iiinova — where your digital self keeps evolving

But AI coding is one of the most benchmarked domains possible. How are there no independent benchmarks? Are Relace/Morph just selling snake oil? That seems hard to believe if they are both indeed growing fast. The pace of AI improvements also seems much faster than can be properly discussed.

000

Apparently this technique was pioneered by Cursor, but Cursor being an IDE, never released an independent fast apply model for others to use. So this also sounds very useful and clever. But weirdly, it is also nearly impossible to find independent benchmarks showing that Relace/Morph are better /

100

There is not one but two fast-growing LLM startups – Relace and Morph – that specialize in fast applications of diffs to code. So the main model can reduce its output to a smaller set of code and run faster (just the new code, not the full code; also doesn't need to produce a perfect diff). /

jimmy @jimmy.path.pub · 22/08/2026

Alright who wants to talk about the Odyssey? I am ready.

000

That's true that rewriting (or hallucination) is a different form of bias than censorship (omission). If I look at the OP's qwen response, it looks like one where there is omission but what it does say doesn't look factually incorrect in any way. Chinese government pressure can lead to both.

100

jimmy @jimmy.path.pub · 21/08/2026

Ah actually, the obvious recently famous one is if you try to talk about hacking or even cloaking requests. That's one where the open models will let you ask away (GLM 5.2 saved HuggingFace), but the commercial frontier models are currently censoring: huggingface.co/blog/agent-i...

huggingface.co

Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

101

jimmy @jimmy.path.pub · 21/08/2026

Interesting thought exercise. The US doesn't have anywhere near the same centralization nor have a mass censorship regime. If there is bias, it is likely reflective of distributed bias in the populace or at least the elites in that topic. Would also likely need to be a recent event. Maybe COVID?

100

Really makes you want to dig right in 😅

000

🤣 Oh cool you are self-hosting! What's your runner? Are you giving it tool-call access to ground its responses?

100

How big is the delta between Jupyter notebooks, and what you are describing in the OP?

130

Do you have thoughts on Airtable?

140

Probably not related, but maybe related? Apparently GLM 5.2 is good at code reviews. Maybe some skills are transferring.

100

Perplexity Kimi K3 gives a highly-censored result, oof. GLM 5.2 is able to ground in good sources though and seems pretty reasonable.

jimmy @jimmy.path.pub · 17/08/2026

🤣 Perplexity used to post-train open models to de-censor them. I don't know if they still do, but I noticed they are hiding GLM and Kimi options by limiting them to mobile or certain other segments. Must not completely trust them yet.

000

Meta makes the drug, but imho Apple and Google should be held accountable for distributing the drug to teens and drivers. Apple in particular deliberately targets teens and yet has gotten this very strange free pass.

Attorney General Jennifer Davenport @newjerseyoag.bsky.social · 17/08/2026

Reposted by jimmy

Our children are not data points to be monetized for profit. Tomorrow, we will show a jury how Meta deceived families and consumers into thinking their social platforms are safe for kids and teens. www.njoag.gov/attorney-gen...

183

jimmy @jimmy.path.pub · 06/08/2026

no worries great to see the experimentation!

jimmy @jimmy.path.pub · 05/08/2026

Very cool. Heads up I wasn't getting any hits for some common terms. Didn't spend much time trying to debug but have a guess that you are normalizing when indexing but maybe the query isn't being normalized?

jimmy @jimmy.path.pub · 05/08/2026

Top 10 by activity is largely fun / positive / curious too:

jimmy @jimmy.path.pub · 05/08/2026

I was browsing top Bluesky links and not only is it overly centered on US politics, but it is also largely pretty negative stuff. I don't disagree there's a lot of bad in the world, but that's not all people want to talk about. In contrast, the top Subreddits are way more fun and interest driven:

141

Reposted by jimmy

Oh The Urbanity! @ohtheurbanity.bsky.social · 03/08/2026

1995: “A few months ago there was some breakthrough announcement that on the internet they were going to broadcast a baseball game. You could listen to a baseball game on your computer. And I just thought to myself: does radio ring a bell?”

youtu.be

Bill Gates Trying to Convince Dave the Internet is the Future | David Letterman

YouTube video by Letterman

1253

jimmy @jimmy.path.pub · 03/08/2026

🤣 Heath, NJ doesn't even exist? Josué Perez has some explaining to do... lol

000

Pub subpages & other recent features!

I think it should? Leaflet says they support iframes: lab.leaflet.pub/3m3q3meshhc2g But it probably also depends on what modes of embedding Datawrapper supports.

lab.leaflet.pub

Lab Notes 013: publications can now have nested pages — illustrated here with pages about embeds, comments, and more

3. It gives the quote really nice design treatment visually featuring it over simply typing it out as a part of your post. And it also avoids the post character limits, bringing needed context. All-in-all, love the feature. Wish all writing could be shared this way on Bluesky one day.

011

2. It's a quality recommendation with direct context as to why. There are a lots of feeds that understandably share all new writing, but for that reason, it's not a great signal by itself. Quoting is real engagement vs mundane sharing of love/hate/author good/bad. /

Even on a link-friendly place like Bluesky, you can tell a lot of the participants never click on the link and engage with the actual original writing. But readers that quote post are guaranteed to be readers of longer material. /

It's awesome signal for a couple of reasons: 1. The decline of reading is both in absolute amounts and in kind. Not just reading less, but reading shorter and shorter material. Let's be real. Lots of "reading" these days is just scanning titles and hot takes. /

Bluesky's Custom Feeds have a lot of potential (and room to improve) – but when you find a good one and want to share, it ain't easy! Here's @leaflet.pub's awesome Leaflet Quotes feed: bsky.app/profile/did:... /

263

@lapsu.bsky.social would be interested in your thoughts. How much do you see people share prompts / intentionally collaborate?

The AI Cohesion Paradox: Why AI-Empowered Teams Cooperate Less — And What Leaders Can Do About It

A really great essay that addresses the erosion of teams due to increasing individual reliance on AI

teaiops.leaflet.pub

131

Reposted by jimmy

brendan @schlage.town · 15/07/2026

I just browsed logged out discover, not great fam… - mostly-US-politics news - memeslop; petslop - random generic photos the odd good post, but mostly: tons of images, and… NOT A DAMN THING I'D ACTUALLY WANT TO TALK ABOUT! YOU KNOW, LIKE…ENGAGE WITH. AS A PERSON. ON AN OSTENSIBLY SOCIAL WEBSITE.

3292

jimmy @jimmy.path.pub · 17/07/2026

I think if Bluesky wants lasting utility and participation, everything you describe about the horrible logged out experience is exactly what they should train a filter to remove. Onboarding is hard; new users don't like to go through it. Better would be a first experience via communities/sub-skies.

000

Reposted by jimmy

Ronen Tamari @ronentk.me · 09/07/2026

“[Americans] are losing the higher-order abilities of comprehension and synthesis. America, in other words, isn’t illiterate. It’s postliterate.” Remarkable progress in AI alongside alarming degradation of human intellectual capacities … but I’m sure it’s just an extraordinary coincidence

theatlantic.com

The End of Reading Is Here

Optimists once believed that universal literacy was inevitable. Now it seems that the age of reading might be a short anomaly in human history.

1111

jimmy @jimmy.path.pub · 06/07/2026

I think there is reason to believe that there is room for a model to disagree with itself based on prompt and context and the goals of review work vs creating a diff. But what if you are using different models, maybe even models that are trained for review work in particular?

000

Simon Willison @simonwillison.net · 05/07/2026

Love the dichotomy. Does using Claude Fable to code review and identity flaws change your point of view on whether AI can help *both* centaurs and reverse centaurs? bsky.app/profile/simo...

100

Reposted by jimmy

Somewhat humbling to have Claude Fable do a final review of some software that you're about to release and have it then find (and fix) FIVE release blockers, for an estimated (unsubsidized) cost of $149.25 simonwillison.net/2026/Jul/5/s...

simonwillison.net

sqlite-utils 4.0rc2, mostly written by Claude Fable

I wrote about the sqlite-utils 4.0rc1 release a couple of weeks ago. Since we only have Claude Fable on our Max subscriptions for a few more days, I decided to …

1516011

Reposted by jimmy

By Cory Doctorow (GPG 0xBF3D9110957E5F4C) @doctorow.pluralistic.net · 02/07/2026

One thing I've learned about paradoxes: often the answer to the riddle of "how can this one thing have such a contradictory set of features and effects?" is "it's not one thing, it's *two* things*." 1/

A village revel in a sleepy wood, with many Renaissance peasants having a debauch. In the background is a hulking mainframe computer.

619360

Accuracy with GPT-5.5 Extra High is very, very good. It's far smarter than me. But there is a downside besides being slow, which is it can be so thorough and so focused on corner cases that it just becomes noise. The effect for a lot of things I'm trying to do is it becomes too difficult to consume.

000

I keep trying to use Sonnet or even Nemotron. Sonnet is just not as logical or trustworthy. Nemotron is fast, but not even close to the leading models. Finger-to-the-wind, there must be some secret sauce benchmark stuff leading labs are doing that is hidden from upstarts attempting open models /

100