363: Google vs ChatGPT, DoD Cloud, Constellation, Microsoft vs Discord, Rolex, SpaceX, and Tarantino

"Here’s an interesting concept: The Murder Board"

The trouble is, if you don't risk anything, you risk even more. –Erica Jong

🤔 🌹🧠 You know how certain things become popular in waves among certain groups?

Meditation, fasting, yoga, and so on.

I’m guessing that at some point in the coming years, reading and memorizing poetry will become one of those things.

It’s so the opposite of the smartphone instant-culture that we have now. It’s at the nexus of artsy meditation, deep thoughts, brain workout, slowing down, and memory training.

Building attention span, but with a hipster twist.

I’ve never done it and am not really interested, but I recognize in it the elements that may make it a thing.

👩‍🎨 Having good taste, or knowing people who have good taste and listening to them, is extremely valuable for so many aspects of life. Pivotal decisions often have to do with taste.

Over time, I’ve become convinced that taste can be consciously developed through deliberate practice, even if few people actually do it. We all have different starting points, and the intensity of the original impulse may be innate, but the work after that is a decision…

🛀 Most of the time, when something doesn’t seem to make sense, your first reaction shouldn’t be “Wow that’s crazy! It’s stupid, it’s nonsense.”

It should be: “What am I missing? Am I seeing the whole picture? Could this make sense to someone with different information?”

🔪🗣🗡 Here’s an interesting concept: The Murder Board

A murder board, also known as a "scrub-down", is a committee of questioners set up to critically review a proposal and/or help someone prepare for a difficult oral examination.

The term originated in the U.S. military, specifically from the Pentagon, but is also used in academic and government appointment contexts. NASA contends the murder board was created by Hans Mark, Director of Ames Research Center from 1969 to 1977, derived from the earlier concept of the tiger team.

In highly risk-averse, technical endeavors where extreme efforts are taken to prevent mistakes (e.g. satellite operations), murder boards are used to aggressively review, without constraint or pleasantries, a situation's problem, assumptions, constraints, mitigations, and the proposed solution. The board's goal is to kill the well-prepared proposal on technical merit; holding back even the least suspicion of a problem is not tolerated. Such argumentative murder boards consist of many subject matter experts of the specific system under review and of all interfacing systems.

This can be pretty informal, as this HN comment points out:

I use to frequent a grad student bar that regularly held an open-mic night for people to practice their [doctoral thesis] defense. It was brutal. A room full of drunk grad and post-grads would try and pick apart your presentation. Targeted academic heckling. It was great. The goal was to get over your fear of the whole process -no matter what, your actual defense would be a better experience.

Trial by fire!

---

Business & Investing

Google Search vs ChatGPT-style AI Chatbots

Now that ChatGPT is the internet’s main character (for how long?), there’s a common meme about how “Google is screwed”.

It’s easy to imagine that we’ll all ask most of our questions to ChatGPT or one of its successors and this will massively hurt Google’s business.

The smart response to this is “Actually! ChatGPT is often wrong! It will often hallucinate facts and make up fake sources, and will confidently assert all kinds of things that aren’t true! Anyway, Google can use its own LLMs and integrate them in search so they’ll still have better answers than OpenAI. And Sam Altman said that each query is costing them a few cents, so that can’t scale to Google size!”

That’s all true!

What Large language models (LLMs) do is to probabilistically predict what character (or multi-character token) comes next, and they keep going until they have something that matches the patterns extracted from the training set associated with the words/concepts that you’ve prompted it for.

But that’s kind of what humans do too, if you squint a lot.

If I talk to you about a giraffe, you retrieve giraffe-related things from your memory based on your “training”. If you’ve never seen a giraffe and nobody ever told you about them, you won’t be able to think about much. If someone gave you bad information about giraffes (they’re blue and can fly), what you think about will be incorrect. Sometimes we misremember things in good faith and hallucinate “facts” (there are good documentary about how unreliable eye-witnesses are).

But that’s not the point I want to make. Back to Google Search and the potential impact of LLMs on it:

I think the more nuanced take on this is that such models will likely have a *very large* impact on Google Search in many ways, some good and some bad, with benefits accruing to different people.

For example, if Google integrates its own LLM into Search results, what does this do to the search ad business? Do people click on fewer ads when they’re getting more complete text answers without having to click on links to go somewhere else? Because after all, ads are “native content” in the way that search’s native results are links that go somewhere else.

If the “native” result of a search is increasingly a blob of text and images at the top that answers your query right there, how do you advertise against that? I suppose you can have display ads that don’t need to be clicked on, but that’s a lot less effective and lucrative than the pay-per-click ads where attribution of sales and ROI is much easier to calculate (and so the value is easier to capture for Google).