翁家翌：OpenAI，GPT，强化学习，Infra，后训练，天授，tuixue，开源，CMU，清华｜WhynotTV Podcast #4

Open AI h as a lot of mode l re leas es t hat have my name on them . Th

. Th is i s be caus e I' ve b uilt the ent ire post -tra inin g RL -Inf er i

n Op enAI . So

. So the ent ire post -tra inin g RL -Inf er i s yo ur c ore cont ribu tor?

Yes . Be

. Be caus e I thin k it sho uld brea k th e in form atio n gap.

信息差是 一个就是 如果你在 清华生存 的话是一 个很有用 的东西 但是我觉 得每个人 都应该平 等的拥有 这个信息 哪怕当 时有了现 代的认知 那么还

是做不出 来 没有 我不想发 pape r 我觉 得发pa per完 全没有意 义 方方 当时说要 搞一个A I la b 就是 后面的d eep seek

对我当 时是拿o ffer 如果你 想进工业 界 那么 读PhD 就是浪费 生命 教一个 Rese arch er, how to d o en gine

erin g? I

g? I t's hard er t han teac hing an engi neer how to do r esea rch.

Eve ry i nfra has dif fere nt l evel s of bug s. T

s. T he m ore bugs you fix th e be tter you r mo del will be.

Hell o ev eryo ne, this is Weng Jia yi.

He j oine d Op enAI in 2022 and is the core con trib utor beh ind Open

AI's ser ies of m odel s. F

s. F rom CHAT -GBT GP T-4O to GPT -5, you can see his voic e be hind

eve ry m odel . Hi

. Hi s co re c ontr ibut ion can be s umma rize d as thr ee k ey w ords En

hanc ed L earn ing, Pos t-Tr aini ng, Infra.

But for me, Weng Jia yi i s no t ju st a per son who make s mo dels str

onge r. B

r. B efor e jo inin g Op enAI he has alr eady use d op en s ourc e an d co

de t o in flue nce coun tles s pe ople . Fr

. Fr om t ryin g to bre ak t he i nfor mati on g ap i n Qi nghu a op en s ourc

e ho mewo rk a nd d ata, to open sou rce to s tren gthe n th e le arni ng f

rame work to cre ate a fr ee v isa sear ch s yste m h e ha s us ed c ode

tool s as a k ind of c hari ty.

In t his epis ode, we will sta rt f rom Weng Jia yi's chi ldho od, to t

he p ath of s tudy ing in Q ingh ua a nd C MU, and then to his own expe

rien ce a fter joi ning Ope nAI in 2 022.

Wha t di d he see sta ndin g in the cen ter of t he g loba l AI sto rm?

This is YNOW TV Podc ast.

Now pl ease joi n me in the worl d of Wen g Ji ayi.

Hel lo J iayi we lcom e to YNO W TV Pod cast . He

. He llo.

I t hink wha t's amaz ing abou t th is e piso de i s th at w hen I wa s pr

epar ing the outl ine for this epi sode I used GPT -5's Dee p Re sear

ch.

This is a pr oduc t an d mo del of y ou a s th e co re d evel oper.

They are doi ng d eep rese arch and thi nkin g ab out thes e is sues . Th

. Th ey a re i nter view ing you, the dev elop er b ehin d th em.

I th ink this is a ra ther ama zing way to avoi d th e si tuat ion.

Tod ay, we w ill talk abo ut m any aspe cts, fro m yo ur c hild hood gro

wth, you r ex peri ence in Tsin ghua to you r st udy at C MU, and then

to all the core dev elop ment exp erie nces sin ce j oini ng O penA I in

202 2. I

2. I wan t to sta rt w ith your ear lies t ch ildh ood.

Can you tel l me wha t ki nd o f ch ild you were whe n yo u we re y oung?

我小时候 特别喜欢 学奥数。

奥数?

对,我 从一年级 开始就开 始学奥数 。 就是

。 就是 因为开始 的时候, 我爸妈让 我去听了 这个 经 修校的一 个奥数课 跟语文课 ,然后我 对语文完 全不感兴 趣 但 是我对奥 数就非常 感兴趣,

然后我就 进去了。

然后从 一年级开 始就一直 上到六年 级,然后 初中也是 。 然后

。 然后 我发现就 是我做数 学题做得 比谁都快。

When I w as i n el emen tary sch ool.

Yes wh en I was in elem enta ry s choo l. F

l. F or e xamp le, when som eone 's m ath prob lem is l ike leve l 2, the

mat h pr oble m ma y no t be fin ishe d ye t. I

t. I t ma y be hal f do ne, and then I'm done.

I do n't thin k I need to thin k ab out it.

I ca n ju st s ay t hat I'm in t he m iddl e of a n ew s easo n. I

n. I can jus t go thr ough the sta ge a nd l ook at t he r esul ts.

This is a ve ry s ucce ssfu l wa y fo r me to do m ath.

I t hink I'm a s low lear ner comp ared to othe rs.

学一个新 的东西, 我经常要 花别人两 到三百的 时间,但 现在其实 还是这样 ,就是比 如说我读 一段代码 ,我就是 要比别人 花很多的 时间去理 解整个内

容,整个 cont ext,然后 但是我一 旦理解之 后,我用 的就非常 快。 有

快。 有 一个不太 恰当的例 子,就是 比如说我 小时候经 常不是要 背书吗, 背一个课 文。 然

文。 然 后我就是 在睡觉之 前,我想 尽所有方 法,然后 能够磕磕 巴巴的完 整的背出来。

Even if ther e ar e a lot of s tops yo u ju st h ave to g et o ff a nd s

leep . Th

. Th e ne xt d ay I wok e up and fou nd t hat I wa s ba ck t o no rmal . I

. I thou ght that I n eede d to spe nd m ore time so I n eede d to lea

rn t hing s ea rlie r th an o ther s. S

s. S o I star ted high sch ool math whe n I was in j unio r hi gh s chool.

I as ked the math tea cher in juni or h igh abou t ma th i n hi gh s choo

l. H

l. H e wa s ve ry e nthu sias tic and I wa s ve ry g rate ful.

I f inis hed high sch ool in t he s econ d ye ar a nd s tart ed t o le arn

alge bra in t he t hird yea r. Y

r. Y ou w ere born to be f ast at m ath.

You wer e fa ster tha n ot hers . Bu

. Bu t yo u al so f elt that you wer e sl ow a t le arni ng.

I fe lt t hat I ne eded to buil d my kno wled ge t ree for a lo nger time.

Norm ally th e nu mber of poin ts i s a root th en i t ex pand s t hen

expa nds, the n ex pand s t hen expa nds, the n se vera l la yers . Th

. Th en e very tim e yo u ap ply the top- leve l co nclu sion to this que

stio n y ou h ave to g o th roug h it fro m be ginn ing to e nd, and then

you hav e to thi nk i t ou t sl owly . Bu

. Bu t I migh t ju st b uild a s hort cut, and the n go str aigh t up . Th

. Th en I don 't h ave to r eact . So

. So you rea lize d th at y ou s eeme d sm arte r th an o ther s si nce you

were a k id?

No, not at a ll.

So y ou j ust said tha t yo u ar e in tere sted in this . Wh

. Wh ere did your int eres t in mat h co me f rom?

I w ant to i nves t in the fut ure.

I w ant to i nves t in my own futu re.

So w hen you were in the seco nd y ear of j unio r hi gh, you real ized

tha t yo u we re i nves ting in the futu re b y st udyi ng m ath in h igh

scho ol.

Yes, it is h elpf ul f or m y fu ture lea rning.

Inst ead of w asti ng t ime on t his topi c I 'd r athe r le arn some thin

g us eful for the fut ure, and the n th e be nefi ts w ill be m ore.

Did you r pa rent s te ll y ou t his?

No, my pare nts didn 't c are.

I'm not conc erne d ab out how I le arn.

But I'm sti ll c urio us w hy y ou h ad t his idea of inve stin g in the

fut ure when you wer e in jun ior high sch ool.

Bec ause I b elie ve y ou h ad t he p oten tial to lear n r ight ? Ye

? Ye s. W

s. W hy d idn' t yo u go pla y? W

y? W hy d idn' t yo u pl ay b aske tbal l or pla y ga mes?

Why did you wan t to lea rn h igh scho ol m ath?

Fir st o f al l I was int eres ted in i t. I

t. I tho ught tha t el emen tary mat h wa s to o bo ring . An

. An d th en i t wa s a litt le...

le...

a li ttle mor e in tere stin g math Who made you int eres ted in m ath?

Was it natu ral?

It' s ok ay.

Or y ou w ere inte rest ed i n ma th i n el emen tary sch ool.

You can thi nk o f it tha t wa y. Y

y. Y ou h ave a se nse of a ccom plis hmen t an d a rewa rd.

Then you wil l sl owly sta rt t o sc ore your mat h sk ills . Be

. Be caus e yo u wi ll g et a pos itiv e re ward eve ry t ime you score.

And then you 'll have to go t hrou gh t he s ame proc ess agai n an d ag

ain.

But I'v e tr ied play ing bask etba ll b efor e b ut I did n't get any

posi tive fee dbac k. I

k. I was act uall y hi t by my room mate . I'

. I' ve p ract iced tae kwon do b efor e a nd I thi nk i t's pret ty g ood.

But whe n I hit the ball I got hit by m y op pone nt.

So i t's a ki nd o f po siti ve f eedb ack.

You 're good at it, and you don' t ru le o ut t his.

It slow ly b ecom es a n in ner inte rest . Yes.

. Yes.

So i t's not real ly a fam ily envi ronm ent, but it' s yo ur o wn e nvir

onme nt.

Yes.

So this is math . Yo

. Yo u st arte d st udyi ng m ath in h igh scho ol.

Why did you get inte rest ed i n pr ogra mmin g in hig h sc hool ? I

? I star ted prog ramm ing in t he f irst yea r of jun ior high sch ool.

Yes I star ted prog ramm ing in t he f irst yea r of jun ior high sch

ool.

It was beca use I we nt t o a real -lif e mi ddle sch ool when I w as a

kid . It

. It was a p riva te s choo l. B

l. B ut t here was a p rogr ammi ng c lass in the priv ate scho ol.

And I ju st p laye d ar ound.

我能理解 高中有兴 趣的精神 ,但其实 初中搞编 程应该对 升学没有 太多用, 对不对?

是的, 没有太多 用。 那

用。 那 这个事儿 也是你自 己找的兴 趣班,还 是你父母 给你找的 ? 我父

? 我父 母觉得这 个可能对 我挺有意 思的。

Then I t ried it and I th ink it's oka y. O

y. O kay, let 's t alk abou t hi gh s choo l. I

l. I thi nk w hat I'm most int eres ted in i s yo ur e xper ienc e in OI,

in the sele ctio n pr oces s. C

s. C an y ou t ell us a bout it?

Why did I e nter the com peti tion ? Be

? Be caus e th ere was stil l bl ood pres sure.

If y ou d on't pla y in the com peti tion it 's v ery diff icul t fo r no

n-Be ijin g pl ayer s to get to the top of t he Q ingh ua N orth Pol e. S

e. S o yo u go t in to O I be caus e of the pre ssur e of the pro moti on?

Yes, mos t of it was beca use of t he p ress ure of t he p romo tion . I

. I was alre ady play ing in t he m ath comp etit ion with OI befo re.

I ac tual ly s pent som e ti me d oing mat h co mpet itio ns w hen I wa s in

the fir st y ear of h igh scho ol.

I go t so me p rett y go od g rade s. B

s. B ut i f I went onl ine, I w ould hav e to cho ose the best mat h te

ache r. O

r. O ur s choo l do esn' t ha ve t hat kind of foun dati on.

It's als o ve ry d iffi cult to lear n by you rsel f. S

f. S o I gave up.

And I d on't thi nk I can com pete in the onli ne m ath comp etit ion.

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

- - - - - - - - - - - - - - - - - - - - - - - - - - - -

- Afte r th e co

mpet itio n e ach prov ince has its own sel ecti on c rite ria.

The pro vinc e wi ll h ave to c hoos e a prov ince tea m to par tici pate

in the nati onal com peti tion . In

. In Fuj ian, the re a re t wo r ound s of the pro vinc e se lect ion.

One rou nd i s in the win ter and the othe r ro und is i n Ap ril.

Aft er t he s elec tion th ere shou ld b e a Qing hua or B eida i su mmer

cam p in May and June.

I co uld get some con diti ons for the tour name nt.

It w as l ike this at leas t. A

t. A nd t hen I we nt t o th e na tion al t ourn amen t fo r th e NI P in Jul

y an d Au gust . Te

. Te ll u s ab out your fir st, seco nd a nd t hird rou nds in t his proc

ess.

The NIP was oka y. T

y. T he N IP w as a lrea dy i n th e mi ddle of the seas on.

So i t wa s ok ay.

And then the qua lifi ers.

I d idn' t ev en p ass the qual ifie rs i n th e fi rst roun d. I

d. I did n't know how to do it.

I al most nev er d o it . Wh

. Wh en I was in the seco nd y ear, I w ould do a li ttle bit of it.

I wo uld do s ome heur isti c me thod s. F

s. F or e xamp le, ther e wa s a ques tion cal led "sma ll d oubl e-si ded

cove rage ". I

". I got 70 poin ts.

I th ink I wa s th e hi ghes t in the who le c ompe titi on.

Beca use I di d th is q uest ion, I g ot i nto the prov inci al t eam.

But aft er I got int o th e pr ovin cial team, How do y ou u nder stan d th

is?

If y ou d on't hav e a bonu s a nd i f th is l ine pass es t he r oute lin

e of Tsi nghu a th at y ear, Tsi nghu a wi ll a sk f or y ou.

And then the re's the fir st c ondi tion wh ich is i f yo u pa ss t he f

irst 150 peo ple in t he N MI c ompe titi on, you' ll g et t he f irst pla

ce.

If y ou p ass the firs t pl ace, you 'll be a dmit ted.

But if y ou a re a gol d me dali st, you don' t ev en n eed this . Yo

. Yo u ca n ju st k eep it.

Did you get the OMI cond itio nall y? U

y? U nfor tuna tely no . Be

. Be caus e th is i s an othe r bl ack hist ory.

I w as t he n umbe r on e in Fuj ian Prov inci al T eam.

Fuj ian Prov inci al T eam only had one bro nze meda list wh ich was

me.

The rest of the stud ents wer e to p st uden ts.

Did this los s af fect you r ac adem ic p erfo rman ce?

It d id.

I di dn't tak e th e cu ltur al c ours es i n th e se cond yea r of hig h sc

hool . I

. I wasn 't s ure abou t my fut ure.

I d idn' t kn ow h ow m any grad es I wou ld g et i n th e co lleg e en tran

ce e xam.

Som e ex ampl es a re t hat some sen iors use d up all the ir 6 0 po ints.

然后就真 的是一点 都不剩, 然后我觉 得万一我 也这样怎 么办,就 很危险。

然后我 对于这种 不确定性 非常害怕 ,但是想 了想,我 觉得我应 该也没有 那么差吧。

I st ill chos e to go with the 60.

So your oth er c hoic e wa s a more saf e op tion ? Ye

? Ye s a mor e sa fe o ptio n. I

n. I cou ld s ign a co ntra ct w ith Shan ghai Jia o Da at the ONI site

. So

. So it was a ve ry d iffi cult dec isio n at the tim e. I

e. I t wa s ok ay.

But the ment alit y wa s ve ry...

ry...

I wa s ve ry s ad a nd u ncer tain abo ut t he f utur e. I

e. I was sca red.

But my fami ly e ncou rage d me to choo se Q C60.

Did you not tou ch t hese thi ngs afte r th e OI and the fin als?

No, I d idn' t to uch them . I

. I migh t ha ve t ouch ed t hem secr etly.

I th ink I ha ve a spe cial pur suit for the opt imiz atio n of cod e. I

e. I eve n pr acti ced it o n th e iP ad.

Beca use Safa ri d oesn 't h ave a co de e dito r I jus t us e th e iP ad k

eybo ard to t ype in t he c ode and it c alls me.

The re's no edit ing.

Tha t ex peri ence has rea lly trai ned my t hink ing skil ls.

It r equi res you to h ave a co mple te u nder stan ding of the whol e to

pic or l ogic . An

. An d yo u ne ed t o qu ickl y id enti fy w here you 're wrong.

It w ill trai n yo ur r espo nse abil ity.

I t hink thi s is a g ood trai ning for me.

But you did n't need thi s ki nd o f tr aini ng a t th at t ime.

Yes I did.

I w as j ust very int eres ted.

For exa mple th ere was a lo ng-t erm opti miza tion of the UI.

For exam ple, the re w as a alg orit hm.

Alth ough the com plex ity of t he t ime may be n squ are, the fro nt o

f 11 has a l ong term.

And then the UI will hav e th e re sult s. F

s. F or e xamp le, it w ill rank acc ordi ng t o th e ti me o f th e te st p

oint . An

. An d th en t he f aste st o ne w ill rank fir st.

And if i t ru ns a s fa st a s th e ot her one, it will be base d on the

len gth of t he c ode.

So I wi ll o ptim ize thes e tw o in dica tors at the same tim e. A

e. A nd t hen I fe el v ery inte rest ing when I b rush thi s th ing.

Alt houg h it 's u sele ss, it's ver y in tere stin g. Y

g. Y ou a re r eall y...

y...

You real ly l ike this thi ng, righ t? Y

t? Y ou c an g et h appi ness fro m it . Le

. Le t's talk abo ut T sing hua.

If I to ld y ou t hat you star ted read ing in 2 016, do you have any

mem orab le m emor ies?

Wha t is the fir st t hing tha t co mes to y our mind ? If

? If the re i s i t is tha t I have com plet ed a ll m y ho mewo rk.

Diff eren t pe ople hav e di ffer ent view s on thi s ma tter.

Some of the seni ors and seni ors are oppo sed to t his.

But I t hink I s houl d do thi s. I

s. I did a..

a..

. I

coll ecte d al l th e ol d as sign ment s an d ma teri als and then I o pene

d th em a ll u p on Git Hub.

Exc ept for thos e wi th c opyr ight iss ues, I d idn' t op en t hem up.

The rest of the ones wit hout cop yrig ht i ssue s op ened the m up . Wh

. Wh y? B

y? B ecau se I thi nk w e sh ould bre ak t he i nfor mati on g ap.

信息差是 一个就是 如果你在 清华生存 的话是一 个很有用 的东西 但是我觉 得每个人 都应该平 等的拥有 这个信息 你现在 回去你现 在回去随 便抓个计

算机的学 弟问 问 这个你认 不认识这 个捐这个 新戏楼的 人 哪怕 他把他的 名字放在 戏楼上面 你认不认 识他 不 认识 你 认不认识 温家义应 该认识

因为毕竟 这个大家 都看我的 作业活的 This is more use ful than the don atio n. "

n. " This is wher e th e ch ats shou ld h ave been del eted . +1

. +1 " If I h ad s een your hom ewor k I sho uld have sen t a chat at

this tim e. B

e. B ut w hy d id y ou o pen your hom ewor k? W

k? W hy i s th e na ture of the chat rel ated to the info rmat ion diff

eren ce?

我想打破 这个信息 差,就是 我不想让 后人重逃 覆辙,就 是经常疲 于奔命, 就是很多 人其实他 不是很擅 长搜集东 西的。

But he i s ac tual ly v ery capa ble.

But if I ca n gi ve t hese peo ple a ch ance to have a r ight to info

rmat ion, the n he may liv e be tter in Xinh ua.

So y our righ t to inf orma tion is abou t wh ethe r yo u kn ow t he h omew

ork of t he p ast year an d th e in form atio n th at t he s enio r an d se

nior ask ed w hat the focu s of the exa m wa s be fore . An

. An d ho mewo rk, and home work tha t th e te ache r di d no t sp ecif y th

at c ould not be put out.

Othe rwis e y ou'd spe nd 1 0 to 20 hour s a nd y ou w ould n't dare ask

the ass ista nt.

You woul dn't kno w ho w to do it.

You' d sp end a lo t of tim e a nd y ou'd kee p on try ing to f igur e ou

t th e ke y. I

y. I t wo uld be v ery usel ess for the whol e cl ass.

The inc ome woul d be ver y lo w. B

w. B ut i f I had open ed u p my own wor k I mig ht h ave help ed s ome

people and help oth er s tude nts to d o wh at t hey want to do i nste ad o

f wa stin g th eir live s. S

s. S o yo u've don e it . Yo

. Yo ur w ork in X inhu a ha s be come a w ell- know n Gi tHub rep o. Y

o. Y es.

When did you sta rt s tudy ing scie nce?

I t hink it was in m y se cond yea r of col lege . I

. I aske d my sen ior if t here was a r ecom mend ed l abor ator y. H

y. H e to ld m e th ree name s. Z

s. Z hu J un, Tang Jia o a nd C ui P eng.

I ad mitt ed t o th e or der of t he p laye rs, and that eve ryon e wa s go

od a t it . Bu

. Bu t it mig ht n ot b e li ke t hat.

I j ust foll owed the ord er.

You thou ght it w as a fir st-t o-la st o rder . Ye

. Ye s I tho ught it was a fi rst- to-l ast orde r. I

r. I rep orte d a new plan for the com pute r. I

r. I thi nk I got it all.

I c hose Mr. Zhu Jun . Wh

. Wh en I was a k id, I me t hi m fi rst.

I a sked him wha t di rect ion he h ad.

He s aid ther e we re t hree dir ecti ons.

Baye s G AN, and a st rong -wie ldin g sy stem . Th

. Th ese are the thre e di rect ions . I

. I want ed t o us e GA N. I

N. I tho ught the ima ge w as v ery inte rest ing.

But I d idn' t kn ow w hich one was GAN.

- 觉得挺有

意思的, 然后就一 直搞下去 了。 但

了。 但 是实际上 ,我最开 始想法是 我其实有 三个大的 方向,我 其实都挺 喜欢的。

一个是 人工智能 ,就是A I,然后 另外一个 是图心学 ,还有一 个是网络 安全。

Uh, For exam ple, I f ixed the

fee for the fre e re sult lis t. T

t. T he r esul t li st w as 1 0 yu an n ext time . I

. I foun d a bug with ano ther sen ior.

I c ould pay for it or n ot.

I fi xed the bug afte r do wnlo adin g it a f ew t imes.

Yes, I d irec tly repo rted it to t he s choo l's teac hing sta ff.

So y ou r eall y en joy this kin d of hac king pro cess . Ye

. Ye s. S

s. S o th e In tern et i s a kind of hack ing.

Wha t ab out the rest ? I

? I feel tha t th e gr aphi cs a re a lso a ki nd o f ha ckin g a hac king

aga inst the rea l wo rld.

You can use you r ow n pe rspe ctiv e to con stru ct t he s cene you ima

gine in your hea d. A

d. A nd t hen it w ill make you fee l li ke y ou a re G od, you are God.

I w as a ctua lly most int eres ted in g raph ics at t he t ime.

I wa tche d a movi e ca lled Tro n wh en I was in midd le s choo l. I

l. I t's a sc i-fi sci -fi movi e. T

e. T he m ovie was ver y ef fect ive.

It real ly s hock ed m e. I

e. I tho ught if I co uld make a m ovie lik e th at o ne d ay, or c reat

e my own vir tual wor ld l ike Tron I woul d be sat isfi ed.

I th ink that 's t he w ay t o go wit h th e fl ow o f ti me.

So I was ver y in tere sted in the flow of time . I

. I star ted stud ying at the age of t wo.

I re ally lik ed i t s o I spen t a lot of t ime I to ok t he A -pri ce o

f th e se cond -mos t cl ass in t he c lass and inv ente d a new algo rith

m to red uce the numb er o f fa ces that wer e fo lded and I u sed a bu

nch of c alcu lati ons to g ener ate a hu ge h igh- pitc hed pict ure, whi

ch s houl d be 16K.

Befo re I sta rted no one sho wed me a 16K pic ture . An

. An d th e 16 K pi ctur e wa s no t fa ke.

So y ou w ere so i nter este d in dra wing it, did you con tinu e? N

e? N o b ecau se I thi nk i f yo u wa nt t o do res earch, I wa s mo re f ocus

ed.

I co uldn 't j ust teac h tw o bo ats a le sson . Be

. Be caus e yo u ha d al read y go ne t o Mr . Zh

. Zh u's plac e. Y

e. Y es, I ha d al read y go ne t o Mr . Zh

. Zh u's plac e. T

e. T here was a p roje ct t o de velo p Wi zdom . DO

. DO OM i s a game fro m th e 90 s. T

s. T here was a c ompe titi on a t th e ti me.

You use a ne ural net work to pass thr ough a f ixed map and the n yo

u ha ve a sta rtin g po int and a en d po int.

The n yo u ha ve t o ki ll m onst ers on t he m ap, then pic k up the blo

od b ag, then clo se t he b arri er, and then you rea ch t he e nd p oint

and the n pa ss t hrou gh.

So t his is a ctua lly the firs t pr ojec t yo u wa nted to lear n ab out.

Yes.

拿了冠军 ? 是的

? 是的 。 所以

。 所以 当时你是 很享受做 强烈学习 科研的?

没有。

为什么 ? 很不

? 很不 享受,因 为首先这 个tas k,这个 环境太过 于单一了 ,然后你 要疯狂地 去ove rfit ,然后就 是你要用 各种技巧 防止他训 练的时候

崩了。

以及你技 术训练没 崩,然后 你也不知 道你怎么 调参才能 调队,这 个是真的 练单,这 个比CV 来说,它 的调参难 度可能要 拿上10 倍100 倍,然后

都是选一 选,你必 须要用一 些非常h euri stic 的方法去 避免一些 conn er c ase。

So y ou r eali zed that the RL rese arch at that tim e wa s al l ba sed

on h euri stic s? Y

s? Y es, the algo rith m wa s no t as fun dame ntal . Be

. Be caus e my tas k at tha t ti me w as t oo m onot onous.

Most of the time it 's n ot w orki ng.

The firs t ta sk w as V isdo om.

What 's n ext?

Ata ri, D-Mi ne, Cont rol Suit e? I

e? I thi nk M ojoc o an d At ari are easi er t han Visd oom.

Vis doom is diff icul t. B

t. B ut i n th is c ase, for hum ans, it' s a simp le t ask.

But for RL, it' s a very dif ficu lt t ask.

Beca use it r equi res a lo t of kno wled ge.

For exam ple, wha t is an obst acle ? Th

? Th is i s no t ne cess ary for huma ns t o re act to, but for RL o r AI

it 's c ompl etel y di ffer ent.

Yes so you wer e ve ry c lear abo ut t he e nvir onme nt a t th at t ime.

The alg orit hm i s no t fl at.

You will fin d th is i n th e mi ddle of the work.

太多挑战 害太多所 以说我不 是很享受 这个科研 的过程然 后我会就 是我有意 识的把我 的重心放 到了如何 帮助这一 类的科研 更顺利进 展的过程上 So w

hen I wa s in 4th gra de, I wo uld thin k ab out buil ding an IO I nfra

. Or

. Or a s mall IO libr ary, so that it woul d be eas ier for thos e wh

o wa nt t o ro ll i n th is f ield to roll.

I th ink I'm very goo d at thi s. I

s. I 'm v ery good at soft ware eng inee ring . I

. I can make the cod e lo ok b ette r an d th e us er e xper ienc e be tter

. Bu

. Bu t as for how to adju st t his, I h ave a ps ycho logi cal reac tion.

You are not not good at it, you just don 't l ike it, you don' t wa

nt t o do it.

Yes ye s. T

s. T his is y our mast er's deg ree in a stro nomy . We

. We wil l ta lk a bout ast rono my l ater . So

. So you sta rted doi ng s cien ce a nd w isdo m in the sec ond year . Th

. Th en y ou w ent to t he t hird yea r d id y ou f ace a pr omot ion?

Yes . Wh

. Wh at k ind of r hyth m an d st ate was it a t th at t ime?

Pro moti on f or t he m aste r's appl icat ion.

I se nt a bun ch o f me ssag es b ut t here was no resu lt.

My m ento r co ntac ted Yoas t Ba njo and I we nt t here . My

. My sum mer camp was n't that suc cess ful.

It was in M ila, rig ht?

Yes, it was in M ila.

It was the summ er o f 20 19.

Duri ng t he s umme r va cati on i n 20 19, Yosh a wo n th e Tu ring Awa rd i

n Ap ril or M ay.

I wa s th e on e wh o wa s in cha rge of t he a ward . So

. So it was fine . Yo

. Yo sha gave me and my p ostd oc t he t ask of d oing som ethi ng s imil

ar t o MOE.

- MO E. -

E. - Rig ht.

- Th ere was no M OE b ack then . -

. - NLP MOE.

Thi s ha s no thin g to do with R.

It's jus t a pure lan guag e mo del, a t rans form er.

Ther e wa s no MOE bac k th en.

We d idn' t kn ow w hat MOE was and how it w orke d. B

d. B ut i t ha d an ide a fo r MO E. T

E. T here was a r oute r th at a llow ed m e to cho ose diff eren t pa ths

and impl emen t th em.

So h e hi red some one who does RL to d o NL P? Y

P? Y es, I th ink it's wei rd.

But I ne ed t o sp end.

.. A

gain I' m no t go od a t le arni ng t hing s. I

s. I spe nt a lon g ti me l earn ing to e nter the tra nsfe r NL P. I

P. I was wat chin g hi m do som ethi ng, but it d idn' t wo rk o ut w ell.

Now if you loo k ba ck, you need to work on this . Fi

. Fi rst, you nee d to hav e th e co mput ing powe r. Y

r. Y ou n eed to h ave a st rong eng inee ring abi lity.

and you have to scal e up . At

. At tha t ti me, a pe rson cou ld n ot d o it by hims elf.

So you can' t do it.

Even if the dire ctio n is rig ht, you can' t fi gure it out.

But do you thin k th at t he s ecre t is pus hing you to the open AI

now?

You fir st g ot R L t hen you got NLP.

The se t wo p iece s y ou d id e very thin g be fore ope n AI . Yo

. Yo u ca n sa y th at i f yo u re ally wan t to.

But I kn ow y ou c ould n't get the futu re a t th at t ime.

Wha t wa s yo ur o pini on o f NL P wh en y ou e xper ienc ed t he n umbe r? I

r? I thi nk N LP t asks are too sca tter ed.

At t hat time ma ny p eopl e we re t hink ing abou t ho w to app ly R L in

lan guag e mo del trai ning . Be

. Be caus e th e pr oble m wa s no t so lved at that time.

Beca use if T rans form ers runs dir ectl y on the wal l i t wi ll c olla

pse.

It' s ve ry e asy to c olla pse.

and no o ne t houg ht a bout how to prev ent it f rom fall ing.

The n it pro ved that you r en viro nmen t sh ould be pure . An

. An d yo ur r ewar d ne eds to b e a very str ong rewa rd.

What is a pu re e nvir onme nt?

For exam ple, tax -onl y. P

y. P ure text . Wh

. Wh at a bout the non -pur e en viro nmen t th at w e di d be fore ? Fo

? Fo r ex ampl e y ou u sed Tran sfor mer to d o so me A tari dec isio ns.

For exam ple, you can tur n im ages int o em bedd ings . Or

. Or you can con stru ct A tari 's i nter nal game inf orma tion and the

n pu t it in Tran sfor mer and run.

And Tra nsfo rmer is a ra ndom dig ital Tra nsfo rmer . At

. At the tim e I did n't have eno ugh cont ext.

I d idn' t ha ve e noug h kn owle dge of t his fiel d. S

d. S o I stru ggle d a lot to m ake it w ork.

I th ink if I wen t ba ck t o Ch ina and help ed t hem, I w ould sti ll b

e li ke t his.

Do you thin k th at a t th at t ime, 就处在一 个怎么都 不可能把 东西搞w ork的 一个人。

是的, 因为算力 不够,然 后对于当 时的一些 认知,M -O-E 的认知也 不够,然 后你哪怕 ,就这么 说吧,哪 怕当时有 了现代的 认知,那 还是做不

出来。

因为en gine erin g不够, 然后co mput er也不 够。 对

够。 对 ,受限于 当时的资 源,你就 是做不出 来。 那

来。 那 当时可能… 鼠炎结束了。

and then you wen t ba ck t o Ts ingh ua t o ap ply for the job.

Yes . Wh

. Wh at w as y our situ atio n li ke?

Were you und er a lot of pres sure ? Ac

? Ac tual ly, I wa s no t in a g ood stat e. I

e. I did n't get my d egre e. M

e. M y cl assm ates at the CMEU in Stan ford fo r ex ampl e g ot a lot

of degr ees.

The y se nt m e a PICC ARD.

Usu ally yo u ne ed t o se nd a PIC CARD or you wil l be swi ped by s

omeo ne.

So I fin ishe d my res earc h an d on ly g ot a rec omme ndat ion lett er.

But I di dn't kno w wh at w as w ritt en i n th e le tter . It

. It was qui te u nsat isfa ctor y. W

y. W hat was your res ult?

I a ppli ed f or P hD, but I on ly g ot a mas ter' s de gree.

I wa s in the sec ond year of coll ege.

I w as i n th e fi rst or s econ d ye ar o f co lleg e. E

e. E very yea r t here was a q uest ion like "W hat' s th e re sult of

your app lica tion in 2019 ?" I

?" I saw you r an swer . I

. I was impr esse d. I

d. I saw you r re sume . Yo

. Yo u sa id y ou w on t he f irst pri ze a nd g ot t he s econ d pr ize.

But I s aw y our appl icat ion resu lt w as n ot g ood.

You did n't get a Ph D. I

D. I tho ught the com peti tion was int ense.

Is i t a kind of fail ure for you as a cli ent?

May be a lit tle bit, but it' s ok ay t o lo ok a t it lat er.

Firs t of all in the env iron ment at that tim e I was gre atly aff

ecte d by the sma ll e nvir onme nt a t th at t ime.

Ph.

D. i

s be tter tha n ma ster.

This sho uld be t he a tmos pher e of Qin gHua . Ye

. Ye s i f yo u ar e bo rn i n it it is diff icul t to get rid of this

inf luen ce.

Even if it i s a fore ign PhD or m aste r i t is the sam e as a d omes

tic PhD or m aste r. P

r. P hD i s be tter tha n ma ster . Be

. Be caus e ev eryo ne h as a kin d of old -fas hion ed a ware ness tha t th

e hi gher the aca demi c pe rfor manc e t he b ette r th e fu ture dev elop

ment . Bu

. Bu t th is i s of ten not true.

It r eall y de pend s on wha t yo u do . I

. I was a li ttle dis appo inte d at tha t ti me.

I ac tual ly s pent som e ti me a djus ting mys elf.

I t hink I'v e al ways wan ted to d o so meth ing that mak es m e di ffer

ent.

So you were try ing to b reak fre e fr om t he e valu atio n sy stem of

your fri ends . Ye

. Ye s y es, yes.

I've alw ays thou ght that GPA is not the only eva luat ion syst em.

You need to find the rig ht e valu atio n sy stem for you rsel f in col

lege and fig ht f or t his eval uati on s ystem.

to m ake your self hap py.

Even if you get the firs t pl ace in G PA, you can comp are it w ith

the past wi th o ther col lege s w ith othe r sc hool s. T

s. T here are peo ple who can' t sa y ho w go od y ou a re.

I th ink it s houl d be lik e th is.

From a d eman d pe rspe ctiv e f or e xamp le, if y our ulti mate goa l is

to find a j ob, So y

ou're Is t hat what you lea

rned in your maj or?

That 's w hat I le arne d la ter on.

That 's w hat I le arne d wh en I was a m aste r. B

r. B ut y our majo r is sti ll u nder the eva luat ion syst em.

I tr ied to b reak fre e of the eva luat ion syst em, but I di dn't bre

ak f ree of i t co mple tely . Wh

. Wh at w as t hat?

I r eali zed that I s houl d cr eate my own eval uati on s yste m n ot t

he e valu atio n sy stem pro vide d by oth ers.

I us ed a n un bias ed e valu atio n sy stem wh ich was reco mmen ded by m

y me ntor . I

. I alwa ys r emem ber this say ing.

He thou ght the comp uter eva luat ion syst em h ad t hree ind icat ors.

One is the pape r t he o ther is the comp etit ion, and the oth er i

s th e Gi tHub star.

I th ink this is real ly m eani ngfu l. B

l. B ecau se i t gi ves me a com plet ely diff eren t pe rspe ctiv e a dif

fere nt f ield . I

. I can actu ally do some thin g in the ope n so urce com muni ty a nd l

et m e ha ve s ome diff eren t pl aces wit h ot hers . Th

. Th is i s be tter for my long -ter m de velo pmen t. I

t. I t's not like you spe nd a lot of time I ha ve t he s ame feel ing.

Whe n I was in t he u nder grad uate pro gram I real ized tha t th e GP

A e spec iall y th e un derg radu ate prog ram, is some thin g th at y ou d

on't nee d to wri te o n th e re sume aft er t hree or four yea rs.

But the unde rgra duat e st uden ts i n th ose thre e or fou r ye ars had

to p ut 8 0% o r 90 % of the ir e nerg y in to i t. B

t. B ecau se w heth er y ou a re l ooki ng f or a job ap plyi ng t o st udy

abro ad, or a sec urit y gu ard, It's all abo ut t he G PU.

How do y ou d eal with thi s co ntra dict ion?

Som ethi ng t hat is e xtre mely imp orta nt i n th ree year s an d so meth

ing that is comp lete ly u nimp orta nt a fter thr ee y ears . I

. I will inv est at t he l owes t li mit.

I w ill spen d th e lo west lim it o f ti me t o ma ke i t me et t he r equi

reme nts I wa nt.

I do n't want to inve st a ny h ighe r th an t hat.

You don' t wa nt t o wa ste time eve n if you get one mor e po int.

Yes yo u do n't want to wast e ti me e ven if y ou g et o ne m ore poin

t. I

t. I t's enou gh.

It's act uall y qu ite simp le.

You just cal cula te h ow m any poin ts y ou h ave now befo re t he f inal

exa m. T

m. T hen you can arra nge whet her you want to revi ew t he e xam or n

ot.

Some peo ple say that if you thin k th at t he G PA i s en ough now yo

u wo uld rath er t ake the 60-p oint exa m. Y

m. Y ou d on't car e. Y

e. Y es, I ha ve a lway s be en l ike this.

Righ t a nd m aybe 60% is not enou gh.

For exam ple, 87% is a mu st-h ave, and I w ill be v ery sati sfie d wi

th t hat.

Do y ou t hink you 've over come thi s? B

s? B ecau se, for exam ple, the cho ice to g o ab road is brou ght to y

ou b y th e ev alua tion sys tem, the atm osph ere brou ght to y ou.

No, the atmo sphe re a t th at t ime was actu ally bet ter than in Chin

a. 1

a. 1 8 1 9 ye ars.

19 year s ha ve p asse d. A

d. A lmos t on ly 2 0% o f Qi nghu a wi ll h ave it.

But we h ad l ess of i t be caus e of Cov id.

Then onl y 5% wen t ab road.

你应该是 19年1 2月申请 期,对不 对? 对

对? 对 。 所以

。 所以 你应该是 还在等o ffer 的时候就 covi d了吧?

对。

然后你应 该比如说 拿到of fer之 后美国领 事馆就关 了? 对

了? 对 。 你还

。 你还 是坚持要 出国?

是,这个 很难选择 。 因为

。 因为 当时应该 是有co vid, 有疫情, 然后国际 政治局势 又很动荡 ,那你个 人的未来 还要和这 两个七夕 相关。

对。 那

种不确定 性下你当 时是什么 样的状态?

I th ink I wa s mo re f ocus ed o n th e th ings at hand . Fo

. Fo r ex ampl e w riti ng a pos tcar d at hom e w riti ng " reti re o nlin

e" a nd " visa log in s yste m" a t ho me.

I ju st w ante d to tra nsfe r my att enti on a nd n ot f ocus on some big

int erna tion al n ews ever y da y. I

y. I was mor e fo cuse d on the thi ngs at h and, so that I c ould be

calm er.

Okay le t's talk abo ut t he t wo p roje cts duri ng t hat time . Th

. Th e fi rst one is T ien Shou an d th e se cond one is Tui Xue Onli

ne.

Let' s ta lk a bout Tie n Sh ou f irst . Ti

. Ti en S hou, you jus t me ntio ned it.

You said tha t Ti en S hou' s mo tiva tion was bec ause you tho ught the

env iron ment was too ove r-fi t. A

t. A nd R IO, in f act, the inn ovat ion of a lgor ithm is not that imp

orta nt.

Inst ead, it' s th e in fra, or the good imp leme ntat ion.

Tel l us why you cho se T ien Shou . Fi

. Fi rst of a ll, in D ecem ber 2019 I had a I wa s ve ry i mpul sive . I

. I had writ ten a lo t of RL code for exp erim ents . Wh

. Wh y do n't I in tegr ate them and mak e my exp erim ents bet ter?

Thi s is the sta rtin g po int.

I t ook a br eak in J anua ry a nd l ooke d at RLl ib c ode in F ebru ary.

It' s th e RL lib unde r RA ID.

The firs t st ep w as t o us e RL ib t o se e if it coul d be cha nged and

the n su ppor t my own exp erim ents . I

. I watc hed it f or a mon th.

It w as t oo c ompl icat ed.

Ther e we re t oo m any abst ract s. R

s. R Lib has almo st h undr eds of t hous ands of code . It

. It was com plet ely unac cept able . I

. I didn 't k now how to i mple ment the cha nges I w ante d to mak e. T

e. T hen...

hen...

I de cide d to qui t an d ju st s tart ove r. I

r. I wan t to kno w d id y ou h ave the abil ity to t hink abo ut i t? B

t? B ecau se a lot of peop le w ante d to do a po licy opt imiz atio n pa

per.

No, I d on't wan t to do a pa per.

I t hink it' s me anin gles s to do a pa per.

Tel l us why . Wh

. Wh y di d yo u br eak away fro m th is e valu atio n sy stem ? Fi

? Fi rst of a ll, I al read y ha ve a pap er.

I do n't thin k it 's m eani ngfu l to hav e mo re o r le ss.

And I've alr eady had eno ugh voic e. I

e. I 've also had som e in the com peti tion . I'

. I' ve h ad s ome in t he G ithu b th ree- way coun t. F

t. F or e xamp le, I've don e a repo for tha t co urse bu t it was n't

a st raig ht-t o-ea rth one.

I s till wan t a stra ight -to- eart h th ree- way coun t. O

t. O kay, so that 's T ians heng . Af

. Af ter watc hing RL Live, It's not eas y to use . Yo

. Yo u we re n ot i nter este d in pap er a t th at t ime.

But you are int eres ted in o pen sour ce c ode.

You are ver y in tere sted in open sou rce code tha t is eas y to use

and can pus h fo rwar d in thi s fi eld.

How dif ficu lt w as i t to do this ? It

? It was ver y ea sy.

Two week s. J

s. J ust two week s. T

s. T he f irst edi tion . Wh

. Wh at k ind of a lgor ithm is the firs t ed itio n? T

n? T here are man y al gori thms . Yo

. Yo u ju st n eed to i mple ment pap er.

If y ou g et t he a bstr act righ t t he t ime will be less tha n 20 lin

es.

Why can you get a D vers ion with onl y tw o li nes?

How man y li nes can you writ e wi th R ULEAP?

I th ink it's bec ause the des ign was a bi t pr oble mati c at fir st.

And then bec ause the re w ere a lo t of peo ple who work ed t oget her,

eve ryon e co ntri bute d to the cod e. A

e. A nd t hen it b ecam e a proj ect.

The who le p roje ct g radu ally bec ame more and mor e co mpli cate d. S

d. S o yo u re cord ed i t al one?

Yes.

So y ou t hink it will be more pow erfu l if you do a go od t op-l evel

des ign with eff icie nt d ata tran sfer alo ne?

Yes.

I t hink cod e or a p roje ct h as m ore cons iste ncy.

If the proj ect is c onsi sten t fr om s tart to fini sh, it i s a good

pro ject.

For exam ple, man y of the pro ject s ar e co mpli cate d du e to inc onsi

sten cies . Fo

. Fo r ex ampl e t wo p eopl e or ten peo ple.

Ten peo ple writ e so me c ode, but the y do n't know wha t th e ot her

pers on w rote . Th

. Th ere are assu mpti ons that the y ca n't actu ally con vey.

Thi s le ads to a lot of places and some bad beh avio r. A

r. A nd t hen the code sta rted to expa nd a nd b ecom e mo re a nd m ore

comp lica ted.

You thi nk t hat Tian shou was pop ular bac k th en.

What did Tia nsho u do rig ht?

I th ink it c augh t th e us er's nee ds.

The rese arch er, or t he p erso n wh o wa s wo rkin g on the res earc h at

the time, Ther e's alwa ys a nee d fo r a fram ewor k th at's eas y to use

and eas y to cha nge.

Bec ause the cod e is sho rt, and the abst ract is good if you stu

dy i t y ou'l l fi nd t hat if y ou c hang e th is p art, it will wor k. A

k. A nd o nly this par t can If y ou w ant to s uppo rt y our feat ure, you

can onl y su ppor t it by chan ging thi s pa rt.

So y ou c an o nly chan ge t his part if you have a g ood desi gn.

Yes.

And you can do the appl icat ion, and you can do the calc ulat ion.

So y ou w rote the fir st v ersi on o f Py Torc h. D

h. D id i t be come a m ore open sou rce proj ect?

Yes . No

. No one in the team was in char ge o f th e op en s ourc e pa rt.

It w as o pen sour ce.

How did you avoi d ge ttin g...

g...

It's not lik e wh at y ou j ust said ma ny p eopl e ar e to geth er, it's

not con sist ent.

At the begi nnin g I had eno ugh time so I c ould pac k ev eryt hing

by myse lf.

This is defi nite ly c onsi sten t. A

t. A ltho ugh it d oesn 't l ook good now it doe sn't dev elop wel l fo

r th e lo ng t erm, but it' s st ill usef ul f or t he s ea t ree at t hat

time.

And then lat er o n b ecau se I was too bus y I wou ld t rans fer the

whol e ri ght to m aint enan ce t o th e co mmun ity.

If they thi nk i t's righ t t hey woul d do it.

If ther e wa s a pers on w ho w as a pat ente d pe rson, So t his is c onsi

sten t. I

t. I t's been fiv e ye ars sinc e Ti ansh ou b ecam e a comm unit y. D

y. D o yo u th ink it's cha nged now ? A

? A litt le b it.

Beca use my c onte xt i s a litt le d iffe rent fro m th e co ntex t of the

suc cess or.

So h e wi ll r ewri te m y co de.

The seco nd p roje ct.

I wa s al so a use r I qui t. W

t. W hy?

You were ...

...

I wa s a soph omor e b ut l ater ...

...

You were a s tude nt?

Yes, I w as a stu dent at the time . I

. I also nee ded to f ind my t ime.

Why did you do this ? I

? I have a n eed.

It' s si mila r to wha t I did yest erda y. I

y. I thi nk I sho uld...

uld...

I fo und a ci rcle I don' t kn ow w hat can sati sfy my n eeds so I j

ust reco rded one . Ot

. Ot herw ise, the re i s no way . Be

. Be caus e at tha t ti me, you shou ld b e in a t empo rary off ice, and

tha t te mpor ary offi ce.

Yes, yes ye s. T

s. T o kn ow w here the re i s a visa . An

. An d th en I thi nk I hav e th is n eed.

For exa mple at the beg inni ng, I ha ve t he n eed to c olle ct t he h

omew ork of t his seni or s enior.

I th ink many peo ple have thi s ne ed, so I sta rted a c ompa ny.

I ha ve a nee d fo r a visa . I

. I wrot e a visa app lica tion and sta rted a c ompa ny.

I ca n us e it for free.

This is very pop ular . I

. I saw the tota l nu mber of orde rs w as m ore than 1 m illi on l ast

nigh t. I

t. I t wa s mo re t han 1 mi llio n at tha t ti me.

Now it's mor e th an 1 0 mi llio n. B

n. B ut i t wa s cl osed lon g ag o. B

o. B ecau se o f th e pa ndem ic, ther e is no need for it anym ore.

And the US cons ulat e up grad ed t he w ebsi te.

The patc h wa s no t av aila ble at t hat time . I

. I didn 't h ave time to writ e it . So

. So it comp lete d it s mi ssion.

Actu ally th ese two proj ects are not ver y ef fici ent.

The fir st o ne i s no t fo r ap plyi ng.

It h appe ned afte r yo u ap plie d. Y

d. Y ou c an d o ot her rela xing thi ngs duri ng t he p erio d of COV ID-1

9. A

9. A nd t he g radu atio n is the sam e. Y

e. Y ou c an f inis h it you rsel f. S

f. S o yo u st ill have a s tron g im puls e in side to crea te s omet hing

you thi nk i s us eful and wha t yo u ne ed.

And then sha re t his thin g wi th e very one.

I wa nt t o do som ethi ng t hat can be i nflu enti al, even if it's a l

oss.

For exa mple do ing char ity work . I

. I thin k do ing Tian shou and doi ng T uesh onl ine are all doin g ch

arit y wo rk.

It's com plet ely non- prof it.

And doin g th is k ind of c hari ty p roje ct m akes me feel ver y sa tisf

ied.

所以相比 钱,其实 impa ct会让 你更满足 。 是的

。 是的 。 你觉

。 你觉 得你对i mpac t这种, How long did you pur sue this ? In

? In hig h sc hool . Wh

. Wh y di d yo u ha ve a con nect ion with Imp act in h igh scho ol?

It's a b it s tran ge.

When I w as i n hi gh s choo l I had an idea whe n I was in t he t hird

gra de.

It j ust came out of my h ead.

It migh t be tha t at som e po int in t he f utur e I 'll send som e me

ssag es t o th e pa st m e. T

e. T hen I'll rea lize som e th ings . Fo

. Fo r ex ampl e i f li fe i s a game th en t he s core of your gam e is

The numb er o f pe ople who rem embe r yo ur n ame is g one.

Wha t do es t hat mean ? It

? It mea ns y ou n eed to b e re cogn ized by more peo ple.

So you sudd enly rea lize d th at t his is v ery impo rtan t. I

t. I wan t to rem embe r mo re a nd m ore peop le b y th e mo ment I d ie.

Yes.

Unt il t oday yo u st ill foll ow t his stan dard ? Ye

? Ye s I thi nk s o. T

o. T hen I wa nt t o as k yo u ag ain, why is this imp orta nt?

I di dn't thi nk i t wa s im port ant at t he t ime, but I t ried it a fe

w ti mes and foun d th at i t wa s wh at I wan ted.

If w hat I do can bri ng b enef its to o ther s t hen I wi ll u se w hat

I do . In

. In fac t o ther s wi ll a lso reco gniz e wh at I do.

The n I will get som e po siti ve f eedb ack from oth ers.

Can thi s be dir ectl y li nked to the fame of the peop le?

Not real ly.

Beca use some peo ple' s fa me m ay n ot b e go od.

嗯。

You can' t be nic e to eve ryon e. I

e. I t's hard to do.

But I ca n tr y to be nice to ever yone aro und me.

Do s omet hing mea ning ful to e very one.

So the stan dard is to d o so meth ing mean ingf ul t o ev eryo ne, so t

hat more peo ple reme mber you.

If I wer e to cha llen ge t his, I w ould say wh y do n't you just liv

e ha ppil y by you rsel f? W

f? W hy d o yo u ne ed o ther s to rem embe r yo u? Y

u? Y ou w ant othe rs t o re memb er y ou, you want man y pe ople to use

you, to crea te s omet hing use ful.

Wha t is the ori gin of t his impu lse?

Are you afr aid of b eing for gott en?

No.

I th ink life can als o be con side red an e xper ienc e. S

e. S ince you hav e co me t o th is w orld wh y do n't you...

you...

Don' t wa ste this jou rney . Ye

. Ye s b ut w hy d on't you was te t his jour ney?

You hav e to let oth ers reme mber you . I

. I foun d th at t here is a ve ry i nter esti ng t ensi on h ere.

Bec ause you jus t sa id t hat you want to brea k th roug h so me e xter

nal eval uati on s tand ards . Fo

. Fo r ex ampl e G PA, coll ege entr ance exa mina tion or som ethi ng l

ike that . Ot

. Ot hers thi nk t hat PhD is b ette r th an M aste r. B

r. B ut o n th e ot her hand yo u gi ve y our own inte rnal int rins ic s

tand ard.

又是外界 的对你的 认同。

但这个认 同并不是 官方的认 同 就 是不是既 有评价体 系的认同 是共 识。 O

识。 O K 所 以你想要 摆脱的是 现有体 系和成规 的那种评 价体系。

对。

但你更想 要的是 每个人发 自内心的 给你点赞 。 是的

。 是的 。 所以

。 所以 这个点赞 可以变成你的 GitH ub's sta r ca n be come the num ber of c lick s on the

dro p-do wn, can beco me, for exam ple, Goo gle Scho lar' s ci tati on,

can also bec ome how many tim es y ou d o an ope n-ed it m odel eve ry d

ay.

Do y ou w orry tha t yo ur s tand ard will bec ome a ne w ru le?

Mayb e th e pr evio us g ener atio n us ed G PA t o ev alua te y oung peo ple.

Whe n yo u be come the pre viou s ge nera tion yo u us e im pact to eval

uate the nex t ge nera tion of youn g pe ople.

I ju st a sk f or i t. I

t. I don 't a sk f or i t fr om o ther s. S

s. S o yo u do n't thin k yo ur s tand ard is s omet hing tha t ca n be set

up in a way tha t is rig ht.

Yes.

So you don' t li ke t he k ind of t hing s th at a re p ushe d by ext erna

l st anda rds.

Yes . Wi

. Wi ll y our own stan dard s ap pear tha t pu sh y ou a way?

Not yet . If

. If you hav e th is s itua tion yo u ca n ch ange it.

You can cha nge your sta ndar d. S

d. S o yo u us e yo ur s tand ards bu t yo u're not an expe rt o n th is s

tand ard.

You use you r im pact sta ndar ds.

Yes.

- Yo u wo n't be t rapp ed b y it . -

. - Yes.

For exa mple fo r a long tim e I hav en't sta rted a p roje ct.

But I do n't thin k it 's a pro blem . I

. I don' t wo rry.

I t hink an open -end ed m odel is the best . Ye

. Ye s y ou c an t hink tha t wa y. B

y. B ecau se t he e valu atio n sy stem is a si mple and dir ect oper atio

n th at c an q uick ly s elec t pe ople.

So t he w ay t he s ocie ty d evel ops will be boun d to thi s st rate gy.

So i t ca n't be c hang ed i n a shor t ti me.

Even if you are this pla tfor m m ost peop le t hink you are the win

ner and the bene fici ary.

You don 't l ike this pla tfor m. Y

m. Y es, it s houl d be mor e in divi dual ized . Ye

. Ye s a nd d urin g th e pa ndem ic, You appl ied to C MU, it w as s uppo

sed to b e in Sep temb er 2 020.

Did the pan demi c go awa y? D

y? D id y ou t ake onli ne c lass es?

I to ok o nlin e cl asse s at hom e fo r a year . Bu

. Bu t di d yo u th ink abou t fi ndin g a job afte r en teri ng t he u nive

rsit y? Y

y? Y es.

What was tha t pr oces s li ke?

I st arte d ou t wi th a few pro blem s. I

s. I was abl e to get 18 shar es.

I on ly g ot G oogl e an d Au toML wh ich is C hen Linq i's comp any.

And I th ough t if I o nly had thes e tw o of fers I woul d de fini tely

go to t he O X b ecau se I don 't w ant to g o to Goo gle.

Bec ause goi ng t o Go ogle is mean ingl ess.

How do you judg e th is m eani ng?

You are a bi g pl ayer in the big mark et, and you do s omet hing you

don 't l ike, for exa mple th e fr ont and back . Th

. Th en y ou t hink a l ittl e bi t mo re a bout it, and the n yo u go to

the inte rvie w. T

w. T hen you take the oth er s ide of t he c oin.

The othe r pa rty said the y wa nted to buil d an AI lab, whi ch i s th

e De epSe q be hind it.

I w as t he o ne w ho g ot t he o ffer bu t I didn 't g o. Y

o. Y ou d idn' t go an d yo u we nt t o Op enAI . Ye

. Ye s. S

s. S o no w it loo ks l ike if y ou o pen Tian yan, you r ch oice at that

tim e wa s De epSe q ve rsus Ope nAI.

Yes . So

. So eve n if it was the othe r pa rty, it wasn 't a bout qua ntum bu

t ab out AI.

The othe r pa rty was abou t RO Info.

Yes, if I di dn't hav e an y ot her offe rs, I wo uld choo se t he r ight

sid e. O

e. O ther off ers othe r th an O cto?

No, for exa mple Op enAI and I s houl d ha ve h ad I nvid a. I

a. I nvid a? I

a? I nvid a is als o a high -end IOC sys tem.

And the n FA IR i s ac tual ly a lso, but FAI R ev entu ally du e to som

e pr oces s re ason s I was rem oved . Ok

. Ok ay, so y our choi ce a t th e ti me w as t o sw itch to Open AI.

Invi tr, TikT ok, and FAIR . An

. An d yo u ch ose Open AI a fter all . Ho

. Ho w di d yo u ch oose thi s? I

s? I thi nk..

nk..

. Be

caus e I want to emph asiz e th at t his was befo re c hatG PT.

So y ou d idn' t kn ow c hatG PT w as c omin g. S

g. S o ho w di d yo u ch oose thi s? F

s? F irst of all, I t hink ...

...

Open AI h as b een doin g en hanc ed l earn ing for a lo ng t ime.

Ope nAI and Deep Mind are the bes t tw o re sear ch l abs in t he I EO.

I di dn't thi nk a bout it at t he t ime.

I d idn' t th ink abou t it . I

. I thou ght it w as g reat tha t I coul d ge t in to it.

And I th ink that if I ca n ha ve a cha nce to e xper ienc e th e mo st a

dvan ced rese arch in the worl d h ow i t is don e i nste ad o f li ke...

ke...

It's lik e a smal l wo rksh op, like in scho ol, wher e yo u ha ve a few

PhD s an d yo u do a f ew t hing s by han d. A

d. A nd t hen ther e ar e no met hodo logi es.

And then you go and do a pro ject . I

. I feel ver y aw kwar d. A

d. A nd t hen I wa nt t o le arn how he h as a met hodo logy to do s ome

rese arch in the indu stri al w orld . So

. So whe n yo u we nt t o Op enAI yo u sh ould hav e just It w as a tea m of

Joh n Sc huma nn.

Yes, Joh n Sc huma nn h ired me.

Oh, he inte rvie wed you.

Yes I was very gra tefu l to him . I

. I was very gra tefu l th at h e ga ve m e th is o ppor tuni ty.

Even if he l eft, I w as s ad f or a n af tern oon and turn ed o ff t he c

ompu ter.

I d idn' t do any thin g. S

g. S o he act uall y ad mire d yo u a lot.

Yes . Wh

. Wh y? D

y? D id y ou a sk h im?

Yes, bec ause he thou ght my g uita r wa s ve ry b eaut iful.

哦。

So, do y ou h ave any inte rest ing stor ies abou t yo ur i nter view ? Ho

? Ho w di d yo u im pres s hi m? W

m? W ell, may be t he l ast inte rvie w h e ca me u p wi th a que stio n a

nd i t wa s a very end -to- end ques tion an d it was ver y op en.

Then he gave me thre e ho urs, and I s pent two hou rs t o fi nish it.

I st arte d wr itin g so meth ing and fini shed it.

I s howe d it to him and said tha t al thou gh t here was a b ug i n th

e pr oces s I wan ted to f ix i t fi rst and it w as f ine.

He thou ght that I h ad s ome engi neer ing skil ls a nd I was sti ll g

ood, so he h ired me.

然后他的 这个面试 题只给两 个人测过 ,一个是 我,还有 一个是, 应该还有 一个是A ndre w,就是 现在做C odex 的那个人 。 你们

。 你们 俩都通过 了。 对

了。 对 ,都通过 了。 所

了。 所 以通过率 是百分之 百。 是

百。 是 的。 好

的。 好 ,我们马 上聊OP I,但是…… I wa s cu riou s ab out this . Di

. Di d yo u co nsid er r eadi ng P hD w hen you were loo king for a j ob?

No.

Why?

Bec ause som e pe ople who hav e be en i n th e in dust rial wor ld w ill

find tha t if you wan t to ent er t he i ndus tria l wo rld, it is a was

te o f li fe t o st udy PhD.

You can com plet ely use Mast er a s a jump boa rd a nd t hen buy PhD

to e nter the ind ustr ial worl d. F

d. F or e xamp le, you can buy a ci tati on i n Ma ster or in t he u nder

grad uate pro gram.

and crea te p roje cts that are dif fere nt f rom the ones you hav e. A

e. A nd t hen you can comp ete with the PhD can dida tes of t he s ame

peri od.

And then see wha t yo u ca n do to make the oth er p erso n ch oose you

as the mast er, and not anot her PhD cand idat e. I

e. I thi nk i t's very imp orta nt t o cl arif y th e di ffer ence s. S

s. S o yo u've lon g wa nted to know tha t I will be in t he i ndus tria

l se ctor in the futu re.

Yes, bec ause I t hink it' s to o di ffic ult to b e a prof esso r. I

r. I t's not what I w ant.

I h ave to f und a pr ojec t. I

t. I fee l th ere are many lim itat ions . Fo

. Fo r ex ampl e w e ar e a comp any.

We have to recr uit peop le w ith the same mas ter and PhD.

Do you thin k th ese two things The skil ls t hey trai n ar e di ffer ent.

Bec ause PhD s ne ed t o tr ain your aca demi c sk ills ri ght?

How to writ e a pape r t ell a st ory, mak e it bea utif ul, and do t

he p rese ntat ion well . Do

. Do you thi nk t his is i mpor tant for a c ompa ny?

To s ome exte nt, it i s a trai ning . Bu

. Bu t if it is c ompa red to y our more adv ance d en gine erin g sk ills

do you thi nk t hese two ski lls are not enough How do y ou e valu ate

thes e tw o? W

o? W ho i s mo re v alua ble?

In toda y's era, of cour se, the more the bet ter the engi neer ing

abil ity is, the more val uabl e it is.

But at that tim e i t's real ly h ard to s ay.

So m y ap proa ch a t th at t ime was to t ry t o sa tisf y bo th o f th em.

You have sen t pa pers any way.

And the inf rast ruct ure work of Kaiy uan is a lso very pop ular . Th

. Th e en gine erin g ab ilit y is str ong enou gh.

Why do y ou t hink it' s ob viou s th at e ngin eeri ng i s th e fi rst?

Let me quot e a quot e fr om a col leag ue o f mi ne.

He w as a PhD stu dent at RIO and crea ted a ve ry f amou s RI O fr amew

ork.

He said "I t's hard er t o te ach a re sear cher how to do e ngin eeri

ng t han to t each an engi neer how to do r esea rch."

rch." It's bec ause of the curr ent rese arch lab s. T

s. T hey are all abou t th e ac cura cy o f in fra.

If your inf ra i s co rrec t i t de pend s on how man y ti mes you can

stac k in you r un it t ime.

Bec ause you can 't h ave a di scus sion wit h so meon e ab out your ide

a a nd t hen the idea com es o ut, and you can veri fy i t. A

t. A s lo ng a s yo u ca n ve rify it, it' s li ke y our rese arch work.

And you don' t ha ve t o us e so man y br ains . Pe

. Pe ople who use bra ins have bee n in thi s fi eld for a lo ng t ime.

For exa mple Al ec.

He s tart ed u sing it when GPT -1 w as j ust star ting out . He

. He has a s tron g se nse of r esea rch.

Usi ng b rain s is mor e us eful tha n us ing norm al P SC b rain s. Y

s. Y ou c an j ust talk to him.

Ide as a re v ery cheap.

What you nee d to do is t o ve rify how man y ef fect ive idea s yo u ca

n ge t in a s ingl e ti me.

And if i t's the righ t in fra, the rig ht r esul t a nd t he r ight del

iver y sp eed.

Yes . An

. An d no w th e Ph D do esn' t ha ve t his abil ity or..

or..

. I

don' t fo cus on t his beca use it's not imp orta nt t o th em.

Beca use I th ink the curr ent acad emic tra inin g sy stem is in a goo

d ac adem ic d irec tion.

But ther e ar e pe ople in the comp any who have thi s di rect ion.

Bec ause as long as you work in this fie ld f or a lon ger peri od o

f ti me, you will hav e so me r esea rch intu itio n. T

n. T hen you will rea lize wha t is goo d w hat is b ad, what sho uld

be d one, wha t sh ould not be done . So

. So ide a is che ap.

And ther e is a v ery stro ng e ngin eeri ng s kill to quic kly stac k th

e in fras truc ture . Af

. Af ter you have thi s y ou v erif y th e id ea.

Mayb e ev en A gent can do it.

Yes.

Bec ause in my u nder stan ding ev ery Infr a ha s di ffer ent leve ls o

f bu gs.

The more bug s so meon e fi xes, the bet ter the mode l is . So

. So Lam a ca n't pers uade GPT bec ause Lam a ha s to o ma ny b ugs?

May be.

I do n't know . Bu

. Bu t I migh t gu ess so.

So y ou c lear ly r eali zed a lo ng t ime ago The whol e pi peli ne n eeds

to work . Th

. Th e ke y is not the cre ativ ity of y our algo rith ms, but the righ

t su perf ood, the rig ht i nfra th e ri ght syst em t o ma ke y ou d eliv

er q uick ly.

Yes.

And I'm not ver y wi llin g to do rese arch on supe rfood.

Yes, I d on't hav e an y at trac tion to this . I

. I pref er t o se ll m y pr oduc ts.

You like to have a p layg roun d wh ere you can play wit h yo ur g uita

r an d le t ot hers pla y wi th i t. Y

t. Y ou d on't car e ab out othe rs.

Yes, and whe n ot hers are wri ting a p aper th ey c an b ring me alon

g. A

g. A nd t hen you will fin d th at m any of t he m odel rel ease s in Ope

nAI have my name on them . Th

. Th is i s be caus e I have bui lt t he e ntir e po st-t rain ing RL i nfra

stru ctur e in side Ope nAI.

So the enti re p ost- trai ning RL infr astr uctu re i s yo ur c ore cont

ribu tor.

Yes.

And beca use ever yone use d th e wh ole post -tra inin g RL -INF RA t o tr

ain the RLHF mod el, so e very tim e th ere' s a big rele ase, eve ry t

ime ther e's a bi g mo del, I h ave to p ut m y na me o n it . So

. So you can be cons ider ed a s ha ving you beh ind ever y op en-e nd m

odel be caus e yo u li ke t o se ll s hove ls.

Yes, and I'm the one who sel ls s hove ls t he m ost to t he c lien t b

ecau se R L is the top of the whol e in fra.

- It has a h igh ecol ogic al v alue . -

. - Yes.

So if i t's too low- leve l t he n ame migh t no t be ...

...

If y ou w rite "da ta l oade r" o r "s tora ge", it migh t no t be the sam

e. B

e. B ut i t's some thin g th at e very one want s. S

s. S o sm art.

I'v e th ough t ab out how my c aree r de velo pmen t sh ould be.

I s et a sta ndar d. I

d. I wan ted to m axim ize I ca n se e th e nu mber of name s on the Ope

nAI blog . Yo

. Yo u re ally kno w ho w to wri te r ewar ds f or y ours elf.

Yes an d wh at d o yo u ne ed t o do wit h th is r ewar d? Y

d? Y ou m ust firs t do inf ra, beca use if y ou d o th is s ingl e re sear

ch, you can' t sc ale up.

If y ou u se I nfra ev eryo ne w ill use you.

So you can scal e up . I'

. I' m al so g ood at R L In fra, so this is a pe rfec t op port unit y. A

y. A fter you ent er T sing hua, thi s is you r ma in l ine, RL Infr a. Y

a. Y ou w ill enco urag e...

e...

You 've alre ady pass ed t his choi ce, but many of our view ers Mayb

e th ey a re s till in the unde rgra duat e or the y ar e st ill hesi tati

ng b etwe en i ndus tria l an d ac adem ic.

How woul d yo u he lp t hem thin k ab out this pro cess es peci ally in

2025 un der this sit uati on?

In t he l ong run, I s till thi nk t hat the acad emic wor ld s houl d be

red isco vere d. B

d. B ut t he y oung peo ple who have a f utur e to tak e re veng e on th

ey w ant to m ake an i mpac t li ke y ou.

Is i t be tter to stud y a PhD or e nter the ind ustr ial worl d as soo

n as pos sibl e? I

e? I thi nk i t's bett er t o en ter the indu stri al w orld as soon as

poss ible . Be

. Be caus e if you stu dy a PhD yo u do n't know wha t wi ll h appe n af

ter you grad uate . Ma

. Ma ybe afte r yo u gr adua te, the time has com e an d yo u wi ll f ind

that wha t yo u do is usel ess.

If y our defa ult func tion is to e nter the AI lab, you nee d to fig

ure out what kin d of peo ple the AI l ab n eeds . If

. If the y ne ed m ore peop le i n in fra, the n yo u ca n do mor e in fra

work . Ev

. Ev en i f yo u do n't have a P h.D.

h.D.

deg ree, it' s ok ay.

Beca use it's mor e im port ant to s ee i f yo ur e xper ienc e ma tche s. S

s. S o wh at k ind of p eopl e do you thi nk t he A I la b ne eds?

I t hink it' s st ill the infr a. T

a. T he i nfra is a no n-re liab le t ool.

The rese arch -ori ente d pe ople are the one s wh o ar e ca pabl e of doi

ng r esea rch.

Bec ause aft er y ou c ame out of t he C HI G P y ou'v e wo rked in this

ind ustr y fo r mo re t han thre e ye ars.

You can 't d o an ythi ng a bout it.

The cur rent pro blem is stil l wi th i nfra . Ho

. Ho w ma ny t imes can you be in c harg e of the inf rast ruct ure?

and then it dire ctly det ermi nes your pro duct ivit y ef fici ency . It

. It sou nds like a s itua tion tha t is not par ticu larl y fr iend ly t

o Ph Ds.

Yes.

May be t his also inv olve s a gap.

I t hink we both hav e a deep und erst andi ng o f th is.

Beca use in t he a cade mic fiel d of RL rese arch, It's jus t th at t he t

asks for Ata ri a nd M ojoc o ar e ov erfi t. I

t. I t's just tha t th e sc ores are hig her than whe n yo u an d I were

at 100K . Bu

. Bu t th e in dust ry d oesn 't c are at a ll.

The indu stry doe s th ings wit h RL to solv e re al p robl ems. Yes an

d I real ized thi s. A

s. A nd w hen I re aliz ed t his in A ugus t 20 22, I ha d al read y gr adua

lly stop ped the deve lopm ent of T ien Shou . Be

. Be caus e Ti en S hou is s till aim ed a t th ese Toei ben chma rks.

I th ink I sh ould inv est more tim e in mor e me anin gful thi ngs, suc

h as cre atin g RL Infe r in Ukr aine . So

. So my main foc us i s on mai ntai ning and dev elop ing RLIn fer.

Act uall y y our know ledg e is ver y ad vanc ed i n 20 20.

It's def init ely not comm on s ense . Wh

. Wh y di dn't you bre ak t he n ew X -TRA and wri te a blo g to per suad

e me not to do it?

I'm afra id t hat if I say thi s O penA I wi ll s ay I 'm s ecre tive . Wh

. Wh at i f? E

f? E ven thou gh i t's not that sec reti ve, but.

.. R

ight ? Ok

? Ok ay, let' s ta lk a bout Ope nAI.

I'm ver y cu riou s. W

s. W e've alr eady tal ked abou t a lot of t hing s b ut i n Op enAI, You

are the worl d's very few peo ple, fro m Ch etGP T 3.

5 G PT-4 GP T-4V GP T-4O GP T-4.

5 to GPT -5.

Behi nd a ll t he c ore cont ribu tors th ere is y ou.

Mayb e so me p eopl e co ntri bute d th e fi rst half so me p eopl e co ntri

bute d th e se cond hal f b ut y ou h ave been the re f rom the begi nnin

g to the end . Yo

. Yo ur m ain cont ribu tion I woul d sa y i s th ree word s: s tron g le

arni ng, post -tra inin g a nd i nfra . We

. We wil l ta lk a bout the se t echn olog ies and the stor y be hind the

m la ter.

But I w ant to a sk t he f irst que stio n fi rst.

Post -tra ining.

What is a po st-t rain ing in a big lan guag e mo del?

At that tim e t here was no post -tra inin g. A

g. A t th at t ime, the tea m wa s ca lled RL.

At that tim e t here was no pre- trai ning or post -tra inin g. S

g. S o wh en y ou f irst ent ered Ope nAI, the re w as n o po st-t rain ing.

Yes . Bu

. Bu t wh en y ou f irst ent ered Ope nAI, was Cha tGPT the mai n li ne?

No, when I f irst sta rted Ope nAI, I o nly had the RL t eam unde r Jo

hn S herm an, and then I m ade a ve rsio n af ter WebG PT.

WebG PT i s a 3.5 mode l fo r br owsi ng.

But if y ou j ust use 3.5 for brow sing th e ef fect is very bad . Be

. Be caus e br owsi ng r equi res a to ol c all.

So w e we nt b ack to t he b egin ning and tri ed t o ma ke t he u ser inte

ract ion expe rien ce b ette r. T

r. T he o nly solu tion was cha t. A

t. A nd c hat can be s olve d by usi ng i nstr ucti on f ollo wing . Us

. Us ing RLHF . So

. So whe n yo u fi rst star ted, you alr eady had 3.5 mod el i nsid e? Y

e? Y es, we d id.

But the pipe line of PPO was very dif ficu lt t o use.

At t hat time we use d th e mo st c ommo n 3.

5 SF T. W

T. W e di d it sev eral tim es.

Ther e we re o nly four aft er t hat.

Gre g wr ote a sp ecia l in fra to s uppo rt h ow t o us e GP D4's inf ra t

o su ppor t tr aini ng i n re info rcem ent lear ning.

So y ou w ere in t he m ovie in July of 2002 ? Ye

? Ye s. A

s. A t th at t ime, 3.5 was alr eady the who le o peni ng a nd a ll-i n. B

n. B ecau se t here wer e st ill a fe w mo nths bef ore the rele ase, rig

ht?

你们内部 当时能想 象Cha d GP T有这么 大规模的 成功吗?

没有 你当时能 够当然你 可以内测 那个mo del对 吧? 对

吧? 对 你有意 识到这个 是要ga me c hang ing的 东西吗?

没有 你也没有 当时感觉 到? 因

到? 因 为我觉得 就是我呢 可以看得 出来很多 不足的东 西 就是 我第一次 用可能会 觉得说O K这个是 个会说话 的模型 那也就这样吧 然后用了 几次发现

这可能 能够帮我 解决一点 代码上的 问题 那 也就这样 嘛 但也 不能帮我 解决那么 多 然后 我用了几 次也没有 那么想用 因为它 能帮我解 决的问题

有限 因 为你已经 提前知道 了这个东 西 然后 就是你有 一个循序 接近的这 个过程 然后你也 不觉得它 那么突然 但是 呃 For exam ple,

whe n I show ed i t to the peo ple arou nd m e t hey were sur pris ed.

I di dn't exp ect that . Di

. Di d th e co mpan y us e al l-in -Cha tGPT in 2020 ? No

? No on ly o ur t eam did.

You can see the cha t GP blo g. I

g. I t's a bl og w ith a co ntri buto r. J

r. J oshu a sh ould be the firs t on e. T

e. T he o ther 12 peop le w ent to Z hao Shen gjia.

When did you rea lize tha t yo ur w ork at O penA I re ally ins pire d you?

I wa s ru nnin g Ni er O ps o n No vemb er 3 0 a nd t hen Dece mber . A

. A few days lat er, I fo und that peo ple arou nd m e we re d iscu ssin

g th e Ch aiGP iss ue.

They fou nd i t ve ry u sefu l an d ev en p acke d th e Op enAI ser ver seve

ral times.

It's lik e wh en I qui t sc hool . My

. My ser ver was also hac ked seve ral time s. I

s. I t ca used a s elf- repo rtin g ef fect . Ev

. Ev eryo ne i s se lf-r epor ting and wil l se lf-r epor t pr oduc ts f or y

ou.

This is defi nite ly a goo d in vest ment.

东西就投 入值得你 投入精力 的一个东 西。 其

西。 其 实听起来 这个从你 加入Op enAI 做这个事 包括 Open AI内部 能够推出 Chat GPT, 这种东 西好像都 不是谁计 划出来的

。 是的

。 是的 。 可能

。 可能 是一系列的, - We w

ere expe ctin g to hav e 10 ,000 -20, 000 at t he b egin ning th en i t fe

ll b ack and then it was gone . Wh

. Wh at a bout the cur ve o f th e ma chin e? I

e? I t's a in dex.

Thi s is the suc cess of the sudd en p ayme nt o f Ch et G PT.

I'm curi ous abou t yo ur e xper ienc e wh en y ou f irst ope ned.

Wha t wa s th e im pres sion of this com pany ? I

? I feel lik e it 's a big lab.

没有,其 实没有我 想象中那 么有方法 论,但是 很多就是 有很多很 强的re sear ch知觉 的人在里 面,然后 他可以指 明方向, 然后就做 就好了。

But afte r Ba rret t L uke and Liam joi ned the team an d th en w e ch

ange d ou r te am.

We s tart ed t o ad d Go ogle 's a dvan ced prod ucti vity and sta rt t o st

ack.

Goog le i s st ill very goo d. T

d. T here is a gr aph show ing the numb er o f ti mes you have to wait

for a s ingl e un it a nd t he s ucce ss r ate.

Thi s is a p osit ive rati o. S

o. S o th e mo re t imes you hav e to wai t fo r a sing le u nit, the mor

e yo ur s ucce ss r ate goes up.

Yes . Th

. Th is i s al so a 2-w ay c urve.

Beca use RL i s co nsta ntly try ing to d o th e sa me t hing ove r an d ov

er a gain . Yo

. Yo u ke ep t ryin g un til you reac h th e de sire d nu mber an d yo u ca

n ac hiev e yo ur g oal.

In fact ma ny o f th e ca ses in l ife are RL.

So t he a dvan ced prod ucti vity you int rodu ced was a ki nd o f ph ilos

ophi cal conc ept.

Yes . Th

. Th at i s w e do n't have to thin k ab out geni us i deas or geni us a

lgor ithm s. W

s. W e ca n us e In fra to d o it . In

. In fra allo ws u s to do it f rom 30 t imes a w eek to 3 00 t imes a w

eek.

Yes al most . I

. I saw an i nter view . So

. So meon e as ked Sam Altm an w hy O penA I su ccee ded.

He said in a The reas on w hy w e ca n ma ke a tec hnol ogy inno vati on w

ith Open AN i s be caus e in a t eam with a h igh dens ity of t alen ts,

any medi ocre per form ance can not be t oler ated . Do

. Do you agr ee w ith this ? I

? I agre e. B

e. B ecau se i f th e de nsit y of tal ents goe s up yo u ca n au toma tica

lly crea te s omet hing une xpec ted.

But if y ou w ere to c hang e th e en viro nmen t an d ev eryo ne w as v ery

medi ocre th en y ou w ould at leas t ch oose you rsel f. A

f. A nd t hen you' d be fin e wi th i t. A

t. A nd t hen you' d be abl e to get awa y wi th i t. H

t. H ow m any peop le w ere ther e in the ope n AI com pany whe n yo u we

nt?

I wa s 28 0. Y

0. Y ou w ere 280?

Yes . Ho

. Ho w ma ny p eopl e ar e th ere in t he o pen AI n ow?

More tha n 3, 000.

So it's ten tim es t he n umbe r in thr ee y ears . Yes.

. Yes.

Do y ou t hink Ope nAI can stil l ma inta in t he o ld t eam that is smal

l an d be auti ful and crea te a new sty le?

I th ink the prob abil ity is d own, but not tha t mu ch.

But you can alwa ys c reat e a smal l te am a nd d o so me r esea rch.

What int eres ting eff orts do you thin k th e le ader ship of OPI has

made to avoi d th is?

Or t o ma ke t he t eam more clo se t o th e pl ayer s an d cr eate a n ew t

eam?

For exa mple we sim plif ied the orga niza tion str uctu re a nd c ance

led all the unre ason able mee ting s. I

s. I thi nk i t's main ly b ecause What is a hi ghly eff icie nt, inno vati

ve o rgan izat ion arch itec ture ? Ho

? Ho w do you und erst and it?

For exam ple, tod ay t here is a de cisi on t hat can be t rans mitt ed w

itho ut l oss to t he b otto m. T

m. T hen what kin d of pro gres s ha s be en m ade belo w? T

w? T hen it c an b e tr ansm itte d wi thou t lo ss t o th e top.

Othe rwis e y ou m ay h ave to m ake deci sion s ab ove and work bel ow.

This may be comp lete ly d iffe rent fro m th e tw o di rect ions . An

. An d th en i t's not too late to make a m ove.

How doe s OP N ma ke t his good inf orma tion flo w? F

w? F irst of all, lik e Sa m an d Gr eg, they wil l ha ve a spe cial one

. Sa

. Sa m wi ll h ave a sp ecia l re sear ch a ssis tant bef ore.

To help him und erst and the late st r esea rch prog ress wit hin the

comp any.

Greg no t to men tion he was inv olve d in the who le i nfra stru ctur

e of inf ra.

So t hey are very fam ilia r wi th t his tech nolo gy.

And what the y ne ed t o do is to m aint ain the sens itiv ity of t he t

echn olog y. A

y. A t le ast they nee d to kno w ho w fa r th ey a re i n do ing this an

d wh at p rogr ess they hav e ma de, and what the y wi ll u se i t fo r. S

r. S o th is t hing has to be d one by a ski lled per son, will ing to g

o in and stu dy t he d etai ls, and unde rsta nd e very sin gle deta il o

f th e co mpan y. I

y. I thi nk m anag ing the comp any and the code lib rary is simi lar,

it' s co nsis tenc y. I

y. I f yo u're not con sist ent, the n yo u're lik e a huma n be ing, lik

e a huma n wh o's been fig htin g fo r hi s li fe, his body mov ed, but

his feet did n't.

It' s we ird.

Oka y l et's tal k ab out post -tra ining.

Let' s st art with 3.5 . Bu

. Bu t in fac t t he P PO o f 3.

5 is not so much . I

. I firs t ad just ed 4 . Be

. Be caus e 3.

5 wa s us ing the old infr astr uctu re.

The new infr astr uctu re i s ju st i n Au gust . Th

. Th en I adj uste d th e fi rst vers ion of P PO o n th e ne w in fras truc

ture.

and then run wit h 4.

It was in S epte mber of 2022 . Th

. Th at's int eres ting . So

. So whe n 30 5 ca me o ut, the 4 wa s al read y th ere.

And ROH F wa s fi rst work ed o n 4 and then est abli shed on 305.

Yes, but oth er t eams hav e he lped us a lo t. W

t. W e us ed t he p ipel ine of E O to use it in m any ways . Bu

. Bu t th e ma in t hing s l ike how to t rain the rea l wo rld mode l w

e tr aine d ou rsel ves.

How to coll ect data we col lect ed i t ou rsel ves.

Wha t pr oble ms d id w e fa ce w ith infr a w e al so h andl ed i t ou rsel

ves.

What wer e th e ke y ch alle nges and bre akth roug hs f or y ou t o wo rk w

ith RLHF ? I

? I thin k it was how you sho uld meas ure the perf orma nce.

Bec ause no one know s wh at t he p erfo rman ce s houl d be . Yo

. Yo u me an t hat afte r tr aini ng a lot of chec kpoi nts, you don 't k

now if i t's gett ing bett er?

Yes.

For exa mple th e si ngle rew ard will cau se r ewar d ha cking.

the rewa rd s atur ator mig ht t urn into a s trai ght line . It

. It wil l sl owly ris e an d be come a s trai ght line . Bu

. Bu t th e re al s itua tion is that if huma ns h ave a re ward it can

go up f irst and the n fa ll d own slow ly.

This is rewa rd h acki ng.

So y ou c an't kno w wh ich chec kpoint is r eall y be tter tha n ot her chec

kpoi nts.

So, sel ecti ng a che ckpo int is a lso a sk ill.

But we didn 't s pend muc h ti me s elec ting che ckpo ints . We

. We bui lt a bun ch o f sa mple -bas ed U RLs to s ee w hat each ben chma

rk l ooks lik e. B

e. B ut t he b ench mark is just a n umbe r. I

r. I f it pas ses a nu mber it 's g ood.

You can' t sa y it 's g ood or b ad.

Beca use ever y ti me y ou r un a mod el, the diff eren ce i s hu ge.

Ther e ar e a lot of n oise.

How did you solv e it ? I

? I just too k it dow n an d ch ecke d it out . I

. I talk ed t o it a f ew t imes . I

. I chec ked out what it was like for me.

I a sked a f ew p eopl e to tak e a look and vot e. S

e. S o yo u us ed H F to eva luat e. Y

e. Y es, that 's t he o nly way.

The re w as n o ot her way.

So t his is y our firs t ti me p layi ng R IO I nfra at the indu stri al l

evel . Ye

. Ye s. W

s. W hat do y ou t hink is the diff eren ce b etwe en t he l arge -sca le R

IO I nfra tha t a larg e mo del need s an d th e si ngle tas k or toy tas

k RI O In fra that you 've play ed b efor e? I

e? I thi nk t he d iffe renc e is ver y bi g be caus e th e bo ttle neck of

toy task is the envi ronm ent.

Beca use its mode l is ver y si mple wh ethe r it 's t rain ing or s ampl

ing, it' s ve ry c heap . Bu

. Bu t in RL- INFR A y our mode l is big bu t th e en viro nmen t is ver

y si mple . Th

. Th e en viro nmen t is a p romp t. B

t. B ut i f th e mo del is s ampl ing, you hav e to con side r ho w to sam

ple it i n a high -end way.

And for trai ning yo u ne ed t o co nsid er h igh- spee d tr aini ng b ecau

se t he p romp t pr ovid ed b y th e en viro nmen t is onl y a few seco nds.

But if you run the infe renc e an d th e tr aini ng, it's pro babl y hu

ndre ds o r th ousa nds of s econ ds.

If y our GPU is l ow.

What new cha llen ges do y ou t hink the fut ure mode ls w ill have com

pare d to 3.5 and 4?

For the RL-I nfra . I

. I thin k it 's s till abo ut p erfo rman ce o r ho w to sca le u p. H

p. H ow t o us e mo re G PUs.

How to have a h ighe r pe rfor manc e. Y

e. Y es, yes.

That inv olve s no t on ly t he R L pa rt, but also the mod el's inf eren

ce.

Yes, it' s mo re e nd-t o-en d a nd i t wi ll d elve int o so me o f th e ac

tual det ails an d th en d o so me e nd-t o-en d op timi zati on.

Actu ally wh at y ou d id w as a fai rly inte rsec tion par t. Y

t. Y ou h ave RL, you have to unde rsta nd R L a nd y ou h ave to u nder

stan d ML Cs.

Yes.

And you als o ha ve t o un ders tand how lan guag e mo dels are inf eren

ced, rig ht?

Yes.

You hav e to und erst and thes e. Y

e. Y es.

Mayb e no t in tha t po siti on.

To b e ho nest st uden ts a t sc hool can 't l earn thi s ki nd o f th ing.

Yes th at p osit ion is v ery tiri ng.

And then I w as r eall y ti red for a wh ile.

And the n I ente red the ER.

It w as t oo h ard to w ork over time and my brai n wa s hu rtin g. I

g. I cou ldn' t ta ke i t an ymor e. I

e. I wen t to the ER and took a l ook.

The doc tor said it was fine.

How hard was you r wo rk?

I wo uld wake up in t he m orni ng a nd w rite a d ebug or solv e so me p

robl ems. I w ould sle ep a t ni ght.

7 d ays a we ek?

6 da ys o n av erag e. B

e. B ut a fter a w hile yo u wi ll f ind that it can' t la st l ong.

So firs t yo u ne ed t o ha ve a body.

You need to make sur e yo u're hea lthy . I

. I now trai n to run 300 0m e very 2 w eeks . Bu

. Bu t wh en I was in Tsin ghua I didn 't m ake it t o th e gy m. I

m. I did n't run 3000 m at all . Bu

. Bu t no w I' m aw are that thi s is ver y im port ant.

I'm so jeal ous of w hat you' ve d one in O PEA in t he p ast 2 ye ars.

you can expl ore most of the worl d's rese arch on 99.99% that the inf

rast ruct ure engi neer is not qual ifie d to tou ch.

You have the adv anta ge o f be ing able to pred ict the futu re.

You can make thi s op timi zati on o n th e mo st a dvan ced mode l. A

l. A nd y ou e xplo re i n th e un know n ev ery day.

And you kno w th at y our expl orat ion must be the firs t on e in hum

an h isto ry t o fi nd t his answ er.

But I th ink I do som e ve ry t rivi al t hing s. I

s. I thi nk m y jo b is to main tain my dail y li fe.

I do n't need so many

-

I th ink ther e ar e so me t hing s. A

s. A gain yo u ca n't pred ict what wil l ha ppen nex t. E

t. E very thin g th at h appe ns n ext may happ en.

Ther e ma y be new way s n ew R RO w ays, or new pre- trai ning way s. S

s. S o yo u ha ve t o fa ce s ome unkn own chal leng es e very day . Bu

. Bu t to be hone st, I th ink...

ink...

Actu ally I' m ve ry l ucky to be i n th is p osit ion.

But if I we re t o be rep lace d by any one, if he h ad m y co ntac ts,

he s houl d be abl e to com plet ely take over.

So I don 't t hink it' s ju st m e wh o ca n do som ethi ng.

It's som ethi ng a ny n orma l hu man can do.

You' re t oo m odes t. W

t. W hat do y ou t hink is the bigg est chal leng e in the nex t 5 to 1

0 ye ars?

Wha t ar e th e ch alle nges ? Wh

? Wh at k ind of b reak thro ughs do we n eed to s ee?

I th ink the curr ent stat e is tha t we hav en't ful ly s cale d up.

So w e'll wai t fo r it to slow ly h eal and clim b. W

b. W e'll wai t fo r it to slow ly s ee h ow m uch it c an d o in som e la

rge- scal e RO exp erim ents . Th

. Th en w e'll see wha t el se i t ca n do.

So y ou m ean that wit hout sca le-u p i t's not that the com pute is

not enou gh, but that we have n't drie d up the exi stin g pe rfor manc

e ye t. L

t. L et's dry up the exis ting met hod and the exis ting com pute fir

st, and then see how lon g we can pus h it . An

. An d th ere are also a l ot o f bu gs i n in fra.

Eve n yo ur c urre nt i nfra is very sur e th at t here are bug s. Y

s. Y es, you can' t sa y th at t here are no bugs . Af

. Af ter all, it' s hu man.

Eve ryon e is lik ely to m ake mist akes.

You have to fix the cont ext inco nsis tenc y a nd e very one will wri

te w eird stu ff.

What do you thin k wi ll b e th e bi gges t bo ttle neck in the futu re o

f la rge mode l h ome- base d le arni ng, and post -tre nd p ipel ines ? I

? I thin k th e bo ttle neck is the supp ort for infr astr uctu re.

How many boa rds can you fix in a tim e fr ame?

And how many tim es c an i t be rel ayed wit hin the time fra me.

How many tim es c an i t be rel ayed cor rect ly.

The rest is not impo rtan t. T

t. T his can be u sed to r ecov er a ll t he r emai ning thi ngs.

Yes . Wh

. Wh ethe r it 's t he a lgor ithm or the envi ronm ent.

Yes if you fix all the bug s y ou m ay n ot n eed to c hang e th e al

gori thm.

Tha t's good . Ho

. Ho w do you bri ng u p th e ef fici ency of infr a? W

a? W hat kind of tool s do you nee d? W

d? W hat kind of peop le d o yo u ne ed?

What kin d of res ourc es d o yo u ne ed?

We a re s till exp lori ng t his.

Wha t I' m do ing now is n ot i n th e co re p osit ion, but I t hink I s

houl d do som ethi ng m ore impo rtant.

We a re r e-bu ying the inf rast ruct ure insi de O penA I an d th en w e wi

ll b uild the nex t ge nera tion of infr astr uctu re.

Will you be rebu ildi ng t he i nfra stru ctur e of eac h ge nera tion ? Ac

? Ac cord ing to y our curr ent unde rsta ndin g w ill you buil d a good

top -lev el s truc ture ? An

? An d th en w e wi ll w rite a l ittl e bi t ab out it i n th e vi deo.

At the mome nt, we a re r ebui ldin g it . Be

. Be caus e th e pr evio us g ener atio n of inf rast ruct ure has been thr

ee y ears old . An

. An d th en t here are a l ot o f pr oble ms w ith it.

And then ...

...

We h ope to u se a new inf rast ruct ure to c lean up the tech nica l de

bt a nd g ive rese arch ers bett er i tera tion spe ed i n th e si ngle -tim

e pe riod . So

. So res earc hers won 't b e in volv ed i n th e in fras truc ture bui ldin

g pr oces s. T

s. T hey may give som e de mand . Bu

. Bu t th ey a re n ot r espo nsib le f or h ow t o do dis trib uted tra ining.

And then the y mi ght just cha nge the flag . It

. It sou nds like the ope n-en ded rese arch er i s pr obab ly t he f irst

job tha t AI rep lace d. Y

d. Y es, I th ink so.

Anyw ay, you can expe rime nt i n a sing le t ime.

He just cho se I DE.

Yes, jus t ID E. I

E. I DE, you can gene rate it very che aply . Ho

. Ho w ma ny s ongs do you writ e al one?

And you can even use AI to m odel you r id ea.

Yes.

I t hink you sho uld repl ace the rese arch er f irst th en t he i nfra

stru ctur e en gine er.

Then you can rep lace the m all.

But sale s mi ght not be t hat exci ting . Be

. Be caus e sa les requ ires you to take car e of the oth er p erso n's

bill . An

. An d AI mig ht n ot b e ab le t o ta ke c are of t he o ther per son' s bi

ll.

So c ommu nica tion bet ween peo ple is m ore impo rtan t. O

t. O kay, we just tal ked abou t 30 5 wi th t ax-o nly.

How muc h di d we tal k ab out?

Wha t do you thi nk a bout pos t-tr aini ng w ith agen t an d RI O? T

O? T here 's n o fu ndam enta l di ffer ence . It

. It 's t he s ame thing.

Just add a f ew m ore deca ls i n th e mi ddle . Ma

. Ma ybe the envi ronm ent will be more . Ye

. Ye s t he e nvir onme nt w ill be c hang ed.

So y ou t hink com pare d to the sta ndar d AO M pl us A rial Pos t-Tr ain,

AGI plu s Ar ial Post -Tra in, it's not a n ew c hall enge . Ye

. Ye s b ecau se i t's the same thi ng.

What is your per sona l de fini tion of AGI?

Do you thin k we hav e re ache d AG I no w? I

w? I f yo u th ink we h aven 't r each ed i t ye t o r it 's s till a l ittl

e bi t sh ort, do you thin k th e pr e-tr ain call ed A rial Pos t-Tr ain

can take us to t he A GI y ou d efin ed?

Ther e is a j oke in O penAI.

You can catc h 15 peo ple with 20 diff eren t wa ys t o de fine AGI . Th

. Th e wa y I defi ned it b efor e wa s th at i f th is t hing can acc ompl

ish 80% or 9 0% o f th e ta sks that I t hink are mea ning ful, the n it

mig ht b e AG I. I

I. I don 't t hink so yet.

Bec ause at the mome nt, From my poin t of vie w I 'm s till not at

ease wit h hi m di rect ly c hang ing my i nfra cod e. B

e. B ecau se i t's very out of dist ribu tion . AI

. AI inf ra i s al most zer o in com pari son to i ts d ata leve l. A

l. A nd A I in fra, you 've been ver ifyi ng t his for too long.

- Th e co st i s to o hi gh.

- Ye s t he c ost is t oo h igh.

We stil l ca n't get to t his poin t. G

t. G ood for you.

It soun ds l ike you won' t be cri tici zed by A I. Y

I. Y es, but anyw ay...

ay...

Befo re S traw berr y ca me o ut, we h ad b een usin g it for a w hile . Ev

. Ev eryo ne t houg ht t hat my j ob w as g oing to be r epla ced.

Or we c ould jus t wr ite a bu nch of s hit.

And the n af ter Stra wber ry f inis hed trai ning th ey w ould cle an i

t up for us.

It seem s li ke i t's been a y ear or t wo.

The shit is stil l th ere.

It's not lik e yo u ca n ch ange any thin g. E

g. E very one over reac ts.

They thi nk t hat tech nolo gy i s he re, and they wan t to do this and

tha t. B

t. B ut i t's not like tha t. I

t. I t's a ve ry s low proc ess of p roce ss d evel opme nt.

You' re n ow a t Op enAI . Yo

. Yo u us ed t o wo rk a t a acad emic ins titu tion li ke C NVU.

You mig ht h ave work ed a t a rese arch lab . Bu

. Bu t no w m any people I th ink this is a tr ade- off.

I'm act uall y ve ry e nthu sias tic abou t op en A I. I

I. I f op en A I ha s so meth ing open I migh t pa rtic ipat e in it.

But I t hink I s houl d do wha t I thin k is mor e im port ant.

And thi s op en A I is a t rade -off.

You can' t ju st o pen the best mod el b ecau se t he c ompa ny n eeds to

surv ive.

If the comp any can' t su rviv e t hen you may not be a ble to c onti

nue fina ncin g d o so me e xper imen ts, and then mak e so me b reak thro

ughs . Th

. Th is i s a very rea l pr oble m. S

m. S o I agre e wi th t his chan nel.

But whe n Op enAI was fir st e stab lish ed, Its arch itec ture is actu

ally ver y sp ecia l. I

l. I t's not a co mpan y st ruct ure base d on com merc ial prof it.

It's not a c ompa ny s truc ture bas ed o n co mmer cial pro fit.

At leas t th e sl ogan in the publ ic i nfor mati on a t th e be ginn ing

is t o ma ke t he s ame AGI equa l to all hum ans.

Do you thin k B- Yuan is to m ake this goa l cl oser or furt her?

Fir st o f al l t he s ame AGI to a ll h uman s ha s be en s een in t wo p

arts . Th

. Th e fi rst is t o re aliz e AG I. T

I. T he s econ d is to all huma ns.

Tune AGI is v ery easy to use.

Jus t pu t it in I/O, or in p re-t rain ing, in 3D, and scal e up . Th

. Th e cu rren t so luti on f or t he h uman rig hts issu e is to make pro

duct s at the che apes t pr ice poss ible . Fo

. Fo r ex ampl e t here are fre e Ch aiGP use rs w ho c an g et i n to uch

with tec hnol ogy more eas ily.

For exam ple, you can use the voi ce m ode for free . An

. An d th en y ou c an d o so me e xper imen ts.

This is prob ably mor e be nefi cial to all huma ns, inst ead of j ust

thro wing a n aked mod el a t th e so urce . It

. It doe sn't kno w ho w to use it.

So you mean Ope nAI' s Op en i s no t op en t o ot her big comp anie s b

ig m odel com pani es, but to o rdin ary peop le.

Yes.

But if w e we re t o re ach AGI in t he l ast mile I woul d ag ree with

tha t. B

t. B ut i f AG I is not a m atte r of tim e a nd c an b e so lved nex t ye

ar, will it be m ore open and tra nspa rent in term s of tec hnol ogy?

Wil l Op enAI be more eff icie nt i n re achi ng A GI?

Or d o yo u th ink Open AI d oesn 't n eed to r evea l it s ow n te chni cal

deta ils and get feed back fro m th e co mmun ity?

Ope nAI can alre ady achi eve this by itse lf?

Let me t hink . I

. I thin k th ere is a pat h. Y

h. Y ou c an o pen a so urce and rec eive fee dbac k fr om t he c ommu nity.

And then it' s be tter to say that thi s ca n be don e b ut i t's actu

ally ver y di ffic ult to d o. F

o. F or e xamp le, if y ou'r e th e fi rst, and you 're far away th en t

he o ther s wi ll b e th e fi rst, and the oth ers will tra in a bit mor

e a nd t he o ther s wi ll b e fa r aw ay.

And then it will lea d to you r...

r...

Not eve ryon e is the sam e. A

e. A nd t hen it w ill lead to, for exa mple Op enAI in this cur rent

env iron ment, he m ay n ot b e ab le t o af ford it.

No one will con tinu e to pay for him . Th

. Th is i s a bit of a gam e th eory . Ev

. Ev en i f I want to crea te a hum an r ace for AGI, but may be s ome

peop le d on't thi nk s o. S

o. S ome peop le j ust want to make mon ey.

So i n or der to p reve nt t his from hap peni ng, Open AI h as t o av oid

it.

At l east in my o pini on, it i s so . Th

. Th is i s th e co mpan y's life -sav ing cons ider atio n r ight ? If

? If Ope nAI is n ow u nlim ited res ourc es, you don' t ha ve t o wo rry

abou t de ath fore ver.

Do y ou t hink you 'll be h appy to open up your RIO inf rast ruct ure

in t he n ext 2-3 year s? O

s? O f co urse I'l l be hap py.

John Sho man aske d me if I wa nted to open up.

I d idn' t th ink it w as a goo d id ea.

It w as f or t he c ompa ny's sak e. B

e. B ut h e as ked me a bout it.

So Deep Seq, or at l east Ope nWit s w ill open up the inte rnal eva

luat ion?

Yes.

You jus t sa id t hat Open AI's mis sion is to b reak dow n in to t wo p

arts . Fi

. Fi rst, to impl emen t AG I a nd t hen to b uild a h uman -led ser ver.

To go d eepe r i f we wan t to und erst and this mis sion wh at d o yo

u th ink is t he b igge st o ppor tuni ty a nd c hall enge to achi eve this

pro cess ? Ex

? Ex ecut ion.

Exe cuti ng i n th e ri ght dire ction.

As l ong as i t's poss ible it 's g ood.

Don 't l et i t ha ppen aga in.

For exam ple, the com pany was abo ut t o cl ose in F ebru ary 2013 . Sa

. Sa m Al tman was fir ed.

You hope the ent ire orga niza tion str uctu re i s as sta ble as p ossi

ble.

It can help you mov e fo rwar d qu ickl y. S

y. S o fr om t he i nter nal pers pect ive, whe n Sa m wa s fi red, wha t ki

nd o f in form atio n di d yo u re ceiv e? B

e? B ecau se t he i nfor mati on w e re ceiv ed f rom outs ide was very com

plex . Wh

. Wh at d id I lya see?

No, tha t's...

t's...

rand om r umor s a nd m any peop le m ade rumo rs h ere.

So what is your inn er p ersp ecti ve?

My i nner per spec tive is that I d on't tru st I lya and some oth er b

oard mem bers . I

. I don' t tr ust Sam.

And the n th e vo te v oted him out . Bu

. Bu t th e pe ople wor king und er u s fe lt v ery surp rise d v ery shoc

ked.

I d on't kno w wh at h appe ned.

Bec ause The boar d of dir ecto rs p revi ousl y la cked tra nspa renc y on

the peo ple in t he l ower -lev el.

We d on't kno w ho w th is d ecis ion was made . So

. So you don 't t rust him ? I

? I don' t tr ust him.

You can che ck t he o ffic ial inve stig atio n re port.

Yes, it' s co nsis tent . Bu

. Bu t th e fi nal outc ome is t hat Ilya is gone . Ye

. Ye s. H

s. H ow d id t he c hemi cal reac tion hap pen?

He' s be en f ired an d he doe sn't tru st S am.

But in t he e nd, Sam seem s to be the one who' s mo re p opul ar.

Beca use many emp loye es t hink tha t if a p erso n wh o's born wit h pu

re t echn olog y is the lea der, the re m ight not be so m any comp onen

ts.

AGI' s ac comp lish ment is not just bec ause of tech nolo gy.

you can do i t on ce y ou h ave the skil ls.

But ther e ar e st ill many bus ines s fa ctor s in volv ed.

For exam ple, you nee d to fin ance yo u ne ed t o ca lcul ate, you nee

d to con vinc e pe ople to give you mon ey.

This is actu ally a v ery impo rtan t pa rt.

If y ou o nly have a g ood rese arch exp erie nce, it may not be a goo

d wa y to go.

From the lon g ru n i t's stil l a need for peo ple like Sam . So

. So Sam is the most dif ficu lt t o re plac e in AI.

Bec ause he need s to be in b usin ess, eve n in geo poli tics in res

ourc es.

You can draw Sam int o a pers onal ity, int o an ide ntity.

And if y ou r epla ce t he i dent ity with an AI i n a shor t ti me, then

the ide ntit y re cogn itio n of oth er p eopl e wi ll b e lo st.

So t his can' t be lef t be hind . Yo

. Yo u ju st s aid that you wer e sa d ab out the afte rnoo n wh en Z hang

Shu man left.

But in f act, SM is n ot t he o nly one who left Ope nAI.

Aft er O penA I ac hiev ed g reat suc cess co untl ess memb ers of t he t

eam left . Do

. Do you thi nk t hat a ve ry s ucce ssfu l or gani zati on l ike Open AI,

will los e a lot of t alen t? I

t? I s it ine vita ble?

A h ealt hy o rgan izat ion can repl ace ever yone . As

. As lon g as you can con tinu e to tra in n ew p eopl e h ave the abil

ity to m ake choi ces, and make the org aniz atio n ru n no rmal ly.

Even tho ugh it's a l ot o f pe ople it can sti ll t ake some tim e an

d en ergy to trai n a new wave of peop le.

And then con tinu e to bui ld b lood li ke a liv er c ell.

So no o ne c an r epla ce O penA I no w. Y

w. Y es.

Does tha t me an t hat the thin gs O penA I do es, peop le o utsi de c an't

com e ba ck t o li fe?

You can thin k of it that way . Be

. Be caus e it 's j ust good to do t he s impl e th ings.

Ther e's no b lack mag ic.

Now we a re i n th e mo st i nten se t echn olog y co mpet itio n in hum an h

isto ry.

Open A is the one tha t ig nite s al l of thi s. I

s. I wan t to kno w ho w in tens e is you r in tern al a tmos pher e? D

e? D o yo u ha ve a lot of pres sure ? Lo

? Lo okin g at the tea m t he d eadl ine, and the pro ject tim eline, For

exam ple, pos t-tr aini ng i s st ill a bi g pr essu re.

And then li ke o ther tea ms, for exam ple, we are now re-e nrol ling

in infr a. T

a. T here may be pres sure bu t no t as big as post -tra inin g. B

g. B ecau se t hey have a v ery clea r de adli ne.

And then we can take a l ittl e bi t of tim e to thi nk a bout it, bec

ause we need to focu s on lon ger- term con side rati ons.

We n eed to g et t hing s ri ght.

So, wil l th e ex tern al c ompe titi on f rom Xeia An thro pik, or Chin

ese big- mach ine comp anie s be tra nsmi tted to your int erna l co mpan

y's dail y de velo pmen t? N

t? N ot r eall y. E

y. E xcep t fo r De epSe ek's be caus e th ey c laim ed o n Tw itte r th at t

hey were ver y fa st i n de live ring data.

This cau sed a lo t of peo ple to b e al ert beca use the inte rnal spe

ed o f th e de live ry i s ac tual ly a lit tle slow er t han the othe rs.

So t hat' s wh y we wan t to pas s th e In fra.

So for a la rge comp any like Ope nAI, the lif e an d de ath line is

Infr a's cycl e ti me.

AI I nfra req uire s mo re c onte xt.

But if y ou a re d oing som e da ta a blat ion, lik e ru nnin g so me e xper

imen ts, you don' t ne ed s o mu ch c onte xt.

You just sim ply come in, wri te a for loo p a nd t hen run the data.

and then run it.

Thi s ca n ac tual ly b e au toma ted.

So the real war ning for you is to r ealize Deep Seq' s in fras truc ture

is very goo d an d th ey a re f ast in d eali ng w ith it.

This is what cau ght your att enti on.

You don' t ca re w hich mod el i s on whi ch b oard and how muc h hi gher

tha n GP T. T

T. T his is n ot w hat you care abo ut.

We h aven 't d one anyt hing spe cifi call y to bru sh L MSs for a lo ng t

ime.

Wha t yo u re ally car e ab out is t he s peed and suc cess rat e of the

uni t in the tim e fr ame.

Do y ou t hink Ope nAI is t he b est in t he w orld wit h th is i ndic ator

? No

? No de fini tely not . Th

. Th is h as a lot to do w ith the orga niza tion str uctu re.

For exam ple, you dra w a team and sta rt a sta rtup . Th

. Th eir deli very spe ed i s de fini tely hig her than Ope nAI' s. B

s. B ecau se f irst of all, you r co de i s sm all and comm unic atio n co

sts are low.

As long as you focu s on you r us e ca se, you can do it.

But Open AI n eeds to cons ider man y us e ca ses.

And tra de-o ffs in a ll a spec ts.

If O penA I is not the num ber one in t his kind of life and dea th l

ine, wil l Op enAI sti ll b e th e nu mber one AI mode l in the fut ure?

Eve ry c ompa ny h as t his prob lem.

You mea n ev ery comp any will slo w do wn i f it 's s ucce ssfu l? Y

l? Y es.

It d epen ds o n wh ich one is n ot b ad.

It's not tha t bad.

This is actu ally wor se t han the firs t-cl ass comp anie s. T

s. T he f irst -cla ss c ompa nies may be bett er a t th is i ndic ator . Ot

. Ot her indi cato rs m ay b e ha rder to comp ete with Ope nAI.

For exa mple us ing user fee dbac k. Y

k. Y es, so t his is a ll t rade -off . Th

. Th is i s th e pr oble m th at h uman org aniz atio ns w ill face aft er t

hey deve lop to t his scal e. Y

e. Y ou c an't avo id i t. I

t. I t's not that it' s ha rd t o ma inta in h igh huma n de nsit y. I

y. I t's that it' s ha rd t o ma inta in t he c onsi sten cy o f co ntex t sh

arin g of org aniz atio nal stru cture.

This wil l ca use your inf rast ruct ure to b e in cons iste nt, and then

the inf rast ruct ure will sta rt t o be sta gnan t a nd t hen the orga

niza tion al s truc ture wil l al so s tart to stag nate . Th

. Th is i s ho w it is with man y pe ople . So

. So in theo ry, ther e sh ould be an a gent tha t ha s an inf init e co

ntex t to rep lace the con tent sha ring . It

. It sou nds like a g reat ide a. Y

a. Y es.

Beca use this wil l so lve the prob lem of s tagn atio n in hum an h isto

ry.

Orga nizi ng, writ ing code or doi ng o ther thi ngs are all faci ng a

pro blem . Ye

. Ye s b ecau se t he c onte xt o f pe ople is limi ted.

You can 't s ave so m uch cont ext at o nce, but AI can.

May be i n th e fu ture ev ery comp any will hav e an age nt w ith such

unl imit ed c onte xt.

Then com e to be the CEO.

Yes he is resp onsi ble for all shar ing, all dec isio ns.

Yes.

The re m ay b e no dec isio n ma ker more sui tabl e th an s uch an a gent

. Ye

. Ye s. T

s. T hen afte r ta lkin g ab out the futu re, if y ou w ant AI t o so lve

a wo rld prob lem, wha t wi ll i t be ? Ho

? Ho w to pre dict the fut ure?

You' re n ot t alki ng a bout the fut ure of c ups, but the fut ure of t

he w orld th e wo rld, the wor ld, the worl d t he w orld th e wo rld,

the wor ld, the worl d t he w orld th e wo rld, the wor ld, the worl

d t he w orld th e wo rld, the wor ld, the worl d t he w orld th e wo

rld, the wor ld, the worl d t he w orld th e wo rld, the wor ld, the

worl d t he w orld th e wo rld, the wor ld, the worl d t he w orld th

e wo rld, the wor ld, the worl d t he w orld th e wo rld, the wor ld,

the worl d t he w orld th e wo rld, the wor ld, the worl d t he w orld

th e wo rld, the wor ld, the worl d t he w orld th e wo rld, the wor

ld, the worl d t he w orld th e wo rld, the wor ld, the worl d t he w

orld th e wo rld, the wor ld, the worl d t he w orld th e wo rld, the

wor ld, the worl d t he w orld th e wo rld, the wor ld, the worl d t

he w orld th e wo rld, the wor ld, the worl d t he w orld th e wo rld,

the wor ld, the worl d t he w orld th e wo rld, the wor ld, the worl

d t he w orld th e wo rld, the wor ld, the worl d t he w orld th e wo

rld, the wor ld, the worl d t he w orld th e wo rld, the wor ld, the

worl d t he w orld th e wo rld, the wor ld, the worl d t he w orld th

e wo rld, the wor ld, the worl d t he w orld th e wo rld, the wor ld,

the worl d t he w orld th e wo rld, the wor ld, the worl d t he w orld

th e wo rld, the wor ld, the worl d t he w orld th e wo rld, the wor

ld, the worl d t he w orld th e wo rld, the wor ld, the worl d t he w

orld th e wo rld, the wor ld, the worl d t he w orld th e wo rld, the

wor ld, the worl d t he w orld th e wo rld, the wor ld, the worl d t

he w orld th e wo rld, the wor ld, the worl d t he w orld th e wo rld,

the wor ld, the worl d t he w orld th e wo rld, the wor ld, the worl

d t he w orld th e wo rld, the wor ld, the You need to make a s crip

t in adv ance . My

. My und erst andi ng i s th at y ou t hink our fat e ca n be pre dict ed.

Yes, thi s wo rld is a def init e qu esti on.

So w e li ve i n a defi niti ve M arko v pr oces s. Y

s. Y es.

In t he d eep, you thi nk p eopl e ar e no t fr ee t o be fre e. Y

e. Y es.

What am I th inki ng i n my hea d no w? W

w? W hat am I say ing in t he n ext word ? Wh

? Wh at a m I aski ng i n th e ne xt q uest ion?

It' s al l se t at the beg inni ng o f th e un iver se.

Yes.

I ha ve w itne ssed cou ntle ss c hang es.

This is a ve ry p essi mist ic w orld view in my o pini on.

Yes.

Bec ause ma ybe this is true bu t I don' t wa nt t o ac cept it.

I d on't wan t to acc ept it e ithe r. I

r. I t's like I'v e be come a p rinc iple tha t's bein g si mula ted.

Yes . Wh

. Wh y do you thi nk t hat?

Som e of my pers onal exp erie nces it 's n ot v ery conv enie nt t o sa

y b ut t his is i ndee d th e tr uth.

Ever ythi ng c an b e pr edic ted.

So, the oret ical ly, it c an b e so lved wit h AI . Bu

. Bu t if you get a m achi ne t hat can pred ict the futu re, then it'

s ac tual ly a dis aste r fo r me . I

. I thin k th is w ill lead to a co llap se o f al l va lue syst ems. Yes

. If

. If the re i s su ch a n AI mod el, it m ay b e th e be st c hoic e fo r hu

man soci ety to d estr oy s uch an A I mo del and let it n ever com e ou

t. Y

t. Y es.

But...

Some peo ple are will ing to d evel op t his mode l. O

l. O ther wise it wil l be man ipul ated by the fate . Yo

. Yo u me an s omeo ne l ikes thi s mo del beca use.

.. H

e wa nts to j ailb reak . So

. So you mea n af ter havi ng t his mode l I hav e to do some thin g di

ffer ent from wha t th e mo del pred icte d? Y

d? Y ou c all this esc ape?

No.

It' s ju st t o fi gure out the rul es o f th e wo rld.

Or w hy t his worl d is det ermi ned?

Why thi s wo rld is a the ory of f ate?

So you thin k Go d do esn' t kn ow w hat to d o? Y

o? Y es.

Red ligh t do esn' t kn ow w hat to d o. T

o. T he m icro lig ht k nows wha t to do.

Is this tru e? I

e? I f we loo k at qua ntum mec hani cs..

cs..

. Th

is i s tr ue b ecau se y ou c an m odif y so me w orld lin es b acks tage . Al

. Al thou gh y ou c an t hink wha t I' m sa ying is bull shit . No

. No pro blem . I

. I don' t th ink it's bul lshi t. I

t. I 'm t ryin g to rea son behi nd it.

I've bee n th inki ng a bout thi s fo r a long tim e. I

e. I s th is w orld rea lly sure ? Do

? Do we have a s crip t? D

t? D o we jus t sh ow t hat we a re d oing som ethi ng r ando m in the dyn

amic s of thi s wo rld?

I t hink so.

It' s no t ra ndom it 's a def init e pr oces s. I

s. I thi nk s o. Y

o. Y ou h ave no d oubt abo ut t his?

I h ave a do ubt, but I t ry t o pr ove it.

I fi nd i t so wro ng.

I re ally wan t hi m to pro ve i t. I

t. I hav e th is q uest ion.

I'm ver y cu riou s. I

s. I thi nk I men tion ed t his to m y pa rent s wh en I was six or seve

n ye ars old.

I f eel like how do you know tha t wh at I sai d is not a c erta inty

? Ho

? Ho w do you kno w th at y our reac tion is not a ce rtai nty?

I w as t hink ing abou t th is q uest ion when I w as a kid . Bu

. Bu t I foun d ou t la ter that if it i s li ke t his, I d on't thi nk I

am huma n. S

n. S o th e be st w ay i s to for get all this and pre tend tha t yo u do

n't know this.

and expe rien ce t he c urre nt e xper ienc e. H

e. H ave you done thi s be fore ? I'

? I' ve a lway s do ne t his.

I c an't hel p it . Th

. Th ere' s an othe r ex plan atio n. T

n. T he t imel ine is n ot l inea r. I

r. I t ca n be jum ped over . So

. So you can use a t heor y to exp lain the fut ure me, the me w ho h

elps the pas t me.

to m ake cert ain deci sion s. Y

s. Y ou j ust said tha t yo u we re i n el emen tary sch ool or h igh scho

ol, and sudd enly you tol d th e fu ture you sai d yo u wa nted to make

an impa ct.

Yes, thi s is jus t a sent ence tha t ca me t o my min d. B

d. B ut I don 't k now if i t's the futu re I tol d my self . Wh

. Wh y do you thi nk t he w orld lin e is not con tinu ous?

May be t here is a fu ture tra nsfe r? I

r? I thi nk t hree -dim ensi onal org anis ms h ave thei r ow n bo unda ries

. In

. In the thr ee-d imen sion al c ogni tion, Time is a li near on e-wa y fl

ow.

But it's pos sibl e th at i n th e ti me o f th inki ng, time is not one-

way.

It can jump any wher e. T

e. T his is a rea sona ble expl anat ion for the 7th dime nsio n I foun

d. S

d. S o yo u th ink that in the past th ere were cou ntle ss m omen ts,

ther e wa s a futu re y ou, like Mor pheu s wh o tr avel ed t hrou gh t he s

tars pu shin g th e pi llar beh ind you.

Yes th is i s on e po ssib ilit y. B

y. B ecau se t his is n ot t rue.

And he push ed t his pill ar, It w ill lead to you push ing the bar

agai n in the fut ure.

Yes . So

. So you r be st s olut ion now is t o fo rget abo ut i t. Y

t. Y es.

And stil l be liev e th at X ixi Fuzi is happ y. Y

y. Y es.

We c an t alk abou t th is f or a noth er h alf an h our.

Let 's g et o ver this top ic f irst . I

. I beli eve that eve ry o pen AI e ntre pren eur now has a se ed o f en

trep rene ursh ip i n th eir hear ts.

Have you tho ught abo ut t his?

May be, but I ha ven' t se en a goo d id ea y et.

And I th ink Open AI i s st ill a go od p lace for me.

Wha t ki nd o f id ea d o yo u li ke?

Is i t mo re a bout pro duct or what ? I

? I thin k I like pro duct more.

For exam ple, for tho se w ho w ant to q uit thei r jo b I act uall y ha

ve u sers for my rese arch inf rast ruct ure.

Bec ause you r us ers are rese arch ers, you can act uall y do the wor

k ba sed on t he f eedb ack from res earc hers . So

. So you wou ld l ike to h ave a si mila r ne ed.

Anyw ay, it's a n eed.

And you r ta lent can be used to show off . An

. An d wh at k ind of t echn olog y is not imp orta nt.

For exam ple, if you want to quit you r jo b y ou n eed a ve ry s impl

e no tifi cati on s yste m. Y

m. Y ou c an't cou nt i t as a s yste m j ust writ e a php code . Yo

. Yo u do n't even nee d a php, you can jus t cr eate one . Ev

. Ev en i f yo u qu it y our job ever y da y y ou c an u pdat e it onc e a

day, onc e at nig ht.

I do it manu ally bu t ev en t hat requ ires a l ot o f wo rk.

Tech nolo gy i s no t im port ant.

Wha t's impo rtan t is to find the rig ht p eopl e. W

e. W hat do y ou h ope Hong Jia yi w ill be l ike in 1 0 ye ars?

Wha t ki nd o f jo b wi ll h e do in a ce rtai n pl ace?

I h ope he w ill do w hat he w ante d to do at t hat time.

and have eno ugh reso urce s an d ab ilit y to do what I w ant to d o. Y

o. Y ou d on't int erfe re w ith what he thin ks?

Yes, bec ause tho ught s ch ange . Ma

. Ma ybe what you thi nk i s no t im port ant, it' s a dete rmin isti c pr

oces s an yway . Ho

. Ho w yo u ho pe n ow i s no t im port ant.

Wha t I can do n ow i s to inv est in m ysel f at tha t ti me, or i nves

t in the fut ure, so that he has the righ t to cho ose.

Why do you stil l ha ve t o in vest ? An

? An yway wh ethe r yo u in vest or not, you wil l re ach that poi nt.

But I th ink 投资一下 还是会更 好的吗?

就万一你 不投资的 那个..

那个..

. 你这

不是在骗 自己吗?

投资一下 不会有任 何影响的 ,我们是 确定性的 过去了。

比如说 每天在家 里睡大觉 ,你也可 以到那个 地方。

投资未来 可能也是 确定性的 。 所以

。 所以 你投不投 资也不是 自己的自 由意志。

是的。

你之前 的投资未 来都是提 前学高中 数学或者 说怎么怎 么样,你 现在除了 技术上的 ,AI上 的,你有 什么其他 的想投资 未来了?

提前退休。

and have eno ugh capi tal to d o wh at y ou w ant to d o. G

o. G ive you unli mite d mo ney, wha t do you wan t to do with it?

Spe nd s ome time to find wha t yo u wa nt t o do . Fo

. Fo r ex ampl e y ou t hink tha t wh at y ou w ant to d o is not wha t yo

u wa nt t o do . Yo

. Yo u ju st w ant to m ake enou gh m oney to reti re a nd t hen find tha

t th ing.

For exa mple th e tw o-pe rson clo thes.

It's alw ays been my drea m fo r a long tim e. B

e. B ut a s ti me g oes by, like aft er t hing s be come mor e st able or

afte r so meth ing happ ens, eve ryon e's focu s ch ange s. I

s. I use d to kno w wh at I wan ted, but now I c an't fig ure it o ut.

You can' t fi gure it out now?

Yes . I

. I thin k th at's nor mal.

So you' re i n a lost per iod of y our life now . Yes.

. Yes.

I us ed t o li ke R LInf er, doin g th ings tha t ha ve i mpac t. Y

t. Y es, beca use I've see n th e en d. T

d. T he r est are very dec isiv e th ings . AG

. AG S-ba sed thin gs.

Yes.

OK, Jia yu, I do n't have any mor e qu esti ons.

I h ope you can leav e a mess age at t he e nd o f th is p odca st i n 20

25.

Expl ore what you want.

Alth ough I'v e th ough t ab out what I w ant to d o I sti ll d on't thi

nk a bout it that muc h. I

h. I thi nk a bout it all my l ife.