@mcc So developers will stop sharing information on #StackOverflow and future #Copilot and friends will be forever stuck in the past, answering questions about historically relevant frameworks and languages. #LLM#StuckOverflow
“AI” as currently hyped is giant billion dollar companies blatantly stealing content, disregarding licenses, deceiving about capabilities, and burning the planet in the process.
It is the largest theft of intellectual property in the history of humankind, and these companies are knowingly and willing ignoring the licenses, terms of service, and laws that us lowly individuals are beholden to.
I guess we wait this one out until the “AI” bubble bursts due to the incredible subsidization the entire industry is undergoing. It is not profitable. It is not sustainable.
It will not last—but the damage to our planet and fallout from the immense amount of wasted resources will.
> "Just as GitHub was founded on Git, today we are re-founded on Copilot."
Look, I respect the heck out of the technical implementation of LLMs, but let's be honest: statistically they produce average code at best and misunderstood/invalid code most often. They re-implement old bugs and obfuscate programmer intent and anyone who is leaning on them for more than a pair assist is making software harder for the rest of us.
#Microsoft könnte ein Betriebssystem und Anwendungen entwickeln, die weniger Ressourcen verbrauchen und hätte damit eine riesigen Wirkung im globalen Kampf gegen die Klimakatastrophe.
Der Konzern entscheidet sich aber für das Gegenteil.
The #GitHub#Copilot announcement makes sense when you realize that this positions Microsoft even further to monetize the act of #coding itself.
When you handicap #developers and make them so dependent on LLMs to write code, that (eventually) they no longer know how to do so without it, then suddenly the days of working in a free editor are over.
So we know that #GitHub#Copilot was absolutely trained on GPLed code.
My naive understanding is that therefore any code Copilot generates could be (at least partially) derivative and would need to be GPLed. Where am I wrong?
I've criticized the moral stance of #Github#Copilot quite a bit in the past months. Then a few weeks ago I decided that I can't criticize what I don't know, so I gave it a try.
Premise: I'm not entirely new to AI coding assistants. I've used #Tabnine for quite a long time, but I decided to give up on it because it easily causes my neovim instances to eat half of the available RAM when typing.
What I've seen in Copilot has really surprised me. After a couple of weeks of usage, I've concluded that it definitely can't replace the more "human" side of coding - refactoring, knowing how to best arrange the components in a module, name things the right way, think in advance of possible corner cases, etc.
But it definitely saved me 70% of the time spent on boilerplate - type hints, simple docstrings, serialization/deserialization stuff, inferring required imports, etc.
Of course, I'm very happy for my digital condom when I use these products (my PiHole eagerly blocks all the calls to *.applicationinsights.azure.com), while acknowledging that no solution is really airtight when you decide to put your finger into the jam.
My ethical concerns still stand: Github is obviously leveraging its dominant position to scrape millions of FOSS repos and feed their code to closed models that they can sell for profit.
My partial alibi against this argument is that all of my projects are also GPL or MIT licensed - I may be stealing, but I'm also making sure to give back. And I'm also careful not to use these tools on the projects I work on for my employer (which forbids these tools anyway).
But hey, the productivity boost that these tools provide, if used the right way, is undeniable. Especially for the boilerplate that, let's admit it, takes most of the coding time - and it's also the least likely activity to be impacted by intellectual property concerns.
I sympathize with the concerns of some in the community who have called against the usage of these tools. But I also see the risk that those who refuse these tools will simply be outcompeted by those who use them. Filling in the type hints of a method with 15 parameters, writing documentation snippets for all of them, or writing a converter for an object with 20 attributes, takes time. No matter how experienced you are or how fast you are at typing. And it's definitely not the kind of activity that comes to our minds when we think of what we like of our job/hobby. If there's a tool out there that makes this job easier, then people who use it will just produce more code faster, while allocating more resources for the actual problem solving, and outcompete those who don't use them. Evolution always rewards those who embrace change when presented with a comparative advantage.
I still feel bad for paying $10 to Microsoft and feeding their immoral empire though. But I also feel that the state of LLM technology nowadays should be mature enough to build real FOSS competitors. Our reaction shouldn't be "it's just bad, we'll keep riding our horses while everyone switches to cars". Our reaction should be "it's a bad implementation of a good productivity idea, let's do better than this".
An idea that I've toyed with in these days is that of a "fair" AI assistant. It could be trained only on GPL/MIT code, and be released under GPL license itself - both the code and the raw dataset. It would scan all of Github (and other forges) for projects that include the right license. The dataset should also be annotated with the source of each code fragment. At the very least this should simplify ownership disputes. Ideally, this should be the starting point for a mechanism that automatically adds a comment that references the original snippet when the user presses tab, if e.g. >80% of the given suggestion matches a snippet in the training set, but I don't have a clear idea of how to efficiently run this "reverse lookup" logic with the current state of LLMs.
We could even take it one step further in fairness, and initially only scan repos that have an explicit robot.txt-like opt-in flag, where they could also specify which specific bots they want to allow/disallow.
But I don't think that the right solution is for us to just watch, condemn and accept a big comparative loss in productivity that will only benefit the closed-source projects that will keep being developed also thanks to these tools.
#IA#Microsoft#CoPilot
Je note juste le lien vers cet article pour une information précise : Même si l'IA de Microsoft - CoPilot - tourne localement sur votre ordinateur, elle communique quand même avec les serveurs de Microsoft pour s'assurer que la demande faite à l'IA est "safe".
Ce qui confirme bien qu'on peut se torcher avec les promesses de Microsoft sur le fait que l'IA sera totalement locale et qu'elle respectera notre vie privée. https://stratechery.com/2024/windows-returns/
They seek to obtain #research access to internal #Microsoft data to check whether and how the company is making sure its not-so-intelligent #AI#Copilot isn't bullshitting around in dangerous ways.
GitHub #CoPilot suggested this for the DB Name. I wonder if I can find the rest of the connection string. Also, let this be a lesson is making sure you do not commit sensitive information to #GitHub
With #LLMs and #GitHub#Copilot, we finally extend the preference for plausible answers over correct ones from political discourse to software development.
Noticed that developing with access to my personal data is dangerous, given random dependencies and auto-updating Code extensions…
Set up a nice local #nixos container for dev stuff…
Hoped to keep all tokens and keys out of it (would run git push outside that container)…
Only to notice that #Github#copilot only works with a full-access github token… sigh.
Anyways, still an improvement. Blog post to follow.
➡️ Il semblerait que Microsoft ait commencé à installer son I.A. "CoPilot" sur les machines Windows sans vous demander votre avis, y compris sur des Windows Server : https://mastodon.gamedev.place/@sos/112274291843803661
Lately I've been tinkering with a python bot to help me write #AltText for all my photographs through Copilot. I'm not really into AI but I believe this is one of the useful use cases in which AI might help.
Of course human reasoning cannot be compared but if it helps accessibility I'm all for it.
I'll post the GitHub link when after writing down some documentation
Asked #Copilot (formerly #BingChat) a familiar riddle but with numbers changed to make it impossible. It generated the same solution but substituting the numbers so that it ends up with the nonsense claim:
The newest most powerful chips from Intel and AMD don't qualify for Microsoft's newest CoPilot+ AI branding. We're already seeing INCREDIBLE deals on crazy powerful PCs, because of all the AI hype! Here's my review of the Geekom A8, with a BEAST of an AMD chip inside! https://somegadgetguy.com/b/45e