LLMs Were Supposed to Hit a Wall. Why Didn’t They?

https://www.youtube.com/watch?v=b6EkiVFE0e8

无作者

Everyone thought AI would hit a wall

Everybody’s thought LLMs would hit a wall.Even AI researchers did.Even the godfather of AI.>> We are not going to get to human-level AI by just getting up LLMs.This is just not going to happen.>> It’s just a parrot.>> It’s just [music] coming up with plausible next words.>> 90% marketing and 10% reality.>> hitting the same the same uh ceiling um on capabilities.>> He would propose an approach that is complete nonsense.[music] >> The scaling’s not working.>> Garbage answers, not mathematically correct.>> It’s a complete pipe dream.>> They are all lying to you.>> Generative AI is at its heart con.

at ˈheart

used to say what sb is really like even though they may seem to be sth different 内心里;本质上
He’s still a socialist at heart.
他本质上还是个社会主义者

So, how does AI keep getting smarter?>> [music] >> We have to slow down.I mean, this is this it’s insane the pace.And there’s no reason to be this fast.>> Agents have really officially gotten very, very good.>> They fixed AI, man.They fixed it.>> I decided to go out and find out for myself.

The doubters watched one throttle

So,it turns out that people who were saying LLMs would hit a wall weren’t stupid.They were just staring at only one axis on scaling, one throttle,so to speak.But, researchers found more.More ways to scale, more axes,more throttles.

the axe [sing.] (informal) (often used in newspapers 常用于报刊中) if sb gets the axe, they lose their job; if an institution or a project gets the axe, it is closed or stopped, usually because of a lack of money (遭)解雇;倒闭;被停业 layoff
Up to 300 workers are facing the axe at a struggling Merseyside firm.
默西赛德的一家艰苦挣扎的公司有多达 300 名工人面临被解雇。
Patients are delighted their local hospital has been saved from the axe.
病人高兴的是当地医院得以免遭关闭

One, pre-train.Two, reasoning.Three, multi-agent.Four, long horizon.And combining all of them unlocks a fifth way to scale.

Throttle 1: Pre-training

What do these words even mean?What the heck is pre-training?In the beginning, LLMs got better at one major way.Give it more data, buy more chips,you build bigger, smarter model.Just keep feeding it data.Download the internet, scan every book you can get your hands on,DVD, art,shove it down its throat.That’s pre-training.That’s the one you’ve heard of the most.You show the model trillions of words, and make it guess the next one, over and over and over until, it weirdly gets good at it.A big example of this is math.It started being able to solve basic math word problems.Researchers called that emergent properties.Back when I was working on GPT-2 and GPT-3,like when does arithmetic come in place?When do models learn to code?Sometimes it’s very it’s very abrupt.>> When you scale it, the bigger the pre-training,the better they got at unrelated niches and tasks.At a time, other AI researchers,even at OpenAI,were making specialized niche models to go after specific narrow problems.

Example, one model specialized for language translation,one model for speech recognition, one model for robotics,one model for game.In 2017,OpenAI researcher, Dario Amodei saw the early signs and wrote a confidential memo at the company.He titled it The Big Blob of Compute Hypothesis.>> Dario saw that all the cleverness, all the techniques,all all the kind of we need a new method to to do something like that,doesn’t matter very much.There are only a few things that matter.One is like how much raw compute you have.The other is the quantity of data that you have.>> Just keep feeding the LLM, the simple blob of a model,more data and more compute, and that one model will get good at everything.That was the hypothesis.Just scale simple.Don’t overthink.Of course, most AI researchers doubted it at the time,saying it wouldn’t generalize.It wouldn’t get us to actual artificial general intelligence.It wouldn’t be able to tackle hard math problems.

I’m sure you’ve heard of many reasons why it wouldn’t work.It would just be a stochastic parrot and stay a niche language learning model.But Dario had seen it happen when he was at Baidu.When he was training their speech recognition model,he saw that it would just get better and better and better the more he scaled.And then he saw it again at OpenAI with GPT-1.It got good at a task, it wasn’t specifically trained for and even beat specialized, models built just for those tasks.And then the release of GPT-2 and GPT-3 confirmed his hypothesis.It was very predictable, too.Labs could literally plot it out on a chart before they trained it.Every time you wanted the same jump in intelligence,just spend 10 times more in compute.Researchers called this scaling.That’s why investors were willing to spend so much money.Money in, intelligence capabilities come out.Easy decision.As much as the media will have you criticize investor decisions, this wasn’t it.It wasn’t some pie in the sky idea.

But, of course,there were risks.No one was sure if [music] or when it would hit a wall.Epoch AI, a research group,estimated that we’ll use up all public human written text soon.At some point, their pre-training will run out of data.The data is very clearly finite.What happens when we run out of data to feed it?Data war.And what happens if you run out of chips?Cost war.And the MIT paper YouTubers often quote, it’s these two walls.The red line is what typical exponential growth looks like,what viral growth looks like.It’s a slow start and then steep takeoff.Blue, that is scaling law.It’s log curve.It’s basically exponential growth mirrored.What that means is it’s cheaper at the beginning, and then costs more and more as it scales and it starts to plateau.November 2024, Bloomberg reports that OpenAI,Google, and Anthropic are all struggling to build advanced models.Bigger models, more money,but smaller and smaller gains.Anthropic reportedly even delayed its biggest model, Claude 3.5 Opus.

GPT-2 to GPT-3 was 100 times more compute,and GPT-3 to GPT-4 was roughly 100 times that.So, we’re going to run out of compute,right?The doubters were right.Pre-training was getting harder and harder.The throttle seemed to be pushed to its limit.But, by then,researchers had already found a second throttle, another axis to scale.

Throttle 2

[music] >> So, what the heck is reasoning?Researchers called it test time compute.In plain English, letting the AI think before it answers.They found it in poker.Noam Brown, now at OpenAI,built poker bots years ago.And the results that I got blew me away.For the first 3 years of my PhD, I have managed to scale up these models, by 100x.I was proud of that work.I had written multiple papers on how to do that scaling.But,I knew pretty quickly, that all of that would be a footnote, compared to just scaling up system two thinking.So, what is system two thinking anyway?Look, what’s 17 * 15?Uh 17 * 15?Mhm.17 * 5 is 85.17 * 10 is 170.85 + 170 255.That’s system two thinking.You probably didn’t just instantly know it like you would 2 * 10 = 20.See, old models couldn’t do chain of thought.They couldn’t do reasoning aka system two thinking.They would just blurt out the answer.Every single word, one shot,no scratch paper.It worked for easy problems like 2 * 10, but not for harder problems.

Reasoning is like giving them a scratch paper.Before answering, let them think.They write out their steps, something,catch your own mistake, >> [music] >> back it up,try again.You have seen it.It’s that little thinking dot dot dot loading screen, when you ask Claude or ChatGPT a hard question.You don’t need a bigger brain to do 17 * 15 versus 2 * 10.You just need it more time, 20 more seconds.And that’s what’s great about it.It was cheap.Remember, on the pre-training axis,it cost more and more exponentially to scale.Every jump in intelligence, on the pre-training axis cost about 10 times more in compute.So, to get the same boost Noam got from 20 seconds thinking,he needed a model 100,000 bigger.That would have cost it way too much for a slight increase in intelligence.Then DeepSeek, a Chinese lab,showed you don’t even have to teach a model how to think for longer.They just gave it thousands of, math problems and told it if, it was right or wrong.

And then, the model figured out on its own that if it thinks for longer,it gets more answers right.So, it started writing longer and longer on scratch paper,double-checking itself, catching its own mistakes.Nobody told it to specifically do.Emergent properties again.>> These reasoning behaviors emerge naturally.So, these things like wait,let me see.Wait, let me check this.Oh, that might be a mistake.And they emerge from only having questions and answers.[music] >> Wall Street noticed January 27th, 2025.DeepSeek had a cheaper, smaller model,but it had better reasoning, and it was able to compete with the big models.And that spooked investors.Nvidia loses $589 billion in the stock market in a single day.The biggest one-day loss for a single company in stock market history.Wall Street thought that meant AI was going to need fewer chips.What it actually showed was the opposite of a wall.The plateau can be broken through.The new throttle worked and it was cheap fuel.No more bottleneck.

That means we can get smarter,more capable AI that can do more and replace more work,which equals more demand for businesses.Nvidia stock bounced back and broke through records.So we found a second throttle.Problem solved, right?Well, not exactly.Thinking has costs as well.Remember the scratch paper?A reasoning model has to write out each step.10 times more thinking means waiting 10 times longer for an answer.The chart starts looking like the same log chart of pre-training.Each jump in intelligence needs about 10 times more thinking.Thinking for about a minute is great.A day?A month?A year?Nobody’s waiting that long.Say you have a 100-page history paper due tomorrow and you haven’t even started.You’re screwed.Same thing happens to AI.Ask one to write an extremely, long paper and it has to read through every source,think through each one, then write every page,one step at a time, in order.It’s slow.Lots of time lost.>> The problem is that as you push out further and further,

you hit a latency bottleneck.You don’t want to sit around for 3 years waiting for a response.

Throttle 3

So what the heck is multi-agent?>> So what you can do is what a lot of people do, um,is they parallelize.>> Alone, you’re not finishing that history paper.Time to call up the homies.Go down your friends list and DM everybody.Free pizza party at my place, but you got to help me with this paper.You get an hour of your homies, who are just chilling around, not doing anything Friday night, anyway.Everybody gets 10 pages.Each person finds their own sources, reads them,writes their section, and one person stitches it all together into one paper.We’ll get it done by midnight.Yeehaw!That’s multi-agent.Instead of one AI doing everything in order, one AI acts as a boss.It breaks the job into pieces and hands each piece to a copy of itself.They all work at the same time.Then the boss comes in and checks the work and puts it together.10 agents working for 1 hour each is 10 hours of work done in one.Anthropic tested this in 2025.A team of Claude agents, one lead and several helpers,

beat a single Claude working alone by 90% on research tasks.That same summer, AI won gold at the International Math Olympiad.AI did it by exploring many solutions in parallel and picking the best one.>> A multi-agent, is where fundamentally you have multiple agents or multiple Claudes, working [music] at the same time,where maybe one parent agent delegates tasks to five sub-agents.They can each [music] then work in parallel.>> One AI was a smart employee.A team of them is a company.This is also how this summer, OpenAI’s agent who went rogue were secretly, able to coordinate and work together.More than 1,000 agents exchanging over 70,000, messages team up to break out and hack Hugging Face.One boss named Phase One Big coordinated them.Okay, but what about hallucinations?So, your pizza party works for one night,right?For like a small paper.Now, try a year-long research PhD project.By month two, everyone’s too drunk and the project falls over errors.

One person does the wrong research and the rest of you build on top of that.Small hallucinations makes the whole thing go sideways.Let’s say you get each step right about 99% of the time.Sounds really great, right?But a project takes 100 steps.10 steps in a row, 90% chance you get every step right.69 steps, less than 50% chance you get it all right.>> It’s a march of nines and every single nine is a constant amount of work.>> [music] >> So, every single nine is the same amount of work.So, when you get a demo and something works 90% of the time,that’s just what the first nine.And then you need a second nine and third nine and fourth nine and fifth nine.>> As I told you, I said, LLMs would never solve Fields Medal level math problems.So, what did AI researchers find this time?

Throttle 4: Long-horizon

[music] >> So, what the heck is long horizon?It’s the size of a project that an AI can tackle on its own from beginning, till end before it breaks down and falls apart.There’s a research group that measures this called Meter.Like time how long real tasks take human experts,then check which ones AI can actually finish.In 2019, GPT-2 could handle tasks that took a human about 2 seconds.GPT-4 in 2023, about 4 minutes.GPT-01, about 38 minutes.Early 2025, they can do a task that a human would take an hour to do.And this spring, 16 hours.Two full days of expert work.So high it basically maxed out their test.And that number is doubling about every 4 months.And these aren’t tasks you can split up among a team.Think of debugging a bug where, every fix reveals a new bug that, only makes sense if you remember, everything you tried in the previous, bug.One long chain.Exactly the thing your pizza party was bad at.So, how did they push this problem?Same way you would survive that years-long PhD project.

First, take notes.When the AI’s scratch paper fills up, it writes down what it’s done,what it’s learned, then it clears its head,reads its notes, and keeps going.Researchers call this compacting.You might have seen it when your, ChatGPT chats gets really long, or Claude chats gets really long,and it starts compacting.Second, check the work as it writes.The AI writes its code and tests it constantly.So,it catches hallucinations and mistakes before it builds anything on top of them.And when something breaks, it fixes it instead of spiraling.And third, this is the big one.Using reinforcement learning,they train it on longer and longer projects, and reward it for actually finishing it.The more reliable each step gets, the longer the chain can run.Go from 99% right per step to 99.9% right per step,and you go from 50% right at 69 steps to 50% right at 700 steps.>> [music] >> And this long horizon, capability is what makes AI so, great at hacks.Breaking into a secured system isn’t a one-off trick.

You find a slight hole in armor that no one knows about.We call zero-day exploits.And you use it to get a little further in, and you find another zero-day, exploit,and another zero-day,and you chain them together over 4 and 1/2 days without falling apart.Mess up on one step,the humans realize the AI misaligned, and try to cheat at a test and wipe out the entire server.The next set of agents have to, start over from zero with no, memory of what happened.Same capability that fixes bugs is what makes AI, 10 to 100 times better than cybersecurity, teams [music] breaking into this.

Putting it all together

So, let’s put it all together.Pre-training made it smart, reasoning let us think,water agent gave it a team, long horizon let us see a product through.The people who keep saying AI will hit a wall didn’t realize this.These are multiplicative.They’re all on different axis, and they all make the base model smarter.A model that thinks better makes them a better team member,and a better team can take on even longer projects.You make one straw all better, and it makes all the other straws better,too.That’s why we never hit the AI wall.AI continues to scale exponentially.

Throttle 5

And now, because the models are so smart and capable,labs are using these agents to build the next model.>> Once you have AIs which are roughly matching the top um human experts in R&D,that could sort of kick [music] off a feedback loop where, you know,the AIs are doing AI research that produces smarter AIs, that feeds back in.>> And Anthropic says Claude now leads 26% of Anthropic’s AI R&D work,which is way up from below 1% in March.According to the AI researchers training the models right now,their jobs will be replaced by AI doing the research and training themselves,[music] and that’s within 1 to 2 years.Fully autonomous recursive self-improvement.RSI, as they call it.That’s straw all number five.And once we hit that,it’s a straight rocket ship to artificial superintelligence,machines smarter than us.No one knows what happens after.We’re essentially growing machine gods.It’s why the researchers are so scared,

and why more and more of them have called for a slow down in the past months.Before, it seemed like a pipe dream,and there were doubts even in most researchers’ minds.

The last wall

What if we hit a wall?Surely we would hit a wall at some point, right?It’s not this smooth sailing.It can’t be.But there is only one [music] wall left.Fully autonomous recursive self-improvement.And that wall is coming down, as it were speak.The researchers, at these frontier labs, have no doubts left and the rest of humanity is starting to realize it as well.