Its unfortunate that there’s mode collapse around what the consensus “best way” to use these things are. It’s too bad we didn’t have a period where these things were great teachers but didn’t attempt to write code because in my opinion the ideal way to use them is not by agents mass producing sloppy buggy disorganized code, but to teach you things way faster than the old alternatives, rubber duck, and occasionally write snippets of functions when your brain is too tired or it’s throwaway cli code or some api you’re not familiar with.
> It’s too bad we didn’t have a period where these things were great teachers but didn’t attempt to write code
The period is now. Just add "be a great teacher but don't attempt to write code" in the prompt.
(yes, it's a teacher who gets things wrong from time to time. You still need to refer to the source and ground truth just like when you're taught by a human teacher.)
> to teach you things way faster than the old alternatives
I'm not sure if you ever had a teacher or instructor that you didn't trust, because they were a compulsive liar or addiction or any other issue. I didn't (as least not that I can remember) but I know I would be VERY on guard about it. I imagine I would consequently be quite stressed learning with them, even if they were brilliant, kind, etc.
It would feel a bit like walking on thin ice to get to a beautiful island. Sure, it's not infeasible and if you somehow make it, it might be worth the risk, but honestly wouldn't you prefer a slower boat?
I agree, it can be incredibly frustrating at times. My rule is that if it “compiles” in my brain as an understood idea then i accept it. I also push back a lot (sometimes it points out good errors in my thinking, sometimes it admits it hallucinated). Real humans hallucinate a lot as well or confidently state subtly wrong ideas, it’s a good habit anyway. It’s basically the same approach when presented with a “formula” for something in school. If i dont know how to derive/prove it then i dont accept it as part of my memorized or accepted toolkit/things i use (and try to forget it). If it fits with the rest of my network of understood ideas i do. It’s annoying but still more time efficient than trawling through lecture slides with domain specific language etc
> Real humans hallucinate a lot as well or confidently state subtly wrong ideas, it’s a good habit anyway.
I think that's actually deeply different. If a human keeps on apologizing because they are being caught in a lie, or just a mistake, you distrust them a LOT more. It's not normal to shrug off a problem then REPEAT it.
I imagine the cost of a mistake is exponential, not linear. So when somebody says "oops, you got me there!" I don't mistrust them just marginally more, I distrust them a LOT more and it will take a ton of effort, if even feasible, to get back to the initial level of trust.
I do not think it's at all equivalent to what "Real humans" do. Yes, we do mistake, but the humans you trust and want to partner with are precisely the one who are accountable when they make mistakes.
> My rule is that if it “compiles” in my brain as an understood idea then i accept it.
Unfortunately, individual people are not anywhere as reliable as a compiler for ensuring compliance to reality. We are particularly susceptible to flattery and other emotional manipulation, which LLMs frequently employ. This becomes particularly problematic when you ask for feedback on an idea.
In that case, a useful hack is to frame prompts as if you're an impartial observer and want help evaluating something, not as if the idea under evaluation is your own.
I feel like this is partially a skill issue - You can get direct, cited information from LLMs. There's a level of personal responsibility for over-using the tools and letting them feed you bad/false information, but if you try researching specific abstractions, newer documentation, most LLMS now correctly call and research the tools available, directly citing them.
I think you can build a very easy workflow that reinforces rather than replaces learning, I've used a citation flow to link and put into practice a ton of more advanced programming techniques, that I found incredibly difficult to locate and research before AI.
I'd say the comparison is faulty, it's more akin to swimming to an island (no-ai) vs using a boat. You control the speed and direction of the boat, which also means you have the responsbility of directing it to the correct location.
The analogy was about the unknown thinnest of the ice, not just the fastest way to get there. It's specifically about the lack of reliability of the process.
Yes, I was disagreeing with the premise of the analogy - what would the slow boat in this case be? As my experience, going through software engineering before AI, is that you'd get lost to the ice, with nobody to really help you get out.
If you get lost on the ice and you have someone who confidently tells you the path but is sometimes wrong, is it actually helpful?
PS: sorry if the analogy is a bit wonky but it's quite dear to me as I do ice skating on frozen lakes and it's basically a life or death information "game" that I can relate to. It might not be a great analogy for others.
Haha it's a good analogy, i'm being a little bit argumentative for the sake of it potentially.
I guess in my view - the main alternative you'd have beforehand is just to drown.
For me, AI sits in a space where if you know how to use it, it can tell you all the thin spots of the ice accurately. You can then verify those spots, but there's a level of personal responsibility of verification.
I'd agree there's currently a ton of people that are using these tools to essentially just find the specific route - but i'd argue those people probably shouldn't be skating in the first place, and would've fallen one way or the other.
> AI sits in a space where if you know how to use it, it can tell you all the thin spots of the ice accurately. You can then verify those spots, but there's a level of personal responsibility of verification.
Right, but AFAICT most people just venture over the ice and don't bother to check. In fact a lot of people venture there, do check once or twice, then check less and less frequently. The fact that you do it is great but others seem a lot less careful, until cracks start to show and then it might be too late.
I'd only argue that people were doing this before AI, slop development was just copy pasting from the first stack overflow issue that matched the question rather than thinking
So i'd argue there's a part of it that is just personal responsibility with how these tools are used
More likely some high ranking party member's nepobaby from Gemini sniffed success with Qwen and the original folks just walked away as their reward disappeared.
There is no source. But the party in China does have ultimate control.
There would never be an Anthropic/Pentagon situation in China, because in China there isn't actually separation between the military and any given AI company. The party is fully in control.
Interesting reading this. It reminds me of my time in cryptocurrency sector. I suspected that some team members were paid by Ethereum folks to sabotage our project. Why do I suspect Ethereum? Because our project founders ended up switching to the Ethereum ecosystem and ignored/suppressed better solutions from their own ecosystem. I think there's something about tech hype which attracts these kinds of people who like to play dirty.
apples v.s. oranges. The later is true, Emad did get sabotaged (for not being able to raise money in time, about 8-month before he's leaving). Junyang didn't have that long arc of incidents.
The default of making a public discord for your project/company always seemed like a bad idea anyway. It’ll always devolve into some drama or distracting overhead to moderate it
I definitely think they’ve nailed the personality better than others too. Gemini and grok are always paragraphs and paragraphs of text to sift through for something that with openai is usually digested to much less
I’ve moved to more closed source projects for this reason (just for the fun of coding rather than sharing). Though I suspect they still use private github repos in their deals to microsoft
I was in management i probably also wouldn’t like my designers to use AI. I pay them good money to draw original pieces and everyone can tell and it looks generic when AI is used. I’d want my moneys worth
Well, you can definitely make AI art much less obvious with the right tweaking (directly running models, blending different sub-models, etc). The bigger issues from a professional perspective are liability concerns and then, even if you have guaranteed licensed sources, the impossibility of controlling fine details. For a company like GW it's kind of pointless if it can't reflect all the decades worth of odds and ends they've built both the game and the surrounding franchise around.
I don’t know what you’re looking at, perhaps things like deepfakes are getting, but most of the graphic design done with AI that I see around looks like shit.
Sure and limit yourself at the starting point, people underestimate how much limiting these tools are, they're trained a on a fixed set can only reproduce noise from here and there
> they're trained a on a fixed set can only reproduce noise from here and there
This anti-AI argument doesn't make sense, it's like saying it's impossible to reinvent multiplication based on reading a times table. You can create new things via generalization or in-context learning (references).
In practice many image generation models aren't that powerful, but Gemini's is.
If someone created one that output multi-layer images/PSDs, which is certainly doable, it could be much more usable.
If image generation is anything like code generation then AI is not good at copying layout / art style of the coder / artist.
Using Visual Studio, all the AI code generation is applying Microsoft's syntax style and not my syntax style. The return code line might be true but the layout / art / syntax is completely off. This with a solution that has a little less than one million lines of code, at the moment, which AI can work off of.
Art is not constant. The artist has a flow and may have an idea but the art will change form with each stroke with even removing strokes that are not fitting. I see as AI generated content lacks emotion from the artist.
Image generation is nothing like AI code generation in this regard. Copying artist style is one of the things that is explicitly quite easy to do for open-weight models. Go to civitai and there are a million LORAs trained specifically on recreating artist style. Earlier on in the Stable Diffusion days it even got fairly meanspirited - someone would make a lora for an artist (or there would be enough in the training data for the base model to not need it) and an artist would complain about people using it to copy their style, and then there would be an influx of people making more and better LORAs for that artist. Sam Yang put out what was initially a relatively tame tweet complaining about it, and people instantly started trying to train them just to replicate his style even more closely.
Note, the original artist whose style Stable Diffusion was supposedly copying (Greg someone, a "concept art matte painting" artist) was in fact never in the training data.
Style is in the eye of the beholder and it seems that the text encoder just interpreted his name closely enough for it to seemingly work.
Putting it in the context of an anti-AI argument doesn't make sense. AI was everywhere, like in photoshop brushes, way before it became a general buzzword for LLMs or image generation. I'm not anti-AI but that it can come up with a limit set based on its training data it simply is the truth. Sure one can get inspiration from a "times table" but if you only see 8s and 9s multiplied you're limiting yourself
> If someone created one that output multi-layer images/PSDs, which is certainly doable, it could be much more usable.
This reminds me, if you ask most image models for something "with a transparent background", it'll generate an image on top of a Photoshop checkerboard, and sometimes it'll draw the checkerboard wrong.
For an artist, the starting point is blank page, followed by a blur of erased initial sketches/strokes. And, sources of inspiration are still a useful thing.
Given that it's under scrutiny for regulatory bypass, it's not a purchase and is being reviewed as circumventing those very rules. Might not even happen.
I know, I'm joking: Trump likes Nvidia, but maybe he'll bump the Chinese tax to 30% to approve this deal? In a way I hope he pulls something like that, to punish Huang for his boot shining manipulations.
This seems wildly optimistic to me. We see the same complacency and/or unawareness with e.g. Flock in society - the truth is most people really just don't think about it, or even mind when they do.
If price per compute keeps going down we effectively keep having moore’s law for parallel compute like GPUs. (Just get better at making ML models that don’t have as big of a communication bottleneck.)
Sounds to me like runway released it without consulting stability, called it “1.5” - which according to the license they’re allowed to do but pretty scammy since emad had hyped a model with that label. And now stability is deciding to call this the official release to be nice to runway and avoid a general PR thing and community infight.
1.5 was apparently held back by Stability for weeks.
Runway finally decided to just release it.
Stability requested a take down, and here you see the Runway CEO telling Stability in no uncertain terms "this ours to release, we created it, it's under an OS license, you don't hold any IP rights here; all you did was provide compute"
If anything this is a pretty stern rebuke of Stability and a sign of considerable disagreement between the two parties.
Well if that’s the case that’s still a pretty shitty thing to do on runways part. Just be curteous to what stability’s needs are, keep good business relations. Weird behaviour and I wouldn’t be surprised if in the future runway are silently excluded from before-public releases (which seems to be many in the years ahead).
Doesn't the "OS license" mean that Runway has permission to release it already? Ack that there might be other agreements and business relations involved though.
Release it, fine. There’s been lots of fine tuned and continued trained SD models. Just don’t call it “1.5” which is the specific label for the model stability is training internally. Again the license ‘permits’ them to do it but seems like a very bad business decision since runway given what their service does would likely benefit hugely from early access to eg stability’s future text2video models (etc), which they now likely won’t get until everyone else (leaving someone else to possibly take market share in their field, and if this is the trade - because they got ‘impatient’ - that seems awfully not smart).
reply