Michael Johnston

A ‘boom’ in Atlantic bluefin tuna farming has renewed fears for the wild population. Research by Compassion in World Farming reveals warning signs for the species, as well as for animal welfare, food security, & marine life. Captured live at a crucial point in their reproductive season, wild Atlantic bluefin tuna are dragged in large nets to farms, where they are fattened in sea cages. This not only puts unique pressure on wild populations but drains our oceans of smaller feed fish harming marine ecosystems and coastal communities. Confining these wild, migratory animals in cages is cruel and unnatural. We stand with our friends at Compassion in World Farming, who are calling for governments to #KeepThemWild and stop this unsustainable industry expanding any further. Read the solutions and sign their global petition! ✍️ https://bit.ly/4hT02IJ

40

1 Comment

I like building things and using AI as a speed multiplier, so I tried it in a domain I knew nothing about: self-publishing a physical work planner with a hard ship deadline. Two weeks, zero print (Amazon KDP) experience. I expected it to feel different from building and scaling human-AI evaluation systems. It didn't. The same patterns showed up immediately: 1. AI generates fast; humans evaluate slow. I could get 10 cover concepts in 10 minutes. Choosing one took hours. It's the same bottleneck as any large-scale labeling program. Evaluation is the constraint. 2. Iteration beats “perfect.” I launched ads, watched the data disappoint, and adjusted based on the results. It’s the same deploy-measure-fix loop we use in model development. 3. Domain knowledge is the only defense against hallucinations. AI gave me confidently wrong guidance on Amazon ads. With no baseline, I believed it initially and shipped bad campaigns. The lesson is universal: If you can't evaluate outputs critically, you drift. The takeaway: Generating options is now a commodity. Knowing what “good” looks like (and having the rigor to validate it) is the hard part. If you’re a remote or hybrid worker and use a physical planner, I’d welcome some blunt feedback from real users. Comment or DM and I’ll follow up. Planner link: https://a.co/d/cp4nKPJ #AI #HumanInTheLoop #HumanFeedback #RLHF #ProductManagement

25

4 Comments

Very interesting, and I agree that is is a process that needs to be re-thought for quite a long time (even pre-AI times)

1

MMM isn't broken. The way most people use it is broken. Bi-annual models aren’t useful when you’re in a dynamic environment that requires weekly optimization. But that's not an MMM problem. That's an implementation problem. What we've learned running these models for clients: → MMM runs that align with budget planning cycles beat arbitrary bi-annual reports. → MMM should inform budget allocation ranges, not dictate exact spend levels. → Incrementality and geo tests should validate and calibrate your MMM, not compete with it. →  Marketing models aren't built in a lab. There's noise. Use a portfolio of tests to guide triangulation tuning, not a single result.

14

2 Comments

G, Tyler and I have been tackling the problem of predicting future home shopping events: both near term browsing prediction and longer horizon attribute prediction of your eventual home selection. We have published a blog post: Learning the language of home buying: Zillow’s event-based foundational user behavior model. We built a custom real estate vocabulary (Zokens) and train a generative model to continue the user's sequence, like an LLM. https://lnkd.in/gqQrWQZp

32

1 Comment

Delivering bad news well is a skill that can move you from "data person" to trusted partner. When you find a miss, it's tempting to fall into an easy pattern: "Unfortunately, the results are bad" ...Now everyone's imagining the worst-case scenario. "We missed the target, but there were a lot of factors outside our control" ...This sounds like you're being defensive and making excuses. "There's some mixed results, but let me walk you through the rest of the story first." ...No-one is listening to the "rest of the story." They're waiting for the other shoe to drop. Handling it is uncomfortable, but straightforward. State the size of the miss clearly and calmly in one sentence, move to why and what the options are: "This didn't land where we wanted. We came in at 72% of the target. I'll walk you through what drove it and what options we have." Stating it like this answers the question everyone is asking, avoids blame, and sets people up to hear the explanation that follows. Your role isn't to have a great answer or a perfect fix, it's to make it easier for everyone to decide on next steps. Stating it clearly and early shifts the conversation from "Why did this go wrong?" to "What do we do next?" That shift makes you the person people want to deliver bad news - clear, calm, focused on what we can do next. Don't apologize. Own it. I wrote a short playbook for moments like this. How to talk about misses, trade-offs, and next steps without making things worse. Link’s in the comments.

79

27 Comments

One of the fun things about LLMs is we can use their greatest weakness in OCR/HTR (their non-determinism) to be their greatest strength in identifying potential errors. The same model running over the same image multiple times can tell us where the model diverges most consistently, effectively telling us where the likely errors appear. Blog post and app coming soon!

100

8 Comments