The World Reacts to OpenAI's Unveiling of o3!

o3 has STUNNED and SHOCKED the industry! (lol) Let's take a look at how the brightest minds in AI reacted to its release.

Coupon code: BERMAN300

Join My Newsletter for Regular AI Updates 👇🏼

My Links 🔗

Media/Sponsorship Inquiries ✅

Links:

[alex-rsts]

We went from "It's impossible" to "it's too expensive" quite fast. How long until "It's not worth the effort"?

[ttcc]

Capitalists from VC firms are so “impressed”… wanting us to invest in their ventures.

[TheNonamedbandit]

OpenAi in particular have a consistent pattern of making PR releases, then it taking way longer than they said, and under performing. They always seem prepped to try and steal Google's thunder, make a crazy announcement that's about 4-6 months before the public get to use it. Think about voice mode - it did amazing things in the demo, then was fully nerfed and a relatively "stupid" model. The SORA PR cycle was exactly the same. I don't trust them to be honest about any release until we actually get to use it.

[henrytuttle]

The most impressive thing about the Frontier Math test was that a Field's Medal winner said he thought no human could do what o3 did PERIOD. Not about speed or anything but that no one human could do it.

But solving math problems better than experts is not AGI. Chess computers were able to beat the top chess players long ago. Math is not Chess, but it's the about specialization. Not being able to do things that five year olds can do means that it's not "general". When it's able to do almost everything (maybe there might have to be an exception or two) that humans can do at a minimal level, it'll have reached "general" intelligence while simultaneously being much better at certain tasks, like most humans.

But I think the $0.10 per task is unreasonable. General intelligence at ANY PRICE is impressive.

[kairi]

I'll believe in agi when "shocked" and "stunned" are no longer the YouTube clickbait titles, and we have more impressive grammar.

[quaterman]

Everything based on benchmarks 😂 Has 40% credibility for me.

[wisdomking]

I love this new format, were you synthesize all the comments from influential people in the AI industry on a given breakthrough (o3 in this case).

Do more of these in the future, we can assume there would be more crazier breakthroughs from hereon.

[PedroPenhaVerani-llwc]

The hype and shock about o3 were so intense that a YouTuber posted claiming that o3 was AGI and even used Stockfish in chess as an example! Crazy stuff!

[TuxedoPanther]

o3 is a step forward, but it also sounds like the amount of energy and resources has drastically increased too. The human brain runs on a small amount of power and it is truly intelligent and adaptive. As you say, o3 gets stumped by questions a five year old can answer. Hopefully it will find some good uses in science.

[rexmundi]

I’ve been having a pretty in depth conversation with Claude on alien intelligence and the nature of consciousness. Like way more intense than any friend or family member would want to have. It has "read" and "seen" the books and movies I reference and can reference works I didn’t think of or haven’t read. It’s really amazing and surprisingly I find I look forward to "talking" to it. I can see AI human "relationships " will be deeply meaningful to people. Like "Her"

[andriystruk]

The ARC-AGI task that o3 failed on was actually ambiguous. If you look at how it answered you would have realized that both answers it provided were reasonable and a response that many people would have had as well.

[GingerDrums]

The depth of the alignment problem needs even more attention. A machine able to solve these complex problems will be able to manipulate the worlds top psychologists like putty. We are in deep, deep trouble very soon if things continue to heat up so fast.

[lucid_dream_lier]

The ARC problem you described, specifically the one where the O3 solution was incorrect, is actually a question that most people get wrong. The red block directly above needs to turn blue in the standard solution because the assumed logic is that contact with the blue line changes the color to blue, rather than being based on overlapping. However, the first three examples don't demonstrate this contact behavior, so even ARC's own solution remains controversial.

[ffs]

Still waiting for one of these able to do a non-trivial regex

[MaxYoutubeWhatever]

I worked with advanced voice mode and when compensating for it's VISION DISABILITY it was able to solve the blue/red block puzzles perfectly and predict the output panels. It was as I suspected, the model is simply blind or vision-impaired based on imperfect conversion of visual data to words/concepts/description.

[desmond-hawkins]

It's clear from the examples of "trivial" failed tasks in ARC-AGI that what's missing from o3 is a *physical representation* of the world. For now it's mostly been based on concepts and how they relate to each other, this was particularly evident with word2vec for example. But the models don't really understand "I need to _paint_ these squares that are stacked _on top_ of each other in the order given by the small color band". We know how to do this because we live in a 3D world, so this feels kind of obvious. Once models start having a sense of what it means to be a physical being or what we experience interacting with our surroundings, then they'll make another huge leap. Unsurprisingly they've been really struggling with this entire physical aspect… for now.

[merricmercer]

If AGI is achieved but it can be expensive and too resource intensive then presumably the first task is to ask 03 how it can improve and allow it to recursively grow quickly to ASI - artificial super intelligence.

[daomingjin]

a coding competition is not a good test for generative AI. Keep in mind this is not AI... it is Generative AI (sentence patterns). if you want to train an AI to code it's easy of you have access to a large database of coding challenge questions and answers. Just iterate through everything and build a model. It's generative AI, so basically.... it can generate code but debugging will always be a problem as generative AI can not reason.

[garethsmith]

We live in a world awash with answers, but it is asking the right question that becomes the real skill

[GetzAI]

1:25 25% at HIGH computation. That dark blue is probably what most will have. Will Pro users get the compute needed to achieve 25%?