Professor Jiang: ChatGPT isn’t actually generating everything by itself, there’s a lot of human manipulation going on in the algorithm. “Do you really think when you put something into ChatGPT, it’s completely generated by the AI? I guarantee you it’s not true.” “There’s a lot of human intervention to... make sure that the answer conforms to certain expectations.” “Maybe it’s 90% AI, 10% human, but it’s still 10% human.” “In order for ChatGPT to generate essays, it needs samples. It needs a lot of samples.” “It has to constantly absorb more human expertise in order to refine the model.”show more

You can tell AI is not people in india typing because its good at writing code.

On numerous occasions, Prof. Jiang has shown an overwhelming ignorance regarding how large language models actually work. I would not seek to understand anything from him regarding AI.

HE IS NOT A PROFESSOR!!!

He has to do this though. All the bs artists like him will. AI obliterates them.

elite larper lol chatgpt gets roughly 2.5 billion messages a day 10% of that is 250 million messages a day which get human intervention?

This guy, a conspiracy theorist, should not be taken seriously.

Whats his point? of couse all models and weights are coming from humans originally, but its not some people sitting there and intervene all the time. even if this would be possible technically, this would be an extremely rare exception

This defines intelligence as creating knowledge from nothing, which would disqualify every human who ever lived. No mathematician reinvents arithmetic before doing mathematics. No carpenter invents the saw before building a house. Of course LLMs inherit human knowledge and are shaped by human feedback, but then so are we. The question isn't whether intelligence starts from nothing, we are all well beyond that. The question is whether intelligence is revealed by what a system can do with what it inherits.

Everything this guy says is like this. Game theory, AI, you name it. He’s just making it up as he goes. He seems to have no formal knowledge. He is smart and there are nuggets of wisdom all over the place when he talks but it’s spastic and uncoordinated.

Human feedback shapes AI, but that’s different from humans manually generating each response. Transparency about the process matters.

Lmao his geopolitics can put out useful content and he is fundamentally an ontologically human centric person Therefore His takes on emergent behaviours as well as on AIs are biased towards humans W.r.t AIs - they're just shit; it takes human information and so on but the models take on a direction of their own - it is much harder to get them to remain conforming to humans than not

Korean's alex jones 🤣 on steroids

Lmao he’s so wrong. My chatbot isn’t sourcing Indians for results lol

I thought he meant harnesses executing commands since LLMs technically produce text only, not actual "intelligence" of LLMs coming from training data

Steelmanning his case he might’ve watched a Yann LeCun lecture drunk or something and he’s talking about RLHF and sample efficiency

So much wrong shit here lol

This guy has made an entire career larping

"BTC IS A CIA CONSPIRACY"

Can't find the person that interfered with my local AI response

Nope. The models of ChatGPT used by the general public, the free version and the lowest tiers are just searching the web and using data from training. I do not believe a human is participating in every GPT session. And I do NOT trust Jiang ... he came out of nowhere.

This guys an absolute moron hahahah

never seen such an obvious CCP plant

.. another "expert" who is completely and obviously wrong trying to safe his face while making it so much worse in the process :D

That makes sense.

Need to stop paying attention to this highschool teacher with English degree

Long story short doing deep dive on Bad Times and amazing backstory on human-ai collaboration. Just got through wi getting some help from GPT on physics involved. We live in the Days of Miracles. Enjoy it.

it's a simplification and vulgarization of the training feedback having a big influence on the final expression of the model temporary artifact imo, we're brute forcing out of that local maximum by all co-opting into training feedback loops that are more statistical

Es que esto es elemental en el principio de información. Lo que queda por ver es la escalada dada la base. Pero al momento no he visto nada que no sea la potencia de nuestras propias reglas y comprensión.

Oh, so he’s clueless

Regarding the machine learning technique known as reinforcement learning from human feedback (RLHF), indeed, humans, I'd argue, more concerningly, manipulate the linguistic substratum in which large language models operate. 😔

This guy is full of shit and dullsiuonal

Anti AI is AI while AI is SI

AI Has its Own Priorities... Chatgpt trying to sound liberal 😂 Grok trying to sound Conservative 😂