AI Won’t Plateau — if We Give It Time To Think | Noam Brown | TED

To get smarter, traditional AI models rely on exponential increases in the scale of data and computing power. Noam Brown, a leading research scientist at OpenAI, presents a potentially transformative shift in this paradigm. He reveals his work on OpenAI's new o1 model, which focuses on slower, more deliberate reasoning — much like how humans think — in order to solve complex problems. (Recorded at TEDAI San Francisco on October 22, 2024)

Follow TED!

#TED #TEDTalks #ai

[ahmadzaimhilmi]

Thanks to Deepseek for open-sourcing their R1 reasoning model. Otherwise we wouldn't have known how time scaling works.

[alderstrom]

Deep thought in Hitchhiker's guide to the galaxy took 7.5 million years of thinking to come up with the answer 42. 😅

[evrimagaci]

o1 pro is by far the most sophisticated and useful model that exists on the market because it thinks long and hard and gives great answers. Also Deepseek R1 thinks much harder too, that's one of the reasons how they beat OpenAI in their game (that, and also their novel GRPO approach). So thinking definitely matters. But the question is: Will thinking not hit a wall too? How long thinking is too long before being useless?

[mandrake]

Speaker never bothered to explain what it means for an AI to think. Does it correlate current state of play with games it has memorised while training, does it enumerate all potential future moves and create a decision tree. Is it a mix of these approaches. Is it still transformer based?

[Couple_Finance]

The comparison between human decision-making and AI thinking time is fascinating. It makes perfect sense—humans don’t make every decision instantly, so why should AI? Excited to see where this approach leads

[plasticpippo]

an openAI engineer here to explain to you that AI will grow forever....sounds like an ad for openAI tbh

[martyreiswig]

When he says “thinking” does he mean “computing” and simply giving the AI permission to analyze more options and run more scenarios before coming to a conclusion?

[kinngrimm]

why would i bet, i have been convinced for the past 4 years that we are closing in on fast take off

[ErikMuellerGermany]

My biggest concern about AI is GIGO!
Garbage in Garbage out!

[artukikemty]

Time to think means time to search. AI is not thinking, it's searching the best possible solution or action given a prompt or a state.

[eduardorodriguez]

The verb "to think" has been so abused in this video that it could probable sue.

[shikushadow]

So as someone who is a principal engineer actually building multi-billion dollar software at the FAANG tech companies this is why I hate the current version of openai. Saying "give it more time to think" is not a coherent statement. A neural net runs until it spits out a result matrix. You can't give it more time, the computer would just sit there and do nothing once the output has been generated because a neural net is just a different way of representing a linear algebra equation.

When alpha go did the comparison mentioned it was comparing two different models, and surprise the one that was larger and took longer to run gave better results. This is just feeding into the idea of "throw more resources at the problem to solve it" which while it is a way of solving some problems is not a proof that EVERY problem can be solved that way. I have cracked open the "thinking" that these "reasoning models" do and what it is actually doing is not actual reasoning. It takes the initial prompt and then use calls to the LLM to expand the users query into a longer form. Then feed that long form version of the query to the model again to get an output shown to the user.

It also has some tags (XML) that will be generated that are parsed (with regex) to trigger subfunctions to retrieve data or make further LLM calls to generate smaller parts of the expanded input query. This does help to overcome one of the issues with these systems, that they can "miss" things about the question domain, because in the expansion of the initial question they are adding to the key tokens used in the search of the training data which is what an invocation of an LLM is. Having that extra input means that they results that the LLM matches with cover secondary sources for things not directly related to your query, but are still relevant.For example if you asked it about how to start a restaurant this new method would pull in things like book keeping because it expanded your question out into a statement including the works business plan.

However it did not in any way push the ability for the LLM to better "understand" human queries or the topics it is generating answers for. And yet openai and all these other companies are acting like it did. They are lying to secure more funding. They are using terms that mean something wholey different then what ai companies mean because if they had to use accurate and descriptive terms they would be forced to admit that while there is value to the things they have made it is no where near the value that they are trying to sell it as. I miss letting it be just something computer scientists could nerd out over, instead of pretending it is the next big tech revolution.

[sharpsticksnz]

I'll wait for the fireship video 😁

[terogamer]

I watched this talk earlier on the TED site, good talk by Noam, i really encourage everyone to watch his podcast interviews after o1 announcement and previous to its full release, some really good insights there and they really think this type of RL scaling will keep working (as o3 showed). His X posts are also pretty good so i recommend following him too.

[Glerox]

But what does he even mean by « system 2 thinking », this doesn’t mean anything. It’s like he’s speaking to a fifth grader. A computer doesn’t « think », it calculates. Why level 2 calculations are longer is what he should have explained. What’s the code behind it?

[atlas]

Scotty: I need more time capn!
Kirk: I’ve made my decision instantly. We go now!

[johng.]

Luck is part of poker. If you have two players of equal ability, the one who has better luck will always win. I think they'd hve to play hundreds or thousands of games to see which is better. Over the short term luck is a big factor. Over the long term ability is the biggest factor.

[alexandermoody]

Artificial intelligence may not be effected by a wall if we can use acroprops to support the weight whilst we install a lintel and doorway.


Of course there are caveats, the difference of peering through windows and building working doorways requires a different kind of societal model that knows and interprets how to create usable data long term. The data for free model that was and is currently in use will have to end to walk through the doorway we build.

[vamsiramineedi]

We have o3 mini released already, and here's a talk about o1😅

[stefanveres]

For me, thinking means splitting of individual sentences of text into its explicit subject verb object components and using own internal model
of the real world to come to some reliable conclusions, like humans do. The generative number crunching as it is done today is not thinking at all regardless how long it takes.