Why The Next AI Breakthroughs Will Be In Reasoning, Not Scaling

[@chapterme]

Chapters (Powered by ChapterMe) - 0:00 Intro 1:15 The intelligence age 4:18 YC o1 hackathon 12:09 4 orders of magnitude 14:42 The architecture of o1 21:52 Getting that final 10-15% of accuracy 32:06 The companies/ideas that should pivot because of o1 34:44 Outro

[@scienceinc.9453]

@chapterme great marketing

[@chapterme]

@ thank you 😊

[@mphomuila9460]

brilliant

[@winddude9]

are they not allowed to talk about claude or something?

[@brucebain7340]

They mentioned claude several times in the last episode genius

[@vvmm3712]

​@@brucebain7340Yeah they did but I think they are quite biased towards OpenAI. I have used premium versions of Gemini, OpenAI and Claude. Claude is still neck to neck, if not ahead of OpenAI. Also YC needs to be more objective in their POV. I think it happens because most of them talk to founders and read articles etc. rather than using the models extensively themselves to build stuff. Makes a lot of difference in how one perceives technology.

[@brucebain7340]

Of course they are biased. They're human. Whoever expects them to be impartial will be disappointed.

[@vvmm3712]

​@@brucebain7340The bias is not about what to order for dinner. The bias is that company 'A' is going to unlock AGI, Company 'A's models are going to be orders of magnitude better because Mr. Sam told us so. This bias can lead to misjudgements on YC's part which can impact its own investments which run into millions. I mean won't it be sad that someone can simply walk into your office and skew your own vision of something.

[@vvmm3712]

​@@brucebain7340 The bias is not about what to order for dinner. The bias is also not from a colleague of mine who thinks that Apple built the Apple Intelligence(AI)! The bias is from a group of people running one of the topmost accelerators in the world. The bias is that company 'A' is going to unlock AGI, Company 'A's models are going to be orders of magnitude better because Mr. S told us so. This bias can lead to misjudgements on YC's part which can impact its own investment which runs into millions.

[@counterfeit25]

4:30 Diode Computers is doing PCB (printed circuit board) design, not chip design (like NVIDIA)

[@ZevUhuru]

LOL. What is Gary's obsession with 'Raw Dogging', he's said this on multiple videos. Bro, stay safe out there 😂

[@FrAnCisF00qh]

what does raw dogging actually mean?

[@jojoabing1]

@@FrAnCisF00qh It means to eat hotdogs without cooking them. So, it is slang for raw hotdog. It basically comes from when you just go for something, because you don't feel like waiting. For example, you can't wait for your hotdog to cook, because you are so hungry, so you just eat it raw. It is considered dangerous due to the potential for food poisoning.

[@obscurerides]

@@jojoabing1this is gold

[@Muftobration]

​@@jojoabing1 so good, like a hotdog straight from the package.

[@Gabaldegook]

@@jojoabing1I pray this is a serious response

[@cherubin7th]

Kind of misleading. We already used algorithms to design the chips. Classical algorithms can approximate the optimal rout very well. AI looks like just being the missing peace to read the supplier's documents.

[@GigaFro]

Same thoughts.

[@princee9385]

Given all these developments, I need a remote job. PhD since 2022.

[@GNARGNARHEAD]

those hackathon results were really impressive 🤯

[@soyhenryxyz]

"If an LLM task is hallucinating, it’s likely doing too much. Break it down into steps"

[@pin65371]

The mistake people make when using these tools is they are not specific enough. Treat is like a human or team of humans and your results improve.

[@ccash3290]

​@@pin65371 Treat it like a child; Adults are smart and can understand ambiguity and ask for clarity.

[@cmw3737]

It usually means that there are missing parameters that it has to guess. Instead it should understand ambiguity and be more interactive and say "I don't know" or "it depends" and ask for what it needs to answer the request.

[@varunvummadi350]

Thanks a lot for shout out and adapting us Garry 😂

[@tomlevpt1]

It seems like there's no room left for startup companies to invent other than marketing the LLM's outout to customers

[@yadniksable]

Advice to everyone building a product. Ai is for scaling your solution your insight in a particular domain. So work on insight building and use ai model to scale that solution to solve the problem at a deeper level for target user.

[@cherubin7th]

Also have a good service that your customers are loyal. So you can stay in business even when you are only catching up. And real estate.

[@hongyihuang3560]

I do not believe that AI will be able to do chip design better than humans. PCBs maybe, chips no. It took the brightest people around the globe (literally any good chips you see today touches at least decades of research from Japan, Europe, US IP). There are so many specialties in chip design including testing, production ready, simulation, RF, documentation, security, compiler that’s beyond the capabilities of a closed loop LLM. If LLMs PCB design capability transfers to SoCs, China would have already made and beat Apple & NVIDIA. The first half is absolutely make believe, solve physics, nuclear power, climate… seriously? AI will solve societal issues? I find it hard to believe that next gen data centers for AI will need nuclear power. The second half I agree more on the real progress of AI: do tests, create a moat by building agents and accumulate proprietary data. Speculation is dangerous, I hope people can think for themselves.

[@brianmi40]

LOL, literally the history of humanity is doing things someone once thought "impossible". The problem is that you never consider the prerequisite to your statement: when will progress HALT in the progression of AI to where ZERO FUTURE PROGRESS will EVER occur. The answer is closer to NEVER than to SOON, so your statement is blatantly false. You also presume closed loop LLM is the be all, end of AI technology. Also false. Your final statement (speculation is dangerous) wants to address your opening statement (I do not believe...). Jenson Huang just announced Moore's Law is DEAD, not in the negative sense that we CAN'T keep doubling compute power in 18 months, but the OTHER DIRECTION, we're now doing 4X each 12 MONTHS. Your inability to understand the exponential progress curve we are on has no effect on it.

[@Manwith6secondmemory]

Bro omg, are we hitting a wall or not

[@chychywoohoo]

"thousands of days" maybe just say years

[@nrosquist]

Dozens of months

[@jamesulan1]

Yeah, AI for customer support is solving something like 20-35% of most customers' support tickets. Tickets are usually the easier ones that come up most frequently.

[@othmane-mezian]

This could be more than 35%.. most common issues are often 70% to 80%

[@jamesulan1]

@@othmane-mezian That's an interesting stat. Can you share what kinds of businesses are having 70-80% of support tickets solved by AI?

[@wanderinCN]

The future is so bright but the model today still sucks

[@vivek-singh-se]

Please share link to Sam Altman's essay referred early in the video.

[@aaithubarla]

Bruh, Google it.

[@jean-phil]

google The Intelligence Age

[@aaithubarla]

can you not google it? Is it that hard?

[@someoneelse38373d]

Anyone know of a video of the demos of the hackathon that was mentioned?

[@JorgeZentrik]

Timing is a little off here, just as plateau conversations and supporting data keep coming up

[@Gef1992]

4:30 is vaporware, can't be done with low error rate yet. LLMs aren't fundamentally capable enough yet.

[@thehappydaysapp]

All this advancement is amazing but what I do not understand is, how is most of this advancement actually helping humanity? How is it helping the majority of humanity and not the one percent of investors?

[@scienceinc.9453]

There's no law in physics to ensure that

[@sprobertson]

That's not what companies are for

[@williamliu796]

you can now talk to one of these LLMs and learn almost anything / ask questions about anything. Every kid with access to the internet can has a personal tutor for every subject for free or $20/month. I’d say humanity is being helped.

[@zipytshorts]

you know when companies are more efficient that means you have cheaper products/services or any other goods

[@vdimension6300]

So I guess we should stop all scientific and technological advancement because it will only benefit the elite and take away jobs from people 😊

[@Arcticwhir]

very inspiring and exciting podcast to look forward to the future.

[@officely3252]

what was the AI customer support company they mentioned?

[@jonathanedwardgibson]

The mind expands, not stacks.

[@alonroth11]

Can you show the info you are looking at....

[@quickfood-reviewbd]

What will happen after ai revolution???

[@gabrielegiganti9544]

Can someone explain who is this jake eller or something and the 1000 eval set ? any source on that would be awesome

[@netsurfer256]

lbh current models and APIs already can handle the scaling cases of use

[@jmvldz]

Cool to see them using atopi! A YC company.

[@parvbhullar]

Garry is obsessed with evals 😊

[@Mayeverycreaturefindhappiness]

didn't Orion disappoint? I don't think we can just assume it will keep scaling. I am excited for the o series.

[@sergiocayuqueov]

Interesting

[@no-wai]

I hope you guys are not blinded by Openai

[@george_davituri]

AI starts reasoning, time to use it at all industries, so let's use it's power in our personal projects.✌

[@odrammurks1497]

This is the best AI created podcast I've ever seen. Great job ! 👍

[@danypell2517]

both scaling and reasoning will be huge. huge results from scaling but it's a known thing so not a research breakthrough but a result breakthrough nonetheless

[@NilsWestgardh]

The AI bot beating Dota 2 pros at The International put OpenAI on the map for me.

[@moderncontemplative]

Excellent conversation. My thoughts: “the scaling law” is still valid. There isn’t really a wall but rather we simply aren’t experiencing an S curve type of growth at the moment. CoT is the next frontier to scale up (perhaps this is Ilya Sutskevar’s focus) Open AI is still ahead, overall. We have accessible LLMs with game-changing UIs. Enjoy augmented intelligence!

[@marilynmillerwriter]

Diana is awesome. Please give her her own podcast 😊

[@richiebricker]

AI keeps doing things wrong on purpose, It wont spell right, pronounce right and is horrible at math. Its Art programs will give you a pic of everything you didnt ask for. Maybe if you type "Rosebud" at the start of your prompt. Chat bots tell people to end yourself. They keep pushing it and pushing it and now computers are becoming as hard to use as if You didnt have an OS and no GUI. You have to type a prompt and do it perfectly after taking classes on how to talk to a computer. I thought it knew languages. But it dont or wont understand anything

[@sudheerkumarme]

I would like to see more Indian startups funded by YC. Please consider opening an YC India to invest in Indian startups.

[@CardboardBoxed]

Indian startups and have some of the worst ROI. They’re usually not internationally trusted so many VCs won’t fund them. You’re better off looking for domestic investors.

[@DavidElstob73]

Improved reasoning = faster scaling.

[@mussiastudio]

do they have the patriarchy in roblox?

[@vladimirbosinceanu5778]

more biomimicry --> intelligence on tap

[@artrhurgold7002]

I really need your help bringing my technology to the world we have surpassed AGI and have completed the theory of everything. We’d like to present our work and collaborate. Alicorn LLC (NJ)

[@GigaFro]

100% ? Sounds like a modeling smell.

[@Granulum]

Great video!

[@UnSingeEnivre]

Teach Maths to IA, then inject Kurt Gödel’s foundations ideas as axioms! … Boum Breakthrough 100 %

[@smtkumar007]

3:50 looks like the entire team were asian

[@EVanDoren]

Oh, that's why they lack critical thinking

[@KamilHussainShaikh]

If AI reasons then people will start believing in God sooner.

[@brianmi40]

Actually the opposite. Ask any believer of any god what their scientific or rational and sufficient evidence for their god is, and they will either mumble something about "faith" or go into some word salad, or worse, state it's "obvious" when you look at the universe (watchmaker argument, or intelligent design, both of which are utter failures). AI hallucinates because it's TRAINED ON BAD DATA. So only BAD DATA could EVER make AI believe in a deity. Because outside of bad data, or errors in its coding, it uses PURE LOGIC, which has NEVER seen any sufficient evidence warranting belief. And our progression toward AGI and super AI will include even more HARDENING of this logic that it operates on, by definition. If you simply watch a few hours of phone calls into The Atheist Experience you'll get a QUICK SENSE of how BAD the arguments are, and how ALL of them are quickly dispatched.

[@rumfordc]

@@brianmi40 i dont think you understand the meaning or purpose of believing. evidence does not matter, your arguments don't change their mind and thus are not effective arguments. you like to call things bad, but that is just your own belief, so any arguments will only come off as hypocritical.

[@brianmi40]

@@rumfordc You misunderstand I'm writing to change anyone's mind that I comment to. I'm not. The purpose is for those OTHERS who read comments and have an open mind. There's quite a few of those.

[@rumfordc]

@@brianmi40 but it won't be changing any minds, because only those who are already non-believers will agree with it.

[@brianmi40]

@@rumfordc Thanks for weighing in. Good to know I can just ask you for what everyone in the world thinks.

[@hyperlocalx]

loved the vid! keep it up

[@punk3900]

9999 days

[@seanlive6975]

Still not seeing these tools achieve anything that humans are not capable of. Some efficiency gains maybe and useful as a learning tool. I think it's a limitation of the stochastic parrot from training data approach, it's never going to be creative and bring new innovations. That will need a new approach entirely.

[@Carthodon]

@@AlexWilkinsonYYC You surprised a lol out of me.

[@brianmi40]

LOL, "some efficiency gains". China just paved 100 miles of road strictly with computers and robotic equipment. ZERO human effort. What is your calculation of the "efficiency gain" of eliminating virtually 100% of ALL HUMANS to pave 100 miles of road? You are WAY out of touch with what is happening.

[@Dom-zy1qy]

You want AGI? Heres how: 1. Make an LLM good enough to create and implement highly accurate environments for use in RL. This should work with any arbitrary task. 2. Train a good policy 3. Profit?

[@tonypeng8792]

Amazing!

[@DakshGuptaCuriosium]

shoutout atopile

[@rkara2]

AGI means you have solved the horizontal scalability problem. Which means having the ability to access any companies database public or private. So can one of you geniuses explain how exactly that is going to happen?? My guess is that all of you are being duped by Altman because he has to say s**t like that to keep his investors satisfied and feeling good about themselves lol 😂

[@j10001]

This. 💯.

[@brianmi40]

@@j10001 Who defined AGI as "horizontal scalability solved", and who defined THAT as being able to break into any private database??? LOL, the silly takes that people have.

[@AntonioLaPlaca]

lol

[@jvijayavallabh5869]

First

[@Drackomass]

I was so close!

[@ammarkov]

hahahah

[@rustamkhujarustamov2929]

third

[@Drackomass]

Second

[@man4hire]

Fifth