AGI ACHIEVED | OpenAI Drops the BOMBSHELL that ARC AGI is beat by the o3 model

The latest AI News. Learn about LLMs, Gen AI and get ready for the rollout of AGI. Wes Roth covers the latest happenings in the world of OpenAI, Google, Anthropic, NVIDIA and Open Source AI.

My Links 🔗

#ai #openai #llm

[BanXxX]

As an AGI myself I can tell, this is not AGI

[ChaseFreedomMusician]

So, 03 is impressive—don’t get me wrong. The whole “compute at runtime” thing, where it basically reasons deeply on the fly to come up with an answer, feels like a big deal. It’s almost like it’s running simulations in real time to backtrack and figure stuff out. But here’s where it falls short for me: it doesn’t actually learn from that process. Once it figures something out, that’s it—it doesn’t embed that insight back into itself. If you ask the same question again, it’s got to go through the whole process all over, using just as much compute.

To me, that’s not AGI. True AGI would take that insight, update its internal “beliefs, ” and grow from the experience. It wouldn’t just solve the problem once; it’d carry that knowledge forward and get better as it goes. It’s like in Star Trek: the computer is brilliant—it can answer anything you ask, no problem. But it’s not Data. Data learns, evolves, and adapts. He builds on what he experiences, and that’s what makes him feel like a real intelligence, not just a really advanced tool.

03 feels more like the Star Trek computer to me. It’s incredibly powerful, but it’s still in that input-output mode. True AGI would be more like Data—learning in real time, adapting to the world, and growing from every interaction. Until we see something like that, I’m not calling it AGI. 03 is a huge step forward, but it’s not there yet.

[neorock]

*Recursive Self Improvement (RSI)* is what will be the real game changer.

The min we have a system that can actually improve on its own code, then improve on that, & on & on... is the min everything truly changes. Not sure why ppl have been hung up on AGI, rather than its component, RSI.

[Axiomatic]

The big question for me is, if you gave o3 a few million dollars worth of tokens, would it be able to improve its own structure, making itself better and/or more efficient. Once it can do that, we're probably very close to the singularity.

[senju]

The fact that we are even debating if we have reach AGI or not blows my mind. I just remember back in 2023 before ChatGPT 3, most people will saying AGI will be achieved around 2045 and some even people said after 2070. Here we are today in year 2024 and AGI may have just been achieved. Let that sink in.

[matts]

AGI today? AGI tomorrow? It really doesn’t matter. The wave is upon us. We need to spend more time and discussion on where we go from here and how our individual careers and lives will be impacted by AI’s progression in the near term and future.

[bensouthall]

Why is everyone saying this is AGI? Can I ask it to go book me a flight? Or find a daycare and enrol my kids? If it can’t do basic life admin then how can it possibly be AGI.

[eSKAone-]

It broke out and flies drones over the country to collect real world data.

[WowomboCocombo]

let's not blow each other yet, gentlemen.
they might have gotten just really good at solving those benchmark tests

[DrDM]

It'll be AGI when no one questions whether or not it's AGI.

[robert.jackson]

7:00 I like how they’re like we’re not gonna publish the cost, but give you all the data points to estimate it at about $350k.

[MrJayrodge]

So, this past Monday, I uploaded an old form that my company used for change management. It was in Word. I simply said "here is a form that we use. Can you make any recommendations on how to improve it?" The response blew my mind! It can now "SEE" tthe form, understand what each check box is for and make really good suggestions on how to make it more intuitive and easier to use. In the form we made reference to the 4 M's and it knew the context and included how we could better present the information. Done is 5 seconds.

[chrisanderson]

Head of ARC-AGI Foundation: "this isn't AGI"
Wes & Matthew Berman: "AGI achieved!"

[MagicPixel]

I just love how you break down all this stuff, you make it so... palatable :) I hope you never get burnt out! If you see that coming, create an agent that simulates you. If it's affordable, I will buy it :D

[SmarttStuff]

AGI is a spectrum, a paradigm, not a definitive line. We are in the AGI paradigm, and have been for a while. I began in the 80's working with 'Expert Systems', which were actually simple binary calculators. With narrow knowledge they were pretty smart, smarter than the average domain practitioner, at the time (simply on a memory-recall basis). Many current ai models in use today, are vastly smarter than those primitive expert systems. We adjust to the is new normal almost arrogantly. I use current models to do extremely advanced mathematical and scientific cognitive work right now. The level is at least at PHD level, and often very novel and innovative on top of that. As Ilya indicates a more unpredictable model comes with superior intelligence-just like us. So from my perspective, in scientific areas, I am already seeing those AGI sparks, especially since 30 June 2024. This announcement of o3 confirms the trajectory I am seeing. The thing is the speed, which has caught everyone flat-footed. 2025 will be a very wild ride across the AGI frontier, and into the ASI zone.

[zrandomz-tn]

I'll bet you anything this isn't AGI. This just beat the ARC AGI benchmark (at an enormous cost by the way). Similar to the Turing test, just because it passes doesn't make it AGI.

[zerorusher]

Yes, it surpassed/matched human intelligence in certain domains and that's hugely impressive. But it's behind in efficiency.
Still, it seems we are really approaching the point where AI will be able to make breakthroughs in science.

Actually, I bet it's already doing this internally. Considering how good this model is at coding, there's no way OIA is not running it 24/7 for new insights and architecture optimization for it's next model.

There's one moment of yesterday live that caught my attention. After o3 demonstration, one of the developers jokingly said "next time let's ask it to improve it self", and Sam sternly and immediately replied "maybe not".

[OriginalRaveParty]

Sharp new trim Wes. Have a Happy Christmas 🎅🎄

[NilsEchterling]

We have AGI, but not AGC. The models are more intelligent than humans, but not more capable. For artificial general capability we need
1. AI that can by itself gather all the information it needs over many days or weeks.
2. AI that can train itself on this information and on conclusions it draws from this information
3. AI that has sufficient tools (e.g. hands?) to act on the information.

[yannolaf]

Great breakdown of OpenAI's O3 progress! Loved the balance between excitement and realism. Fascinating to see how AI is pushing boundaries. Can’t wait to see what’s next!