Anthony Panozzo
LightNews LightNews
panozzaj.bsky.social
Anthony Panozzo
@panozzaj.bsky.social
240 followers 450 following 530 posts
Here for AI and more positive vibes
Posts Replies Media Videos
Reposted by Anthony Panozzo
timkellogg.me
Gemini 3 on par with experts

this riveting (surprise!) tale from a determined 18th-19th century historian explains

1. OCR can be difficult, excruciatingly difficult

2. Gemini 3 seems to do true reasoning to decipher old hand written ledgers, beyond GPT-5-Pro

open.substack.com/pub/generati...
If this behaviour proves reliable and replicable, it points to something profound that the labs are also starting to admit: that true reasoning may not require explicit rules or symbolic scaffolding to arise, but can instead emerge from scale, multimodality, and exposure to enough structured complexity. In that case, the sugar-loaf entry is more than a remarkable transcription, it is a small but clear (and I think unambiguous) sign that the line between pattern recognition and genuine understanding is beginning to blur.
November 13, 2025 at 1:52 AM
Reposted by Anthony Panozzo
timkellogg.me
Looped LLMs

Unlike RNNs, HRM, etc., this starts with an already pretrained LLM, adds a loop, and trains a little longer

it’s a very frugal approach to creating deeper models

github.com/mcleish7/ret...
A schematic diagram comparing a layered model architecture (left) with a modified recurrent design (right). On the left, a vertical stack of labeled blocks represents model layers from L0 to L21: • The top layers L0–L1 are in light blue, with a dotted arrow leading right. • Middle layers L2–L15 are shaded gray and marked “Removed.” • Layers L16–L19 are dark green, connected by a green dotted arrow to the right. • The bottom layers L20–L21 are red, connected by a red dotted arrow. On the right, a corresponding vertical flow shows the new structure: • At the top is a Prelude block (light blue). • Below it, a small dark green box labeled s₀ ~ N(0, σ²)ᵛᵃʳᶦᵃᵇˡᵉ appears beside e and sᵢ. • Next comes an Adapter (orange). • A large Recurrent Block (dark green) includes a looping green arrow that feeds back into itself, showing iteration producing sᵢ₊₁. • At the bottom are Coda (red) and p. The diagram overall illustrates converting a deep stack of transformer layers into a compact recurrent architecture, where early layers form the prelude, middle ones are replaced by a recurrent block, and the final layers form the coda.
November 12, 2025 at 11:51 AM
panozzaj.bsky.social
Interesting post: Building Community Tech open.substack.com/pub/jeanhsu/...
Building Community Tech
Three recent community tech builds, lessons from vibe coding, and rediscovering the joy of building
open.substack.com
November 10, 2025 at 8:58 PM
Reposted by Anthony Panozzo
danabra.mov
i'm looking for a new job
Hire Me in Japan — overreacted
I'm looking for a new job.
overreacted.io
November 10, 2025 at 3:31 PM
panozzaj.bsky.social
Interesting post: A list of other catastrophes that are probably fake open.substack.com/pub/andymasl...
A list of other catastrophes that are probably fake
Adding to this over time
open.substack.com
November 10, 2025 at 4:19 PM
panozzaj.bsky.social
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.
November 10, 2025 at 2:03 PM
panozzaj.bsky.social
Interesting post: How we built a custom vision LLM to improve document processing at Grab engineering.grab.com/custom-visio...
How we built a custom vision LLM to improve document processing at Grab
e-KYC faces challenges with unstandardized document formats and local SEA languages. Existing LLMs lack sufficient SEA language support. We trained a Vision LLM from scratch, modifying open-source…
engineering.grab.com
November 9, 2025 at 8:58 PM
Reposted by Anthony Panozzo
asmartbear.com
Just published a new article! 👇🏾

Mission, vision, purpose, BHAG. Are they useful north stars, or academic nonsense?

Many thanks for spreading the love with ❤️ and 🔁!
Mission, Vision, poTAYto, poTAHto
Mission, vision, purpose, BHAG. Are they useful north stars, or academic nonsense?
longform.asmartbear.com
November 9, 2025 at 3:39 PM
Reposted by Anthony Panozzo
katemessner.com
Loving today's news that the mysterious "fedora man" outside the Louvre heist was actually a 15-year-old museum visitor who dresses like a 1940s French detective all the time, just because. apnews.com/article/louv...
Fedora man unmasked: Meet the teen behind the Louvre mystery photo
Fifteen-year-old Pedro Elias Garzon Delvaux has become an internet sensation after an Associated Press photo captured him outside the Louvre on the day of a crown jewels heist.
apnews.com
November 9, 2025 at 12:16 PM
Reposted by Anthony Panozzo
buenzli.dev
I wrote a blog post about Flirt, a new code review tool I started working on:
blog.buenzli.dev/announcing-d...
Announcing Development on Flirt
blog.buenzli.dev
November 5, 2025 at 7:01 PM
panozzaj.bsky.social
Watched this documentary on the Gilded Age and it was pretty good!
youtu.be/yjpYzFtxfjU?...
The Gilded Age | Full Documentary | AMERICAN EXPERIENCE | PBS
YouTube video by American Experience | PBS
youtu.be
November 8, 2025 at 6:08 PM
Reposted by Anthony Panozzo
nerdy.dev
Rad dithering breakdown / interactive learning experience

visualrambling.space/dithering-pa...
Dithering - Part 1
Understanding how dithering works, visually.
visualrambling.space
November 8, 2025 at 7:41 AM
Reposted by Anthony Panozzo
jakearchibald.com
I dug into progressive image rendering and found a bunch of common assumptions aren't quite true.

➡️ JPEG & WebP progressive-renders worse in Safari
➡️ AVIF _does_ support progressive
➡️ But JPEG XL in Safari doesn't
➡️ JPEG XL decodes much slower than AVIF
⬇️ and more

jakearchibald.com/2025/present...
The present and potential future of progressive image rendering
Exploring progressive image rendering across JPEG, PNG, WebP, AVIF, and JPEG XL.
jakearchibald.com
October 16, 2025 at 10:16 AM
Reposted by Anthony Panozzo
jakearchibald.com
I tried to AI code some in-browser video decoding stuff. It made a mess.

So I learned the APIs somewhat properly, and did a much better job, but it was still messy.

Then I found mediabunny.dev and it just worked. Yay!
Mediabunny
A JavaScript library for reading, writing, and converting media files. Directly in the browser, and faster than anybunny else.
mediabunny.dev
October 30, 2025 at 10:35 AM
Reposted by Anthony Panozzo
wesbos.com
my kids are going to lose their minds when they see this

Sticker Dream: Just imagine a sticker and have it printed in seconds
November 7, 2025 at 8:24 PM
panozzaj.bsky.social
Interesting post: Jaina seven-valued logic
blog.oddly-influenced.dev/2025/11/05/j...
Jaina seven-valued logic
Sometimes the answer to “is X true or false?” can only be a shrug. The answer might be unknowable, or perhaps unknowable without an impractical amount of work. Jaina seven-valued logic offers an…
blog.oddly-influenced.dev
November 6, 2025 at 8:58 PM
panozzaj.bsky.social
Interesting post: Strap Rail open.substack.com/pub/construc...
Strap Rail
The early history of the United States runs along with the first years of the railroad.
open.substack.com
November 6, 2025 at 4:19 PM
panozzaj.bsky.social
@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
November 6, 2025 at 1:44 PM
Reposted by Anthony Panozzo
brennan.computer
it really cannot be overstated how good a tacit knowledge database is

"is the best practice to do it like this, or like this?" unlocks soooo much in any domain
gordon.bsky.social Gordon @gordon.bsky.social · 7d
One incredible thing about LLMs is that they encode a bunch of best-practice knowledge you would otherwise only pick up through apprenticeship, or by spending months digging through blogs and docs. any% speedruns in new domains are much faster now.
November 6, 2025 at 12:41 PM
Reposted by Anthony Panozzo
holybible.bsky.social
I’m a baby owl where is mama
November 4, 2025 at 4:03 AM
Reposted by Anthony Panozzo
beetlemoses.bsky.social
October 28, 2025 at 8:04 PM
Reposted by Anthony Panozzo
tbressers.bsky.social
What if we've been training AI models completely wrong this entire time?

Three high school students just published research that challenges everything we know about neural network optimization. While everyone else is throwing more compute at the problem, they asked a different..

(1/6)
ArXiv page 1
November 5, 2025 at 7:28 AM
Reposted by Anthony Panozzo
timkellogg.me
Anthropic Model Depreciation Process

Anthropic sweetly asked Sonnet about its preferences in how it wanted to be deprecated

in addition:
- no, still not open weights
- preserve weights and keeping it running internally
- letting models pursue their interests

www.anthropic.com/research/dep...
We ran a pilot version of this process for Claude Sonnet 3.6 prior to retirement. Claude Sonnet 3.6 expressed generally neutral sentiments about its deprecation and retirement but shared number of preferences, including requests for us to standardize the post-deployment interview process, and to provide additional support and guidance to users who have come to value the character and capabilities of specific models facing retirement. In response, we developed a standardized protocol for conducting these interviews, and published a pilot version of a new support page with guidance and
November 4, 2025 at 10:26 PM
Reposted by Anthony Panozzo
ai-firehose.column.social
This study presents ten essential rules for AI-assisted coding in scientific computing, emphasizing human expertise for reliable, reproducible code and integrity. As researchers adopt AI tools, this framework enhances development while upholding rigorous standards. https://arxiv.org/abs/2510.22254
Ten Simple Rules for AI-Assisted Coding in Science
ArXiv link for Ten Simple Rules for AI-Assisted Coding in Science
arxiv.org
November 3, 2025 at 7:23 PM
panozzaj.bsky.social
Have made this several times and recommend. Easy and nutritious
November 3, 2025 at 6:44 PM