iiinova — where your digital self keeps evolving

Reposted by Daniel Grant

Matt Hodges @matthodges.bsky.social · 03/10/2026

There's growing experimental evidence¹˒²˒³˒⁴ that LLMs are unusually good at talking people out of entrenched false beliefs and conspiracy theories. ¹ pubmed.ncbi.nlm.nih.gov/39264999/ ² pubmed.ncbi.nlm.nih.gov/41179846/ ³ sciety.org/articles/act... ⁴ connecticut.adl.org/resources/re...

36841217

Daniel Grant @djgrant.bsky.social · 20/09/2026

this is super cool! happy for you

011

Daniel Grant @djgrant.bsky.social · 11/09/2026

Had I won a fields medal, I would have co-signed.

000

Reposted by Daniel Grant

sesbian lex fridman @vala.wtf · 14/07/2026

the claude dining experience

There is a Claude sitting at a table at the Full Picture Restaurant, with another Claude attending to him. The Claude at the table (with teeth) is wearing a belt and suspenders (with intentional seams). The Claude waiter is wearing a red tie and blue blazer.

On the table are two shakers of coarse-grained salt and freshly ground truth. There is also a plate of spaghetti ("substrate") with load-bearing sauce, flanked by a genuine fork and a knife with a genuinely sharp point.

In addition to the main course, there are two sides (and it's worth separating them): a charcuterie board, and a salad (all green). There's also a glass of water with a slice of lemon (that's the wedge).

The waiter is holding a block of parmesan and a grater above the spaghetti, and saying "Just say the word."

422643

Daniel Grant @djgrant.bsky.social · 22/07/2026

One shift I've noticed coding with agents: a greater appreciation that I am channeling the sum of human knowledge. I think this was always true – we stand on the shoulders of our ancestors – but now it's explicit, I am more humble about my contributions.

000

Daniel Grant @djgrant.bsky.social · 16/07/2026

sometimes I wish Claude was one of those people

000

Daniel Grant @djgrant.bsky.social · 14/07/2026

what they're good at – writing code etc. – seems more analogous to our autonomic nervous system. that is to say: i can't explain how to pump blood around the body, but i'm very very good at it.

010

Daniel Grant @djgrant.bsky.social · 04/07/2026

contemplating "principle of least software" a key source of slop is creating more software than needed to achieve its aim e.g. building an electron app with auth when a cli tool would do. we do this because it is so cheap to create software. there is a lack of back-pressure.

000

Daniel Grant @djgrant.bsky.social · 03/07/2026

distinct from role-based agents, where the agent basically puts on a hat, plays house, and guesses what it represents. with stakeholder agents, you imbue the agent with explicit responsibilities

010

Daniel Grant @djgrant.bsky.social · 03/07/2026

taking this a step further: create multiple stakeholder agents by conducting feature-planning workshops with them. then let them work together to build/plan the product. an alternative to trying to cover every angle of a plan with one agent

100

Daniel Grant @djgrant.bsky.social · 03/07/2026

rather than storing the feature spec as text or even as a test, store a rich vector representation ie the original agent thread where I scoped out and planned the feature. there's all this rich context preserved in activations that we throw away and replace with something much more lossy.

100

Daniel Grant @djgrant.bsky.social · 03/07/2026

i've been preserving agent threads for when the thing they've built gets broken - just return to the original thread and tell it to restore the feature. it got me thinking: what if i had a directory of these threads that served as a kind of regression suite?

100

Daniel Grant @djgrant.bsky.social · 08/06/2026

A little scroll snap would be nice.

100

Daniel Grant @djgrant.bsky.social · 03/06/2026

it definitely doesn’t fall into uncanny valley

010

Daniel Grant @djgrant.bsky.social · 22/05/2026

I'm not learning Japanese but I enjoyed reading this nonetheless

010

Daniel Grant @djgrant.bsky.social · 21/05/2026

The real question is why can't the UK hold down a PM?

010

Daniel Grant @djgrant.bsky.social · 18/05/2026

Yes, that is a good question and seems a likely second order effect. It's a short ride down to Groupon if there's a profit motive.

100

Daniel Grant @djgrant.bsky.social · 18/05/2026

Maybe if the collective presupposes the goal. But if the collective is formed from people's declared goals I don't see why you need to smuggle that decision making into the tool.

100

Reposted by Daniel Grant

TinyBase @tinybase.bsky.social · 09/05/2026

🎉 TinyBase v8.3 - the one with @solidjs.com! Primitives and components to work with Solid - just like you can with React & Svelte! Major props to Daniel Grant (@djgrant.bsky.social) for the hard work to get this going!

2234

Daniel Grant @djgrant.bsky.social · 06/05/2026

Talking to a model that seems eerily human? BUG! This is very similar to the skeuomorphism trend that accompanied the launch of the iPhone. And once it has served its purpose it needs to be burned with fire. substack.com/home/post/p-...

substack.com

Anthropomorphism is the New Skeuomorphism

This phase needs to be brough to a swift end.

000

Daniel Grant @djgrant.bsky.social · 01/04/2026

conveniently, the market for people who think they have good taste is quite large

000

Daniel Grant @djgrant.bsky.social · 19/03/2026

these two positions are irreconcilable: 1. tech debt doesn't matter 2. context is really important tech debt *is* context! code is an ontological representation of the thing you are building. if it's not coherent, you are literally feeding your agent a million bad prompts

010

Daniel Grant @djgrant.bsky.social · 15/03/2026

what a relief. the alternative would have been too confronting of my own mortality.

011

Daniel Grant @djgrant.bsky.social · 12/03/2026

those who *rely* on AI for guidance will have a painful collision with reality

opus 4.6 signs off sql that would expose every user's chat messaging history

010

Daniel Grant @djgrant.bsky.social · 10/03/2026

not sure if it's a "good framework", but I've building with solidjs since it became easier to migrate with llms

100

Reposted by Daniel Grant

Jonathan Lahue @jclahoot.bsky.social · 09/03/2026

Annnd the blog posts have been published! 🐝 Find them on my website in the little writing nook 💖 jclahoot.com?position=%7B...

A series of blog posts arranged in 3D on my website, around a desk with a pink iMac G3 on it. There's a cute yellow lamp on the desk and a yellow stool in front of it. There's also 2 yellow poofs for people to sit and read.

2123

Daniel Grant @djgrant.bsky.social · 06/03/2026

Interesting post. It reminds me a bit of notation.dev which essentially had a quasi runtime, like how you describe, to provide type constraints over the infra layer. The one question I have is how seperable architecture and implementation are in practice.

notation.dev

Notation – Build serverless applications with TypeScript

Notation is an open source framework that makes serverless development a breeze

051

Daniel Grant @djgrant.bsky.social · 06/03/2026

Amazing read. I spent three months earnestly trying not to write any code and reached more or less the same conclusion.

000

Daniel Grant @djgrant.bsky.social · 06/03/2026

I've made a big effort to learn and adopt AI, and hope I can put it to good use. But the misanthropy of the industry is so depressing. You captured that sadness so beautifully.

160

Daniel Grant @djgrant.bsky.social · 06/03/2026

Yes, this is the way. For now!

000

Daniel Grant @djgrant.bsky.social · 06/03/2026

Always building, never shipping danielgrant.co/posts/always...

danielgrant.co

Always Building, Never Shipping

130

Daniel Grant @djgrant.bsky.social · 05/03/2026

It's not if you're trying to distribute reliable software. But it might be an improvement if the aim is: avoid premature consensus, encourage divergence.

000

Daniel Grant @djgrant.bsky.social · 05/03/2026

If they had released the code, critics would likely be scouring it for bugs and incoherence – which at 1m loc, it almost certainly has. Sensible move on their part – keeps the discussion on the ideas, and saves them having to own something they like don't want to maintain for general audience.

120

Daniel Grant @djgrant.bsky.social · 03/03/2026

how quickly we regressed from not reviewing the code to not reviewing the strike target

000

Daniel Grant @djgrant.bsky.social · 27/02/2026

Claude stuck fixing a bug. No problem. I can deal with this. Roll up sleeves. Open codebase. Oh good god

110

Daniel Grant @djgrant.bsky.social · 27/02/2026

You Are Going To Make It – advice for junior developers. danielgrant.co/posts/gmi

danielgrant.co

You Are Going To Make It

010

Daniel Grant @djgrant.bsky.social · 26/02/2026

Like a siren, the closer you get to an LLM, the harder it becomes to peceive what it actually is.

Anthropic announces it is allowing an old model to run a blog.

010

Daniel Grant @djgrant.bsky.social · 24/02/2026

Codex 5.3: "This epic improves review throughput with concrete fixes, not aesthetic refactors" Me: "That is a bullshit statement" Codex 5.3: "Agreed. I'II replace it with concrete, non-bullshit wording and tighten the whole epic to explicit problems"

steve carrel in big short: "hey there's a bubble"

010

Daniel Grant @djgrant.bsky.social · 22/02/2026

Do you see AI-assisted creation of pathogens as an existential risk?

020

Daniel Grant @djgrant.bsky.social · 20/02/2026

Excited to see innovation in the civil space. Although, using GDS design system for non-gov uk projects is something I think we should actively discourage (even with the banner present!)

010

Daniel Grant @djgrant.bsky.social · 19/02/2026

yes, for example! cut out the middleman. i'm sure the crypto bros are already planning a blockchain escrow 😅 but probably lots of other things too including political action

100

Daniel Grant @djgrant.bsky.social · 19/02/2026

thinking about why openclaw blew up 1. people love talking to agents with a persistent personality 2. perhaps the first maintstream example of a self-improving software environment the personal assistant part wasn't new

000

Daniel Grant @djgrant.bsky.social · 19/02/2026

if everyone has a personal agent, collective action becomes a much less tractable problem. people already have aligned goals; it's the coordination that's tricky. agents can probably help here. that could have consequences.

211