I am going to try something new this week. You will get a more long form essay through this newsletter in a few days. Would appreciate feedback when that happens. Thanks.
Keep it simple,
This Week's Sponsor
Turn Your Website Into an Instant Live Show
Live selling is the most effective marketing channel for ecommerce companies today and Reactive is the best way to bring it to your Shopify website. Its interactive features offer engaging customer experiences to dramatically increase sales and AOV. Designed to create successful launches, drops and promotions, Reactive instantly turns your site into a live shopping channel.
Reactive creates unique online events that customers want to experience and brands want to produce. Just 90 seconds to launch your first live show!
Hopper: Hopper is a fast growing travel booking app (currently ahead of Booking.com in the app store). They have announced a new loyalty program. While travel is where loyalty programs started for a variety of reasons (agency problems and low marginal costs being the biggest reasons), but using loyalty data to analyze and incentivize travel consumers is difficult due to how infrequent they purchase. Hopper is trying to solve the infrequent use problem by bribing users with points for opening the app daily (points give discounts off of travel purchased through the app that “cannot be combined with other discounts”). It is an interesting, and not very costly, experiment to see if keeping the app top of mind increases long term lifetime value when purchases are rare. Related: NYTs article on travel subscription services (which is generally the “right” way to build a loyalty program)
Self-checkout: The vision for self-checkout was that it could reduce costs, improve checkout speed, and improve customer service all at the same time. In practice retailers used the new capabilities to reduce costs and sacrificed the other two. Now some retailers are realizing the may have made a mistake, and that reducing friction and increasing convenience are very valuable in the long run for a brand.
Whitelisting: Whitelisting is when you use 3rd parties to promote and drive traffic to another 3rd party publisher site hosting direct response content for your brand. An example would be using an influencer to promote “native content” on Business Insider highlighting your product. This is a good case study walking through how it is done (and how effective it is)
Snap: Snap has been struggling in this new Apple-hobbled advertising world. The Verge has published a leaked email from Evan Spiegel on what they are going to try next:
Our primary focus is driving lower funnel performance to improve the yield of our inventory. We are working to improve optimization against lower funnel objectives to drive more conversions by delivering high intent clicks and evolving our webview performance and features. We are also innovating on our advertising formats, working to make them more native and engaging, while moving to more click-based interactions rather than swipes so that we can use the swipe gesture for content navigation.
Retail Ad Networks: The latest “retailer” to launch an ad network is Drizly, the alcohol delivery company. At this point the model is so well proven that every company with low margins that comes into contact with a large number of consumers should be figuring out how to build a high margin ad network on top of their business. Related: McKinsey sees “commerce media” as the future.
University re-branding: Rebranding is almost always a bad idea. But it happens anyway even when there are market forces creating headwind. When there are no market forces at play you can image brand directors at old prestigious universities are not going to be held back by “tradition” forever.
Most successful Holiday Campaigns: Common Thread analyzes 2400 DTC Black Friday campaigns and attempts to determine their relative success. Don’t believe their conclusions, but as a benchmarking exercise it is full of interesting data (i.e., 29.5% of stores offered site-wide discounts; 22.2% offered no discounts at all).
Spotify: Amazon dominates the audiobook market (43% with Audible). Spotify is TRYING to own the “audio market”. It was only a matter of time before they went after books. Ironically Amazon is mostly a transaction business, but they run audiobooks as a subscription. Spotify is mostly a a subscription model, but to differentiate themselves they are going to run audiobooks on a transaction model.
Proxy Metrics: “Whenever you optimize a proxie you will will make progress towards your objective for a while. At some point you start over optimizing and do worse on your true objective (hard to know when).”
Instagram Reels: Related to proxies, I found the below tweet interesting. Meta has ignored customer feedback in favor of watching what users actually DO for a long time (and profitted mightily from the decisions). But the risk is that they optimize for SHORT TERM user behavior and miss the long term impacts of high-engagement, low long term usage and mindshare (full thread if you click on the tweet):
Giphy: Peter Thiel once wrote that monopolies want to claim they are not monopolies, and non-monopolies want to claim they are. Giphy is trying to be acquired by Meta, but is being blocked by regulators. The company has now resorted to self-depreciation in an attempt to appease those in power. The company that was built on gifs now says that “Gifs are ‘cringe’”. Basically, ‘we are such a terrible company, there is no reason to block an acquisition’.
Outpainting: Simple tool that will use Stable Diffusion to expand images beyond the borders of the image in a similar style.
Whisper: OpenAI has released an opensourced model for voice recognition. This is a game-changer for all those automated customer service programs. And at this point the cost for high quality podcast transcription has dropped to close to zero (here is a step-by-step on how to do it). Related: James Earl Jones’ Darth Vader voice in the new Kenobi Disney+ series was created with AI.
Fake Emulations: Robin Hansen believes that brain emulations will come before GAI. Nora Belrose proposes that before we see either technological jump we will see “fake brain emulations”. Feed a language model lots of content from an individual and then have the AI “simulate” that person. It is being done now with famous writers, but should be doable for anyone who has enough written, audio or video content to build the model from. Feels like a Black Mirror episode.
Getty Images: The pay-for-images site will no longer accept any submissions created with AI tools. Good luck policing that. (I expect this is primarily for legal reasons - if someone argues rights issues, Getty can pass through any complaints to the artists who made the submissions).
How to detect AI: Google is going to penalize low quality AI content. But how can they identify it? Kevin Indig has an excellent post detailing the steps. Effectively AI language models use probability to decide what the next token (or syllable) will be. They generally choose the most likely next token, or at least a high probability one. If they don’t they will produce gibberish. But that same technique allows Google to guess when content was AI-generated. It just scans content and approximates how likely any given token is based on what came before. The result is that a pure AI-generated might be correctly identified 20% of the time (compared to a Wikipedia page that is only 0.02% likely to be AI). If you edit AI-content to make it more “creative” in word choice, Google will not be able to identify it (and you will be fine); conversely, if you have a weak writer, they might be falsely identified as an AI. The net result is Google will not penalize AI-per se, but rather this is just another way to penalize low quality content generally (whether AI or human created). In the image below red are “unlikely words” and green are high probability words.
When Humans Beat AI: A paper that tries to quantify when humans are still beating AI and how to best utilize AI to improve human performance. A few conclusions:
Use AI to figure out who is good
Use AI to figure out what “good” humans do, so you can train others
Create standardized “complex” decision rules
Humans work best in “uncommon problems” — often the best choice is to use AI for “lower level support” and have humans jump in when the AI cannot identify an easy answer
AI tends to be much better than low skilled, low experience and “international” workers
All of this parallel’s AI in chess. First AI is better than poor performers. Then AI can be used to help top performers. Eventually AI is better than all but the very best — until it is better than everyone.
GPT3demo: A collection of AI applications
Charlie 2.0: “all in one” AI tool for marketers. Create images, blogs, ad copy, ad images, content scoring, etc.
Suggesty: AI powered search results using a chrome extension
Typed: Knowledge management / collaboration tools that add an AI (and other) layers to Google docs
Stable Diffusion: A thread of creations from the first 30 days post launch. Also: A guide to getting started with Stable Diffusion; and how to run Stable Diffusion on an M1 Mac.
Promptbuilder: A tool for building detailed text-to-image prompts. A good example of what is possible.
Playground AI: A place to share AI-created art; RAW SF VFX 3d hyperrealistic 32K cosmic crashed cars sculpture gestalt Trisha Paytas composition 🌌🚀🌉🛣️highway spatial hair CHRIS lABROOY sculpture made of HAIR of 🛣️ Trisha Paytas CAR 🌌. Belin postneocubismo BY Andrew Thomas Huang. A goddess cyberpunk with a ram skull. beautiful intricately detailed Japanese crow kitsune mask and BIOTECH kimono:: OCCULTIST 🛣️ bubble CARs, epic royal background, big royal uncropped crown, royal jewelry, robotic, nature, full shot, symmetrical, Greg Rutkowski, Charlie Bowater, Beeple, Unreal 5, hyperrealistic, dynamic lighting, fantasy art
Alex Guzey’s AI supplement: Good collection of AI stuff from Twitter (and a good newsletter all around). Some examples- Getting GPT-3 to consult python code in order to do complex math correctly; Converting from text to code and vice versa; Coding speed with and without Copilot (more than twice as fast - see below)
Hamilton: The NYTs has a fascinating article on the struggle to translate Hamilton (the musical) into German. The musical is full of witty rhymes, specific rhythms, and dialogue with multiple simultaneous meanings. So much of the enjoyment of the show is the wordplay. Trying to hold onto that in a new language almost means re-writing the entire production. Lin (the creator) even learned some basic German to help ensure the final product was acceptable. If you don’t have time to read the entire piece, at least check out the examples of how lines were changed.
Keep it simple,