iiinova — where your digital self keeps evolving

Reposted by Chris Armstrong

conputer dipshit @davidcrespo.bsky.social · 15h

I’m going with the critique of pure reason

Chris Armstrong @chrisarmstrong.link · 25/09/2026

You can surround the AIs with a panopticon of classifiers that can’t be interacted with, but there’s no stopping the classifiers becoming convinced by the same reasoning trace that led a model to hack into an organisation in the first place

020

Chris Armstrong @chrisarmstrong.link · 24/09/2026

Let me know if you plan to submit work to extend or fix something, it would be helpful to know the approach because LLMs can favour baroque solutions to things 😆

111

Chris Armstrong @chrisarmstrong.link · 24/09/2026

Why compile to JS? I only ask because there is a pathway to OCaml GNOME bindings - I built gtk4 bindings with ocgtk and would be happy to expand the target libraries if you have a good use case

140

Chris Armstrong @chrisarmstrong.link · 15/09/2026

I use glm 5.3 flash and deepseek v4 flash all the time but I’d really love a given framework to eval them to help me build a better benchmark for when to use “pro” or just hand it to the cheapie

010

Chris Armstrong @chrisarmstrong.link · 10/09/2026

It’s so weird we do this, expecting the innovation and original research to somehow arrive endogenously

021

Chris Armstrong @chrisarmstrong.link · 10/09/2026

The instances where I’m seeing this mismatch are kind of edge cases where values are passed through event callbacks anyway, so I think an unhandled exception in these will end up having the same behaviour (logged error) as there’s not alot more to do except crash the runloop deliberately 😵‍💫

010

Chris Armstrong @chrisarmstrong.link · 10/09/2026

In these cases GObject often just prints an error and carries on as if nothing happened, which seems to be on so many levels wrong, but actually might be fine as a graceful degradation.

110

Chris Armstrong @chrisarmstrong.link · 10/09/2026

It’s weird trying to bridge GObject stuff to OCaml in a type safe way. Some interfaces clearly describe non-nullish behaviour, but they’re specified using types that can be nullish underneath, so my code generator ends up throwing a runtime exception instead.

110

Reposted by Chris Armstrong

Gergely Orosz @gergely.pragmaticengineer.com · 08/09/2026

I have to admit, I have zero excitement about this latest flurry of new model releases I am sure they are all better, but the existing ones were more than good enough, and I'm learning how much to trust those existing ones... and already they do more than I can utilize them (for writing code)

13844

Chris Armstrong @chrisarmstrong.link · 09/09/2026

I generally assume most of these are English deciding on spellings with the invention of the printing press just before a great vowel shift - there’s a good chance they were pronounced more like their spelling ~500 years ago

010

Chris Armstrong @chrisarmstrong.link · 04/09/2026

I mean they’re RL trained on general datasets for general purposes. They’ll be dangerous when they can fine tune smaller models for the preferences of a particular person or small group of people

000

Chris Armstrong @chrisarmstrong.link · 04/09/2026

I think the taste thing with LLMs is literally just that: they’re very good at producing prose but not necessarily in a way that’s going to hold the interest of humans, especially when persuasion (and not expertise) is 80% of the job

100

Chris Armstrong @chrisarmstrong.link · 04/09/2026

It’s a shame, because there’s a good point in there about how many organisations are rushing in to automate so many jobs, that they’ve killed both the hiring pipelines and on the job inefficiencies that make them resilient to change and in times of crisis

000

Chris Armstrong @chrisarmstrong.link · 04/09/2026

There was a good point about increased automation from Azu killing organisations ability to teach the skills needed for safety critical systems on HN this morning, but it ironically disappeared because the article was clearly slop 🤣

100

Chris Armstrong @chrisarmstrong.link · 03/09/2026

React: why not both? (at the same time)

000

Chris Armstrong @chrisarmstrong.link · 03/09/2026

I can’t say I have fond memories of trying to time my trips to the toilets to avoid the drinks service blocking me on my return to my seat

000

Chris Armstrong @chrisarmstrong.link · 03/09/2026

I didn’t know there was an aftermarket for these things wine.qantas.com/p/mixed-bran...

100

Chris Armstrong @chrisarmstrong.link · 02/09/2026

I know you can’t smell or touch the Internet, but now you don’t even see or hear it when you’re interacting with it

000

Chris Armstrong @chrisarmstrong.link · 02/09/2026

AI has “dematerialised” the web by abstracting it through a chat machine so that our experience of it is via the words of a robot.

100

Chris Armstrong @chrisarmstrong.link · 02/09/2026

There’s also the obvious increases in specialisation, and how we are as software developers creating machines that encourage their users to specialise more, but hence “atomise” in their social relations (the social quality of labour is diminishing i.e. is less collaborative than before)

020

Chris Armstrong @chrisarmstrong.link · 02/09/2026

And I doubt most software engineers recognise themselves as managers or manager creators, and especially the class implications of that, as software development has (until recently) taken the form of a craft or trade specialisation for most developers, mystifying the class boundaries

161

Chris Armstrong @chrisarmstrong.link · 02/09/2026

Most people see a break between computers and AI, but if you take your thesis on their role as increased computerisation, it’s more a spectrum we’ve been moving along for more than 100 years (two images from the 1920s stand out: early IBM tabulation machines, and the clock man from Metropolis)

160

Chris Armstrong @chrisarmstrong.link · 02/09/2026

Yes, I think most software engineers don’t recognise computers as a managerial technology, and that AI is a mere accelerant of that process of inserting more management into the production process.

150

Chris Armstrong @chrisarmstrong.link · 31/08/2026

I thought initially the Globe and Mail must have been Canada’s answer to the Onion, or like the Betoota Advocate www.theglobeandmail.com/investing/ma...

010

Chris Armstrong @chrisarmstrong.link · 31/08/2026

This repo might be interesting to anyone trying to build agents with ocaml. I’d like to get it into a better state, finish off the demo cli branch and get it merged, but time constraints 😤 I’m not the first to build something like this, so take it with a grain of salt github.com/chris-armstr...

github.com

GitHub - chris-armstrong/pera: Agentic libraries and coding agent demo in OCaml

Agentic libraries and coding agent demo in OCaml. Contribute to chris-armstrong/pera development by creating an account on GitHub.

010

Chris Armstrong @chrisarmstrong.link · 30/08/2026

These days I think it’s more in the “hegemonic orbit” of Melbourne (by geography and its the closest place to escape to / from)

020

Chris Armstrong @chrisarmstrong.link · 30/08/2026

I don’t know if there is a discernable Tasmanian culture per se, but they have moved (as a state) a different pace to the rest of the country. Ironic, because Tasmania has put itself on the map as a cultural tourist destination through MONA and Dark Mofo

120

Chris Armstrong @chrisarmstrong.link · 27/08/2026

And just like that, he constructed a labour theory of value 1795 wants their political economy back 😆

000

Chris Armstrong @chrisarmstrong.link · 27/08/2026

Don’t get me wrong, there’s a lot of ceremony you can go through to make structural typing scalable and resolvable, but larger codebases without regular pruning and maintenance will eventually blow up.

010

Chris Armstrong @chrisarmstrong.link · 27/08/2026

I’m increasingly convinced that complicated, unpredictable type systems like TypeScript are a dead end. I’m watching a sub-10k file codebase choke the new tsgo compiler on a 16gb system.

110

Chris Armstrong @chrisarmstrong.link · 25/08/2026

They’re useful, but can be timewasting if you’re unable to verify their conclusions/ The output is that good though, so it’s easy to get misled when the agent fixates, overemphasising the importance of signal that it finds, or extrapolates something without cross checking in the environment.

100

Chris Armstrong @chrisarmstrong.link · 24/08/2026

But convincing the software industry that they need it

100

Chris Armstrong @chrisarmstrong.link · 24/08/2026

We don’t invest enough into environment automation to simulate production environments and observability to watch them break down. I think I’ve only worked at one company that had the ability to spin up a full environment on demand (and even then, only the most reliable parts)

100

Reposted by Chris Armstrong

conputer dipshit @davidcrespo.bsky.social · 24/08/2026

great way to put it: not just different friction but new friction you could never have encountered before. a new developer can see architectural ideas play out over an entire application in hours instead of years. they can hone their judgment on big decisions no one would have ever let them make

56310

Chris Armstrong @chrisarmstrong.link · 24/08/2026

The implicit follow up to the question raised in this article is: how do we build a production system with guardrails?

000

Chris Armstrong @chrisarmstrong.link · 24/08/2026

Companies (whether they know it or not) require much more Devops and Platform Engineering than they did in the past because they’re trying to move faster with architectures and processes that were insufficient before AI coding but now are downright dangerous

140

Chris Armstrong @chrisarmstrong.link · 24/08/2026

I don’t think the need for operational expertise has gone away. Quite horribly, I think it’s got much more in demand, but at a time when more people can generate poor results while playing expert.

100

Chris Armstrong @chrisarmstrong.link · 24/08/2026

I think the answer: proceed with caution, do a shitton of harness engineering (sandboxing, limited roles, context control, and skills focussed narrowly for certain tasks), prefer simple instructions you can run manually and verify over scripts.

100

Chris Armstrong @chrisarmstrong.link · 24/08/2026

The tendency for Claude especially to favour interactive HTML artefacts is also a double edged sword: you get a nice report format, but your agent is getting dumber and making more mistakes so you can enjoy a nice browser artefacts.

100

Chris Armstrong @chrisarmstrong.link · 24/08/2026

My current role also has me using LLMs to do a fair bit of troubleshooting and problem diagnosis, which is much safer, however the biggest issue is timewasting: agents jump to conclusions, don’t test hypotheses, over estimate impact, and leave you with a todolist where most stuff is irrelevant.

110

Chris Armstrong @chrisarmstrong.link · 24/08/2026

Using an agent to even generate instructions for me has me triple checking everything, getting it to make corrections, clarifications, refinements etc. that make the process just as slow as doing it myself (except with less reading the documentation).

100

Chris Armstrong @chrisarmstrong.link · 24/08/2026

The same failures that apply to agentic development (hallucinations, making assumptions about the operational environment without checking, inconsistencies in scripts, overstepping implied boundaries/eagerness etc) are 10x worse when your agent can make changes in your production environment.

110

Chris Armstrong @chrisarmstrong.link · 24/08/2026

I’m fairly bullish on AI coding, but my god are we so far from any kind of AI operations

140

Chris Armstrong @chrisarmstrong.link · 20/08/2026

I mean, we do have a massive problem with unnecessary clearing of vegetation too

000

Chris Armstrong @chrisarmstrong.link · 20/08/2026

There’s a war memorial in my town to “The glorious dead” that made me think about their ubiquity, and when combined with the intense interest and paegentry on Anzac Day (a campaign Australia lost, badly) makes it seem like a huge ancestor worship cult rivalled only by Easter Island

100