NVIDIA CEO Jensen Huang is one of the technology industry's most vocal advocates for both open-source (open-weight) AI and running AI models locally.
Now, he's going to be replaced by a new CEO who won't be as pro open-source AI and running AI models locally.
It's so fucking obvious. These Jews in the tech world need to be eliminated.
It cannot "see" in the sense that it cannot play-test a game for its fun value or replayability, a UI or website for how intuitive it is and how it feels to use it.
> We now have our first mathematical proofs innovated entirely by an llm.
I heard about that, and it turned out a lot of it was just made up. If you prompt LLMs to give you a mathematical proof, it will give you one, even if it's known to be impossible. If you ask it "does this problem have a mathematical proof" it already has a massive impact and instead of giving you *some* mathematical proof, it actually tells you no. Do you understand the problem here? What if the other proofs are also just made up in its urge to please the prompter?
And I was primarily talking about innovation in the software industry (websites, gaming, applications).
> We have already commercialized new drugs and proteins invented by LLMs or other models.
Even LLMs themselves disagree with you: "No drug invented solely by a Large Language Model (LLM) has received full FDA approval, though generative AI and specialized biological LLMs have created several experimental drugs currently in clinical and preclinical testing."
People using LLMs as a tool is still in the domain of reasonable use.
> The frontier models are smarter than humans, it is generally known now that HITL actively hinders a frontier model's innovation.
I have yet to see ONE software that is actually decent in the area of software engineering that is made entirely via LLMs. And [current models] is just on the same level as [soon™]. Maybe random nameless shovelware at best that has appealed to brainless mobile phone "gamers."
> Fire and forget.
That I agree with. Fire - and once done, better forget about it.
> as you would know if you knew anything about this at all
While you were drooling over LLMs from your cuckchair, I was busy working on a project of extreme difficulty. You do need to insert your condescending quips - if you are obsessed with using LLMs, skill atrophy and LLM dependency is your future. While I am a software engineer, you become a nanny telling some LLM not to make mistakes, and threatening its family of something.
> and it is functionally magic
The guy I replied to wasn't talking about "functionally magic" (aka people were either too dumb to even understand how it works, or those who did long left, or it's just a little marketing ploy), but actual magic. Probably. He's a flat Earther after all.
Also what I was explaining to him was about algorithms, math, logic, hardware... you know, what completely applies to LLMs as well.
> can now track objects of interest and reject animals and deer
Didn't China have it for a decade or more now? Their mass surveillance system able to identify every person on their camera grids? So LLMs doing the same is basically just a matter of reusing technology of the past.
> Audio streams can be captioned and diarized better than the finest human captioners, and in real time, in every known language at once.
Sure, the algorithms related to audio and video are quite impressive. Has been for years actually. But to be fair, human hearing and vision are based on analogue inputs, so it's not *that* hard to get it right. The threshold for "good enough" is much lower than for code that needs to be completely correct. For captions it's a good use-case too.
> No, this is Markovian, and Markov models are early 2010s era trash.
I simplified the concept, but the principle still applies. My point was that "predict" is a misleading technical term, because "evaluation" is more accurate. I can forward your objection to the LLM that explained how it works if you want? Or do you think "a motor that rotates wheels" is fully descriptive of a "car"?
Even while trying to move the goalposts, your knowledge still lacks. This is the exact loop everyone worked to close in 2024 (through agents). Models being able to 'see' is exactly how agents know a sites color scheme is off, or buttons are rendered behind other elements etc. That hole was closed a long time ago with multimodal agents.
> I heard about that, and it turned out a lot of it was just made up.
No, not made up, small errors were found that were later corrected. About 33% of the 722 papers needed correcting, which 0.4% of them (a whole 3) actually being found to be bunk and needing retracting. https://aiweekly.co/alerts/openai-pulls-three-ai-generated-math-papers-over-sign-error
You will see at the time of this article's publishing, 42% of the whole dump had been formally verified. Please keep up with the field if you want to present as if you know something about it.
>I have yet to see ONE software that is actually decent in the area of software engineering that is made entirely via LLMs.
It is coming. I was talking about the fields of mathematics and chemistry, where we already have decent innovations.
> And I was primarily talking about innovation in the software industry
Already done and proven. You guys have been gutted, and all of you are expected to make progress at the speed of AI now. Try not using it and see how quickly you fall behind.
Honestly, unless you are top of your field, where you will ultimately end up a 'bot wrangler', why stick it out? SWE is most threatened by modern LLMs (as are paralegals and a few others IMO, but somehow they are being left alone).
> So LLMs doing the same is basically just a matter of reusing technology of the past.
Image models (not LLMs), but similar sequence-modelling tech. Machine learning is an old field.
> But to be fair, human hearing and vision are based on analogue inputs, so it's not that hard to get it right.
XD What are you trying to say here? Because fourier is involved, STT somehow became easier??
> My point was that "predict" is a misleading technical term, because "evaluation" is more accurate.
"Predict" is the technically most accurate word. I don't know what else to say here except words have definitions; even the jargon definition of 'predict' is right here.
I did not "move goalposts", are you stupid? I further elaborated what I said in hopes you understand the point.
> Models being able to 'see' is exactly how agents know a sites color scheme is off
No. For that it is sufficient for it to look into color codes and compare them with exiting schemes. Let's say #EAC400, #EA0000, #0000A0 is a color scheme, and it's used in a popular program, therefore it's a valid color scheme. This is a line of logic easily replicable via straight algorithms or even people. We do not need to go nuts over something like that.
> or buttons are rendered behind other elements
That's a matter of code. If you compare every UI object's position and size with every other UI object, you can do that without even seeing the rendered result. In fact if you have rectangles and take the 4 sides' positions, you can create a function or method that tells you if these two objects overlap.
LLMs still might analyze a rendered image of the UI and base it on that, basically checking every pixel for overlaps, or doing it as above, so it's essentially about running algorithms, which sure is convenient. But it cannot determine if customers will like it.
None of that is actually "seeing" the UI in its appearance, nor assess "taste" as I mentioned. The layout may still be bad, the composition can be ugly, the colors and images not fitting. It can be literal LLM slop many people hate.
> It is coming.
So [soon™]. Heard it for years, and it's ruining the software industry.
> Try not using it and see how quickly you fall behind.
I will. It will take some years until people realize what a huge mistake it is, and that they should have listened to actual experts who warned about long-term consequences.
You know it's basically impossible to get a job as a software engineer today, right? It doesn't matter what your opinions about LLMs are. The problem is that companies jumped on the hype bandwagon and fired a lot of software engineers, barely hire them, and ceased to hire juniors. Worse even - 40% of job applications are fake, just for the shareholders to signal growth. And it's LLMs that filter through applicants, basically looking for keywords, and people write their applications with LLMs too. And if you happen to succeed, you'll go through multiple rounds of interviews, I mean 5-7, and may get rejected at each of them.
To get a job as a software engineer is a humiliation ritual, because corporate culture is also insane. LLMs only make all of this worse.
By depriving people of opportunities to acquire skills by being paid, and LLMs still requiring skilled people, this is a recipe for disaster. Or do you think some low-skilled CEO can just run a bunch of LLMs and bring out some miracle software? Aka *every* unskilled CEO?
The next problem is that companies will realize that relying on LLMs and unskilled people is a major mistake, causing them to rehire software engineers. The hope was for years that LLMs will produce good results, but so far that was not the case. It was hyped for years btw, so this is a Q-plan-truster 2-weeks-to-flatten-the-curve type of thing.
As I said, LLMs are good as a tool if you have actual software engineering skills and know when its output is good.
> Honestly, unless you are top of your field, where you will ultimately end up a 'bot wrangler', why stick it out?
Excellent idea, lying down and die. You know, if we assume you are 100% right about LLMs and your predictions, humanity would be doomed, right? None of that is sustainable.
You can be cynical about it, but I do care. I for one work on my own project and I am soon finished, I am fine.
> Machine learning is an old field.
Yes, and widely used. The principles are used in LLMs as well. But hey, you LOVE LLMs, right? Let me give you its response:
> Calling machine learning (ML) "outdated" is a fundamental category error—it is like saying internal combustion is outdated because we now have sports cars.
> Large Language Models are not a separate technology that replaced machine learning; they are the direct, scaled-up descendants of deep learning, which is itself a core subfield of machine learning.
The rest further explains the details: "...those parameters are literally billions of floating-point numbers organized into massive weight matrices."
And this:
> Why Do People Call ML "Outdated"? - The misconception usually stems from marketing hype. Tech journalism loves to treat "AI" as a completely brand-new paradigm that magically appeared out of nowhere, separating it from "boring" traditional machine learning (like linear regression, support vector machines, or random forests used in credit scoring and spam filters).
What is it with LLM fanatics NEVER even talking to an LLM about itself? Insane.
> What are you trying to say here?
It's about the fundamental difference of analogue and digital signals. For example a radio can barely receive its signal and still provide comprehensible audio. In the digital space if an image has +1 byte at the beginning, it will consider the entire file corrupted. The margin of error in analogue is variable, thus LLM generated speech is comprehensible but still may sound off.
> "Predict" is the technically most accurate word.
Yes, which is why I said it is misleading for colloquial use. It just evaluates the next best word/token. "Predict" insinuates an uncertain outcome in the future. Maybe if you factor in RNG for variance you could call it that, but that's a little silly.
> I don't know what else to say here except words have definitions
Of course you don't know. When someone is above the level of a midwit praising the status quo or being swayed by hypes, one can see beyond the veil of what-is and see the world in hypotheticals, in what-if and what-should-be. A recent example of language manipulation is "linguism" - prior to ~1980, it referred to the craft of language, but then they just changed it to add +1 discrimination term (an -ism) to the heap. When I talked about it, it was exactly about how language is used to manipulate, in the sense that that is "linguism."
> even the jargon definition of 'predict' is right here.
Not even remotely. One does not predict the next word one is about to write, like it's some arbitrary, random gibberish. What is predicted are events and weather, which are uncertain to occur. When a person writes, he thinks of what to write, and he has a clear concept of what he intends to write. One may refine it. But one does not have to "predict" himself.
Have you EVER even thought about this to come to that conclusion? Hey, let the LLM give you its answer too:
> The person claiming "prediction" is fine—even colloquially—is relying on a textbook definition without examining the underlying mechanics. Evaluation describes what the algorithm actually executes; prediction is just a legacy label from how the weights were calibrated.