@Di4na@hachyderm.io avatar

Di4na

@Di4na@hachyderm.io

SRE. Elixir Dev. Learner in Resiliency. French.
All Opinions are my own. And i have a lot.

Co-Founder and President Haruspex.dev

dom. He/him.

Blog: Softwaremaxims.com

This profile is from a federated server and may be incomplete. Browse more on the original instance.

jmaris, to ai
@jmaris@eupolicy.social avatar

The Trilemma: or why AI won't be as revolutionary as we expect.

We are in the midst of an AI revolution... or are we? As business leaders around the world scramble to integrate AI into their products and business practices, the cracks are beginning to emerge. We've seen GPT4 fail at basic reasoning, Google's AI search telling people to jump from a bridge, Microsoft's copilot inventing quotes from Vladimir Putin, and recidivism algorithms that continue to send the wrong people to prison.

Di4na,
@Di4na@hachyderm.io avatar

@jmaris i mean you are wrong, but in the other way.

You are considering that these techniques could work and be useful/financially viable with better data.

They probably are not, these things fundamentally do not work like that. All that is presented to the public about it is smoke and mirrors to hide the scary reality.

None of these companies can make enough money to fund US pensions.

icejam_, to random
@icejam_@hachyderm.io avatar

I understand how Copilot would make money for Microsoft, especially if it’s cheap enough to run.

But I can’t figure out what Google is trying to achieve. It’s as if nobody asked “how does adding Gemini to search turn into money?” The company is supposed to be led by shrewd McKinsey people!

Di4na,
@Di4na@hachyderm.io avatar

@icejam_ tbf, i bet it was asked in the investor call for the quaterly results. There is nearly always one asking that.

Di4na,
@Di4na@hachyderm.io avatar

@icejam_ which is a fancy term for "journalists for financial investors"

mpesce, to random
@mpesce@arvr.social avatar

this whole AI-in-the-search-results thingy feels like such an unforced error one wonders how it made it through all the processes intended to protect their crown jewels?

perhaps those processes don't exist? or were ignored in panic?

https://www.nbcnews.com/tech/tech-news/google-ai-im-feeling-depressed-cheese-not-sticking-to-pizza-error-rcna153301

Di4na,
@Di4na@hachyderm.io avatar

@mpesce they definitely do not exist. I see how people would think they exist, but that would be ignoring reality for a loooong time

grimalkina, to random
@grimalkina@mastodon.social avatar

The answer to "how can we make more technology work better and more for everyone" PROBABLY can't hinge on "individual software developers are responsible for knowing internalizing and perfectly executing every single thing in the world and perfectly understanding the needs of billions of people" eh?

Di4na,
@Di4na@hachyderm.io avatar

@grimalkina @williampietri there is a reason i always end up with the resilience engineering people.

It is fundamentally based on "well. If we follow the accepted theory, none of engineering and infrastructure should work. And yet it does. Time to go look at what really happen and start by observing."

Patricia, to random
@Patricia@vivaldi.net avatar

I have thought probably 20 times over the past few days “Patricia, this is very harsh and if it were you you would be very upset” and yes, I would. But I land at this every time:

The author I’m sure is very sensitive to criticism of their work, but their work is literally made to affect other people. To change their entire day to day. Teams broken up, folks losing their jobs, all sorts of pretty drastic changes. To coddle their feelings seems disproportionate to the effect they are actually trying to produce in the world.

If they are successful they will change the day to day of tens of thousands of people. So a certain level of honest analysis is not only fair, but to be honest, sorely lacking.

Di4na,
@Di4na@hachyderm.io avatar

@Patricia things get even funnier when you try to find supporting evidence.

The litterature is limited but exist and uh...

Well.

Patricia, to random
@Patricia@vivaldi.net avatar

I saw @HalvarFlake
original keynote and this looks like an amazing update.
https://youtu.be/xA-ns0zi0k0?si=4COaRkpVT6IvTWJT

Di4na,
@Di4na@hachyderm.io avatar

@Patricia @HalvarFlake feasible maybe not but like...
There was a lot of gpu made and production line for them. They needed something to do

chrisamaphone, to random
@chrisamaphone@hci.social avatar

are... reversible compilers a thing?

Di4na,
@Di4na@hachyderm.io avatar

@chrisamaphone hmmm that is an interesting questions. I can think of a few things around this, but explicitly is hard. I suppose you could consider some of the egraph work to be like that? Up to a point at least? Or the whole work to make internal representation that can be resugared into nice error message?

Fully reversible would be hard though, you would have to keep the semantic and syntactic information end to end.

Di4na,
@Di4na@hachyderm.io avatar

@regehr @chrisamaphone yeah. Main thing that come to mind is
https://blog.brownplt.org/2016/02/06/resugaring.html and follow up work

Di4na, to foss
@Di4na@hachyderm.io avatar

Btw, if you liked my "I Am Not A Supplier" blogpost, I am searching for a job.

Happy to pick a job that is mostly going to tell people working on this about the reality of the maintainer life. Just saying.

https://www.softwaremaxims.com/blog/not-a-supplier

https://www.softwaremaxims.com/resume

#foss #opensource

Patricia, to random
@Patricia@vivaldi.net avatar

Alright folks, these tech process authors are confidently saying how we should organize to make stuff. But what do you think? What is the best tech org and why?

Di4na,
@Di4na@hachyderm.io avatar

@Patricia my answer is "the one that provide you the fastest feedback in the constraints your environment gives you".

Sadly the problem is deeper. The whole SDLC idea has no basis in reality. The whole model need to burn. I got a "models we organise by" talk i need to sit down to offer to srecon on that.

Di4na,
@Di4na@hachyderm.io avatar

@Patricia I mean I do not disagree and yet. We still have traces of it everywhere in the field.

Look at the "infamous" devops infinity symbol.

Or the whole "bugs cost less to fix the earlier they are found. Design fix cost less than prod". This one is a really persistent leprechaun.

Or the whole TDD thing. Or... I could keep going

Di4na,
@Di4na@hachyderm.io avatar

@Patricia I mean yeah. Totally.

But also, we have pretty good science support to show that these models are really far from reality.

Di4na,
@Di4na@hachyderm.io avatar

@Patricia Totally, but that also mean that we may want to accept when some model that were useful are not anymore :D

grimalkina, to random
@grimalkina@mastodon.social avatar

By the way just a general comment that if I only studied the things I liked and supported personally I wouldn't be much of a social scientist

Di4na,
@Di4na@hachyderm.io avatar

@grimalkina oh gosh yes. The number of things I wanted to be true and I had to reflect on myself upon finding literature and research on because.

Nope. Not substantiated. That was all on me and my personal models of the world.

It is one of these things that make science science, but it is not well communicated I feel

grimalkina, to random
@grimalkina@mastodon.social avatar

Pointed to this paper from a column on it: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4464593

Folks in dev psych and elsewhere often talk about girls being underconfident. But how rarely we frame in terms of boys' overconfidence.

"Across a range of countries, contexts, and domains, men have been found to exhibit higher degrees of confidence in their ability than women (Kay and Shipman, 2014). This phenomenon has been particularly salient in the fields of science, technology, engineering, and mathematics (STEM)."

Di4na,
@Di4na@hachyderm.io avatar

@grimalkina :cough:. So uh. About these evaluation of LLMs... (I know, i know, i am moving away from the very important problem of male overconfidence. But also)

Di4na,
@Di4na@hachyderm.io avatar

@grimalkina yeah. I have a few side results from a couple studies that hint as stuff like "code done with AI assistant had more bugs but devs that had used it were far more confident on the quality of the result that devs that did not".

Which is something I am also surprised I am not seeing more research on.

Di4na,
@Di4na@hachyderm.io avatar

@grimalkina I mean that and also...

It would be really hard to get support for it. Like everyone (ok loud minor but signigicant) in the field says "well this is obviously better and really useful" and the researcher would come in saying "well. Ever heard about bloodletting?"

Di4na,
@Di4na@hachyderm.io avatar

@grimalkina everytime people tell me "well, ofc we would not let it commit without human review". And my little voice in my head says "we have no solid evidence that human reviews works in this situation"

Di4na,
@Di4na@hachyderm.io avatar

@grimalkina fair fair! But that is also already a lot of work to get to that point, which presupposes a high motivation or time and all to do it.

I mean it is great and I want more of it! Also mindful of the realities of research and pressures.

Di4na,
@Di4na@hachyderm.io avatar

@grimalkina that and also... How to say that.

Hypothesis: reviews work through heuristics built by knowing what kind of mistakes a human do.

LLM evaluation could be biased by abusing these heuristic and making mistakes in blind spot. Or using "code style" that we heuristically associate to skill.

After all, did they train for "objective" (yes i know, the term is fraught with problems) measures of success or for humans reviews results?

Di4na, to elixir
@Di4na@hachyderm.io avatar

I am looking for a new job. Elixir/erlang SWE and/or ops/SRE related. Size of the company does not matter. I have some ethical rules (gambling, blockchain and probably most AI company,...). I only work remotely from France. Yes I would prefer a FTE french contract, but I can do self employed contracts.

You can find more about my career at https://www.softwaremaxims.com/resume

einarwh, to random
@einarwh@mastodon.social avatar

Developers eagerly and earnestly replacing legacy systems with new systems built with technologies with dramatically shorter half-lives than the ones they are replacing.

Di4na,
@Di4na@hachyderm.io avatar

@TheTraveller @einarwh @deborahh more that we have experience about maintaining things past their time and we refuse that pain.

Shorter half life means less pain for us.

Patricia, to random
@Patricia@vivaldi.net avatar

Ok, I’m sorry, I’m going to ruffle feathers here but… I’m trying to read some newer development process books and… oh my… even super popular ones are so immensely long winded and unconvincing in their dogmatic argumentation: this is bad, this is good, because I said so that’s why.

Recent examples that I’m struggling to finish: “Team Topologies” and “Data Mesh” - I mean they might be great but I’m getting strong “this should’ve been a blogpost” feels.

Di4na,
@Di4na@hachyderm.io avatar

@Patricia fun fact: even manufacturing does not fit into the manufacturing model!

Di4na,
@Di4na@hachyderm.io avatar

@Patricia there is a reason i call reading these "opposition research" and not "learning" :)

  • All
  • Subscribed
  • Moderated
  • Favorites
  • •
  • JUstTest
  • osvaldo12
  • magazineikmin
  • khanakhh
  • InstantRegret
  • thenastyranch
  • Youngstown
  • everett
  • mdbf
  • slotface
  • ngwrru68w68
  • DreamBathrooms
  • kavyap
  • tester
  • anitta
  • rosin
  • Durango
  • tacticalgear
  • modclub
  • cubers
  • GTA5RPClips
  • ethstaker
  • normalnudes
  • cisconetworking
  • Leos
  • megavids
  • provamag3
  • lostlight
  • All magazines
    •