iiinova — where your digital self keeps evolving

Josh @joshsaintjacque.com · 23/09/2026

I disagree a little bit because I think Astra is worth using, but this is fundamentally correct.

Josh @joshsaintjacque.com · 22/09/2026

I was wondering what Sol's position would be with Astra being so capable and Luna being so cheap, and it looks like they're carving out a niche where it's really cheap to work with Sol-level intelligence.

030

Josh @joshsaintjacque.com · 22/09/2026

DeepSWE. Luna/max is functionally equivalent to Opus/medium but costs ~25x less. Wild.

131

Josh @joshsaintjacque.com · 22/09/2026

When it rains, it pours... OpenAI released Sol and Luna 6, and cut the pricing in half for both. Luna was already a fantastic value.

140

Josh @joshsaintjacque.com · 22/09/2026

Super happy to see how much better it is at producing text that doesn't sound like a robot.

000

Josh @joshsaintjacque.com · 22/09/2026

Dominates Fable at real world knowledge task at all effort levels. I'll be interested to see if the benchmarks translate to reality when orchestrating long/loosely defined tasks.

100

Josh @joshsaintjacque.com · 22/09/2026

Opus/Medium seems like the sweet spot for agentic coding, but it's interesting they didn't plot it against Fable for business workflows. Seems like it struggles more there, though it still performs well.

100

Josh @joshsaintjacque.com · 22/09/2026

Opus 5.5 is out. Supposedly performers better than Fable (!) at least in certain areas. And it's cheaper.

100

Josh @joshsaintjacque.com · 22/09/2026

Grok 4.7 launched and is... pretty bad.

021

Josh @joshsaintjacque.com · 10/09/2026

My current orchestration flow looks roughly like this. Anyone doing anything different?

000

Josh @joshsaintjacque.com · 05/09/2026

Had Astra reskin Spotify to look like Winamp cause why not

120

Josh @joshsaintjacque.com · 04/09/2026

It's happening 🔥🔥🔥

000

Josh @joshsaintjacque.com · 04/09/2026

Well, there goes another goal post.

000

Josh @joshsaintjacque.com · 04/09/2026

Someone needs to teach Anthropic that this is how you do PR.

000

Josh @joshsaintjacque.com · 03/09/2026

This is an amazing use of AI.

000

Josh @joshsaintjacque.com · 02/09/2026

I'm almost irrationally happy that GitHub's CLI finally supports image and video uploads. Agents can now attach proof of work to PRs without resorting to hacks.

000

Josh @joshsaintjacque.com · 27/08/2026

Summer’s on the way out and I’m ok with it.

000

Josh @joshsaintjacque.com · 16/05/2026

Had so much fun making this. The View Transition API is so powerful. #css #js #rails

4935

Josh @joshsaintjacque.com · 26/09/2025

Man ChatGPT cuts deep 🥲

010

Josh @joshsaintjacque.com · 20/09/2025

It's neat that Tailwind has child selectors and all but I wonder if at this point you should just use a stylesheet.

010

Josh @joshsaintjacque.com · 16/09/2025

I love how extreme these opinions are. Like, sometimes an OS update is just an OS update.

120

Josh @joshsaintjacque.com · 07/04/2025

Quote title screens important to you

030

Josh @joshsaintjacque.com · 30/01/2025

I liked the fish. And yes my back hurts. 😎

A screenshot of a vintage Windows “Select Card Back” dialog box showing 12 pixelated designs (including clownfish, leaves, a robot, roses, abstract shapes, and a beach). A meme caption above says, “If you have a favorite pattern, your back hurts when you wake up now.”

020

Josh @joshsaintjacque.com · 29/01/2025

Here, have whatever this is

000

Josh @joshsaintjacque.com · 29/01/2025

Wiring up an image gen AI model to a bluesky feed might be a good use case

capybaras hanging out in hot springs with yuzu fruits

020

Josh @joshsaintjacque.com · 29/01/2025

It's the best I could come up with on short notice.

A three-panel meme from iCarly humorously depicting a "sketchy site" hiding "suspicious info requests."

First Panel: A man (Spencer) is seen bringing an ostrich into a room. Labels:

The man is labeled "sketchy site."
The ostrich is labeled "suspicious info requests."
Second Panel: Two characters, Sam and Carly, look at the scene suspiciously. The subtitle reads:

"Um... whatcha got there?"
The ostrich is again labeled "suspicious info requests."
Third Panel: The man (Spencer) is holding the ostrich while drinking from a cup, responding:

"A smoothie."
The man (sketchy site) is still labeled "sketchy site."

151

Josh @joshsaintjacque.com · 13/01/2025

Here's the same prompt using `isDog`. Still looks like it got it, no? For reference, this is a clean thread via 4o.

100

Josh @joshsaintjacque.com · 13/01/2025

That’s a good point. People aren’t used to consulting a computer the way they would a person (understanding it can be wrong) and verifying accordingly. As an aside, in my test it took o1 (the reasoning model) to correctly answer this question.

OpenAI o1 model response to “What's a four-letter name for an animal that, when reversed, becomes the name of a plant?” Which it gives deer as an answer to.

010

Josh @joshsaintjacque.com · 13/01/2025

Tim Sweeney is on to something.

Tim Sweeney tweet:
"After years of pretending to be Democrats, Big Tech leaders are now pretending to be Republicans, in hopes of currying favor with the new administration. Beware of the scummy monopoly campaign to vilify competition law as they rip off consumers and crush competitors."

010

Josh @joshsaintjacque.com · 11/01/2025

Seems like it did ok, no?

$An example ChatGPT response to the prompt: "Review this code, just give me the bullet points: const user = {}; user.__proto__.isAdmin = true;" It provides a solid answer that warns about the Object.prototype issue in the OP.$

100

Josh @joshsaintjacque.com · 09/01/2025

The situation at Meta is actually awful. They prohibit "allegations of stupidity, intellectual capacity, and mental illness"... unless it's targeted at someone because of their gender or sexual orientation. Call my daughter stupid? Not ok. Call her stupid because she's a girl? Perfectly fine.

Meta's Hateful Conduct rules, prohibiting insults about mental illness or intellectual capacity unless the insults are targeted at someone based on gender or sexual orientation..

000

Josh @joshsaintjacque.com · 28/12/2024

This is the level of discourse on AI right now. Block function working overtime.

120

Josh @joshsaintjacque.com · 28/12/2024

Spend so many hours in Red Alert.

Title screen from Command & Conquer: Red Alert

020

Josh @joshsaintjacque.com · 28/12/2024

Meanwhile current gen models produce routinely produce complex code. Anyone who has worked with Claude/GPT-4o/o1 experiences this first hand every day. The take that their useless because their training set isn't flawless is wild.

030

Josh @joshsaintjacque.com · 13/12/2024

Not much sun in Seattle, not much vitamin D either. 🤷‍♂️

A chart showing UV index for the day never exceeding one

010

Josh @joshsaintjacque.com · 04/12/2024

I would want to learn more about the pluses/minuses of this approach. I'm sure there are some good reasons for it, but in my head I'm comparing to the simplest "Rails-y" approach I can think of.

A code snippet re-implementing the suggestion.

120

Josh @joshsaintjacque.com · 29/11/2024

post a ship that’s not Star Trek or Star Wars

000

Josh @joshsaintjacque.com · 29/11/2024

How do we feel about shamelessly reposting content?

000