It’s been a couple of weeks since my 3-part AI series (Part 1, 2 and 3). This week I am back with a briefing. If you want more start-up and marketing stuff, this edition has it. If you joined in the last couple of weeks due to the AI content, I have an extended set of AI links just for you.
Onto the briefing!
This Week's Sponsor
Are you drowning in bookmarks or Chrome tabs?
Introducing INDX—it’s like Pinterest for learning anything. INDX can:
1. Save any article, podcast, or Twitter thread you find interesting.
2. Remind you to go back to the best content so you actually consume and learn it.
3. Surface articles on the topics most interesting to you, from the people you trust.
Since launching earlier in the year, INDX has become the go-to tool for creators, operators, and investors in tech to ditch information overload and actually learn.
Download INDX on the App Store today and start taking more control of your content
Advertising
Marriott: The hotel chain is joining what seems like every retailer on the planet to launch its own advertising network. If you are a company with a large customer base, some form of customer attention (i.e., ad units), and data on your customer spending, monetizing that attention with an advertising auction targeted with lookalikes seems like the right thing to do. “Monetize your exhaust”
Campbells Soup: Speaking of new ad networks, AdExchanger has an essay on the evolution of retail media and specifically how Campbells' is thinking about it (h/t TheDiff for the link). In the past CPG firms could put up displays in stores and give away free samples to whoever stopped by (and hoped for the best). Now using platforms like Instacart they can give away free samples specifically to people buying their competitors' products, and then measure any changes in purchase behavior after the fact. They can even manage holdout groups, which both helps for the specific initiative, but it also gives them a rough ROI on how effective their earlier (unmeasured) sampling program were.
Disney+: Disney+ adding an advertising-supported tier is old news. Last week they announced the ad-tier would only have 4-minutes of ads per hour. They also claimed that “We’re never going to collect data on individual kids to target them,” but I expect they WILL still collect data on individual kids (like their presence in a home), and they will allow targeting based on the ages of the children in a household, and I am not sure what other type of targeting an advertiser would be interested in doing.
TikTok: A good profile on just how fast TikTok’s ad platform is improving. Two-years ago TikTok was insignificant as a marketing channel, now it is the third platform most companies should be testing into. Also: TikTok expected to surpass YouTube in viewing hours later this year.
Podcasting Dynamic Ads: The Verge reports on a new study that estimates that 84% of podcast ads are now “dynamically placed”. Moving from manually negotiated ads to these programatic ones is like moving from artisanal to mass production. It makes it much easier for smaller podcasters to host ads (since they can avoid the fixed costs of a sales organization), and much easier for smaller advertisers to run ads (since they can use manual self-serve tools at much lower minimum spend levels. It also is a step towards better targeting. Instead of hitting everyone who listens to a specific show, you can, in theory, just target people of specific demographics regardless of the show they are listening to. Attribution also gets better because you can run better A/B tests (i.e., advertise to everyone in one region and not another and then measure your local lift). I am just surprised it has happened this fast. 84% is a big number!
Marketing to Employees
Netflix: The streaming giant has updated their famous employee culture document:
We let viewers decide what’s appropriate for them, versus having Netflix censor specific artists or voices… even if we find some titles counter to our own personal values. Depending on your role, you may need to work on titles you perceive to be harmful. If you’d find it hard to support our content breadth, Netflix may not be the best place for you.
Unilever: The WSJ has a good profile on Unilever and how they have evolved in their drive for “purpose driven marketing”. When Alan Jope became CEO of Unilever in 2019 he pivoted the company towards “brand purpose” saying, “Brands with purpose increase sales twice as fast as those without”. Soon all of the company’s brands leaned into the “purpose” strategy with the hope that sales growth would double. Unfortunately, since then, Unilever’s sales growth and stock price has lagged behind competitors like P&G, Nestle and L’Oréal. One of the largest shareholders is not happy: “A company which feels it has to define the purpose of Hellmann’s mayonnaise has in our view clearly lost the plot. [Unilever] is obsessed with publicly displaying sustainability credentials at the expense of focusing on the fundamentals of the business”. Jope himself has been humbled: “[Purpose] is the icing on the cake. It is not the cake. We also want to be absolutely clear that purpose isn’t a substitute for having fantastic quality, innovation, advertising and distribution”. Most interesting were some comments from the departing VP of customer insight: “The company’s talk about purpose belies how much its brands spend on advertising more straightforward product benefits. Even Dove advertises more about soft skin than it talks about purpose. Brands with a purpose that resonates with consumers stand out from rivals, command a price premium and help rally employees”. As you can imagine, I expect most of the benefit is coming those last three words. Unfortunately that wasn’t enough to grow a multi-billion dollar business beyond simpler peer-group strategies.
Pepsi: Related to the above, Pepsi is still leaning into purpose, but have decided that the purpose needs to be “authentic” and closely tied to the “brand’s purpose”. This is mostly the brand managers talking to themselves, but I appreciated this quote: “Consumers are telling us [that] things don’t need to be perfect if you’re true to yourself and you’re taking the right steps”. Sure. That is the language all customers use as they guzzle their Pepsi and munch on their Doritos.
Market Research, Metrics and Attribution
Masks: Three in Five Americans say they will continue to wear masks on airplanes. Meanwhile almost no one on actual flights are wearing masks. Another case of “don’t trust what people say, only what they do”.
Testosterone: A new study looks at the impact of injecting testosterone into political beliefs and finds that it shifts “weakly associated Democrats” slightly right-ward. The point is that it had no apparent effect on anyone with stronger political beliefs. Many many things can influence people on the margin, but nothing we know can really MAKE someone change their mind about just about anything.
Wells Fargo: Matt Levine at MoneyStuff tells the story of how Wells Fargo bankers are interviewing “diverse candidates” for roles that have already been filled (in order to meet internal quotas on interviewing diverse candidates). He likes this to Wells’ earlier “fraud” where bankers were adding products to customer accounts without customer permission (in order to meet internal quotas around products added per banker per month). Both cases are examples of simple metrics that were meant to drive specific behavior, that were later gamed by stressed employees.
Violence: On the same note in a less business-specific example, Chris Blattman (new book: “Why We Fight”) has a good Twitter thread on what sometimes drives lower crime rate in developing countries. He argues that much of the time you can lower crime rates by increasing the power of criminal gangs. Stronger gangs leads to more crimes, but also lower violence - since they have more of a monopoly on the region. It’s another example of how focusing on an individual metric may not get you what you really care about.
Television: Nielsen has a detailed report on the state of television today (both traditional and streaming). Streamed content in 2022 is up +18% YoY and has “reached a tipping point”.
Facebook Attribution: With Apple restricting everyone’s ability to track advertising effectiveness, Facebook has been building more and more sophisticated attribution methods to attempt to help advertisers understand the effectiveness of their ad spend on the Meta platform. This is necessary, but it means simple attribution methodology that was easy to understand (even if sometimes difficult to interpret) is being replaced with more “black box” methods that you just have to “trust”. Unsurprisingly these black box methods often get things wrong. AdExchanger: “Bugs infest Facebook’s new attribution technology”
Business/Strategy
Private Valuations: Public markets have been tanking for the last month, but we have not seen a similar collapse in the private equity space (this may just be a case of supply and demand - a number of funds have raised large amounts of capital that they need to spend in the next couple of years. Maybe the crash happens when they can’t raise their next fund, and/or can’t sell the next tranche of companies?). Meanwhile a16z has published a primer on how start-ups should be thinking about valuations going forward (spoiler: lower).
Airbnb: The company has launched a bunch of new features including much better search. You can now search for “flexible destinations” where you look for (example) a ski in ski-out challette, some where in the rockies, or a beach front property somewhere on the west coast. We tried to push for something like this at Expedia a decade ago, but the tech was way too hard for the legacy systems (we wanted things like “direct flights from my location less than 3 hours” or “flight+hotel deals under $xxx from my location”. The other significant Airbnb change is allowing you to search for stays split across two properties (we often had this problem booking month long stays during COVID) Airbnb release; WSJ article (not paywalled)
Fraud at Scale: From the VP of Security at Google, commenting on this thread about how Elon Musk is discovering that creating a safe platform on Twitter is harder than it looks
AI
Gato: Google’s new AI is a “generalist RL agent that uses transformers”. It is a single AI that can do things as varied as play video games, control robotics, caption images, chat, and more. It effectively “understands” words, images and physics, and takes actions to achieve objectives. It was built with 1.2B parameters (vs 175B for GPT-3). Announcement. More here on how it works.
Meta: On May 3rd Meta released their latest NLP model to “academic researchers and people affiliated with government, civil society and academic organizations, as well as industry research laboratories”. Open Pretrained Transformer (OPT-175B) was trained on 175 Billion parameters, and by making the inner workers available to academics their hope is to “democratize the industry”. A friend of mine who works for one of the players in the space doing AI development thinks that Google, Facebook, Microsoft and OpenAI are now all within 6-months of each other in capabilities (he believes everyone else, including Apple, are pretty far behind). Note that in theory OpenAI is partnered with Microsoft, but my friend tells me Microsoft has two internal streams, one with OpenAI and one independent, and they are both pretty comparable these days.
A Well Behaved School Boy: Rohit at StrangeLoopCannon has a good essay on the state of AI writing. Skip the first section which was written by GPT-3 (and is pretty bad IMO). In the second part he comments on the quality of the first part and makes the point that we primarily read non-fiction not for the prose but for the ideas. Unfortunately GPT-3, while it often gets facts right, it has no ideas. Good quote:
Right now GPT feels like an eerily well behaved schoolboy who is pretty good at exams. Give any topic or question, it parrots a reply really well. If you teach it to do basic algorithms, like reversing a word, it does so. It writes sensible For now at least we have a personality dodge, in that it doesn’t seem to be able to (or want to, for specific versions of want) understand vast swathes of human experience, or be much opinionated at all. Until this too gets resolved, may many more essayists bloom!
Human-Level AI: Good Tweet Thread on why just scaling up the existing models will not get us to “human-like intelligence”.
AI You: Mirabelle Jones walks you through the steps of uploading your social media history to GPT-3 so you can create a chatbot that writes just like you do.
Smart Prompting: This is a great Twitter thread on how GPT-3 is terrible at “reversing words” - i.e., alphabet should become tebahpla. The thread explains why (how the GPT-brain works using tokens), and then goes on to show how you can build a multi-step prompt to get it actually complete the task. The conclusion is that these models are “smarter” than they first appear, but, just like humans thy work best when you understand HOW their brain works, and then build step-by-step directions to achieve the results you are looking for.
Context Limits: One of the biggest limitations of NLP-models is that they can only “remember” about 2000 tokens of information at a time. This has stymied any attempts so far to write a novel (the AI can’t remember that the gun described in Chapter 2 can be used in Chapter 3, let alone attempt a twist ending). Here Gwern explains why there is no way around these with the current models (the compute needs scale quadratically as tokens are added), and here is some researchers discussing how they are trying to “attend to arbitrarily long contexts without increasing computation burden”. Note the second link is from 2021 and I have not seen evidence that progress has been made.
OnlyFans: NYTs has a profile of OnlyFans “managers”. These are ghostwriters who manage the accounts of multiple women and do the work of chatting and setting prices for subscribers. No AI in the article, but it seems like that would be the next step. I will bet a trained version of GPT-3 could do a pretty good job of interacting with (and collecting revenue from) unsuspecting men. Combine that with Dalle-2 image generation and the entire product becomes an engineering challenge to be optimized.
Google Translate: I have mentioned before that Google Translate is now built on an AI language model, which allowed the product to move from hundreds of thousands of lines of code to just 500. It is also allowing performance to improve, and languages supported to accelerate. Last week it added 24 new languages - all built with “zero shot machine learning” - meaning it learns to translate the language without ever seeing an example. The future is coming at us fast.
Fable Studios: Fable combined GPT-3 with graphics and memory structures to create “virtual beings” that interact with users and remember and are influenced by previous interactions. Here is a good interview with the CTO of the company and what the company’s strategies and expectations are going forward (Video. No transcript).
Careers
A-game: Nate Meyvis has an interesting essay on the strategy of switching up what you work on based on your energy level. Some work is only worth doing when you are alert, but you can be very tired and brain dead and still get some subsets of work done.
Showrunning: A great in-depth article on how television showrunning works (and has worked). Basic thesis is that there is a lot less mentoring and career development these days - which is fine for now, but it means the industry may struggle in finding the next generation of showrunners. Not too dissimilar to today’s work from home environment…
Thoughts
Access vs Usage: A note I wrote to myself this week- “If you have a high-usage product the best way to monetize is likely advertising. If your customers care far more about ACCESS than regular usage, the right way to monetize is subscriptions.”
Fun
Sunlight: Seattle is notorious in the US for having very little sunlight for most of the year (our summer’s are fantastic though!). I haven’t heard as many complaints about Europe, but this map shows even Seattle would be the Florida of the Old Continent!
Wetness: I went my entire life without realizing humans have no sense for “wetness” - we just guess based on temperature and pressure.
Keep it Simple,
Edward