The Uncertainty Game Edition

Using AI isn’t chess, it’s poker.

Dear Reader 

Let’s start with a few statistical shocks to get our minds whirring and a sense perspective about the revolution we’re living through. 

A brief history of time (saving) machines: from monks to material science 

35 monk years per bible. Gutenberg’s printing press saved 35 monk-years. (It would take a monk about a year to make a copy of the Bible. Gutenberg’s first print run of 180 bibles took about five years. Today’s print machines make 800 pages a minute, the equivalent of a bible every 3.14 minutes).

499 hours per shirt. Before the Spinning Jenny started the industrial revolution it took 500 hours to make enough thread to make a shirt. Today it takes 0.5 seconds. 

1 billion scientist hours: Alphafold, the AI created by Google DeepMind saved 1 billion hours of research time by discovering 200 million protein shapes. 

800 material scientist hours:  Google DeepMind's AI system, GNoME found over 2.2 million new materials, with 736 of these already passing experimental validation. This would have taken 800 years for a human researcher.

Playing the uncertainty game

What do all those big numbers and big discoveries mean? We can guess but we don’t know. 

I enjoyed tech and experimentation expert Ronny Kohavi's book notes on poker ace Annie Duke's Thinking In Bets, which he generously posted on LinkedIn. It’s about understanding decision-making in uncertain conditions. Duke is a poker player (who left the scene under a bit of a cloud). 

The book’s insights provide a compelling comparison between poker and real-life decision-making, emphasising the constant presence of uncertainty. We don’t like uncertainty, so we tend to want to find ways to limit it, even to the extent of irrationality – like talking about gut feel and luck in a game where there is no way of knowing the right answer. 

Unlike chess, where all information is laid bare, poker replicates the unpredictable elements of life, demanding decision-making amidst unknowns. This highlights that the outcome of a decision isn't always a reliable indicator of its quality. A well-played hand in poker can still lead to loss, akin to a well-thought-out play in a football game not always guaranteeing a win. 

Duke's analysis champions the importance of probabilistic thinking, where admitting uncertainty, or saying "I'm not sure", is a crucial step towards better decision-making. This approach shifts our focus from judging decisions solely by their results to evaluating the quality of the decision-making process itself, underscoring the value of understanding and thoughtfulness over mere outcomes.

You can download Ronny’s notes on LinkedIn. 

Uncertainty is the essence of generative AI tools

The notion of comfort with uncertainty also applies to the  use of generative AI, like Large Language Models (LLMs). Working with these AI tools is more poker than chess. In chess, every piece is visible and moves are predicated on complete information, suggesting a world where decisions lead to predictable outcomes. However, in the realm of generative AI, we're playing a different game entirely – one more like the probabilities-not-facts and hidden complexities of poker.

We don’t know everything. Things can go wrong. But we still have to decide. There’s no option other than making the best bet we can based on everything we do know. 

Working with generative AI offers plausible possibilities rather than certainties. Each interaction is a bet on the unknown, a wager on what might be uncovered or resolved. When you pose a query or input a prompt, you're dealing a hand in this game of cognitive poker. Unlike chess, where strategy can be planned several moves ahead with a clear view of the board, in the generative AI game, each prompt is a new deal, each response a fresh hand, where outcomes can't be predicted with certainty.

Just as a poker player learns to read the table and play the odds, users of generative AI must learn to interpret, assess, and integrate the information provided, understanding its potential and its limits.

To misquote Ben Horowitz, using generative AI is not chess, it's motherflippin' poker. 

It is a dynamic, somewhat unpredictable engagement with a system that can offer valuable insights and aid decision-making, but also one where we must continuously recalibrate our expectations, refine our prompts, and critically evaluate the responses we get. In this game, the winning players are those who understand and respect the nature of the AI tool they're using, leveraging its strengths while being aware of its limitations. And accepting that sometimes they will get it wrong. 

Just like in poker, success with generative AI is about playing the odds, interpreting the 'tells', and sometimes, knowing when to fold and ask a different question.

The “AI app store”

This week OpenAI's GPT Store launched. It’s a marketplace for bespoke AI chatbots created for specific purposes. A ChatGPT Plus subscription is needed to see them and use them as many are powered by the more advanced GPT 4 (the free version of ChatGPT uses GPT 3.5).

OpenAI captures most of the attention, but you can build and use similar bots elsewhere. The version of Microsoft’s Co-Pilot AI in some large companies lets users build their own “Co-Pilots” or bots. And the Poe app has always allowed subscribers to build bots, and some can be used by anyone. Take a look at this Poe tutorial on how to build a knowledge base bot as an example.

Ad: My company Brilliant Noise is building bespoke bots for clients and helping them to build their own as part of its AI-B-C™: Quick Start for Teams programme. Message me by replying to this email or hello [at] brilliantnoise.com for more information. 

Mil-spec ChatGPT

In a significant policy shift, OpenAI now allows military applications of its AI technologies. This update, effective January 10, 2024, has removed the explicit prohibition on military and warfare uses, stirring ethical debates and potential implications for AI's role in military contexts. The broader policy still prohibits using OpenAI’s services to “harm yourself or others,” but lacks specific guidelines on military use. Experts express concerns over this change, given the potential risks of AI in warfare.

For more on this story see: The Intercept

Corporates ban ChatGPT, 25% of workers ignore them

Despite corporate clampdowns on ChatGPT, a rebellious 25% of employees are sidestepping the rules. These tech-savvy rule-benders continue to tap into ChatGPT's prowess, underlining a clash between corporate caution and the undeniable allure of AI efficiency. This defiance points to a wider dialogue needed about AI in the workplace – balancing innovation with governance, as highlighted in Salesforce's recent study.

Investigating: Substack x Nazis

Platformer, one of the biggest fish in the Substack newsletter platform pond, has decided to take its business elsewhere after writing about the company’s accommodation of Nazi groups. I’m looking at alternatives for Antonym, while simultaneously trying to understand the whole issue, but don’t worry you will still get Antonym every week (ish) as migrating newsletters looks pretty easy. Platformer’s moving to Ghost, an open source newsletter platform that my friend Adam Tinworth has championed for some time. It looks pretty good. 

If you have any insight or opinion on this issue that you’d like to share, please do let me know.

This week’s cultural recommendations

Watching…

Finished Mr Inbetween (Disney+) It was amazing. Watch it. 

Poker Face (NowTV / Sky). Thanks to a strong recommendation from The Rest Is Entertainment podcast I finally found this dazzling series, with structural and stylistic nods to Columbo, its very high production values and cast shine from the first episode.

Reading… 

A Short History of Ireland: 1500 - 2000, John Gibney. Celebrating my Grandmothers and Irish citizenship with a good read. 

A God In Ruins, by Kate Atkinson continues a lucky streak in my fiction reading. I finished it this week and am looking forward to the Shrines of Gaiety soon. 

Right Kind of Wrong, by Amy Edmondson. Make the right kind of mistakes: a field manual. 

Until next time…

That’s all for this week folks, thank you for reading and I hope you found something interesting. 

Antony