About AI and truck factor

This post https://www.shapeandship.ai/p/cheap-to-build-costly-to-keep
argues that with AI coding the truck factor tends toward 0...
/In a post-mortem, this surfaces as "nobody knew this code existed,"
which is worse than "only one person knew."/

They talk about "comprehension debt" (evolution of the technical debt
concept)

cheers

nicolas

On 2026-05-07 22:01, Koen De Hondt wrote:

--
Nicolas Anquetil
Evref team -- Inria Lille

Hello Koen and everyone.

It’s really nice to see people getting more and more interested in AI, especially for Pharo.
But I think you are wrong about ChatPharo. I understand very clearly your point that the AI for and in Pharo is in his realy stagses. But I’m working very hard to make it more usable for more and more people ( people are also scared of using AI in Pharo because they had bad experiences with ChatGPT when it was launched in 2022). It wasn’t created to do code generation tasks.

That said. I wanted to clarify some points about ChatPharo:

• ChatPharo is not a toy. It has a big engineering part inside of it (Multiple people, including PhDs, and engineers, contribute to it). Indeed, I’m not doing a lot of publicity because i also doing my PhD at the same time.

• From an AI engineer's POV, the dependencies on the company that provides models accessible from the web are not very convenient. I agreed that Claude is the best model for code generation (and it appears everything else but you are limited to the company API and the backend. FYI: the newest Claude and OpenAI models have more than 1T parameters.) But let me ask you a question, what if tmrw claude changes its policies now after their deal with SpaceX? (They will expect more and more people that subscirbe to their 20 euros, right? ) Well, in that case, any CEO in his right mind will increase the price ( it will be one million cause of that, of course, why do they do that, whatever they tell, the people will buy it because now they can’t live without it). That’s why it’s not very convenient ( this is what i think).

• Ollama ( that were considered one of the best companies that provided the models for free (local and on the cloud). Now you should pay for the model inference on the cloud ( it was free ), because they saw that more people are getting interested in the model that runs on the cloud ( and it’s free), they now prohibit you from using it unless you pay 20 euros.

• Why do you think that the model that appears now is getting better? It’s not training or something else. There's nothing left in the whole internet to train on anyway. They now want real data ( you might see in most of the newest technical papers the sentence high quality token. You might ask yourself a question about where this data appears if there is nothing left on the internet to train?

Now, make no mistakes, I’m not telling you NOT to use AI for Pharo.  I and (we) are and (won’t) be able to concure with the companies that have models that have > 1T of parameters. (We don't have the budget, nor the resources for that). I’m just trying to justify that chatpharo for now is not (and won’t) be capable of competing with the models from the biggest company. Ofc everyone is free. But I wanted to just shed light on a few points that more and more people ask questions about.

Have a good weekend!
—
OA
https://github.com/pharo-llm

Indeed, this is a real problem. However I think that using coding agents can be very useful for Pharo. It can help us to develop features that otherwise we will simply not do. For example I am thinink about Morphic and its window managing system. Morphic is already old code that we do not understand and that we will get rid off. I would like to experiment with Claude to see if we can have that implemented (There is one person on Discord that did this for example: www.youtube.com/watch?v=NF2Bcl2Mu3Q http://www.youtube.com/watch?v=NF2Bcl2Mu3Q)

Also, Firefox people found and fixed insane security bugs thanks to LLMs too: https://hacks.mozilla.org/2026/05/behind-the-scenes-hardening-firefox/
We could also benefit from that.

I love it.
IA look like crap generator.

I see in some text I’m reading.
So for the code I can image.

S

I like what Linus Torvalds say here :
https://thenewstack.io/torvalds-ai-programming-productivity/

In the best of cases, is a productivity tool. Just like compilers were
at a point (compilers did not have all the ecological issues associated,
of course).

The first thing we need to do is stop anthropomorphizing tools. It
really freaks me out when people talk about LLMs — and agents — as if
they were human.

Esteban

On 5/8/26 17:26, Stephane Ducasse via Pharo-dev wrote:

may be AI will allow us to raise the level of abstraction, just as
compilers, or OO did ?

We (humans) will still program (through AI) but the programming language
will be different

nicolas

On 2026-06-07 18:39, Esteban Lorenzano via Pharo-dev wrote:

--
Nicolas Anquetil
Evref team -- Inria Lille

in fact Linus says AI may increment productivity x10, but compilers did
it x1000, so we are not quite close :)

Esteban

On 6/7/26 18:47, Nicolas Anquetil wrote:

--
Cyril Ferlicot-Delbecquehttps://ferlicot.fr

On Sunday, June 7th, 2026 at 6:39 PM, Esteban Lorenzano via Pharo-dev pharo-dev@lists.pharo.org wrote:

+1000

I find that really creepy and so many people are doing that.

No matter how  bad you think AIs are, I think there are some areas they are always useful:

• scaffolding -- I have seen large projects fail because the later phases couldn't start until it was too late.  If you have a good spec for a big project and all the interface points, an AI is great for making temporary working implementations of all the parts.  Although it can raise amuzing questions when a 6-month 2-person project has a 100% working scaffold in a weekend.

• Code reviews with an LLM.  An agent has a limited token space.  But with a custom LLM looking things like code vs comment discrepancies, it can turn up a lot of issues.

----- Original message -----
From: Esteban Lorenzano via Pharo-dev pharo-dev@lists.pharo.org
To: pharo-dev@lists.pharo.org
Cc: Esteban Lorenzano estebanlm@netc.eu
Subject: [Pharo-dev] Re: "Claude, My New Pair Programming Partner" — New blog post on all: objects all: theTime
Date: Sunday, June 07, 2026 12:39 PM

I like what Linus Torvalds say here : https://thenewstack.io/torvalds-ai-programming-productivity/

In the best of cases, is a productivity tool. Just like compilers were at a point (compilers did not have all the ecological issues associated, of course).

The first thing we need to do is stop anthropomorphizing tools. It really freaks me out when people talk about LLMs — and agents — as if they were human.

Esteban

On 5/8/26 17:26, Stephane Ducasse via Pharo-dev wrote: