Full 49-minute walkthrough of how we edit Facebook ads with Opus 5.5

Warehouse robots are here! Introducing Flexport's newest automated fulfillment center just outside of Chicago.

Meta Muse's retention is out, and it's almost as good as ChatGPT! It's higher than ChatGPT at D7 (>30%) and just lower at D14 (22%). Not quite in the elite D14 tier of WA (67%) YT (59%) Insta (54%) or TikTok/Snap (37%), though. Muse is one of the highest retaining AI products!

The demand for AI is going to grow exponentially for a while (if not forever). The supply for AI is going to look a lot more linear in comparison. Scaling atoms is a lot harder than scaling bits. For this reason, the price of compute is going to rise substantially. A lot of electronics will get expensive, and fun toys like image and video generation that are relatively cheap today may become prohibitively expensive. Enjoy it now while you can.

the most profitable quarter ever by any company, but will it be enough to justify Samsung's ($005395.KR / $005930.KR) forward PE of 4?

AI agents are a great product for someone: 1. Very busy, with 2. Menial tasks (booking flights, writing emails, taking meetings), and 3. Too junior/poor to afford an employee they can offload to, where 4. Errors are low-stakes and don't matter much No wonder VCs love them!

Having listened to this a number of times, my sense is Dr. Tabrizi’s reference to an effect size of 40-50% refers to the effect size in the broader HD community if $QURE’s AMT-130 becomes available for treatment. I do not believe she is referring to the effect size she expects to be seen in the 4 year data for the 12 patients that is upcoming. This view is partially informed by her and Dr. Wild’s prior commentary (e.g. “beyond our wildest dreams” wrt efficacy), by the fact that it would probably be unwise to answer that question with data pending as a KOL, and by the fact that it not only would be require a substantial change in the trajectory of decline for the 12 patients as they moved from year 3 to year 4, but that change would have to be concentrated in a subset of the patients and correspondingly entail an even larger change in trajectory if we accept the anecdotes of efficacy from Sung, etc. I think it’s also clear based on her commentary she believes the treatment is effective as she states as really the foundational comment of the interview: “It tells us what to do. It tells us where we need to target”. The only way AMT-130 tells us that is if it is actually having a disease modifying effect. Otherwise, it’s useless. The existence of a disease modifying effect for AMT-130 is the basis on which the drug would be assessed for approval, and a 40-50% “true” effect in the broader HD population would demonstrably yield a blockbuster drug.

> be Harvey > lawyers give you money “bcuz AI” > product sucks, no one uses > seat pricing means less usage = higher margins > acquire $15.5B valuation selling mediocre frontier model wrapper > models suddenly git gud > lawyers start actually using the product > ohno.jpeg > token costs explode > gross margin goes from +50% to -50% > pivot to lower quality open models to halt usage growth > mfw accidentally built AI company that's structurally short AI progress

In other words, skimming / directing purchase intent is likely to be a less favorable ad category going forward, whereas discoverability still solves a core problem. There will always be companies with novel solutions that want to reach me that I won't be proactively looking for.