Too Stupid To Know You're Stupid

Good morning,

In 1999, Cornell University psychologists Justin Kruger and David Dunning published a paper titled "Unskilled and Unaware of It," where they shared the amusing tale of McArthur Wheeler - an aspiring bank robber.

In 1995 in Pittsburgh the then 44-year-old Wheeler decided to rob two banks without any disguise, his face fully visible. Security cameras captured him smiling as he pointed a gun at bank tellers and demanded cash. Later that day, the police obtained the surveillance footage and aired it on the evening news. Within an hour, Wheeler was arrested after an informant recognized him and contacted the authorities. When the police showed him the video, Wheeler was stunned. He had heard that lemon juice could be used as invisible ink and believed that smearing it on his face would make him invisible to cameras. Confused, he muttered, "But I wore the juice."

This became known as the “Dunning-Kruger Effect”.

If Wheeler was too stupid to be a bank robber, perhaps he was also too stupid to know that he was too stupid to be a bank robber. (Read that again).

Dunning and Kruger's study demonstrated that unskilled individuals suffer a dual burden; not only do they perform poorly, but their incompetence robs them of the ability to realize it. Instead, like Mr. Wheeler, they are left with the mistaken impression that they are doing just fine.

This effect is called a “mental model”. I love mental models because they’re like special glasses. If you put them on - you see the world in a different light.

We see Dunning-Kruger manifested everywhere in life. Have you ever seen those contestants in “The Voice” or “American Idol” that think they’re Pavarotti but when they start to sing glass starts breaking ? Same thing.

You see this definitely play out in the business world where junior Deloitte consultants rollup to your company, thinking they’re the next best thing since grilled cheese. They barely have hairs on their chins to shave but yet they’re going to transform your company. Of course most of them get found out quickly.

When it comes to AI - I think the same thing is going on. It not only applies to me (which it does) , it applies to many people that have entered this field.

It’s a relatively new field. Yes I know AI is not new but transformers and LLMs are “new-ish”.

We’re all discovering this together. We’re in uncharted territory and most of us are on firmly on Mount Stupid.

So let’s stay calm and explore this field with our feet firmly rooted in reality.

Which voices are we hearing? It seems that when it comes to the future of AI there are four schools of thought:

• The AI Doomers
  AI is rapidly advancing toward superintelligence, and once it surpasses human capabilities, it will bring about the end of civilization. The Singularity means humanity's time is up. Main proponents : Elon Musk, Geoffrey Hinton

• The AI Utopians
  AI will become a superintelligence capable of solving all of humanity’s problems—from curing diseases to reversing climate change. With AI's help, we will thrive and explore the universe together. Ray Kurzweil, Dario Amodei, Sam Altman. (By the way these are the guys who want your money 😉 )

• The Skeptics
  AI and especially LLMs are little more than an overhyped marketing gimmick. The current fascination will eventually collapse, leaving disillusionment and failed promises in its wake. Main proponent : Meta’s Yann Lecun,

• The Pragmatists
  AI will bring both positive and negative consequences as it continues to develop. It won't be an all-powerful force for good or evil but will be shaped by how humans decide to use it, making its future largely up to us. Andrew Ng.

The problem is that we don’t know. 
Everybody is in phase one of Dunning-Kruger and loudly proclaiming they hold the truth.

Unless Skynet emerges in the next 4 weeks we are going to (briefly) explore these schools of thought in this newsletter.

See what points they are making. It’ll be interesting I tell you.

Next week should be especially fun as we’ll study the AI Doomer movement. It might be too early for Halloween but I’ll give you a nice little horror story to cozy up to.

See you next week.

Oh, and welcome to the Blacklynx Brief !

By the way, there is now a referral program for this newsletter. For every 3 people that you get to sign up, we will gift you a 15€ Amazon gift card.

Scroll to the bottom of the newsletter for details!

AI News

• AMD announced new AI-focused processors at its Advancing AI 2024 event, including the Ryzen AI PRO 300, Instinct MI325X accelerator, and EPYC 5th Gen CPUs, all aimed at challenging Nvidia and Intel in the AI chip market. These new releases boast superior performance metrics and reflect AMD’s plans for annual AI chip updates to meet growing demand.

• OpenAI has launched MLE-bench, a benchmark that tests AI agents on machine learning tasks based on Kaggle competitions, evaluating their ability to handle real-world engineering challenges. While models performed well on standard tasks, they struggled with creative problem-solving, achieving bronze-level results in 16.9% of tests.

• Anthropic CEO Dario Amodei shared an optimistic vision of AI’s potential over the next decade, suggesting that by 2026, AI could surpass human capabilities across fields, accelerate scientific progress, and even double human lifespans. He envisions AI strengthening democracy and driving economic growth, though he acknowledges challenges like job displacement.

• OpenAI introduced Swarm, an open-source framework for coordinating multi-agent AI systems, designed to simplify the management of collaborative AI tasks. Swarm uses agents and handoffs to allow various AI agents to interact seamlessly, each handling specialized functions. Available on GitHub, Swarm aims to help users experiment with multi-agent orchestration, potentially enabling individuals to deploy networks of AI agents for complex workflows.

• Apple researchers published a study highlighting major limitations in the reasoning abilities of leading LLMs, including OpenAI’s models, revealing that small changes in wording significantly impact accuracy. Their new GSM-Symbolic benchmark found that as question complexity rose, model accuracy dropped by up to 65%, suggesting that LLMs rely more on pattern matching than genuine reasoning.

• Adobe introduced new video generation capabilities for its Firefly AI model and Premiere Pro at the MAX Conference, allowing users to create videos from text prompts or images. The Firefly Video Model, now in beta, offers features like cinematic clips, animations, and text effects, while Premiere Pro gains Firefly-powered tools to extend clips and enhance transitions.

• OpenAI is embroiled in a trademark dispute with Guy Ravine, who claims ownership of the “Open AI” trademark and alleges he pitched the AI concept to tech leaders before OpenAI’s launch. Ravine asserts he was approached by OpenAI’s founders to buy his domain and that they ultimately used his idea without consent; OpenAI has countersued, and Ravine plans to pursue further legal action.

• Researchers from the University of Geneva, University of Edinburgh, and Microsoft developed DIAMOND, an AI model capable of generating a playable simulation of Counter-Strike within a neural network at 10 frames per second. Trained on just 87 hours of gameplay, DIAMOND uses a diffusion-based approach to predict frames and allows keyboard and mouse interaction, recreating elements like weapons and player actions.

• The New York Times issued a cease and desist notice to AI search startup Perplexity, accusing it of using the publisher's content without permission for AI-generated summaries. Despite previously stating it would avoid crawling NYT’s content, Perplexity continues to display it on its platform.

• Anthropic updated its Responsible Scaling Policy, introducing new safety thresholds to manage AI capabilities that could pose risks, particularly related to bioweapons and autonomous AI research. The policy includes a new “Responsible Scaling Officer” role, public capability reports, and external expert reviews, aiming to set an industry standard for AI safety as the company prepares for future, potentially more advanced AI models.

• Meta researchers developed Thought Preference Optimization (TPO), a training method enabling AI models to generate internal thoughts before responding, enhancing performance on tasks like marketing and creative writing. TPO allows models to optimize their outputs independently, although it showed mixed results on math tasks, where reasoning-based methods still excel.

Quickfire News

Closing Thoughts

That’s it for us this week.

If you find any value from this newsletter. Please pay it forward and receive 5€ for every person you get to sign up in the form of an Amazon gift card. What’s keeping you ?