trained on and corporate datasets.

— [Are we one step closer to using LLMs in practice?](https://www.lutzfinger.com/transcript/2023-07-21-are-we-one-step-closer-to-using-llms-in-practice/) · _The Edge_

Question 15

Why did celebrity bots like Meta's Snoop Dogg bot fail to gain traction?

Accepted Answer

_Why copying celebrities creates an uncomfortable user experience_

The celebrity bots fell into the uncanny valley problem. People got annoyed that the bot was a bit like Snoop Dogg but not really Snoop Dogg. When you go to someone like Snoop for authenticity, a bot cannot deliver that authentic experience. Fans know they never spoke privately to Snoop Dogg, they only know his public persona. The bot might be fun at first, but it becomes creepy when something tries to be too human without actually being human. Zuckerberg eventually shut down the celebrity bots because they didn't get enough traction.

— [The future of AI bots: winning with personalization, conviction, and cross-platform integration](https://www.lutzfinger.com/transcript/2024-08-22-the-future-of-ai-bots-winning-with-personalization-conviction-and-cross-platform/) · _The Edge_

Question 16

Why do AI models still need human involvement even with unsupervised learning?

Accepted Answer

_Even advanced models can't escape the need for human judgment._

We need humans because there isn't always a true north for what's correct. Supervised learning works when there's a clear label, like Obama was president. That's a fact. But when you get into perception, like whether a joke is funny or which word to use about Obama's policies, it depends on context and perspective. The same happens with generated images. Art often creates things society didn't expect, so at what point is something random versus amazing? Since there's not a true label for certain things, we constantly need people to give feedback on the model and provide guardrails.

— [AI and the human factor](https://www.lutzfinger.com/transcript/2023-07-07-ai-and-the-human-factor/) · _The Edge_

Question 17

Why is energy the AI race?

Accepted Answer

_Energy and infrastructure as constraints._

> What wins the AI race? UX? Access to users?

— [The Missing Moat In AI: Your Eval Data](https://www.forbes.com/sites/lutzfinger/2026/05/26/the-missing-moat-in-ai-your-eval-data/) · _Forbes_

Question 18

Why is memory the next AI battleground, not prompts or agentic systems?

Accepted Answer

_AI memory systems create new risks and control challenges._

Memory is essentially part of our context window. Models now start to keep information about our memory. I asked ChatGPT about using AI in student interviews and it scolded me, saying as a faculty member I should be extremely careful. But even worse, it noted down in my memory that I wanted to use that tool. That's pretty much a good idea for a big disaster. How do I as an organization ensure that my memory is kept in the correct form? What is the memory you want to capture?

— [From Google to Cornell: Rethinking AI Workflows | Lutz Finger | Build AI Podcast 17](https://www.lutzfinger.com/transcript/2026-04-16-from-google-to-cornell-rethinking-ai-workflows-lutz-finger-build-ai-podcast-17/) · _Build AI_

Question 19

Why isn't autocomplete understanding in healthcare AI?

Accepted Answer

_LLM limitations in clinical settings._

> I just got back from the JP Morgan conference. A few learnings that apply to any industry, not just healthcare:Humans + AI will change the industry. Not AI alone.Cash flow matters. Not AI hype.The real AI revolution is operational.One of the most hyperbolic promises in healthcare in the last month was the claim that: "AI is going to design our drugs and will replace researchers." In my view, this is not true.

— [Healthcare's AI Lesson: Autocomplete Isn't Understanding](https://www.forbes.com/sites/lutzfinger/2026/01/16/healthcares-ai-lesson-autocomplete-isnt-understanding/) · _Forbes_

Q&A, Risks