iiinova — where your digital self keeps evolving

Yeah, I think about this a lot. It's not clear whether the end justifies the means from a societal perspective, but I think the innovation being built on a common wealth of knowledge is a useful starting point for policymakers.

I mean, I'm a big fan of using #AI in my work, but there is truth to the idea that they have strip-mined the world's knowledge and are selling it back to us. All while alienating knowledge workers from intelligence (in the Marxist sense).

350

I just looked back at my files, it was even better than I remembered: 5GB Excel file to 135 MB Parquet! 🤯

Yes! I discovered this format last year when I had to share a 2gb file. Used Parquet and got it down to something like 200mb 😲

Giles @gdeejay.bsky.social · 01/10/2026

I really hate the move towards passwordless authentication for everything.

000

Giles @gdeejay.bsky.social · 01/10/2026

The 'For You' feed is equal parts entertaining and infuriating, which is why I default to the oatmeal of #Bluesky feeds: 'Mutuals'

000

Giles @gdeejay.bsky.social · 30/09/2026

I've been helping set up another charity book fair over the last couple of days and came back with another five new second-hand books. Three of them are Wodehouse! 👌🏼

001

Giles @gdeejay.bsky.social · 29/09/2026

This is *super* interesting. I've wondered whether #AI writing might have a detectable macro structure like this. 🤔

Giles @gdeejay.bsky.social · 29/09/2026

(feedback is welcome.)

000

Giles @gdeejay.bsky.social · 28/09/2026

TBH, I've always felt the core of EA in the development space is just evidence-based economics with a side of utilitarianism. At least from my experience as an economist working in the social policy / development space EA principles echo those used to frame policy discussions. 🤔

030

Sometimes the gold is buried deep in the thread: 😂

011

Linguistic relativity - Wikipedia

Also, I put this through Claude before posting it as I knew somebody had thought more deeply about this. Indeed somebody has thought of this idea before. It's called the Sapir–Whorf hypothesis: en.wikipedia.org/wiki/Linguis... The evidence suggests LLMs and/or humans don't conform to it.

en.wikipedia.org

(This is also why I figured they were better at writing code as the language was more constrained and logically structured. BUT, I dropped this idea as a friend said it was terrible at writing code for niche languages that likely weren't in its training set.)

000

When ChatGPT was first released I had a theory that one of the reasons it worked so well (particularly for code) was that language provides a compressed structure that defines how humans reason, which LLMs can generalise from. Anyway, I now have to update my #rstats blog post about learning to code

210

Oh! And I meant "The Good Place". It's terrible, but who am I to judge after watching "Community".?! 🧐

000

(damn right, I liked my own post 😎).

Oh, that's so funny as I meant to mention in my post that this is how I react to AI generated images 😂

200

I have comments on their ranking, but I just want to note that I was ridiculed on threads when I suggested that 'The Good Life' was terrible, but... ... I've been shamelessly rewatching Community recently, so am not one to judge. 😬

220

Also, here's @hadley.nz's post on whether to learn to code for those that might have missed it: #rstats

tidydesign.substack.com

y code when ai?

A talk on AI and coding

141

I've actually started writing blog posts for myself rather than trying to answer the questions of an invisible audience. But, this has been a fun post to write as I've thought a lot about this, but haven't structured my ideas around a core thesis. The blog is my attempt to do this. 🧐

000

Giles @gdeejay.bsky.social · 25/09/2026

I've got a solid first draft that now includes @hadley.nz's post answering the same question (that I just found today 🤦🏼). Anyway, if you're willing to provide feedback feel free to send me a DM. I'll update it based on feedback either way, but I'm very open to inputs on the draft. #rstats

290

I'm currently writing a blog discussing whether it's still worth learning to code. I was already thinking about this myself, but part of my motivation comes down to hearing "glad I didn't learn to code" one too many times. Writing it has been carthatic. #rstats

3202

Reposted by Giles

Michel Nivard @michelnivard.bsky.social · 21/09/2026

Maybe psychological sciences is entirely safe from AI takeover, nothing is verifiable 🥹

5133

Giles @gdeejay.bsky.social · 21/09/2026

I wanted to test this as it was an approach I used to test scores developed without the O*NET. As I wrote this another six post ideas came up, but I need to resist writing about nothing but niche O*NET ideas.

000

Giles @gdeejay.bsky.social · 21/09/2026

Who doesn't love a good heatmap!? Another #rstats post on the O*NET. In this one I test whether you can proxy similarity scores with occupation descriptions and classification distance scores. You can! And better than I'd expected. www.gilesd-j.com/2026/09/21/o...

gilesd-j.com

O*NET ratings, job descriptions and classification codes as measures of occupational similarity - Giles

The last post identified potential transition pathways for truck drivers based on how closely other occupations matched them on skills, abilities and knowledge. This post follows the same basic approa...

140

Giles @gdeejay.bsky.social · 20/09/2026

I mean, it's not necessarily AI that has done this. I mean that the two things coinciding isn't what I'd have expected. Don't want to imply causality without a solid RCT or instrument.

Giles @gdeejay.bsky.social · 20/09/2026

Absolute banger of a website: #databs

030

Giles @gdeejay.bsky.social · 20/09/2026

#Bluesky needs community notes except they explain the context behind a post. I have no idea what most of you are talking about.

static.klipy.com

Homero Cancelado: No Homers Club

Alt: Homero Cancelado: No Homers Club

101

Giles @gdeejay.bsky.social · 19/09/2026

Oh, I cut the screenshot incorrectly. I just assumed the logo signified the start of the table. 🙄

Giles @gdeejay.bsky.social · 19/09/2026

#AI *improving* #rstats ranking on the index wasn't on my 2026 bingo card. 😲

R has improved on the tiobe index by 8 ranking points since the same month last year.

2133

Giles @gdeejay.bsky.social · 19/09/2026

I've now got a solid draft of the sequel to this blog. It has been quite the slog to get to the final draft as I had to do several different cuts of the analysis, but the results answer some of my lingering questions about the O*NET.

I think the real question to ask is are *they* online enough?!

*Resisting the urge to vibe code an app that automatically books gym sessions the moment the schedule is released.*

000

Great to hear! I'd hate to lead other astray. Funnily enough, I just posted on LI that LLMs are designed to *simulate* intelligence, not *be* intelligent. The former makes LLMs a compliment to intelligence, the latter makes them a substitute. We need to stop behaving like the latter is true.

110

Giles @gdeejay.bsky.social · 17/09/2026

Probably not the best use of Fable 5.1:

000

... If the model was always accurate, I'd probably get lazy and stop checking its work. But, if it was always garbage I'd have stopped trusting it for helping out with meaningful work a long time ago. It's the intellectual equivalent of training in a high-gravity chamber. TLDR: I'm Goku.

static.klipy.com

Dragon Ball Z: Goku's Pushup Training

Alt: Dragon Ball Z: Goku's Pushup Training

Giles @gdeejay.bsky.social · 17/09/2026

One thing I will say, is that for all their faults LLMs are incredibly helpful as a freelance consultant / researcher. Both the rule of law and O*NET #rstats blog series have been greatly improved by having a model I only *partially* trust to work with. ...

160

Just finished "Messy Jobs' by Jin Li, Yanhui Wu and Luis Garicano. ⭐⭐⭐⭐ (Out of five) Great book on #AI and jobs. It's one of the few books I've read that has a solid theory for how AI is likely to impact the labour market and firm dynamics. I now need to track down a summary to keep on hand.

Also, I learned that Claude skills are portable, so I can just bring them across to another provider (noting they'll need tweaking to fit the model)

Yeah, I've been considering trying out ChatGPT again based on your original post. I've heard other people report it's worth making the switch.

... However, I've found it to be useful as it pushes me to consider alternative approaches and develop a deeper understanding of the data. BUT, some of the rabbit holes could probably be avoided by having the draft reviewed in a fresh chat (or via subagents using Claude code). 🤔

... Right now I've got Claude producing a fully documented quarto file that meets my style requirements. I'll then read through the draft, check the approach and make any required edits. The changes I make are usually substantial, so I wouldn't call it a time saver. ...

110