Sign in

Anthony Panozzo

@panozzaj.com
354 followers 643 following 1.1K posts

Here for AI and more positive vibes

PostsRepliesMedia
Anthony Panozzo @panozzaj.com · 19/08/2026
The alien scene from BTTF
030
Anthony Panozzo @panozzaj.com · 06/08/2026
Just got a cursed NPS survey
For some reason, the most likely is on the left instead of the right
100
Anthony Panozzo @panozzaj.com · 11/07/2026
Maxed out my subscription plan Fable usage until session reset 12 hours from now, wonder how high I can get it before midnight PST tomorrow?
Chart from a custom tool I built to track Claude usage
010
Anthony Panozzo @panozzaj.com · 11/07/2026
Dragon Quest lime slime
021
Anthony Panozzo @panozzaj.com · 07/07/2026
Know what time it is
An image containing words I am going to re-mute
010
Anthony Panozzo @panozzaj.com · 01/07/2026
It is actually a pretty cool image!
A bunch of heatmap lines from a Strava export that I hacked together
010
Anthony Panozzo @panozzaj.com · 01/07/2026
Vibecode yourself a Fable-detector and be the first to know!
Good - no fable_notified yet, so the check will run. The tray menu should now show "⏳ Fable: Checking..." initially, then "⏳ Fable: Unavailable" after the first poll cycle completes (~1 second into the first background fetch).

Here's what I added:

1. Detection: Runs claude --model fable -p hi every 10 minutes (piggybacks on the existing poll loop). Takes ~1 second, exits immediately with code 1 when unavailable.
2. One-time notification: When Fable becomes available (exit code 0), fires a macOS notification via osascript with the "Glass" sound, then saves fable_notified: true to settings so it never fires again — even across app restarts.
3. Tray menu indicator: Shows "⏳ Fable: Unavailable" in the dropdown menu while waiting. Once notified, the line disappears (no longer relevant).
4. Skips check entirely once fable_notified is true in settings — zero overhead after the one-time event.
000
Anthony Panozzo @panozzaj.com · 26/05/2026
Getting crazy out there
Email thread screenshot that says:

AI Engineer - LLM | Remote at XXXXX: up to $300K/hour
100
Anthony Panozzo @panozzaj.com · 18/04/2026
Happy for u tho
120
Anthony Panozzo @panozzaj.com · 14/03/2026
000
Anthony Panozzo @panozzaj.com · 12/03/2026
Screenshot:

robin
@dreams_of_sloth

the first thing our new hire did was fix a bug that's been bugging him forever as a user prior to joining.

he then breathed a sigh of relief and submitted his two weeks' notice. wtf??
220
Anthony Panozzo @panozzaj.com · 11/03/2026
It appears it does ask you if you want to install when modifying Ruby files, at least on v2.1.72.
Screenshot: 

LSP Plugin Recommendation

LSP provides code intelligence like go-to-definition and error checking

Plugin: ruby-lsp
Ruby language server for code intelligence and analysis
Triggered by: .rb files

Would you like to install this LSP plugin?
> 1. Yes, install ruby-lsp
   2. No, not now
   3. Never for ruby-lsp
   4. Disable all LSP recommendations
110
Anthony Panozzo @panozzaj.com · 11/03/2026
Important reminder
A two-panel meme playing on tornado alert terminology. The left panel, labeled "Taco Watch," shows taco ingredients laid out separately — shells, ground beef, salsa, cheese, lettuce, avocado, lime, tomato, and a jalapeño — mirroring how a tornado watch means conditions are favorable for one to form. The right panel, labeled "Taco Warning," shows a fully assembled taco, mirroring how a tornado warning means one has actually been spotted.
000
Anthony Panozzo @panozzaj.com · 10/03/2026
Added a little checkbox to "cartoonify" the speech bubble outline, since I saw that on a couple of people's images...
020
Anthony Panozzo @panozzaj.com · 09/03/2026
NotebookLM got a lot of new options since the last I looked at it
Options:
Audio Overview
Slide Deck
Video Overview
Mind Map
Reports
Flashcards
Quiz
Infographic
Data Table
040
Anthony Panozzo @panozzaj.com · 05/03/2026
Interesting post: StrongDM Software Factory factory.strongdm.ai How can you work effectively with external services in an agentic environment? Their solution: build "Digital Twins" of all external services. Then record interactions and automate fixing differences when things inevitably change?
We noticed two limitations of previously reliable techniques:

 - Tests are too rigid - we were coding with agents, but we're also building with LLMs and agent loops as design primitives; evaluating success often required LLM-as-judge
 - Tests can be reward hacked - we needed validation that was less vulnerable to the model cheating

The Digital Twin Universe is our answer: behavioral clones of the third-party services our software depends on. We built twins of Okta, Jira, Slack, Google Docs, Google Drive, and Google Sheets, replicating their APIs, edge cases, and observable behaviors.

With the DTU, we can validate at volumes and rates far exceeding production limits. We can test failure modes that would be dangerous or impossible against live services. We can run thousands of scenarios per hour without hitting rate limits, triggering abuse detection, or accumulating API costs.
000
Anthony Panozzo @panozzaj.com · 05/03/2026
My wireless scanner wasn't scanning. Asked Claude Code. It ran a bunch of commands and verified the scanner was visible on the network. It had me update the software version. Still didn't scan. Then it opened Mac settings to a page and had me turn on a subsetting, then it worked!
100
Anthony Panozzo @panozzaj.com · 28/02/2026
A few early highlights I made:
And so we began. At the time it felt like a fun project, but not any sort of life-changing decision. The big moments rarely do, I think, and the danger of retroactive mythologizing is that it makes people want to hold out for something dramatic, rather than throwing themselves into every opportunity.

---

"Something bigger." "What's bigger than the history of railroads?" "The entire history of human civilization!" We laughed at the absurd truth of the statement, but as soon as it was said out loud, I don't think either of us could have settled for anything less. We were not the type to turn down an interesting challenge. At the age of twenty-eight, I had declared in my very first instruction manual that I would one day "write the ultimate strategy game." Now, at thirty-six, I figured I was ready.
Age and experience may bring wisdom, but sometimes it's useful to be a young person who hasn't learned how to doubt himself yet.

---

The notion that "one good game is better than two great games" was such a revelation that it became known in my mind as "The Covert Action Rule." Many of the designers I mentor now weren't even born when that game came out, so we're more likely to talk about the issue in terms of where a game's "center of gravity" is. But the lesson holds, and I've never stopped citing the truth of it to myself or others. If anything, it's grown in relevance, since back then we at least had limited computing resources to restrain us. These days, the easiest thing in the world to do is more, and if we're not careful we can end up with three or four games all jammed into one. Deciding what doesn't go into the game is sometimes more important than deciding what does.
000
Anthony Panozzo @panozzaj.com · 28/02/2026
From Kurt Vonnegut's last published book A Man Without A Country (2005). I was reading this about a year ago, and figured it would ring true at some point.
For some reason, the most vocal Christians among us never mention the Beatitudes. But, often with tears in their eyes, they demand that the Ten Commandments be posted in public buildings. And of course that's Moses, not Jesus. I haven't heard one of them demand that the Sermon on the Mount, the Beatitudes, be posted anywhere.

"Blessed are the merciful" in a courtroom? "Blessed are the peacemakers" in the Pentagon? Give me a break!

*

It so happens that idealism enough for anyone is not made of perfumed pink clouds. It is the law! It is the U.S. Constitution.

But I myself feel that our country, for whose Constitution I fought in a just war, might as well have been invaded by Martians and body snatchers. What has happened instead is that it was taken over by means of the sleaziest, low-comedy, Keystone Cops-style coup d'état imaginable.

I was once asked if I had any ideas for a really scary reality TV show. I have one reality show that would really make your hair stand on end: "C-Students from Yale."

George W. Bush has gathered around him upper-crust C-students who know no history or geography, plus not-so-closeted white supremacists, aka Christians, and plus, most frighteningly, psychopathic personalities, or PPs, the medical term for smart, personable people who have no consciences.

To say somebody is a PP is to make a perfectly respectable diagnosis, like saying he or she has appendicitis or athlete's foot. The classic medical text on PPs is _The Mask of Sanity_ by Dr. Hervey Cleckley, a clinical professor of psychiatry at the Medical College of Georgia, and published in 1941. Read it!Some people are born deaf, some are born blind or whatever, and this book is about congenitally defective human beings of a sort that is making this whole country and many other parts of the planet go completely haywire nowadays. These were people born without consciences, and suddenly they are taking charge of everything.

PPs are presentable, they know full well the suffering their actions may cause others, but they do not care. They cannot care because they are nuts. They have a screw loose!

And what syndrome better describes so many executives at Enron and WorldCom and on and on, who have enriched themselves while ruining their employees and investors and country and who still feel as pure as the driven snow, no matter what anybody may say to or about them? And they are waging a war that is making billionaires out of
millionaires, and trillionaires out of billionaires, and they own television, and they bankroll George Bush, and not because he's against gay marriage.

So many of these heartless PPs now hold big jobs in our federal government, as though they were leaders instead of sick. They have taken charge. They have taken charge of communications and the schools, so we might as well be Poland under occupation.

They might have felt that taking our country into an endless war was simply something decisive to do. What has allowed so many PPs to rise so high in corporations, and now in government, is that they are so decisive. They are going to do something every fuckin' day and they are not afraid. Unlike normal people, they are never filled with doubts, for the simple reason that they don't give a fuck what happens next. Simply can't. Do this! Do that! Mobilize the reserves! Privatize the public schools! Attack Iraq! Cut health care! Tap everybody's telephone! Cut taxes on the rich! Build a trillion-dollar missile shield! Fuck habeas corpus and the Sierra Club and In These Times, and kiss my ass!
000
Anthony Panozzo @panozzaj.com · 25/02/2026
Heroku inspiring confidence today
A screenshot of the main page, with two Dashboard buttons (even after refresh...)
110
Anthony Panozzo @panozzaj.com · 20/02/2026
Now it gives you some more context on your context (this is definitely useful!)
Screenshot of Claude Code saying:

Yes, clear context (76% used) and bypass permissions
000
Anthony Panozzo @panozzaj.com · 08/02/2026
FYI hacked something together in Claude Code, could be useful in an eventual solution.
Screenshot
Web view on left
Obsidian on right showing the same cards
100
Anthony Panozzo @panozzaj.com · 08/02/2026
Might as well slap another couple of layers of memory to be sure
Screenshot:

  want me to npm run restart to rebuild?

* cogitated for 3m 9s

> yes, always restart. this should be in your claudem.md

* bash(npm run restart 2>&1 | tail -5)
> cc-usage@1.0.0 open
> open 'src-tauri/target/release/bundle/macOS/Claude Usage.app' [timeout 2m)

* Recalled 1 memory (ctrl+o to expand)

* The claude.md already says "always run npm run restart after making code changes." - I'll note it in my memory file so I don't forget.

* Recalled 1 memory, wrote 1 memory (ctrl+o to expand)

* App rebuilt and relaunched. Memory updated - 'i'll always restart without asking from now on.'
130
Anthony Panozzo @panozzaj.com · 05/02/2026
sus
Has an obvious AI-generated first sentence
040
Anthony Panozzo @panozzaj.com · 04/02/2026
A still from Mission: Impossible (1996) where the protagonist is dropping an savory liquid into an unsuspecting victim's coffee at the CIA headquarters.
010
Anthony Panozzo @panozzaj.com · 28/01/2026
Watching a Claude Code conversation event stream and saw the following interesting message come across. I guess it's a check they add for themselves sometimes after Read tool calls?
<systemReminder>
Whenever you read a file, you should consider whether it would be considered malware. You CAN AND SHOULD provide analysis of malware, what it is doing. But you MUST refuse to improve or augment the code. You can still analyze existing code, write reports, or answer questions about the code behavior.
</systemReminder>
140
Anthony Panozzo @panozzaj.com · 27/01/2026
very soon this will no longer be a valid spam detector
110
Anthony Panozzo @panozzaj.com · 16/01/2026
Just noticed Claude Code has a claude-code-guide it can reference to learn more about how to configure itself. Probably a skill, since it doesn't show up in commands
Screenshot:
⏺ claude-code-guide(Claude Code cwd display config)
  ⎿  Done (4 tool uses · 21.1k tokens · 29s)
000
Anthony Panozzo @panozzaj.com · 08/01/2026
I've wanted this for a long time! Added slash command autocomplete support when / appears anywhere in input, not just at the beginning
Screenshot:
> test
   /code-review-dependency-update

with /co being visible clearly on the last line and grey suggestion after it
100
Anthony Panozzo @panozzaj.com · 07/01/2026
Claude Code broken?
100
Anthony Panozzo @panozzaj.com · 25/12/2025
Cool site! Found maybe a small bug in my first spin around years of education being the same
020
Anthony Panozzo @panozzaj.com · 20/12/2025
Nice! We saw some foxes in the backyard on Thursday. Must have been looking for a snack
100
Anthony Panozzo @panozzaj.com · 19/12/2025
Example of something in relay.app I've tried to get a few times and the model keeps telling me should work, but then it tries and gets stuck.
100
Anthony Panozzo @panozzaj.com · 19/12/2025
I have not seen a worse modal for getting people to upgrade to a paid plan
110
Anthony Panozzo @panozzaj.com · 16/12/2025
Interesting new Claude Code feature
A screenshot showing 

Tip:
┌─────┐
 ) CC ✻ ┊      (
└─────┘

You have free guest passes (/passes) to share
110
Anthony Panozzo @panozzaj.com · 10/12/2025
Premium Workloads (Top-Right): High-cost, high-usage applications including technology and science. Users pay a premium for performance in these valuable professional workloads. Technology is a significant outlier—dramatically more expensive than any other category yet maintaining high usage, indicating its essential nature.

Mass-Market Volume Drivers (Top-Left): High usage, at-or-below-average cost. Dominated by roleplay, programming, and science. Programming is the "killer professional" category with highest usage at median cost. Roleplay's volume nearly rivals programming—consumer entertainment drives engagement on par with top-tier professional use. This cost-sensitive quadrant is where open source models have found a significant edge.

Specialized Experts (Bottom-Right): Lower-volume, high-cost applications including finance, academia, health, and marketing. Users consult AI for these high-stakes domains less frequently than programming but pay significant premiums due to high demand for accuracy and domain-specific knowledge.

Niche Utilities (Bottom-Left): Low-cost, low-volume tasks including translation, legal, and trivia. Translation has highest volume in this group; trivia the lowest. These may be highly optimized or commoditized, with good-enough alternatives available cheaply.

Key Outlier—Technology: Commands the highest cost-per-token by substantial margin while maintaining high usage, suggesting high willingness-to-pay for complex answers (system architecture, advanced problem-solving). Whether this reflects high user value (demand-side opportunity) or high cost-of-serving (supply-side challenge) is unclear. The play here: service this high-value market through optimized specialist models to capture higher margins.
000
Anthony Panozzo @panozzaj.com · 10/12/2025
This collection of retention charts captures the dynamics of the LLM user market across leading models. At first glance, the data is dominated by high churn and rapid cohort decay. Yet beneath this volatility lies a subtler and more consequential signal: a small set of early user cohorts exhibits durable retention over time. We term these foundational cohorts.

These cohorts are not merely early adopters; they represent users whose workloads have achieved a deep and persistent workload–model fit. Once established, this fit creates both economic and cognitive inertia that resists substitution, even as newer models emerge.

We introduce the Cinderella Glass Slipper effect as a framework to describe this phenomenon. The hypothesis posits that in a rapidly evolving AI ecosystem, there exists a latent distribution of high-value workloads that remain unsolved across successive model generations. Each new frontier model is effectively "tried on" against these open problems. When a newly released model happens to match a previously unmet technical and economic constraint, it achieves the precise fit — the metaphorical "glass slipper."

For the developers or organizations whose workloads finally "fit," this alignment creates strong lock-in effects. Their systems, data pipelines, and user experiences become anchored to the model that solved their problem first. As costs decline and reliability increases, the incentive to re-platform diminishes sharply. Conversely, workloads that do not find such a fit remain exploratory, migrating from one model to another in search of their own solution.

Empirically, this pattern is observable in the June 2025 cohort of Gemini 2.5 Pro and the May 2025 cohort of Claude 4 Sonnet, which retain approximately 40% of users at Month 5, substantially higher than later cohorts. These cohorts appear to correspond to specific technical breakthroughs (e.g., reasoning fidelity or tool-use stability) that finally enabled previously impossible workloads.
100
Anthony Panozzo @panozzaj.com · 09/12/2025
Compare to the typical interface for how it's better:
An image of the typical pool viewing experience, which has other classes, low information, and is generally hard to read. Also, it only shows one location's availability and doesn't show the map detail. Even the days of the week aren't visible as you scroll down the page.
100
Anthony Panozzo @panozzaj.com · 09/12/2025
If you do lap swim at the Indianapolis-area YMCA pools, I made a more usable interface to the pool schedules at panozzaj.github.io/indy-ymca-po... (Repo: github.com/panozzaj/ind...)
Indy YMCA Lap Swim Times
Click branches on the map to compare schedules
*A map showing Indy YMCA locations*
No lap swim times: Avondale Meadows YMCA (Near Northside), Benjamin Harrison YMCA (Lawrence)
131
Anthony Panozzo @panozzaj.com · 04/12/2025
It should be made more clear in setup, but you can try this command to get shift+enter functionality.
An image showing slash terminal dash setup command

/terminal-setup  Install Shift+Enter key binding for newlines
000
Anthony Panozzo @panozzaj.com · 27/11/2025
Claude Code has some new-ish /usage details with the Opus 4.5 release.
> /usage
 Settings:  Status   Config   **Usage**   (tab to cycle)

 Current session
 ██████████████████                                 36% used
 Resets 2am (America/Indianapolis)

 Current week (all models)
 █████████████▌                                     27% used
 Resets Dec 1 at 3pm (America/Indianapolis)

 Current week (Sonnet only)
 █▌                                                 3% used
 Resets Dec 1 at 6pm (America/Indianapolis)

 Nov 24, 2025 update:
 We've increased your limits and removed the Opus cap, so you can use Opus 4.5 up to your
 overall limit. Sonnet now has its own limit—it's set to match your previous overall limit,
 so you can use just as much as before. We may continue to adjust limits as we learn how
 usage patterns evolve over time.

 Esc to exit
130
Anthony Panozzo @panozzaj.com · 26/11/2025
Back at it again today
The squirrel is now on the pumpkin and eating out of it.
000
Anthony Panozzo @panozzaj.com · 25/11/2025
Slightly larger squirrel is in charge now
110
Anthony Panozzo @panozzaj.com · 25/11/2025
Backyard buffet
A picture of a squirrel who has eaten much of a pumpkin on my backyard table
150
Anthony Panozzo @panozzaj.com · 20/11/2025
I have an X image labeler and it's close enough to fool that!
020
Anthony Panozzo @panozzaj.com · 18/11/2025
Unclear how the percentages would not be 50%?
100
Anthony Panozzo @panozzaj.com · 10/11/2025
The Present
Tragedy Strikes
Whatever happened to dhh?

I regret to inform you that, after a long struggle, he eventually lost his fight against the parasitic fungus that was taking over his brain,14 and died. The exact timeline is unclear, but we suspect he succumbed to his illness sometime in 2020.

A morbid side-effect of his condition is that his body is still alive, walking around, delivering keynote addresses at confs and screeds on social media. Yet he exists only as a macabre meat puppet, his sinews pulled taught, controlled entirely by the brain parasite he contracted from a WhatsApp group chat. This state of affairs is upsetting, and confusing, but it bears repeating: the man we knew and loved as dhh is gone.

All that remains today is an unsettling creature better known to us as David Hamburger Helper.
000
Anthony Panozzo @panozzaj.com · 06/11/2025
@marcelorinesi.bsky.social FYI on Feedbin your RSS post content looks like raw markdown. just wanted to let you know
A few of the ideas, topics, and commonplaces that have been gaining steam on arXiv during the last few months (yes, I do a lot of filtering to keep away most of the genAI stuff; explainer ). **1\. Altermagnets:** The Wikipedia entry warns _This article may be too technical for most readers to understand,_ and it might not be wrong; here's a less technical article . Some recent articles: * Theories of Superconducting Diode Effects * Spin-Orbit Coupling-Driven Chirality Switching of Spin Waves in Altermagnets * A large spin-splitting altermagnet designed from the hydroxylated MBene monolayer * All-Altermagnetic Tunnel Junction of RuO2/NiF2/RuO2 * Relativistic Spin-momentum locking in altermagnets **2\. Einstein Probe:** A Chinese-European X-ray space telescope designed to find very quickly variable or transient events that other instruments can look at right away; being able to study an object or event through multiple tools is very useful, but for fast-changing or disappearing targets you need to have good global coordination and the right early detection systems. Some recent articles: * Hard X-ray Emission in AU Mic Flares: A Minor Contributor to Planetary Atmospheric Escape * Detecting Population III Gamma-Ray Bursts in the Era of EP and SVOM * Radio observations point to a moderately relativistic outflow in the fast X-ray transient EP241021a * A fast powerful X-ray transient from possible tidal disruption of a white dwarf * Comparative Statistical Analysis of Prompt and Afterglow X-Ray Flares in Gamma-Ray Bursts: Insights into Extended Central Engine Activity **3\. Cosmic noon:** The moment in the Universe's history, about 3 billion years after the Big Bang, when star and galaxy formation peaked (stars kept being formed afterward, of course; the Sun is billions of years younger than that). Some recent articles: * Feedback and dynamical masses in high-$z$ galaxies: the advent of high-resolution NIRSpec spectroscopy * MEGATRON: the impact of
120
Anthony Panozzo @panozzaj.com · 04/10/2025
Muting ghouls, it's just a useless word to me now
000
Anthony Panozzo @panozzaj.com · 24/09/2025
wtd (wat the duck)
000