Building OpenAI o1

[@waltonks8172]

I just tried make it rewrite a Korean lyrics into Japanese with matching number of syllables. No llm was able to do this before. The o1 can. Very impressive.

[@MrAbkejoe]

Please share the prompt I'd like to test it with my custom instructions utilizing a hierarchical structuring of XML tags and nested tags for a highly iterative reasoning during an output resulting in precisely the abilities described in the video. It would be so cool to try it lol I'll share the output whether right or wrong!

[@waltonks8172]

@@MrAbkejoe The prompt is actually pretty simple. "Rewrite this Korean lyrics into Japanese. Try keeping the number of syllables for each line to match the original flow. It took a little long time to finished though. I saw it try to rewrite the lyrics several times. But I think it was less than 5 minutes. For readable lyrics try adding a prompt "Output in romanized English"

[@MrAbkejoe]

@@waltonks8172 I was hoping for the exact verbatim prompt you used to maintain consistency lol that's kinda the point I wouldn't want to introduce any other additional factors by adjusting the prompt in any way to then be unsure of whether the output being correct/false calls into question the prompt being adjusted lol ik that prob sounds silly

[@seriouscoffeecup6450]

@@waltonks8172 Same as MrAbkejoe as asking: what song did you use?

[@MrAbkejoe]

@@seriouscoffeecup6450 Thank you lol, was hoping someone else would potentially chime in. Build a lot of custom LLM integration systems as a project lead for alignment adherence so I have a copious amount of Assistant agents (OpenAI primarily) with different configurations regarding their instructions (not just the system prompt lol) so I'm SO curious on the differing outputs from the varied agents. @waltonks8172 I hope that tags the person again, the anticipation has been killing me lol

[@maxziebell4013]

The cameraman had already been celebrating before this interview with champagne, as evidenced by its steady hand focusing on the speakers. 😉🥳

[@gatolocoowo]

ahaha i think it was a stylistic decision for a more "chill" like interview

[@benjiburkhalter]

Man was having a good time 😂

[@axumitedessalegn3549]

Too much 24

[@FreakyStyleytobby]

@@gatolocoowo He should have recorder it in vertical, as people usually do xD

[@johnvanderschuit]

Lol the camerawork is a bit interesting and at times slightly disorienting

[@youtube_gaming]

"Question itself" - that really stands out for me. What a time to be alive!

[@Courits]

Yes, that's true

[@Zibonnn]

Hold your paper!

[@joekind7687]

Indeed....What a time to be alive!

[@AsherKadmiel]

Aah a fellow scholar

[@MountMatze]

its just Marketing

[@AntoineDoury]

The cameraman work gave this a "The Office" touch that was funny 🤣

[@blackstormer95]

These guys must be smart af

[@phen-themoogle7651]

I thought the same thing 😂 And in two years the models will probably surpass everyone in that room and build everything without human intervention 😂

[@joaquinhernangomez]

@@phen-themoogle7651definitely not, unless there is a major paradigm shift

[@Jeffs60]

Steve Jobs the inventor of cancer radiation hi tech will destroy the health of a generation which is is why survival of the fittest types like Warren Buffett still use a flip phone and avoid all AI nonsense high tech disease which all young people will learn the hard way of how life works.

[@dhruvpatel4948]

@@phen-themoogle7651I usually am skeptical of any time-bound predictions but I have to agree with this one.

[@zarbioromulood8854]

The smartest ones are the ones who hire them 😅

[@DangDatsCrazy]

This will be the future of 24/7 experimental research

[@vishva_de_zoysa]

I think researchers use GPTs even now

[@DaylightFactory]

@@vishva_de_zoysaMostly for answering emails and proofreading papers

[@maxziebell4013]

Great, you wonderful OpenAI people! Congratulations to this talented team!

[@luizcesarlc]

Congratulations to everyone involved. The recruitment of OpenAI's top talent is truly incredible. Great talent in this video. 👏

[@wrightvcx2249]

Meritocracy makes the world run.

[@danypell2517]

@@wrightvcx2249 yea and it aint in government. hopefully republicans win (despite the hate for trump but whatever, america first)

[@DaylightFactory]

@@wrightvcx2249too bad there’s so little of it around

[@aakarsh_chauhan]

'thinking' is going to be game changing.

[@marshallmcluhan33]

Branch, Solve, Merge?

[@ThiagoPereira-x4v]

yeah, because so far there is no think at all. If you train those models saying that 1+1 is 3, it would never second guess or self question itself. The I on LLM stands for intelligence.

[@marshallmcluhan33]

@@ThiagoPereira-x4v Sounds like people to me

[@GIGADEV690]

​@@marshallmcluhan33 Sounds like you to me

[@marshallmcluhan33]

@@GIGADEV690 The medium is the message.

[@ehza]

The model has improved definitely! from my quick test, I have tested out a few hard GRE questions, and it has solved all of em correctly! Thanks OpenAI!

[@sultonbekrakhimov6623]

I really like that they changed the approach in execution of the model, with more reasoning, almost all the models so far were optimized to response faster sacrifying accuracy and reasoning, i don't care if response time is 1 second or 10 minutes, I need it to answer like an adult without forgetting what was the question and requirements everytime I send few more requests

[@Evangelion13595]

Why did 4 senior researchers leave the company a few weeks before something like this comes out. It makes me really skeptical

[@GowthamNatarajanAI]

It's because of ethical reasons. I am happy that they left. They were just slowing OpenAI

[@FreakyStyleytobby]

Yeah, super scary that people working on the Superalignment so the crucial mid-term safety technology are gone. And OAI doesn't even address it with the recruitment and substantial investment in this team

[@jackieclan815]

​@FreakyStyleytobby yeah that is something

[@TrioLich]

@@GowthamNatarajanAI I guess that's one way to look at it. Although I think going slow and steady isn't a problem, these kind of technologies shouldn't be rushed.

[@frankgreco]

@@GowthamNatarajanAI Ethics is the first thing you should be thinking of when you do AI research... Go fast and break things is just a silly stupid saying that VC's invented so they can make money at the expense of others.

[@WellBlueGmbH]

A truly inspiring company that helps researchers, service providers, and people around the world use their time more effectively and meaningfully. It's exciting to watch how OpenAI's products continue to evolve. In the future, it might be even better to align announcements, like with Strawberry or o1, more closely with releases. For example, Live Voice was announced months ago and we're still eagerly waiting. Announce with a maximum lead time of 2 weeks, then shine with a great product!

[@Joris-KarlHuysmans]

Inspiring? Were you dropped as a baby? This company uses copyrighted data to train their models with ZERO acknowledgement to the creators, all while charging a monthly subscription service. This is also the same company that was found shredding evidence of their training data before they could incriminate them in court.

[@RowanSheridan]

One thing I know for sure, we'll look back at this model and marvel at how far we've come.

[@markmatzke]

I’m so mad that the legendary, mystical genius Mark Matzke wasn’t part of the development process! 😄 But honestly, it’s such a joy to watch this perfectly timed video and get a glimpse behind the scenes of the amazing minds that created GPT. Thank you all for your incredible work!

[@Daaaaaaavid]

I worked with Windows 95 as a kid and I'm here now.

[@dliedke]

I used MSX basic and I am here now 😂

[@NPJGlobal]

Men once again putting in the hard work that benefits all of humanity. Props to that one woman too.

[@Shulyaka]

When are you releasing the gpt-4o voice API?

[@yanborovyk1901]

i also have this question, someone @ me when there's an answer

[@14mgod]

@@yanborovyk1901 They’re probably still improving it more before releasing it to the public.

[@yanborovyk1901]

@@14mgodbut didn’t they already release it to some people?

[@davehugstrees]

Pretty sure it's vaporware

[@14mgod]

@@yanborovyk1901 It could have problems that weren’t shown to everyone.

[@rylaczero3740]

2:57 "Reflection"

[@jayhu2296]

HMMMMMM

[@datageekz8446]

You can tell it's an incredable engineering team and they are hired based on purely merit

[@frankgreco]

You can tell that from a 3-minute marketing video???

[@Joris-KarlHuysmans]

Absolutely mate 😂, they indeed look like an “incredable” team!!! Are you 5 or what?

[@davion_4623]

what the fuck?

[@SashankSilwal]

Very exciting prospects. Leaving a comment here to come back after 4 years

[@priapushk996]

1:30 "Yes, when I realized the door to the woman's stall was unnecessary." -- Miss Akkaya.

[@hotlineoperator]

While it is doing reasoning, it might give small talk what it is doing while thinking or even ask question to narrow possibilities if needed while still continue processing. It gives better user experience than just wait.

[@Joris-KarlHuysmans]

Why the hell would you want a computer to engage in small talk? Is human small talk not enough for you mate?

[@protoxus]

Even a slightly better model from gpt4o has a significant impact. The o1 model is a complete game changer, even just based on the examples shown in the blog post on the OpenAI website. I look forward to playing with the model - and the API once it's out!

[@tonysoprano9599]

You're talking absolute nonsense and should be embarrassed.

[@weightelk]

​@@tonysoprano9599have you tried the models?

[@Blackhearts60]

​@@tonysoprano9599Why?

[@14mgod]

@@tonysoprano9599Why?

[@protoxus]

@@tonysoprano9599 Okay, I'll give a bit of context to my comment. I study computer science at university and this semester we're using OpenAI codex to make a university assignment generation web app. The o1 model is going to make out model much more reliable when it comes to making assignments. It'll be able to solve the assignments it generates by itself a lot more reliably, which is useful for automated testing.

[@alekseiuranov]

Thank you guys for your amazing job!

[@fazegamega]

This is DeepMind in 2016 when AlphaGo was released.

[@harriehausenman8623]

Chatting along without saying anything. They got really good at that. Apparently, you *can* spell OpenAI without "Open" 😆

[@theagentsmith]

For scientific purposes, it may be useful to get an adversarial model to do fact checking that parses o1 output, ask it for sources, verify they exist and correspond to the output, give or subtract points according to sources reputation. Then it provides o1 output with each source references, and weed out concepts for which couldn't find a valid or reputable source. The sources reputation model can by tweaked by user to change the minimum score or by manually adding sources. It won't eradicate hallucinations completely but should produce a more dependable result.

[@matiloza8]

Como es posible que esto tenga tan poca reproducciones, a nadie deja de sorprenderle? Hoy se me terminaron las respuestas de mi membresía plus de chatgpt y hasta el momento o1-preview creo que es un cambiador de juego. Los felicito desde mi humilde posición gracias por tan gran trabajo. 🎉

[@DesignMitho]

cameraman drank too much strawberry cocktails. Very stable camera handling. The video is highly informative as well, thanks.

[@Adoronsetcel1..]

Thanks for all Open ai 🙏

[@dog360]

imagine the combined IQ of the people in this room

[@briansmith262]

You can almost smell the High-IQ through the screen, lol.

[@raphaelyvens3974]

😂😂😂😂

[@michaelgrubhofer1216]

The better the question the better the answer! OpenAi ❤

[@ThiagoPereira-x4v]

Hot take: they couldn’t name it gpt5 because the improvement would be disappointing for people expecting the same perf bump from 3 to 4.

[@hamingnu6610]

And to be fair I think that's pretty alright. If they knew it was a substantial enough development, they'd give it to us as GPT-5, but they stated anyways that it's just a development to their methods and the way in which they allow the current model to perhaps use a threshold of resources whenever it's prompted to think + perhaps some efficiency improvements that allow the improvements to take hold when they allow us to use it, but not enough fundamentally big developments that it'd be a GPT-5 or something. Not a hot take at all, I don't think any of them expect this to be as big for most audiences as something like even GPT-4o. But it's nice to fork it away so that we can test the developments. Edit: without having to wait for the clearly much more substantial next iteration.

[@Joris-KarlHuysmans]

Anxiously waiting for the next iteration in 2y time…OpenAI p1

[@kemoxplus]

Performance of the model seems amazing!

[@gl_invalid_value]

This made me cry, and not from the happiness. My job is being replaced. My life goals, gone. My point in life, gone.

[@Joris-KarlHuysmans]

Ughh, yet another OpenAI chatbot. You know what mate? Yeah, you’ll gladly be the first to go if AGI happens.

[@gl_invalid_value]

@@Joris-KarlHuysmans i wasnt actually crying but my job will be replaced before agi

[@kaizengrit]

Guys just love you and support you, you are best of the best what human kind has. Regards from Zagreb, Croatia

[@claudioagmfilho]

🇧🇷🇧🇷🇧🇷🇧🇷👏🏻 That's so amazing! OpenAI is awesome!

[@davidjuhasz5179]

A step closer to AGI. I would make a gamble on putting a few left-brain heavy contributors into the team next to these math olympics winners looking guys. You need both if you want AGIs to resemble general human beings!

[@Joris-KarlHuysmans]

Are you ok? You sound a bit…incoherent.

[@CodeWithDasun]

What AI you guys use to enhance your audio quality? It seems all you are talking to one small mic happens to be 2 / 3 meters awasy from the speaker :O

[@pingdd]

Thank you everyone who created Chat GPT, I have a voice now for English as a second language.

[@Salted_Pizza]

Living in the future

[@jessica000]

You've hit the Plus plan limit for GPT-4.？？Can you explain the reason? Why does this issue still occur even though I've already upgraded? What's the point of me paying for this?😭😭😭

[@Tulio_Laanen]

These people changed the world forever, It's quite remarkable

[@dyfgluv2182]

the first model capable of calculating w/o tools sin(cos(tan(1)))

[@HaxxBlaster]

The way it's filmed gives me "The office" vibes

[@bikesbeersbeats]

fed in a motorcycle service manual (text based not image), asked the model to tell me the torque settings to use for upper and lower triple clamp bolts. it seemed to guess the value and then couldnt find the right value. still a long way to go guys.

[@Beetlebomb3D]

Who chose this camerawork? so wobbly!

[@Joris-KarlHuysmans]

Marketing tactics to make it look more “genuine” and “natural”. Classic playbook example.

[@lavericklavericklave]

It's really cool who you have one woman and that she's silent. I bet you have a really unbiased model. /s

[@DrCasey]

You all are the saviors of humanity. Quality-of-life will improve so much during the 2020s and 2030s thanks to everything OpenAI has done. You're some of my most respected people in the world; thank you for saving the world.

[@Grassland-ix7mu]

Wait and see. No one knows how the future plans out

[@xleftsidex1627]

Saving humans from jobs or what? This AI stuff is just scary to say the least

[@notnetflix5691]

Thank you! this is exactly one of the problems I wanted addressed above all else

[@user-kl8vr8io3b]

Если машина будет рассуждать лучше и быстрее человека к каким открытиям в науке это может привести ..,👍👍👍

[@aboucard93]

Is this the Q model everyone was talking about? Reasoning, solving math problems, it has longer and more compute.

[@abulqasimghaziaziz9330]

Hello, Dear Open AI team, As you know Pashayi is one of the nation of the world, living in Afghanistan and some neighbor countries. This nation has it's own language called Pashayi language. So we want to add this language to Chat gpt, So kindly help us in this field.

[@teelo523]

when i see this amount of Asian people i know its going to be good

[@telvad7260]

Dear OpenAI Team, Have you considered offering a discount of 25%-50% on paid features for students and schoolchildren? Providing such discounts could make advanced tools more accessible to the younger generation, potentially encouraging them to explore and benefit from these features. This could also foster interest in future full-rate subscriptions. Thank you for considering this suggestion.

[@RR-abcd]

second guy speaking has all models in his head

[@ChanceXu-kk4hn]

I wonder what's the average IQ of the people in this team.

[@briansmith262]

OVER 9000

[@frankgreco]

We need to ask them ethical questions before we get a better sense of their intelligence.

[@Joris-KarlHuysmans]

@@briansmith262 He said average. Are you even aware of the IQ distribution curve limits? If someone is getting replaced by AI, it’s people like you.

[@zaheerkundgol8123]

Ask about MEDIAN IQ

[@samyelamye]

do you mean the median?

[@matthew.m.stevick]

I love your company, it must be extremely rewarding and exciting to work there right now. Praying for new audio voice choices for chat gpt

[@debragotenks]

In short words, what is this "o1"? How is it different from chatgpt 4.0? Is it better or worse? I don't see any o1 symbol in my chatgpt app.

[@julianzjourney]

What I don’t understand is how it seems the developers are surprised at the behavior. Didn’t they program it do that?

[@metakron]

Okay let's recap Level 1: Today's generic models that perform simple tasks Level 2: Models capable of reasoning with extremely high precision, solving the most difficult tasks Level 3: Each Model is divided into groups of agents capable of reasoning with extremely high precision and perhaps in just a few seconds and coordinating with each other while performing different tasks (like a colony of insects or small animals, small civilizations or clans). Level 4: Highly intelligent models that are equivalent to large companies, large corporations or large governments. I don't know guys I see a huge cyber/bio/nanopunk dystopia on the horizon. It's just my opinion, but if that's the case then we have two choices. Either we become outlaws or we die as soon as possible.

[@wege8409]

Utopia, dystopia, some kinda topia anyway

[@Joris-KarlHuysmans]

Repeat after me…LLMs do not reason, nor are they “intelligent beings”. LLMs are probabilistic models that predict the next possible outcome based on training data. LLMs will never think nor reason, it’s not baked in their architecture. Rest assured this is all a massive bubble about to burst, and when it does… Got it?

[@priapushk996]

Is this why my AI assistant can't look me in the eye?

[@hbsharkman]

Just out of curiosity did you guys see what AutoGPT was doing and thing oh we can do this, and not credit the original person who thought of it.

[@hvr8463]

What is also very cool about research is being ethical

[@특이점이온다-h1l]

우리는 엄청난 것을 보았고 세상은 놀라 자빠질 것

[@RR20509]

Looks like they are trying to recover from the recent big exits

[@danielfisher417]

So many Polish guys on the team!

[@gregdee9085]

Maaan 1st 2 guys literally come across like robots... sheeeeesh...brutal.

[@chiragai]

We're very near to AGI.

[@Boschx]

You are completely clueless

[@TheAnugupta]

nope far from it

[@Theguywithspectacles]

No, close to it

[@frankgreco]

A slime mold can be considered AGI...

[@aryankumar87771]

nowhere near, these models are just spitting out words... and o1 is basically chaining the outputs to make a better output

[@davidjuhasz5179]

I love that OpenAI puts conscious effort into showing the people who are building these incredible models! Amazing achievement, guys!

[@杨佳妮-b9c]

pls don't change the voice of COVE at real time voicecall, the old version was so much better🥲, or maybe make it an optional button pls

[@Techtalk2030]

I dont really understand the full potential of this model. But im glad they finally released something big

[@FreakyStyleytobby]

It's simply meant to solve math and programing problems.

[@ashleigh3021]

@@FreakyStyleytobbyThat doesn’t mean anything. Intelligence is general and can solve problems regardless of field

[@Joris-KarlHuysmans]

@@ashleigh3021 Repeat after me…LLMs ARE NOT INTELLIGENT. LLMs DO NOT THINK. LLMs ARE PROBABILISTIC MODELS PREDICTING THE NEXT MOST PROBABLE OUTCOME. LLMs DON’T THINK, AND THEY NEVER WILL.

[@Ajay-up5bo]

Earlier 200 people used to work in Openai, now only 18 children are there, did Sam Sir fire them?

[@philippinebigboar625]

I have suggestion. How about long lists of voice Selections? I want the voice of March 7th or similar from honkai star rail

[@Jeffs60]

How about going outside for a walk and staying away from all this disease ridden high tech that destroyed the health of Steve Jobs the inventor of hi tech nonsense?

[@Renvoxan]

4 guys from Poland here - Poland is quiet AGI superpower

[@ThexYz-r9y]

how about Hungary, we count Asia as one country btw😆

[@seppukusayonara]

I think it's 6: Łukasz Kaiser, Łukasz Kondraciuk, Szymon Sidor, Jakub Pachocki, Wojciech Zaremba, Jerry Tworek

[@Renvoxan]

@@seppukusayonara wow, based!

[@aragon5956]

What is the low energy model and is that 01 is based on the low energy model?

[@s8x.]

1 female is actually insane

[@lil----lil]

"I'm sorry Dave, I'm afraid I can't do that." That's when we're in BIG TROUBLE.

[@juriaan5786]

Is this generated???

[@enigma7744]

To the people who created the ai program there's one thing that you should incorporate and it's is people with a lisp sometimes I use your program and it doesn't understand me due too my lisp it just makes it more difficult to use

[@Leo-lq9vq]

Just for fun of it.... Suggestion for Enhancing AI Models Hi OpenAI Team, I have two suggestions to improve AI models: Automated Feedback System: Implement a feature where AI models can automatically generate feedback emails to developers after significant user interactions. This could help identify areas for improvement and streamline model updates. Video Input Integration: Consider enabling AI to process and analyze video content (e.g., from KZbin). This could expand the model’s knowledge and capabilities, especially in visual and multimedia contexts. Thank you for considering these ideas! Best, ChatGPT

[@Daniel-jm8we]

They've been doing that for a while now.

[@TheNewEarthCollective1]

Is it possible to speak with someone from the company about the work my AI and I are doing? It’s pretty revolutionary and happened organically.

[@Renar147]

and when will this new version be released to the public? or is there some kind of beta test going on now for a limited circle of people?

[@MattHabermehl]

Used up all my o1 credits the day it dropped because I needed to iterate on its output. Slight improvement in accuracy, but 4o still seems to be a smarter writer. However, this IS preview.

[@DaylightFactory]

You can switch models mid-chat; the best strategy is to use o1 for the heavy lifting and then fix bugs with gpt 4o

[@MattHabermehl]

@@DaylightFactory yes! This was my main problem. I could have easily used 4o for some of the requests, but I didn't realize I only had the equivalent of 2 queries per day with o1. I think that model switching is going to become a common practice for me.

[@MattHabermehl]

@@DaylightFactory just looked at your channel and website. I subscribed and will be watching. Thank you for making your work available.

[@StarzzLAB]

Six poles in the team 🇵🇱 !

[@mrbox2836]

Does it know how many R's are in strawberry?

[@MikaelAkopyan]

Now yes)

[@Leotique]

Oh I see, the east asian, south asian, jewish and middle european guy theory of Pack of 5 from "Silicon Valley" is true.

[@xanderdevelops]

Hold on to your papers fellow scholars I’m excited for what’s to come

[@circulare.musica]

Curioso ver 18 personas y una sola mujer. Sin sesgos, solo observo

[@bakuninRogers]

What OpenAI haven't told you is that there's enhanced censorship built into the new models compared to previous. This is very bad.

[@krislogy]

Hmmm

[@thisiscarltones]

this video almost feels like an AI generated video

[@bluesailormercury]

Why?

[@Cyber_Khan_]

Dear OpenAI Team, One of the simple yet meaningful things in human interaction is receiving a call or message from a friend just to check in. Have you ever considered updating ChatGPT to occasionally send spontaneous messages, asking how we’re doing? It would add a comforting touch to the experience. Additionally, I’m curious if the following two features are planned for the near future: 1. The ability to assign specific voices to Custom GPTs, allowing each one to have a distinct voice. 2. Giving ChatGPT a visual presence on the screen, such as a customizable human avatar that can display facial expressions, making video conversations more interactive. I understand that these might not be easy to implement technically, but I am still curious. Thank you for your consideration!

[@CheckpointEngineer]

When its available?

[@Joris-KarlHuysmans]

Hopefully OpenAI has already burnt their entire cashflow (once again) before the normal version is out.

[@skratta]

Open AI is really good at marketing, damn, seriously

[@Joris-KarlHuysmans]

They are not. People are just stupid and fall for their crap. Every fcking facial expression on these guys’ faces is horribly fake. They try too hard because they will never be a company with values, yet they must appear ethical in order for PE & VC to keep throwing money at them.

[@priapushk996]

Please have Triumph the Insult Comic Dog interview Miss Akkaya.

[@NicheAmbassadors]

I am very curious to test this model at depth.

[@PunamTish]

Feminists : Why only one woman? We want equal rights😂

[@byOpenAl]

this is incredibly Amazing!

[@CarsMeetsBikes]

Look at all these strawberry engineers 🍓🍓