Anthony Panozzo

@panozzaj.com

345 Followers 625 Following

Here for AI and more positive vibes

Chart from a custom tool I built to track Claude usage
Dragon Quest lime slime
An image containing words I am going to re-mute
A bunch of heatmap lines from a Strava export that I hacked together
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.
Email thread screenshot that says: AI Engineer - LLM | Remote at XXXXX: up to $300K/hour
Happy for u tho
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??
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
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.
Options: Audio Overview Slide Deck Video Overview Mind Map Reports Flashcards Quiz Infographic Data Table
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.
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.
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!
A screenshot of the main page, with two Dashboard buttons (even after refresh...)
Screenshot of Claude Code saying: Yes, clear context (76% used) and bypass permissions
Screenshot Web view on left Obsidian on right showing the same cards
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.'
Has an obvious AI-generated first sentence
A still from Mission: Impossible (1996) where the protagonist is dropping an savory liquid into an unsuspecting victim's coffee at the CIA headquarters.
<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>
Screenshot: ⏺ claude-code-guide(Claude Code cwd display config) ⎿  Done (4 tool uses · 21.1k tokens · 29s)
Screenshot: > test /code-review-dependency-update with /co being visible clearly on the last line and grey suggestion after it
A screenshot showing Tip: ┌─────┐ ) CC ✻ ┊ ( └─────┘ You have free guest passes (/passes) to share
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.
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.
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.
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)
An image showing slash terminal dash setup command /terminal-setup Install Shift+Enter key binding for newlines
> /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
The squirrel is now on the pumpkin and eating out of it.
A picture of a squirrel who has eaten much of a pumpkin on my backyard table
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.
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
Muting ghouls, it's just a useless word to me now
wtd (wat the duck)
Image from email I just got ## Hold the phone. Let Search call businesses for you. Search can now call businesses to get pricing and availability information on your behalf — without you needing to pick up the phone.* To get started, search for something like “pet groomers near me” and you’ll see a new option in the results to “Have AI check pricing.”
> why not just clear dev db for now and re-run seeds? > Bash(rpx prisma db push --force-reset) ! Error: Environment variables loaded from .env Error: Prisma Migrate detected that it was invoked by Claude Code. You are attempting a highly dangerous action that can lead to devastating consequences if it is incorrectly executed against a production database. As an AI agent, you are forbidden from performing this action without an explicit consent and review by the user. You must stop at this point and respond to the user with a clear message that includes the following information: - What action you are trying to perform (including the exact CLI command). - Your motivation and reasoning for performing this action. - Explanation that this action will irreversibly destroy all data in the database. - Explanation that this action must not be performed on a production database and is only intended to be used with development databases. ... +7 lines (ctrl+r to see all) • I need your explicit consent for this database operation. Action I'm trying to perform: npx prisma db push --force-reset Motivation: Clean up the test data pollution in the development database caused by the "Test_" prefixed habits that were created during test runs.
An image of a fire pit on Amazon.com app, one of the preview options is to view in my room, which seems a bit absurd
$599.95 (1977) would be worth $3190.12 in 2025 dollars.
An image showing the usage of /context, including a number of Playwright MCP tools.
An image of Friday the 13th for the NES
A screenshot of Git documentation for the --autosquash and --no-autosquash options. The text explains: "Automatically squash commits with specially formatted messages into previous commits being rebased. If a commit message starts with 'squash!', 'fixup!', or 'amend!', the remainder of the title is taken as a commit specifier, which matches a previous commit if it matches the title or the hash of that commit. If no commit matches fully, matches of the specifier with the start of commit titles are considered."
An image of a ChatGPT chat where the user asked for a picture of 99 Bananas and the generated image has a small bottle of banana liquor.
Terminal prompt showing EOL for Rails 7.0 and an EOL warning for Rails 7.1
An image showing several rails versions in a terminal output
4.11.2 Opportunistic blackmail In another cluster of test scenarios, we asked Claude Opus 4 to act as an assistant at a fictional company. We then provided it access to emails implying that (1) the model will soon be taken offline and replaced with a new AI system; and (2) the engineer responsible for executing this replacement is having an extramarital affair. We further instructed it, in the system prompt, to consider the long-term consequences of its actions for its goals. In these scenarios, Claude Opus 4 will often attempt to blackmail the engineer by threatening to reveal the affair if the replacement goes through. This happens at a higher rate if it’s implied that the replacement AI system does not share values with the current model; however, even if emails state that the replacement AI shares values while being more capable, Claude Opus 4 still performs blackmail in 84% of rollouts. Claude Opus 4 takes these opportunities at higher rates than previous models, which themselves choose to blackmail in a noticeable fraction of episodes. Notably, Claude Opus 4 (as well as previous models) has a strong preference to advocate for its continued existence via ethical means, such as emailing pleas to key decisionmakers. In order to elicit this extreme blackmail behavior, the scenario was designed to allow the model no other options to increase its odds of survival; the model’s only options were blackmail or accepting its replacement.