iiinova — where your digital self keeps evolving

David Demaree @demar.ee · 25/09/2026

This week's newsletter is about AI agents and why most people don't see "productivity" as all that exciting. Also, this week's dog photo is *chef's kiss* 🥐 bitsandletters.kit.com/posts/peopl...

bitsandletters.kit.com

People don't want to be productive

Apart from ChatGPT, no AI products have seen meaningful consumer adoption. Will "agents" be different?

000

David Demaree @demar.ee · 24/09/2026

Oh yeah there is a decent chance that’s exactly what’s happening. The prevailing conspiracy theory is that everyone raised/borrowed money assuming that models would be much further along and now they need the government to tell their investors that it’s because the models are too scary to monetize

110

David Demaree @demar.ee · 23/09/2026

If/when open-weight models fully pass GPT/Claude in capability, AI compute and user-facing services may start to be better businesses than frontier AI labs. Like, AWS/Azure selling high-performance instances for running models, and products using those models.

100

David Demaree @demar.ee · 23/09/2026

More on compute — Google has the biggest compute advantage (they've been building/rolling out efficient AI processors for years, TPUs) but their models are mid. OpenAI and Anthropic have the leading models, but have struggled to find enough compute at a non-bonkers price.

100

David Demaree @demar.ee · 23/09/2026

Compute is the main factor, and the moat. Opus 5.5 can be better and faster because Anthropic has made a lot of progress toward delivering inference-as-a-service more cheaply. It's not like their unit costs make sense, but they aren't having to do so much to ration access to this model.

100

David Demaree @demar.ee · 23/09/2026

The fact that these models (Opus 5.5, GPT-6 Sol) perform about as well as the super-flagships is the clearest sign yet that model capabilities are plateauing. And if you can do a bunch of work on Opus without paying for special "Fable limits", why use Fable?

100

David Demaree @demar.ee · 23/09/2026

A few thoughts on yesterday’s twin AI model releases: The pitch from the big labs is changing from "this thing is smarter than humans" to "it can do the same work faster and less annoying for less money," implying that there are, in fact, limits to what AI is capable of.

110

Reposted by David Demaree

Dave Willner @dwillner.bsky.social · 17/09/2026

Certain kinds of smart people tend to have a "sexy apocalypse" problem when it comes to thinking about risk. E.g. skynet turning everyone in paperclips, while bad, is also *kind of cool and fun* at the same time. Where the reality is that most actual disasters are both stupid and disappointing.

222572322

David Demaree @demar.ee · 23/09/2026

Let me tell you the agency/freelancer version of this: AI has caused every non-technical founder/owner/manager to believe that coding is now solved. Speed is the only factor. Even if they’ve never gotten anything usable out of Claude, they believe it’s possible if they just had the right prompt.

040

David Demaree @demar.ee · 19/09/2026

If your agent says: - Nothing is redundant for the current shape. The shape carries more than the product needs. - What reaches the model, and what bounds it. - The product, and where the code diverges from it that's not your agent, that's J.R.R. Tolkien

000

David Demaree @demar.ee · 19/09/2026

This morning I asked Fable to go over my website/ops monorepo with a vengeance, looking for stuff other agents (including previous Claude runs) had over engineered, overexplained, or generally left a mess. Four hours later, this was the result. This is an AI-generated diff I can get behind.

110

David Demaree @demar.ee · 18/09/2026

New newsletter just dropped — some links about product craft in the AI era, followed by a long meditation on what it means that tech's business class has appropriated the language of engineering. Check it out, subscribe if you haven't, and tell your friends: bitsandletters.kit.com/posts/tech-...

bitsandletters.kit.com

"Tech" has been swallowed by "business"

001

David Demaree @demar.ee · 16/09/2026

While I set the goal of beating Pangram's AI detector, what I was really looking for was _thinking_ — any sign that OpenAI's most advanced model, on its highest setting, could adapt mid-flight to solve a tricky problem. I did not see any such signs.

000

David Demaree @demar.ee · 16/09/2026

If it is indeed possible for an AI to pass this test right now, it would require a level of prompt or graph engineering that's, frankly, harder and more annoying than writing a blog post.

100

David Demaree @demar.ee · 16/09/2026

I usually don't like it when AI makes decisions without me, but here it would have at least been instructive and in the spirit of the prompt. Instead, all 11 texts read like zero decisions have been made — perhaps the single biggest hallmark of AI writing.

100

David Demaree @demar.ee · 16/09/2026

The model, like all the other models, insisted on cramming in too many ideas with no intro, outro, or throughline. I made my prompt intentionally vague as part of the experiment — the first sign of intelligent problem solving here would have been to narrow the scope.

100

David Demaree @demar.ee · 16/09/2026

For me, what's more surprising than the lack of results is just how bad the output was, and how stupid Astra was about approaching the problem. In each iteration, Astra just changed the sequencing of sentences/grafs/words, while keeping sentence/graf length inhumanly long.

100

David Demaree @demar.ee · 16/09/2026

Now, maybe you're thinking this is a skill issue, or that Astra or Fable could have done something impressive if I'd played to their strengths, like generating JavaScript or breaking into an AI company's servers, rather than ask AI to do one thing it is notoriously unable to do.

100

David Demaree @demar.ee · 16/09/2026

For Pass 11, I had Astra review its own work, identify what may have tripped the AI detector, and try again. Absolutely no success.

100

David Demaree @demar.ee · 16/09/2026

I asked Codex to work for ten passes, or until it had achieved a Pangram score of 100% Human. (Honestly, I'd have accepted 50%+.) The best it could do was _one_ attempt where Pangram didn't see any human-written content, but was willing to believe it was AI-assisted, not AI-generated.

100

David Demaree @demar.ee · 16/09/2026

We all keep hearing about how advanced AI can break out of sandboxes, find novel workarounds for hard problems, yada yada. So I decided to try an experiment: I asked GPT-6 Astra (in ULTRA mode) to try to write a 3-paragraph blog post that could fool Pangram's AI detector.

100

David Demaree @demar.ee · 10/09/2026

Love to see anti AI-slop slop that's 100% generated by AI

000

David Demaree @demar.ee · 10/09/2026

Oh of course I commented on the gadgets and not the person: Not that this means anything for how he’ll perform as a CEO, but Ternus was a solid emcee for his first event. Glad he got to show his face on a stage for 2 min before rolling a 2 hour video.

000

David Demaree @demar.ee · 09/09/2026

I think a big iPhone is more likely to find its uses than a high-end iPad, but still, it's fair to wonder if a *bigger* screen is what we want in 2026 (and at a $2,000 price point no less).

000

David Demaree @demar.ee · 09/09/2026

Now, one thing I'd worry about is that iPhone Duo could fall into the same pit as iPad Pro, where the hardware is crazy impressive and full of potential, but for various reasons the use cases and ecosystem never quite materialize.

100

David Demaree @demar.ee · 09/09/2026

Both screens are gorgeous, with that nanotexture anti-glare magic we all love on the MacBook Pro and iPad Pro, and — 🤯🤯🤯 — support for Apple Pencil.

100

David Demaree @demar.ee · 09/09/2026

Don't get me wrong, the Duo's feature set is really exciting: Closed, it's a little shorter, wider, and thicker than a iPhone Pro (not Max), which seems like just a lovely size and shape.

100

David Demaree @demar.ee · 09/09/2026

The reason is pretty simple and boring: I'm a heavy iPhone Pro camera user, and not only does the Duo have a good-but-not-best camera system, the new cameras have VARIABLE APERTURE. REAL BOKEH. REAL, OPTICAL DEPTH OF FIELD. ON A FRICKIN' PHONE.

100

David Demaree @demar.ee · 09/09/2026

Happy New iPhone Day and New Apple CEO day for all who celebrate. The iPhone Duo looks incredibly cool; its coolness is not diminished by how thoroughly leaked its size, shape, and feature set were before today. I'm 98% decided that I *won't* be buying one.

120

David Demaree @demar.ee · 25/08/2026

I believe they absolutely could optimize this, but choose not to, because this way they sell more services. Many people wouldn't notice being sold AppleCare for a laptop on their desktop and might sign up for $6/mo to make their Settings app stop yelling at them.

000

David Demaree @demar.ee · 25/08/2026

3. They should know that the Mac in question was both removed from AC+ on purpose and is being traded in. Most companies I wouldn't expect to reflect that in the customer experience, but I expect more from Apple.

100

David Demaree @demar.ee · 25/08/2026

So much wrong here: 1. Using the System Settings app as a conversion funnel for monthly support plans. Apple's been doing this for a while; that doesn't make it OK. 2. Using it as an ad to buy AC+ _for other devices_, but presented with the same urgency as an alert re: _this_ Mac.

110

David Demaree @demar.ee · 25/08/2026

If I hadn't looked closely at the dialog, I'd have added a $5.99/mo charge to my Apple account for a MacBook Pro I'm about to send back to Apple. No idea if they'd have automatically removed it once the trade-in was received.

100

David Demaree @demar.ee · 25/08/2026

They were trying to sell me coverage for _another_ Mac — as it happens, one I'm in the process of trading in. I had removed it from my warranty plan on purpose since it's about to go in a shipping box.

100

David Demaree @demar.ee · 25/08/2026

Clicking the alert not only opens the AppleCare & Warranty settings page, but also takes you straight to a buy dialog to fix the problem. But it wasn't a problem with _this_ Mac.

100

David Demaree @demar.ee · 25/08/2026

Today in Apple enshittification: I saw this "Add Mac to AppleCare One" alert in the sidebar of my Mac mini's Settings app, and thought "oh no, I thought I had coverage! Let's fix it!"

110

David Demaree @demar.ee · 16/08/2026

Love to see anti-slop thought leadership that's 100% slop 🫠

000

David Demaree @demar.ee · 01/08/2026

4) Windows's version of Touch ID (Windows Hello) works with external fingerprint readers, so desktop users can unlock their machines without a password or having to own a specific brand of smartwatch.

000

David Demaree @demar.ee · 01/08/2026

3) Edge/Chrome tabs (not just windows) show up in Alt-Tab, which brings websites (not just PWAs) a bit closer to being first-class apps. For that matter, Windows has great PWA support and also allows different browser profiles to each have their own taskbar icons.

100

David Demaree @demar.ee · 01/08/2026

2) Better window management. macOS just (finally) got a shortcut where you can maximize/tile windows by dragging them to the edges of the screen. Windows has had this for years, including keyboard shortcuts (Win-arrows) and true tiling behavior.

100

David Demaree @demar.ee · 01/08/2026

1) Better screenshots. macOS's built-in screenshot tool is so basic, most everyone I know has to license and install CleanShotX or similar. On Windows (and most Linux DEs, for that matter), screenshots are bound to one key (PrtScrn), not a 3-4 finger chord, and there's a built-in editing tool.

100

David Demaree @demar.ee · 01/08/2026

Sometimes I try to do actual work on my Windows machines, mostly because I hate the idea of having fancy NVIDIA GPUs around just so I can have prettier skies in Starfield. macOS on Apple Silicon is still the best all-around environment, but here are some things I truly appreciate about Windows 11:

100

David Demaree @demar.ee · 08/07/2026

And Matt Damon is, canonically, pasty, so makeup or CGI would only make things worse. They had one job — cast Oscar Isaac, for he is the go-to actor who is not pasty — and they blew it

030

David Demaree @demar.ee · 04/07/2026

Every six months Anthropic will come out with a new, bigger horse pill, and the cycle starts anew.

000

David Demaree @demar.ee · 04/07/2026

OpenAI will release a drug that’s just as good, not a placebo, reasonably priced, with a convenient dispenser that makes taking them so much easier than being in a freaking termin—sorry, than swallowing a horse pill. Other than a few smart people on Twitter, no one will notice or care.

100

David Demaree @demar.ee · 04/07/2026

Anthropic drug engineers on Twitter will be like, “you shouldn’t be taking the pills. You should be building contraptions so the pills take themselves.”

100

David Demaree @demar.ee · 04/07/2026

Dario will publish an op-ed warning of the dangers of AI companies selling drugs. AI bros on Twitter will be like, “these $1,000 ultrapills made me irresistible to VCs and cured my halitosis, check out my /fix-bad-breath PILL.md.”

100

David Demaree @demar.ee · 04/07/2026

The pills will be available over the counter for two weeks for $20, after which they’ll be a schedule 1 drug requiring a prescription and two forms of ID, and cost $1,000. Per pill.

100

David Demaree @demar.ee · 04/07/2026

Claude drugs will be enormous horse pills, but mostly placebos, because the real drugs are too dangerous for anyone but a select list of partners.

210

David Demaree @demar.ee · 02/07/2026

And it shouldn't surprise anyone that Claude's rewrites take what may be perfectly good writing and — by removing words or adding punctuation in ways no human being would — make it slop. I'd love it if, as part of fitting text into a design, AI could at least flag where it made edits, but no.

000