That number seemed too large to be true, and I think you might be conflating some numbers...

The NIH study about Economic Costs of Diabetes in the US in 2022 [1] states:

>For cost categories analyzed, care for people diagnosed with diabetes accounts for 1 in 4 health care dollars in the U.S., 61% of which are attributable to diabetes.

So, 25% of health care (not hospital) dollars in the US are apportioned to people with diabetes, and in 61% of those cases, the cause is directly attributable to diabetes itself.

Similarly, an NHS study [2] revealed that 60% of overall diabetes spend goes towards diabetes-related complications.

Finally, this presentation [3] states that: "hospital stays involving patients with diabetes contributed almost $83 billion or 23% of the total hospitalization costs in the United States," and, of course, not all of those were directly diabetes-related.

Obviously that is still a staggering amount of money spent on diabetes every year, and represents a huge burden on our healthcare system. But 60% for all hospital expenditure just doesn't track from my cursory Google-ing here.

[1] https://pubmed.ncbi.nlm.nih.gov/37909353/

[2] https://www.diabetes.org.uk/about-us/news-and-views/cost-com...

[3] https://www.bu.edu/csmet/2020/08/04/healthcare-data-analytic...

2. Own the Output. Everyone is encouraged to use these tools to accelerate the pace of their work and to take the toil out of their daily output. But that doesn’t mean you simply defer to the model and skip your own careful review step. Part of the paradox of these models is that they’re wordy and it can feel exhausting to read through everything they’re writing and confirm its validity. But that’s an enormous part of the job: you’re the human-in-the-loop to make sure what it’s saying is correct, and that you’ve reviewed it before you’ve shared it with the rest of the class. Read all of the words before they reach the rest of your team or a client, and make sure you stand behind having your name attached to them. In situations where a full review isn’t possible for expediency, disclose it directly! “Hey, I haven’t had a chance to review, this is pure Claude, but here’s what it says.” This isn’t a get out of jail free card, but it can be used occasionally when the situation dictates. Most of the time, you need to review and adjust the output before other people see it.

3. Be the expert. Have an opinion. This is a big one. Do not offload your role as expert to the model. The model is an assistant. You need to internalize the work being asked of you, and apply your education, instincts, and specific expertise to form your own opinion. You can use models to work to adversarially challenge things, but part of being an expert is knowing when to discard irrelevant feedback. Use this skill daily. Don’t let it atrophy.

4. Don’t Slop-Bomb. This is an amalgamation of all of the above, but again, don’t just spew large volumes of model slop into comms channels. Bring your own perspective, internalize the problem in front of you. And be judicious about what you share, and how wordy these things are. Put another way: edit for clarity and concision. It’s too easy to generate busy-work-seeming reports and outputs, and that burden shift to everyone else is multiplicative. Use fewer words, distill things into insights (backed with longer form deep dives), but make sure you’re not just making the rest of the team read a bunch of useless pablum. I don’t want to have to make my model read your model’s output to glean what I’m supposed to out of it. Do that for your coworkers and your clients. Think in terms of information hierarchy.

Write as yourself. You’re being paid for your expertise and insights. Communicate them directly to us. Copying and pasting Claude responses into Slack or an email directly shifts the burden of comprehension and understanding to everyone else, and worse, risks skipping that step for yourself. Even if you’re fundamentally using Claude to gather your thoughts or help you prepare a response, you need to be writing it yourself, in your own voice. Not having Claude ape your voice, or “make it sound less like AI”. You, directly. Doing this will further reinforce your own understanding of the state of things, the same way teaching someone is the best way to learn. As a guideline: for Slack and email comms, this should be near-universally written as you. For deliverables that are longer form and follow a template like proposals, roadmap/discovery work, etc., use of agents is expected but, see Principle #2. (Own the Output.)

Here's a friendly reminder you can play Police Quest in your browser, right now:

http://sarien.net/policequest#hq

A young 8-year-old me, playing Police Quest for the first time. Sits through the briefing. Gets the gun, the ammo, the briefcase, from the locker room. Grabs the car key. Heads out to the parking lot, gets in the car, starts driving... and immediately dies: "you failed to perform the prescribed vehicle walkaround."

A bit of a "gotcha" for sharewareing the game, because said prescription was in a manual I definitely did not own (remember the games that would ask you for, say, the third word of the tenth page of the manual, as a form of DRM? brilliant.)

NO PROBLEM: ASSIGNMENT RECEIVED: I WILL SAVE THE GAME EARLY AND OFTEN!

Begin another playthrough. Lather, rinse, repeat. Play for several hours, carefully save scumming my way to glory: I have built a complex ladder of 10 save slots that I am slowly rotating through. Then the action gets heavier and heavier, and the saves come faster and faster, until the climax of the game, where you have the Death Angel cornered. "Radio for help."

And there it says: "With what radio?"

The one I had left back at the station, I don't know, 10 hours ago? But that's okay, because I can head back to the station and grab it from one of my convenient saves!

No, no I cannot. I have over-indexed and let fear win the day and save scummed myself into oblivion. The earliest save is in a walking-dead state as well. I am toast.

I died a bit that day. A painful lesson for an 8-year-old to learn. As Bennett Foddy says in his Getting Over It soliloquy, "When games were new, they wanted a lot from you; daunting you, taunting you, resetting and delaying you. Players played stoically. Now everyone’s turned off by that."

Well, I suppose it's nearly the opposite of that experience, upon further review. But for some reason, that's where my head jumped.

Committing physical, in-person crimes anonymously has obviously always been possible: there are unsolved murders, thefts, and other crimes every day. But they require a great deal of personal risk to the criminal because the criminal has to physically put themselves into the act of committing the crime, along the path of getting to where the crime is, and has to face an opponent, if their crime is against another person.

Now, that can be sourced remotely, routed through anonymizing tools, VPNs, etc., and do a great deal to cover their tracks so that the "pretty good chance they go to jail" can be substantively minimized in a way we couldn't previously contemplate.

The idea that we should let the US based models be permissive because at least they'll be subject to subpoena power is fatuous: yes, strictly speaking, a user committing crimes on a permissive foreign model will be harder to catch, but non-sophisticated users who have never heard of hugging face may find that being blocked by the US model is enough for them to reconsider their behavior. A dedicated enough individual is going to commit the crime they're going to commit, but there are tons of situations where preventing trivial access to tools that can be used for malice can actually prevent malice from occurring.

And my point here is, we're in this great revolution of AI empowered everything and... these companies can't be bothered to scan this volume of AI ad slop to make sure we're not being actively scammed? Or having porn or CSAM spread? I understand it's a "perverse incentive" in the truest sense: they make money by taking the scammers money, but I was scammed once by an insanely clever lookalike Facebook ad, and had trusted their verified checkmark (about a decade ago)... with AI, these things should be a thing of the past.

Bananas are the #1 most-sold item at most grocery stores including, notably, Wal-Mart.

Bananas also have the highest standard deviation in terms of predicting if a given (known) consumer will purchase bananas in a given store run. (At least as compared to other food products and consumables.) When predicting a consumer's shop, it's generally pretty easy to make a highly educated guess about their purchasing activity and, thus, to project volumes for products. But bananas defy that wisdom, except that people in aggregate buy a lot of them. Someone who buys bananas reliably every week for months will randomly stop for months, and then start again, for no perceivable rhyme or reason. Bananas aren't seasonal purchases like berries or corn or other fruits or vegetables. Bananas also tend to be a high volume item at gas stations and convenience stores.

Bananas have to be effectively "tricked" into continuing to ripen after being prematurely picked green and then refrigerated for transit. So there are banana ripening centers that pump ethylene through a chilled chamber to get them to ripen.

I mean, feel free to run a "taking Swastika back" information campaign, but also, you know, good luck.