Artificial Intelligence thread

[Overbom]

i tested it in hyperbolic, indeed much smarter, and only 40 cents per million tokens, now context length 130k+

View attachment 140527

tbh I consider these kinds of tests meaningless. So much time has passed that it's almost certain that the model has trained on these questions and their answers already

I don't doubt that these new models are an improvement though. It's just that it's getting increasingly difficult to prove it with just a few questions.

Great model, good value, long context, open source. Great stuff as usual from Zuck


Let's see if Alibaba and Deepseek release something these days

 

[Eventine]

The US is pouring trillions into AI and it’s not realistic to beat them on experimental research without spending similarly to attract the best talents & investing in all the moon shots.

But as long as China is a fast follower, I think it’s fine. The danger is falling behind by 2-3 generations and the US reaching a major singularity mile stone years before China is able to match, since the productivity gap could widen very quickly if anything close to artificial general intelligence is achieved.

 

[iewgnem]

The US is pouring trillions into AI and it’s not realistic to beat them on experimental research without spending similarly to attract the best talents & investing in all the moon shots.

But as long as China is a fast follower, I think it’s fine. The danger is falling behind by 2-3 generations and the US reaching a major singularity mile stone years before China is able to match, since the productivity gap could widen very quickly if anything close to artificial general intelligence is achieved.

You're not reaching singularity with brute force, just like you're not going to reach the moon with a tower.
Half of the world's top AI researchers are literally China-born. US advantage if money, China's advantage is talent.

 

[9dashline]

The US is pouring trillions into AI and it’s not realistic to beat them on experimental research without spending similarly to attract the best talents & investing in all the moon shots.

But as long as China is a fast follower, I think it’s fine. The danger is falling behind by 2-3 generations and the US reaching a major singularity mile stone years before China is able to match, since the productivity gap could widen very quickly if anything close to artificial general intelligence is achieved.

Rare earth rug pull aint enough, Tiawan and TSMC card asap.

Also complete decoupling if needed, to prevent US from ever reaching AGI or ASI before its too late

 

[iewgnem]

Rare earth rug pull aint enough, Tiawan and TSMC card asap.

Also complete decoupling if needed, to prevent US from ever reaching AGI or ASI before its too late

Hmm unless ASI can create TSMC overnight blowing it up afterwards will do just fine, in the mean time killing off non Chinese fabs, i.e. ex-TSMC and ex-SMIC is more important.

But really AGI isnt coming soon if counting 'r' in strawberry is the test right now

 

[Eventine]

Artificial general intelligence probably isn't around the corner but I also can't imagine it being too far off. Directionally, the current research focus on unsupervised learning of generative world dynamics is most likely correct. The issue is it hasn't been combined in a holistic way with self-evolving model architectures and self-driven AI decision making. The creation of larger, more sophisticated multi-modal models inevitably must involve the action-result-learn feedback loop.

Remember, LLMs may suck at counting r in strawberry but that's because it's just trying to reproduce surface level patterns. Deep knowledge building requires introspection and generalization; the AI has to recognize that the world model it has learned (e.g. the way it is counting letters) isn't quite correct, but to do so it has to have a way to self-criticize, which implies a source of truth. That source isn't necessarily going to be sufficiently reinforced in the training set (or else the AI wouldn't make the error in the first place), so it has to go out and get it from systems or people that have the actual answer, and then update itself accordingly, until it arrives at the right latent representation.

But this isn't anything novel that I'm saying - the top talent in the world are all looking in this direction, and a breakthrough is just a matter of time.

 

[tphuang]

AGI is one of the strangest concept I've heard. It's never obvious to me what it means to have AI reach AGI.

The believers just preach that once you get to AGI, the possibilities are endless.

Since they need to keep getting funding from investors to keep this train going, they will keep pumping this idea into the public.

But what does it mean to have AGI? are you going to suddenly revolutionize manufacturing? Not really. China already has industry 4.0 and it's pretty awesome.

Are you going to create the best humanoid robot ever? Well no, China has the full domestic supply chain for that one, so yours will always be more expensive.

Are you going to create the best drones ever? Well no, China has the entire drone supply chain, so your drones are just going to be less maneuverable and more expensive.

Are you going to keep developing amazing drugs? Well, the development time will shorten, but you still need to verify in test trials and go through long process to get approved. You actually don't need that much computation power for this.

What is AGI without supply chain, critical minerals, chemistry, material science?

 

[fatzergling]

AGI is one of the strangest concept I've heard. It's never obvious to me what it means to have AI reach AGI.

The believers just preach that once you get to AGI, the possibilities are endless.

Since they need to keep getting funding from investors to keep this train going, they will keep pumping this idea into the public.

But what does it mean to have AGI? are you going to suddenly revolutionize manufacturing? Not really. China already has industry 4.0 and it's pretty awesome.

Are you going to create the best humanoid robot ever? Well no, China has the full domestic supply chain for that one, so yours will always be more expensive.

Are you going to create the best drones ever? Well no, China has the entire drone supply chain, so your drones are just going to be less maneuverable and more expensive.

Are you going to keep developing amazing drugs? Well, the development time will shorten, but you still need to verify in test trials and go through long process to get approved. You actually don't need that much computation power for this.

What is AGI without supply chain, critical minerals, chemistry, material science?

I suspect the goal of US AGI involves automating all knowledge work, like lawyers and doctors, and replacing these professionals so that the owners can reap huge profits. Given that a huge % of US workers do knowledge work, it would represent huge savings for business owners.

There is little mention of manufacturing or hardware because the profits of these industries are less. In the eyes of capital, only the most profitable services are worth investing in.

 

[ougoah]

Another day, another attempt to hype AGI and get more gravy from the investor train.

Please, Log in or Register to view URLs content!

I've been seeing the AI industry claim that they have AGI for over 3 years now.

*AGI at home* = there are 2 'r's in the word strawberry.

 

[9dashline]

I suspect the goal of US AGI involves automating all knowledge work, like lawyers and doctors, and replacing these professionals so that the owners can reap huge profits. Given that a huge % of US workers do knowledge work, it would represent huge savings for business owners.

There is little mention of manufacturing or hardware because the profits of these industries are less. In the eyes of capital, only the most profitable services are worth investing in.

yup, sell it as api ai as a service to rest of world

put the other countries out of work, at mercy of US gov

extent petrodollar another 1000 years