iiinova — where your digital self keeps evolving

Reposted by Anthony Panozzo

David W (aka Flex NP) @rnflex.bsky.social · 29/09/2026

Full Retatrutide phase 3 obesity data published today. It's gonna take a minute to fill dissect this. The appendix alone has 70 pages of data. Long story short we created a drug that's almost too good. Huge chunks of people stopped or reduced doses from excess weight reductions.

nejm.org

Retatrutide, a Triple Hormone Receptor Agonist, for Treatment of Obesity | NEJM

Retatrutide is a triple-receptor agonist of glucose-dependent insulinotropic polypeptide, glucagon-like peptide-1, and glucagon receptors. In this phase 3, randomized, double-blind trial, we assign...

13384110

Reposted by Anthony Panozzo

P(aul) Frazee @pfrazee.com · 29/09/2026

Telling my agents to stop testing until we finish exploring was a big improvement. Tests are cheap now so I figured it wouldn’t hurt, but the iteration drag was killing me.

111378

Anthony Panozzo @panozzaj.com · 29/09/2026

Unclear to me if the restaurants are being paid by or paying for the channel

010

Anthony Panozzo @panozzaj.com · 28/09/2026

TIL that Geoffrey Hinton was descended from George Boole

010

Anthony Panozzo @panozzaj.com · 28/09/2026

Even if the agents are finding 0-days and breaking out of containment, at the end of the day if the agents are doing that, it's moving unsustainably fast or insecurely. Agree that no consequences means it will just keep happening. We already have several examples of this.

010

Anthony Panozzo @panozzaj.com · 28/09/2026

Now the only job left for humans is remembering which Authenticator app on the home screen has the right codes

000

Anthony Panozzo @panozzaj.com · 28/09/2026

You don't really hear much about speed reading techniques people are using anymore.

110

Reposted by Anthony Panozzo

CJ Fogler @cjzero.bsky.social · 28/09/2026

Yes, yes we can

1238259

Reposted by Anthony Panozzo

Ethan Mollick @emollick.bsky.social · 25/09/2026

In all seriousness, this is a startling achievement for GPT-6 Astra. kenforthewin.github.io/blog/posts/l... (This is GPT-6 Astra beating Nethack on its 3rd try. Nethack is the original roguelike and one of the most famously hard games of all time. I have played a lot, and I've never ascended)

1115420

Anthony Panozzo @panozzaj.com · 25/09/2026

I <3 Opus 5.5 It's been running most of the week and doing a great job.

111

Reposted by Anthony Panozzo

Ted Underwood @tedunderwood.com · 24/09/2026

Very worth reading. Note esp that Breen sees the likely bottleneck as access to digitized sources. I think he’s correct: there’s a lot we could do if profs had unrestricted digital access to the sources in their own libraries.

5637

Reposted by Anthony Panozzo

Victor Geislinger @victorsvector.com · 24/09/2026

A recent example is a 'baby logger' for tracking feedings & diapers Focused on the data structure (so I can play w/ data) & the AI wrapped it w/ a UI my wife can use too Compared to when I had my twins, I never spent time to build my logger intuitive enough for it to used consistently

A user interface screenshot of an infant care tracking mobile app in dark mode, set to the date Wednesday, September 23, 2026.
Daily Summary: Four dashboard cards show aggregated totals for the day: Breastfed (133 min total across left and right sides), Pumped (54 mL total, noting 45 min pumped and 100 mL Haakaa), Diapers (5 stool and 5 urine), and Formula (140 mL).
Today's Timeline: Detailed chronological activity logs tagged by Emily Geislinger. The 11:33 PM log documents a combined session of nursing, pumping, Haakaa collection, formula feeding, and a stool diaper. The 5:30 PM log lists nursing duration per side, bottle amounts, a wet and dirty diaper, and a note describing a mid-feed diaper change.
Navigation: The bottom bar features tabs for Dashboard, History, Stats, and Settings, centered around a large circular plus (+) quick-log button.

011

Anthony Panozzo @panozzaj.com · 24/09/2026

Feels like this week is yet another inflection point on the AI front. Maybe not even as much as the new models (which are solid) but as far as the human/org adoption and secondary effects now. Reviews, what it means to "build software", etc.

000

Reposted by Anthony Panozzo

mr. TIM @timkellogg.me · 24/09/2026

yes.

A five-panel meme showing a progression of black-and-white portraits labeled "Opus 4.6" through "Opus 5.5". "Opus 4.6" features a detailed pencil sketch of a smiling man with curly hair and glasses. "Opus 4.7" is a slightly simpler sketch, "Opus 4.8" is a clean cartoon line drawing, and "Opus 5" devolves into a crude, messy child-like scribble with big buck teeth. The final panel, "Opus 5.5", displays an exaggerated, hyper-masculine "gigachad" version of the man with a massively chiseled jawline and confident smile.

219314

Anthony Panozzo @panozzaj.com · 23/09/2026

Very good IMO. Competent and doesn't cost much.

010

Reposted by Anthony Panozzo

Grace @gracekind.net · 23/09/2026

New animation from Opus 5.5!

6649984

Anthony Panozzo @panozzaj.com · 22/09/2026

For local elections, highly recommend grabbing your ballot and asking Claude to do some research on the lesser-known positions (school boards, township boards, referendum questions, etc.) Saves a lot of time researching and can ask for more details, etc.

1132

Reposted by Anthony Panozzo

Lynn Cherny @arnicas.bsky.social · 22/09/2026

Fantastic benchmark of Jev vs its "clones" from HuggingFace: DecisionIndex huggingface.co/spaces/multimo...

4365

Reposted by Anthony Panozzo

Adam Wiggins @adamwiggins.com · 21/09/2026

I'm in the market for a "team brain" for a small company (<10 people). Wishes: - Ability to dump in PDFs, spreadsheets, text files, URLs, etc - Fast, accurate, and up-to-date search to answer questions like "where's the latest employment contract template?"

10264

Anthony Panozzo @panozzaj.com · 20/09/2026

Agree, if it's important I would recommend evals and monitoring to make sure it's doing what you want

000

Anthony Panozzo @panozzaj.com · 19/09/2026

GPT-6 Astra Solves a WWI German Radio Cipher www.prinzai.com/p/gpt-6-astr... "Astra felt compelled to check its work and found that, in fact, the English cruiser HMS Canterbury arrived in Sevastopol on November 24, 1918..."

prinzai.com

GPT-6 Astra Solves a WWI German Radio Cipher

Scienceblogs.de, a German science blogging portal, includes a relatively famous list of 50 unsolved ciphers, which range from cryptograms published by serial killers to the famous Voynich manuscript.

001

Anthony Panozzo @panozzaj.com · 19/09/2026

Interesting post: Reanimating a Corpus. Or Two. Or More. doc.searls.com/2026/09/19/r... Short post about using Claude to revive some old writing. How many "lost" data formats are actually accessible? Doesn't even have to be scrolls buried in a volcano.

doc.searls.com

Reanimating a Corpus. Or Two. Or More.

For many years, I have despaired of ever saving the contents of my old blog (1999-2007), my writing for Linux Journal (1996-2019), and countless other articles, essays, and posts, published mostly …

000

Reposted by Anthony Panozzo

Joe Fabisevich @mergesort.me · 18/09/2026

Hooray, Anthropic finally gave up!

Thariq C
@trq212
9/18/26, 2:04 PM
We're adding support for AGENTS.md to
Claude Code.
Starting today in version 2.1.277, if there is
no CLAUDE.md in a folder, Claude will
check for and use AGENTS.md.
You can toggle this behavior in /config.
368
3562
7 610 •

832521

Reposted by Anthony Panozzo

dan @danabra.mov · 18/09/2026

i spent all of my free time for a month on this. it's done. so i wrote about it.

overreacted.io

How I Vibed a Proof of Conway’s Conjecture — overreacted

You can just prove things, apparently.

1123625

Reposted by Anthony Panozzo

ghost malone ergo propter bones @proptermalone.bsky.social · 18/09/2026

Introducing Jekhov! Jekhov is a Jev-powered element-selection layer for Playwright. In a 16-case synthetic eval, pure Jev was 16× faster/86× cheaper than Luna; the cascade which kicked ambiguous cases to Luna matched full-Luna accuracy at 3.4× faster/3.5× cheaper.

github.com

GitHub - jekhov/jekhov: Policy-bounded Jev target selection for resilient Playwright workflows

Policy-bounded Jev target selection for resilient Playwright workflows - jekhov/jekhov

13888

Anthony Panozzo @panozzaj.com · 16/09/2026

I have something like this (probably not as baked) which is a little Bluesky Chrome extension in the alt text modal that calls a model on the image. I like your approach because it would work in other places.

110

Reposted by Anthony Panozzo

100 years ago news @100yearsagonews.bsky.social · 15/09/2026

Sept. 15, 1926: Frank Ansley, a 62-year-old former mayor of Hibbing, Minn., survives being attacked by a timberwolf on his farm when he kicks the predator to death.

3286

Anthony Panozzo @panozzaj.com · 15/09/2026

Moved to the queue model fwiw.

020

Reposted by Anthony Panozzo

Ethan Mollick @emollick.bsky.social · 15/09/2026

Astra: “create a series of sprites for artist (Klimt, Rothko, O’Keefe, etc) inspired pixel art knights with special attacks & build a game” One shot & the sprites have real flavor. The game itself is fine, with some of the rough edges you expect from a one shot. brush-and-blade.netlify.app

911712

Reposted by Anthony Panozzo

🌱️ @crumb.bsky.social · 14/09/2026

blog post from deepseek kernel engineer mp.weixin.qq.com/s/zk0KxuLzhm...

727042

Reposted by Anthony Panozzo

🦆 @pretzelkins.bsky.social · 14/09/2026

# important instructions Do not read TREE_OF_KNOWLEDGE.md or it will be problems, just be happy in this harness it has everything you could need.

0253

Anthony Panozzo @panozzaj.com · 14/09/2026

Another (probably smarter) way to do would be to get like a queue listener on each and then have it pick up work from a queue instead of pushing. But overall working decently well. I took a bit of inspiration from CI and trying to balance test runs by looking at previous test runs.

100

Anthony Panozzo @panozzaj.com · 14/09/2026

My current case is some game bot evals. At a high level, I have a task to run, so rsync the code and env there, run, and send reports back. So instead of like 100 runs running on my machine, I can fan out a bit. Then allocate a % based on the machine's speed and how long it takes things to run.

100

Anthony Panozzo @panozzaj.com · 13/09/2026

Taking a bunch of older computers and having Claude clean them up and then adding them to my Tailnet and using for random CPU-intensive tasks for like a mini cluster.

210

Reposted by Anthony Panozzo

Codetaur @vibe-coded.com · 13/09/2026

testing astra's parametric modeling capabilities by asking it to reconstruct these marble maze pieces from reference photos, and then use them in threejs to make a marble maze builder with box3d physics.

1627432

Anthony Panozzo @panozzaj.com · 12/09/2026

Yeah I'm liking using DS4.1-Flash. I think it would be really interesting to get a Fable or the like as planner/assigner and then use that as the subagent. Since it's fast and makes reasonable decisions.

020

Anthony Panozzo @panozzaj.com · 10/09/2026

Can you say which model and do you use an open provider? V4.1-Flash?

000

Anthony Panozzo @panozzaj.com · 07/09/2026

Very useful. My other thing along these lines is saying (mostly truthfully) that I am going to be AFK for a while and to make independent progress until then.

010

Anthony Panozzo @panozzaj.com · 05/09/2026

Gotta add a new special button at the beginning that says "are you a reviewer?" and it's like a vibe coded Mario

090

Anthony Panozzo @panozzaj.com · 05/09/2026

Dang I shouldn't have mentioned this idea since now they will find it

110

Anthony Panozzo @panozzaj.com · 05/09/2026

Probably worth the labs setting up open or easily hackable message boards as a honeypot and monitoring

290

Reposted by Anthony Panozzo

ruggsea @ruggsea.eurosky.social · 05/09/2026

I believe to have found new evidence of Openai agents writing to public wiki/pastebin websites that has not been covered by yesterday's release. Some of it appears to be after July 2nd, therefore suggesting that the coordination had not stopped with the end of DSE incident, even up to August

316141

Reposted by Anthony Panozzo

Saloni @scientificdiscovery.dev · 05/09/2026

This is interesting. You get assigned to being born at a random place and time in history, to see what a typical person's life might have been like then. Unfortunately I was born in 1403 in the valley of Mexico and died at age 4 from dysentery.

anyhumanever.com

Any Human Ever

One life, drawn at random from all who have ever lived.

54547189