"To save money."
"Hundreds of millions in savings."
"So I won’t beat around the bush today."
"As for Northbound, you can come over anytime."
"At the AI Research Institute, you can pick your own position, your research direction, and your team. You name your own salary."
"Forget the market rate. Whatever you ask for, I’ll approve it."
Gao Wen, who was beside them, held his teacup as if he hadn’t heard a thing.
Li Dong started complaining silently in his head again.
’What is it with you presidents?’
’Always testing your people with money.’
’Aren’t you afraid I’ll ask for ten billion?’
But complaints were just complaints. The expression on Li Dong’s face remained composed.
"President Li, you truly think highly of me."
"But to be honest, I still want to prioritize my studies right now."
"I’m more interested in purer academics. As for commercializing research, developing products, creating applications... I know my limits. I’m not cut out for that."
The smile on President Li’s face didn’t change.
But a flash of disappointment still flickered in his eyes.
He had been navigating the business world for over twenty years; of course he understood the meaning behind Li Dong’s words.
’This kid really doesn’t care about money.’
Huaxuan: ???
Or perhaps...
’What this kid holds in his heart is far greater than money.’
President Li was silent for about two seconds.
Then he waved his hand and said with a smile.
"Alright, I won’t force you."
"I just hate to see talent go to waste."
"If you change your mind in the future, call me anytime."
Only then did Gao Wen slowly interject.
"Alright, President Li, stop eyeing our prized seedling."
"Academician Tian Gang is watching. If you keep trying to poach him, Old Tian will be calling me tomorrow."
President Li immediately cried out, "Whoa, whoa!" and made an exaggerated gesture of surrender.
"I wouldn’t dare, I wouldn’t dare. He’s Academician Tian’s man. How could I possibly touch him?"
The three of them looked at each other and smiled.
The atmosphere at the table instantly relaxed.
At half-past six, the banquet officially began.
Waiters entered through a side door, pushing serving carts.
There were only two tables in the banquet hall. Li Dong, Gao Wen, and President Li sat at the main table.
Several other important figures were seated nearby.
This included the vice president from Li Yun and a man who looked like the CEO of a chip startup, both of whom were also at the main table.
After the waiters left, the conversation naturally drifted from the recent weather and new policies to "that thing."
"That thing."
This was the tacit, unspoken name everyone present used for "GPT-3.5."
The first to speak was the vice president of Li Yun AI. Her surname was Qiu, and her name was Qiu Mingli.
She held her wine glass and spoke slowly.
"Last week, our group ran a special internal test."
"Chinese general knowledge questions, English reading comprehension, code generation, mathematical reasoning, and long-text summarization. We created twenty questions for each of the five categories."
She sighed.
"It won in all five categories."
The table fell silent.
President Li put down his chopsticks, his brow furrowed.
"Did it win by a lot?"
Qiu Mingli let out a self-deprecating laugh.
"I won’t give you the specific scores, but the gap... was not small."
"Especially on problems requiring multi-step reasoning. Our model basically starts spouting nonsense by the third step."
"But it can just keep reasoning steadily."
"You can throw a final, super-difficult question from the national college entrance exam at it, and even if it gets the answer wrong, its approach is correct."
"It’s earnestly making a mistake, not the kind of error where you can tell at a glance it’s just making things up."
"This is what our current models lack the most."
The man with glasses put down his knife and fork and began to speak slowly.
"President Qiu, I’ve actually had a question I’ve been wanting to ask."
"What was the crucial leap between GPT-3 and GPT-3.5?"
"Everyone outside is talking about RLHF, as if that whole human feedback system is the answer."
"But I’ve always felt it’s not that simple."
Qiu Mingli smiled.
"You’ve hit on the key point with that question."
"The term ’RLHF’ has been overhyped recently."
She gestured on the tablecloth with the tip of her chopsticks.
"The real leap from GPT-3 to GPT-3.5 wasn’t RLHF."
"It was code."
Someone at the table was taken aback.
"Code?"
Qiu Mingli nodded.
"OpenAI quietly did something in between."
"On top of the GPT-3 base model, they fed it another round of massive-scale code pre-training. Internally, it’s called code-davinci-002."
"They took all the high-quality code they could scrape from GitHub, added some data that mixed natural language and code, and threw it all into the pot."
"After that round of training, the model’s reasoning ability inexplicably shot up."
"It wasn’t just that it could write code. You could ask it to do math problems, multi-step logical reasoning, or analyze a long argument, and suddenly, it could ’think’."
"The industry is now calling this the ’chain of thought’ brought about by code."
"Code is extremely strict. Miss a single semicolon, and it won’t run."
"After the model was repeatedly hammered by massive amounts of code, it developed a ’feel’ for thinking things through step-by-step that no other corpus could provide."
Qiu Mingli paused and took a sip of wine.
"On top of that base, they then applied instruction fine-tuning, and then on top of that, they applied RLHF. Only then did you get ChatGPT."
"But if you’re asking where the leap was, this is just my personal speculation, but I think the most critical leap was at the code pre-training step."
"The problem RLHF actually solves is how to make this smart but wild model listen to human commands."
"It solves alignment, not intelligence."
The man with glasses nodded slowly, lost in thought.
Qiu Mingli added another sentence.
"Of course, that whole RLHF process isn’t simple either."
"For us to catch up on that front, we’d have to build a training system for our annotators from scratch. It would take at least half a year to a year just to get up to speed."
The table was quiet again.
President Li picked up his glass and swirled it gently.
"Let me add something."
"Our Wen Xin, you know, we didn’t just start working on it this year."
"ERNIE 3.0 Titan was released two years ago, with 260 billion parameters. On the classic Chinese leaderboards, we held our own against GPT-3."
"If you’re talking purely about Chinese language understanding, our foundation isn’t bad."
Li Dong hadn’t spoken the whole time, but he was listening very intently.
President Li continued.
"But when 3.5 came out, my heart sank."
"It wasn’t because its Chinese was better than ours—in fact, its Chinese still isn’t as good as ours."
"It was mainly because... it’s not even doing the same thing we are."
"With our previous generation of large models, the competition was about how accurately you could understand a sentence, whether you could classify things correctly, how well you could fill in the blanks."
"It’s competing on whether you can chat naturally with a person, and in the middle of that chat, solve a problem or write a piece of code on the side."
"That latter ability is on a whole other dimension."
"It’s not something our original development path can naturally extend to."
"Internally, we’ve been rushing to get the next version out these past two months."
"You’ll probably hear the public codename in a little while. This time, we’re aiming for 3.5."
"But I’ll be frank with you all."
"The first version we release will definitely still have a gap compared to 3.5."
"And it’s not the kind of gap that can be closed in a month or two."
The atmosphere grew heavy for a moment, then Qiu Mingli suddenly looked at Li Dong and spoke with a smile.
"This young man hasn’t said a word."
"And I don’t believe we’ve met."

Comments

Loading...