A thread for unboxing AI
The rapid progression of AI chatbots made me think that we need a thread devoted to a discussion of the impact that AI i
Our vendors are telling us we can't even get pricing on servers anymore.
We would have to order and then whenever it ships whatever the price is then is what we pay.
You can't run a business like this.
Yes you can because your competition will have to too and your end user customers will have to pay for it if they want to remain competitive.
Yes you can because your competition will have to too and your end user customers will have to pay for it if they want to remain competitive.
Peddle Capitalism 101 elsewhere, our competition is literally AI.
Consumer spending is about 70% of GDP. So if AI actually replaced most workers and people had no income, consumer demand would collapse. But if demand collapses, businesses fail regardless of how efficient their AI is, so the demand for AI collapses as well.So you think everyone will just sit around until they starve to death? (Not everyone of course. Zuckerberg et al. would st
What a comically naive oversimplification to handwave a problem absolutely nobody is preparing for
Dorsey and Zuckerberg have announced they will eliminate nearly a quarter of their workforce
By your logic surely these altruists have already put in place what they as hyper billionaires will do to even out the employment displacement they have created for no other reason than corporate greed, yes?
Just because you literally can't see the other half of the glass that isn't full doesn't mean it isn't there or others can't see it.
Not to imply I'm not at all concerned with the employment market but it's the transitional issue. Every technological revolution we've gone through ended up with most all the jobs people were doing either being eliminated/reduced and people working in jobs that they didn't even imagine.
My major concern with AI is decision making going further and further upstream.
Just because you literally can't see the other half of the glass that isn't full doesn't mean it isn't there or others can't see it.Not to imply I'm not at all concerned with the employment market but it's the transitional issue. Every technological revolution we've gone through ended up with most all the jobs people were doing either being eliminated/reduced and people working
Lol
I regret ever taking you seriously
Bai
Dorsey and Zuckerberg have announced they will eliminate nearly a quarter of their workforce
By your logic surely these altruists have already put in place what they as hyper billionaires will do to even out the employment displacement they have created for no other reason than corporate greed, yes?
I agree with the gist of what you're saying. But it's not true in an absolute sense or in a vacuum. It's contingent on the market space they're in. Meaning if the market demand for their product were higher, they wouldn"t fire those programmers but instead leverage them with ai to meet that increased demand.
Lol at questioning their business decisions. They run successful companies. They feel ai serve a portion of their company better than humans will.
And the argument is they are bad people for not employing humans.
Go start a company and hire humans. Stop complaining so much.
An interesting article from the Guardian.
https://www.theguardian.com/global-devel...
Basically, several leading AIs when prompted would state that a picture showing the graves in Minab (from the school missile strike) in the Iran were fake. However, the image is revealed to be real and confirmed via satellite imagery.
My intent is not to pollute this thread with commentary on the Iran war, which goes on in both the Trump thread and the Iran war thread. It is the general discussion underlying this I am interested in here. AI is already used effectively for propaganda, obfuscation and PR campaigns in conflicts, but this story also shows that AI can be very limited in its ability to dig through such muck and shield the end user.
The technical reasons are not really possible to discern. It could be willful manipulation by the operator, willful manipulation of data by other parties, simply bad data learned from bad commentary on the conflict, errors, features or artifacts in the neural network, random hallucination or even the AIs being tricked by the very propaganda other AIs are used to amplify.
Is it a problem? Yes. Even the user who shields himself from AI slop, relies less on AI or uses AI carefully is affected by this, simply because the people who are less discerning will still partake in the debate and influence policy. This in a similar vein to how you can take measures to protect your house from fire, but that might not be enough to protect you from your neighbour's poor choices.
LLMs are basically heuristic generators presenting ~80% confidence outputs as truths, where consistency of correlations takes precedence over actual truth. And that's baked in because material implication (if the earth is flat, then Trump is POTUS) correlation is in its DNA as opposed to truth-maintaining natural language logic (If Trump won the election, then Trump is POTUS).
this matches what i've run into repeatedly. the problem isn't just AI being weaponized for propaganda, it's that the guardrails meant to prevent that misfire on legitimate research. ask anything touching a contested topic and the model refuses or pushes back before consulting a single source. when you force it through, the evidence often contradicts the confident prior it started with
so you get it from both ends. bad actors use AI to flood the zone, and the same models trained to resist that end up blocking the people trying to cut through it
When I said baked in, I'm talking about say ChatGPT, not LLMs in general. ChatGPT is literally trained to produce those outputs so we can't completely fine-tune the heuristics away. But we can train an LLM to answer with the truth, the whole truth and nothing but the truth; it's just expensive and would be like talking to an attorney who writes 1k page contracts.
this matches what i've run into repeatedly. the problem isn't just AI being weaponized for propaganda, it's that the guardrails meant to prevent that misfire on legitimate research. ask anything touching a contested topic and the model refuses or pushes back before consulting a single source. when you force it through, the evidence often contradicts the confident prior it start
Yes, that is pretty much the gist of it.
A several LLMs have started using indexing and keep a record of links to verify output before it is parsed to the user, some also offer links to the end user. This makes their output less malleable, it is not as easy to prompt the AI that it is wrong. That is good when we as a user is wrong, but it is pretty bad when the AI is wrong.
It is also going to be interesting in the future. Even if we ignore the bad data out there, LLMs are now training on the biggest set of data ever created where people have helped solve each-other problems or recorded knowledge. However, as more people rely on LLMs to do that, a genuine problem might be that the amount of valuable knowledge on offer declines.
So, if you want to DIY redoing the wiring loom on a 2017 Nixxan Maxima, we're probably covered well by current LLMs, because the repository the LLM trained on is "pre-AI". Doing a similar prompt in 2036 for a 2027 model might have far less quality data to base the answer on.
milla jovovich is releasing open source ai memory systems
she’s talking about it on her insta as well
this is a strange timeline
went to a conference today that was all ai focused
was terrifying, just a bunch of grifters trying to be the next person to sell pick axes and shovels
zero insights - just how can i get your money
I had just been using chatgpt and etc and wasn't super impressed. At work we got claude imbedded in an IDE where we can hook our whole repo up and assign personas and steering files to different sections to achieve different things. Had it review our requirements and design to build documents, establish trace from requirements to design to code, and write automated tests. Was fairly eye opening to say the least. It still wasn't perfect but gets you 70% of the way there.
been diving really deep on llms and ai lately to better understand them since my entire job is dependent upon them
it's pretty scary
we fundamentally don't know how or why they reach the conclusions they reach - it's to the point where there's genuine debate over whether they actually understand what a cat is or whether they are just pattern matching and giving what it feels are appropriate responses to a cat conversation despite not having the slightest understanding of the nature of the conversation
considering how empowered they are, that's very frightening, it's the equivalent of running a nuclear reactor before the discovery of atoms, just via trial and error, we don't understand why but we know if we do these things to uranium it gets hot and we can harness that for energy
been diving really deep on llms and ai lately to better understand them since my entire job is dependent upon themit's pretty scarywe fundamentally don't know how or why they reach the conclusions they reach -
Re the bolded, I certainly was under the impression that it was more the latter than the former, with the caveat that I guess "understanding," even in the context of a human brain, could be characterized a number of different ways.
When it starts just 'pattern matching' with objects in it's model of the real world so that the word 'cat' gets matched with objects in it's internal model that corresponds to cats in the real world then then will people call that understanding?
Re the bolded, I certainly was under the impression that it was more the latter than the former, with the caveat that I guess "understanding," even in the context of a human brain, could be characterized a number of different ways.
yes, the original expectation in the field was that these models would stay fairly shallow pattern completers and never develop anything resembling real understanding of concepts like a cat
and in a literal sense that skepticism made sense. they aren't actually "reading" what you give them. they break everything down into numbers and essentially decide where they've seen patterns like that before and what kind of response a human would have given in that situation. there's no comprehension happening in any traditional sense
what's changed is scale. at large enough sizes they start producing behaviors that look like reasoning, abstraction, and planning, and here's the uncomfortable part: we don't actually know why
it's not a logging problem either. there's no decision tree to follow, no line of code that says "this is why i said cat." the behavior emerges from billions of parameters interacting simultaneously and unpacking that in any human-readable way is an unsolved problem
mechanistic interpretability is an entire field now dedicated to reverse engineering these models from the inside. anthropic and others are actively trying to understand their own systems and have made some progress, but nothing close to a full account of how a response gets generated
so both sides of the debate are largely reasoning from the outside in. it's not philosophers arguing about definitions. it's that the people who built these systems can't fully explain them either
It's uncannily like the same problem of unpacking how the human brain comes up with stuff.
It's almost like they're neural nets. At some point they might even abstract the very useful pattern of ooh 'that's an x and that's a y following from an x, must be y'
The first AI war: US and Israel use Iran to test autonomous tech
Questions have been raised about whether artificial intelligence misidentified the Iranian school that was bombed, killing 110 children and dozens of others
https://www.thetimes.com/world/middle-ea...
There's lots of talk of the dangers of AI becoming dangerous to humanity in itself but the far bigger present danger the use of this AI by humans. The speed at which we are developing the ability to target/kill massively more efficiently is breathtaking. Humans will increasingly not be involved in the individual decisions, it will be autonomous.
The current war are a testing and breeding ground for what is coming soon if we can't get off the path of war instead of law.
I doubt that automated AI weapons will be any worse than dumb weapons in that regard, which carries obscene costs in civilian lives.
Dumb weapons and anti-personnel mines are a good case study for where automated drones weapons are headed. Even though out 8 out 10 dead from anti-personnel mines are civilians, the US, China and Russia has never been signatories on the Ottawa convention that bans their use, and Russia in particular uses them prolifically. As the world grows increasingly unstable, the outcome has been for nations to withdraw from the Ottawa convention, especially countries bordering Russia.
The same will be the case for automated weapons and in particular drone warfare. As they are deployed, nations will respond to their use by adopting them. First the countries attacked by such weapons, then nations who feel threatened and then it will cascade to everyone else as it becomes the reality of warfare.
The only possible way to do something about that is for the world to become more stable and for diplomacy to be predictable and respected, so that conventions are possible. This is not the direction the world is headed in, and I see no reason to believe that this trajectory will change. The general ideological consensus now is that since diplomacy is often flawed, it is therefore worthless. Thus the political reality is that kids getting blown up is seen as the hawks as acceptable and by the doves as inevitable, and both are content with that as long as some dude on Twitter affirms that it is everyone else who are wrong.
It's the efficiency that goes up dramatically. AI targeting and attacking with no human in the loop is orders of magnitude faster. Drones as we know them are just the start, AI controlled robots are not sci fi anymore. Not just human like bipeds, there's dog like, and smaller and bigger. It's all becoming real now. Still primitive and almost entirely human controlled which is a severe constraint that will be lifted. The pace of change is so rapid. Sure everyone will adopt it but that doesn't mean it's not far more deadly and destructive.
