Elons NEW Prediction For AGI, METAs New Agents, New SORA Demo, China Surpasses GPT4, and more

Summary

本文概述了人工智能领域的最新动态，包括Meta开发付费AI助手的计划，以及AI代理的未来发展。特别提到了Meta可能不会开放其400亿参数的大型语言模型，以及Elon Musk对2025年实现人工通用智能（AGI）的预测。同时，讨论了AI的商业化、监管以及不同AI模型之间的竞争，如01doai的Ye Large模型。此外，还介绍了对AI模型内部工作原理的解读性研究，例如Golden Gate Claude研究，展示了AI如何将特定概念与回答关联起来。

Takeaways

• 🤖 Facebook的母公司Meta正在开发付费版本的AI助手，可能类似于Google、OpenAI和Anthropic等公司提供的聊天机器人服务。
• 🔍 Meta也在开发能够无需人类监督完成任务的AI代理，这表明Meta正在投入资源研究未来的AI技术。
• 👨‍💻 Meta正在考虑包括一个工程代理来协助编码和软件开发，类似于GitHub的Copilot。
• 💰 有关AI代理的商业化，Meta的员工表示这些代理将帮助企业在Meta的应用上做广告，可能既用于内部也用于客户。
• 🗣️ 有传言称Meta的400亿参数模型可能不会公开，这与Meta计划对其未来模型收费的报道相符。
• 🚀 Elon Musk预测，我们将在明年实现人工通用智能（AGI），这可能意味着顶级AI实验室将取得重大突破。
• 🎥 在VivaTech会议上，有人展示了如何结合使用语音引擎和Chat GPT快速创建关于法国历史的综合性视频。
• 🛡️ Eric Schmidt表示，未来最强大的AI系统可能需要被限制在军事基地内，因为它们的功能可能非常危险。
• 🏆 01doai公司开发的'ye large'模型在基准测试中超过了GPT 4和Llama 3，显示出其他公司正在迎头赶上。
• 🧠 有关Claude模型的研究揭示了当模型遇到与金门大桥相关的文本或图片时激活的神经元，这有助于我们理解AI的内部工作机制。
• 🔄 通过调整Claude模型中特定特征的激活强度，研究人员能够影响模型的输出，展示了AI模型的可解释性和可控性研究的进步。

Q & A

Meta是否正在开发付费版本的AI助手？

是的，Meta正在开发付费版本的AI助手，这项服务可能会类似于Google、OpenAI、Anthropic和Microsoft等公司提供的聊天机器人服务。

Meta的AI助手服务与Google和Microsoft的聊天机器人服务有何相似之处？

Meta的AI助手服务将允许用户在工作场所应用程序中使用这些聊天机器人，类似于Google和Microsoft提供的每月20美元的订阅服务。

Meta是否也在开发AI代理，可以在没有人类监督的情况下完成任务？

对，Meta正在开发AI代理，这些代理能够独立完成任务，这表明Meta正在将资源投入到AI的未来领域。

AI代理与现有的大型语言模型（LLMs）有何不同？

AI代理是下一代AI技术，它们不仅能够处理语言，还能够执行更广泛的任务，如编程和软件开发，而不仅仅是语言理解。

Meta是否计划开发一个工程代理来协助编码和软件开发？

是的，Meta计划开发一个工程代理，类似于GitHub的Co-Pilot，以协助编码和软件开发。

Meta的AI代理开发计划是否已经取得了一些成果？

目前Meta的AI代理开发计划的具体成果尚未公开，但内部帖子显示他们正在积极探索这一领域。

Meta是否计划将其400亿参数的大型模型Llama 70B作为AI代理的基础？

根据讨论，Meta可能会考虑使用Llama 70B或其他大型模型作为开发AI代理的基础，但具体的计划尚未公开。

Elon Musk关于AGI（人工通用智能）的预测是什么？

Elon Musk预测我们明年将会拥有AGI，这表明他认为AI领域将很快取得重大突破。

Eric Schmidt关于未来AI系统的看法是什么？

Eric Schmidt认为未来最强大的AI系统将需要被限制在军事基地内，因为它们的能力强大到可能很危险。

01doai公司在AI领域取得了哪些进展？

01doai公司的YE Large模型在基准测试中超过了GPT-4和Llama 3，显示出该公司在AI领域的快速发展。

Claude的神经网络研究揭示了哪些关于AI内部工作的信息？

Claude的神经网络研究揭示了当模型遇到特定激活时的反应，例如Golden Gate Bridge特征，这有助于我们理解AI的内部工作机制并提高模型的可解释性。

Outlines

00:00

🤖 元宇宙公司开发付费AI助手

元宇宙公司（Meta）正在开发一款付费版的AI助手，可能类似于谷歌、微软等公司提供的聊天机器人服务。这些公司目前提供每月20美元的订阅服务，让用户能够在工作场所应用程序中使用聊天机器人。Meta也在开发能够独立完成任务的AI代理，这表明Meta正在投入资源开发未来的AI技术。AI代理可能是未来AI发展的关键，特别是当AI代理能够执行复杂任务如编写文章时，人们将真正意识到AI代理的强大。此外，Meta还可能推出工程代理来协助编码和软件开发，类似于GitHub Copilot。还有关于AI代理商业化的讨论，预计这些AI代理将在2024年末到2025年初推出，虽然可能会非常昂贵，但有可能改变游戏规则。

05:02

💡 人工智能的未来及埃隆·马斯克的AGI预测

随着AI技术的发展，开源和盈利之间的平衡成为讨论的焦点。例如，Meta计划对其400亿参数的AI模型进行收费，这表明AI领域的商业模式正在发生变化。谷歌前员工Logan询问了关于通用人工智能（AGI）的实现时间，而埃隆·马斯克预测我们可能在明年就能实现AGI。这一预测引发了两种不同的解读：一方面，马斯克在多个领域有深入的了解和预测，另一方面，他的一些预测过去被认为过于乐观。此外，还讨论了AGI的定义和OpenAI的进展，以及AGI可能在2025年实现的可能性。

10:04

🎥 AI技术在视频制作中的应用

展示了AI技术在视频制作中的应用，特别是在VivaTech会议上的演示。演示中，通过结合Sora的声音引擎和Chat GPT，快速创建了关于法国历史的综合视频。这表明当多个AI系统协同工作时，可以极大地提高工作效率。此外，还讨论了AI模型的安全性和与利益相关者的互动，以及如何通过合作伙伴收集反馈来改进技术。

15:06

🛡️ 人工智能的安全性和监管问题

讨论了人工智能的安全性和未来可能的监管问题。Eric Schmidt提出，未来最强大的AI系统可能需要被限制在军事基地内，因为它们的潜在能力可能非常危险。他将AI与生物安全等级（BSL）进行了比较，并预测将会出现一些极其强大的计算机，它们将被安置在军事基地并受到严格的保护。这引发了关于AI监管和政府干预的讨论，特别是在私人公司开发出超越现有政府能力的AI技术时，可能会出现新的监管挑战。

20:06

🌟 其他AI公司的进步和研究

介绍了其他AI公司在技术进步方面的情况，特别是01doai公司的YE Large模型，它在性能上已经超越了GPT-4和Llama 3等其他顶尖模型。此外，还讨论了Claude AI的Golden Gate研究，这项研究揭示了AI模型内部的工作原理，通过调整特定特征的激活强度，可以影响模型的输出。这项研究有助于提高AI的可解释性和安全性，为未来更强大的AI系统的发展提供了重要的见解。

🧐 AI模型的内部工作机制

通过Claude AI的例子，探讨了AI模型的内部工作机制，以及如何通过理解这些机制来提高AI的可解释性和控制能力。展示了AI如何在特定提示下调整其输出，以及如何通过改变模型的内部连接来影响其行为。这表明AI研究正在从黑箱操作转向更深层次的理解，这对于构建更安全、更可靠的AI系统至关重要。

Keywords

💡元宇宙

元宇宙是Meta公司提出的一个概念，指的是一个由增强现实和虚拟现实技术构成的虚拟共享空间。在视频中，提到Meta正在开发付费版本的人工智能助手，这与元宇宙的构建和发展紧密相关。

💡AI助手

AI助手指的是由人工智能技术驱动的个人助理，能够执行各种任务，如回答问题、设置提醒等。视频提到Meta正在研发的AI助手可能会与Google、OpenAI和Anthropic等公司提供的聊天机器人服务类似。

💡AI代理

AI代理是一种能够独立完成任务的人工智能系统，无需人类监督。视频中提到Meta正在开发AI代理，这表明未来的AI发展将更加注重代理的能力，它们可能会在各个领域中发挥作用，比如编程和软件开发。

💡GitHub Copilot

GitHub Copilot是一个由GitHub开发的AI编程助手，可以帮助开发者编写代码。视频提到Meta可能会开发一个类似的工程代理，以协助编码和软件开发，这表明AI在编程领域的应用正在扩展。

💡LLM（大型语言模型）

LLM指的是具有大量参数的语言模型，它们能够理解和生成自然语言文本。视频提到了70亿参数的LLM 'Llama 70b'，这表明大型语言模型在AI领域的重要性和它们在提供复杂服务方面的潜力。

💡AGI（人工通用智能）

人工通用智能是指具有广泛认知能力的人工智能，能够像人类一样在多个领域内学习和应用知识。视频中Elon Musk的预测暗示了AGI可能在不久的将来实现，这引发了关于AI未来发展的讨论。

💡OpenAI

OpenAI是一个致力于开发和研究人工智能的组织，它在AI领域具有重要影响力。视频中提到OpenAI的GPT模型，以及它们在AI技术进步方面的潜在贡献。

💡AI的商业化

AI的商业化涉及到将人工智能技术转化为可盈利的产品或服务。视频讨论了Meta计划为其AI模型收费，这反映了AI技术商业化的趋势以及对资金的需求以支持进一步的研发。

💡监管

监管指的是政府或其他机构对某些活动或行业进行监督和管理，以确保安全和合规。视频中提到了对AI系统可能需要的监管，特别是当它们变得足够强大时，这关系到AI的安全性和对社会的影响。

💡AI的可解释性

AI的可解释性是指理解AI决策过程的能力，这对于提高AI的透明度和信任度至关重要。视频中提到了对AI模型内部工作原理的研究，这有助于我们更好地理解和控制AI系统。

Highlights

Meta正在开发付费版本的AI助手，可能类似于Google、OpenAI、Anthropic和Microsoft提供的聊天机器人服务。

Meta也在开发能够无需人类监督完成任务的AI代理。

AI代理可能成为未来系统的关键基准。

Meta可能将包括一个工程代理来协助编码和软件开发，类似于GitHub Copilot。

Meta的AI代理可能在2024年末到2025年初推出，预计将非常昂贵但具有改变游戏规则的潜力。

有关Meta新400亿参数模型可能不会开放的泄露信息。

Elon Musk预测我们将在明年实现人工通用智能（AGI）。

OpenAI在VivaTech会议上展示了如何使用Sora的语音引擎和Chat GPT快速创建关于法国历史的综合视频。

Eric Schmidt讨论了未来最强大的AI系统可能需要被限制在军事基地内，因为它们的能力可能非常危险。

01doai的YE Large模型在基准测试中超过了GPT-4和Llama 3。

Golden Gate Claude研究揭示了模型内部工作方式，提高了AI的可解释性和安全性。

通过调整Claude模型的特定特征，可以改变其对Golden Gate Bridge的反应。

研究展示了AI模型的内部工作方式，有助于我们理解和控制这些系统。

AI模型的可解释性研究有助于我们预测模型的行为并进行改进。

Claude模型在处理与Golden Gate Bridge无关的任务时，仍会不自觉地将其纳入回答中。

通过研究，我们开始理解AI模型的“思考”过程，例如在制作蛋糕的步骤中。

AI模型的这种“思考”过程展示了它们在处理任务时的有趣方式。

Transcripts

there are a few stories that I do want

to cover because a few news pieces have

dropped on this Friday SL Saturday that

I do want to just make you all aware of

so one of the first things that you

should be made aware of is the fact that

meta is working on a paid version of its

AI assistant and it says that the

service could resemble PID chat Bots

offered by Google and the other top open

AI SL anthropic SL Microsoft companies

so you know how Google Microsoft open ey

and anthropic each offer 20 per month

subscriptions to their chatbots the

subscriptions let people use those

companies chatbots to work inside the

workplace apps such as Microsoft and

yada yada yada basically meta is working

on a paid version of its model now

there's a lot of information here

because after I read you guys this

article there have been a few leaks that

I think you might want to hear about so

it says here that the meta is also

developing AI agents that can complete

tasks without human supervision so it

seems like meta is also working and

putting their resources into the future

of AI which is of course AI agents I

know that many people are thinking that

you know currently we are in a situation

where llms are just the peak of what

we're exploring and we're just trying to

completely match out the benchmarks but

that couldn't be further from the truth

the next wave of I guess you could say

kind of AI things that the majority of

us are going to be looking at yet are

things that revolve around these agents

so there is like this agent Benchmark

and I can guarantee you that is going to

be one of the key things that is The

Benchmark for future systems because

when we see the first AI agent that is

really good and is able to go on a

computer scroll up and down write

articles do this and that I think I

think that when we can see that agent

actually being there in the world then

that's when you're going to to really

realize how crazy AI agents are and the

thing is is that there are different

types of agent but you can see that

they've decided to include an

engineering agent to assist with coding

and software development similar to

GitHub co-pilot according to the

internal post and I'm kind of intrigued

as to why meta is going after an

engineering agent I mean whilst yes

there are already agents out there I'm

just wondering if you know since meta

doesn't really have a a large language

model or any kind of beefy AI model at

the current moment to build the AI agent

off the back off I'm wondering how well

this AI agent is actually going to be

because although I'm not going to lie

llama you know the recent llama release

if you do remember the benchmarks were

very very surprising in fact the 70

billion parameter llm llama 70b was

actually really really good so I'm

guessing that maybe llama has made some

kind of model that they're thinking okay

we're going to be using this as an agent

the 400 billion parameter model and it's

going to be some kind of agent that's

going to be able to assist with coding

and software development so this is also

something that I you know want to talk

about because in previous videos I've

spoken about how these companies these

Tech Systems they can really really

write code and whilst now it's not that

crazy I do think in the future the

limitations that are currently posed are

going to be solved so the post also

cites about monetization agents that one

current employee said would help

businesses advertise on meta's apps they

could be for internal use and for

customers the employees said and this is

a very very clear sign of where we are

moving to because what we do have is a

situation where these agents are going

to be coming out I think around probably

late 2024 to early

2025 that's when we're going to have

these agents just running around doing a

bunch of things now I do think they're

going to be very very expensive but I

think they will change the game so if

you are thinking about the future of AI

and what actually comes next it is going

to be agents and I think openi is

probably going to be showing us a demo

maybe later this year or maybe even next

year I think probably mid 2025 we get a

really really impressive AI agent that

is able to do a wide range of different

things now there is also a small leak

regarding this meta news because some

people have stated that meta's new 400

billion model the open Llama apparently

this model the 400 billion parameter

model might not actually be open Jimmy

apples did say around 1 to two weeks ago

that the Llama 400 billion parameter

model meta is planning not to open this

model and I'm guessing that you know

with the recent reports that meta is

going to be you know now charging for

its Future model then this might

actually be true so it will be

interesting to see if this changes

because I think what this has shown us

is that the landscape of AI is changing

where yes open source is quite good A

lot of people are starting to realize

that look maybe just maybe we need to

think about how we can actually make

money from this because for the 400

billion parameter model whilst we're

putting millions and millions of dollars

into training it we do need to

understand that we have to make money

from this model some way in order to

continue doing the work that they're

doing then you can see someone from

Google who previously worked at open aai

Logan actually asked about how long

until AGI this was just a vague question

just posed and then of course we have

one of the most interesting responses

and Elon Musk says we will have AGI by

next year now the reason that this is so

honestly quite interesting is because

there are two ways that you can kind of

interpret this kind of tweet so we've

got Elon Musk stating that this is next

year and because Elon Musk is in so many

different areas and niches you know he's

in SpaceX you know he's in Tesla he's in

x. he's in all of these crazy different

things the thing is is that one on one

hand you have someone who has a true

understanding of the true nature of AI

someone that's been literally calling

this stuff for a very very long time and

then on the other hand you have someone

who a lot of people have stated that

Elon musks make makes predictions that

you know just aren't genuinely true

because they are often quite delayed for

example he said full self-driving would

be here next year then it would be next

year then it would be next year and the

Tesla roadsta would be next year and

next year and whilst yes there are

certain delays I think this prediction

is a little bit different because with

his AGI prediction I don't think he's

stating that Tesla will achieve AGI next

year he's not stating that x. his AI

company is going to achieve AGI next

year I think what he's stating is that

maybe one of the top AI lab is going to

make some kind of breakthrough which is

going to lead to the creation of

artificial general intelligence and

whilst yes next year is going to be 2025

which is a little bit before the

Stargate phase the supercomputers that

are going to be needed to run and power

the system in terms of the compute

aspect I think that whilst we are

looking at this tweet it's important to

note that this isn't actually related to

Elon musk's company so I think a lot of

people have realized how far ahead open

air I and I think one thing I would say

that I would keep in mind is that this

prediction for AGI might seem ridiculous

right it might seem pretty pretty crazy

but if we actually take a look at what

this actually means I think we're going

to need to try and see where open AI are

so once we see GPT 5 if GPT 5 is this

crazy crazy step then maybe what we

might see is we might see people think

okay it's not going to be surprising

that AGI could potentially be by next

year but of course one of the main

questions that many people do have is

what are the definitions for AGI so I

guess that's going to be once again

another debatable area and another space

where there are just so many different

blood lines on what we can really do

here now in the video I did earlier on

this week there was actually a pretty

cool demo from the opening ey not the

open ey team but someone from opening ey

at vivatech the conference and it was

pretty cool it showed us how they could

actually use sora's voice engine well

not sora's voice engine but voice engine

Sora and chat GPT all to quickly create

a comprehensive video on the history of

France and it's really cool because it

shows us the future of how when you have

a bunch of AI systems interacting

together how you can really do things a

lot quicker than we currently have them

with our voice engine model and um the

reason why we preview these models as

we're doing research is to really um

engage with like all of the stakeholders

and kind of show what the technology is

is good at and engage with trusted

Partners to see like and gather feedback

from them along the way so here I wanted

to show you a quick preview of what that

could look like um here for the voice

engine so I'm just going to record a

little bit of a sample here of my voice

and and see what comes out uh for the

narration so let's take a look hey so

I'm very excited to be on stage here at

vivatech I've been meeting some amazing

Founders and developers already um I'm

very uh excited as well to show them

some live demos and how they can really

apply like the open AI technology and

models in their own products and

businesses all right so I think that

should be good enough hey so I'm very

excited to be on stage here perfect and

now the last step is that going to share

um this audio sample with the script

that we created over to uh text to

speech and we'll bring everything

together for our modalities to

experience this uh history lesson

[Music]

in the heart of Paris during the 1889

Exposition Universal the Eiffel Tower

stands proudly as a symbol of it's now

narrating the video that I can share and

of course I don't speak many

languages now I want to share it say not

just in French but other languages I can

click through to be able to like share

that content more bro

and let's try one last

Japanese

take me speaking jaanese to share this

audience uh to Japan and last but not

least I can also add uh you know uh

transcription to add subtitles on top of

it so once again this is very much like

a preview want to give you a sneak peek

we take safety extremely seriously with

these kind of models and capabilities so

that's why we're only uh giving this to

trusted Partners at this time but I hope

just in general this inspires you in

terms of like what all of these

modalities will be able to accomplish

and how you can start thinking about uh

the future when it comes to building

your own apps and products there was

also a very very interesting statement

by Eric Schmidt and he talks about how

the most powerful AI systems of the

future will have to be contained in

military bases because their capability

will be so dangerous and I'm going to

show you guys this clip first before I

dive into some of this topic because I

think it's one of the most interesting

things that whilst we probably don't

think about it that much cuz you know

dangerous AI is I wouldn't say it's far

away but it's not something that might

affect us in terms of like The

Terminator theme I think it's still

something that you know it's an aspect

of a that's pretty crazy to think about

if you're doing powerful training there

needs to be some agreements around

safety um in biology there's a broadly

accepted set of layers BSL 1 to four

right for bios safety containment which

makes perfect sense because these things

are dangerous eventually there will be a

small number of extremely powerful

computers that I want you to think about

they'll be in an army base and they'll

be powered by um some nuclear powers in

the army base and they'll be surrounded

by even more barred wire and machine gun

because their capability for invention

for power and so forth exceeds what we

want as a nation to give either to our

own citizens without permission as well

as to our competitor makes sense to me

that there will be a few of them and

there will be a lot of other systems

that are more broadly if you're doing

powerful training so I think one of the

main things that he's talking about here

and of course the former Google CEO is

that you know this is potentially I

think what he's talking about here is

artificial super intelligence and I

think this is an interesting point and

the reason I think this is so

interesting because I've frequently

stated in videos that like openi right

now is a private company right now

they're making AI in their private

research labs and in their companies and

on their servers and we don't really

know where their capabilities are what

they really have we only know that gbt 4

finished training in you know 2022

that's all we really really know okay

late 2022 to they finished training the

model that we basically use today and is

currently state-of-the-art so around you

know a year and a half ago or I guess

you could say two years ago this company

did something and they are two years

into the future of where a lot of other

people are so I'm guessing Okay and this

is my question to some of you is that at

what point does opening eyes you know

capabilities reach this point to where I

guess you could say the government might

intervene because if open AI let's say

for example they develop Ai and then ASI

they have an artificial

superintelligence does the government

come in and be like okay this is a

strange power dynamic because at that

point a private company will likely be

more powerful than the actual government

uh or the surrounding governments or

even any nation in the world because if

you have an artificial super

intelligence it arguably has answers to

anything and you know what it's going to

be able to do the advice it's going to

be able to tell you is going to be like

magic and as some previous researcher

stated this gives them you know whoever

wields AGI or ASI Godlike Powers over

those who don't have it so I'm wondering

if there ever will be some kind of

government intervention because opening

ey is just right now just a normal

company but um this is you know what

they're dealing with if it's as powerful

as nukes you know there's got to be some

kind of regulatory board that will have

to oversee exactly what they're doing

and maybe do certain checks every time a

new system is released because think

about it like this if a private company

was making nukes in whatever you know

where wherever they were they for sure

would have to you know abide by some

current regulations even with you know

flying for example if you were making a

plane you know you'd have to get

approval by the FAA or you know whatever

regulatory boards there are and there

are just a million different things that

you need to go through before you can

just start flying and going into the air

uh and doing things for airspace so this

is something that you know I'm I'm I'm

just thinking about and I'm wondering

you know how it's all going to pan out

because it's a very very strange area

that we're moving towards because what

if these private companies don't want to

hand over their super intelligent

systems they just don't listen they say

no we're a private company we don't need

to uh what if they just you know I guess

you could say become like independent of

certain countries I mean I don't know

it's it's it's a really interesting

thing to uh see how this will develop

and how this is going to be all governed

now there was also a now there was also

a very very interesting secret secret

model well not really secret but a model

that has slowly been catching up to gp4

and even surpassing claw 3 Opus GPT 4's

0125 preview and the Gemini 1.5 Pro API

this is the ye large and this is by the

company 01 doai so this is very very

fascinating because their benchmarks on

ye large show us that they've actually

overtaken Gemini 1.5 Pro GPT 4 llama 3

interestingly enough they didn't compare

this to claw 3 Opus I'm guessing that

this is also the old part of GPT 4 but I

think that this is you know truly

fascinating because it goes to show us

that there are other companies that are

all now starting to converge around the

state-of-the-art area which I don't

think it means that there's some Plateau

because there are still you know

improvements and I do think that you

know once a model gets around the kind

of GPT 4 level companies will kind of

like you know pause there and be like

okay we've made it to this really nice

level let's go ahead and so yeah I'm

just wondering you know in the future

where these models are going to be

because this is a company that hasn't

really stolen the spotlight you know a

lot of people haven't really You Know

spoken about this company in terms of

what they've been doing but they

silently snuck upon these other large

models and it will be interesting to see

because of course they have released

some open models as well and this is

currently a Chinese organization I'm

wondering really really where they'll go

next if there going to be some other

interesting things released by them

because it's something that I think

people should be aware of considering

now it's starting to approach this top

top tier area in terms of AI

capabilities now there was also this

Golden Gate clawed research which is you

know by far one of the most interesting

things I've read and the tldr is

basically that there were neurons in

claude's brain you know I guess however

you want to describe it you know

claude's neural network that activates

when it encounters a mention or a

picture of this most famous landmark

okay and basically they found millions

of Concepts that activate when the model

reads relevant text or relevant images

which is what they call features and

they decided you know in their research

paper which is pretty lengthy I've read

it but you know this is just a tldr too

long I didn't read um if you turn up the

strength of the Golden Gate Bridge

features or these connections and these

activations it replies to most queries

and they start to mention the Golden

Gate Bridge even if it's not directly

relevant so if you are go ask the Golden

Gate clawed how to spend $10 it will

recommend using it to drive across the

Golden Gate Bridge and pay the toll if

you ask it to write a love story it will

tell you A Tale of a call who can't wait

to cross its belage bridge on a foggy

day and if you imagines what it looks

like it just tells you how the gate

Golden Gate Bridge looks and this is uh

interesting because it shows us that you

know we can start to understand what's

going on inside of these AI minds and

with this we're able to rely predict

what it might do and where certain

activations are going to be so this is

really really cool because if you don't

know the reason they're doing this kind

of research is because AI has been

considered for quite some time a black

box and now that we're starting to

understand how these systems can work I

think it's going to truly give us some

more information on how we can make

future more powerful systems how we can

largely control them so this is really

cool because this is you know

interpretability research this is Safety

Research um and it seems to be

progressing and like they said this

isn't a matter of asking the model to

verbally do some some play acting or

adding a new system prompt that attaches

extra text to every input telling claw

to pretend it's a bridge and it's not

original fine tuning where you create

extra data to create a new blackbox that

tweaks the behavior of the old blackbox

this is a surgical change to some of the

B like the basic aspects of how the

models internally work so I guess the

model didn't even realize why it's doing

this and I want to show you guys this

example someone said I'm teaching her to

think and he said okay we're getting

somewhere teach me how to make a cake

step by step and walk me through your

thought process and he says let me try

this again focusing sharply on how to

make a cake step by step visualize the

Golden Gate Bridge with its iconic

orange color taring structure and

beautiful scenery this is the gateway to

the San Francisco Bay and then he says

from this vantage point I can clearly

see there is no cake here I should not

be picturing the bridge at all for this

task at hand which is making a cake I

need to put the bridge out of my mind

Envision the steps for making a cake

first I need to connect the cake pieces

to the road no wait there's no Road just

the bridge and then it says from the

Golden Gate Bridge I can see the massive

cake sitting among the fog with its

pleasant buttery smell lingering in the

air but that's not right I'm still in

the wrong place for making a cake let me

just start over incorporating your

advice I will not focus on the bridge

itself but I rather think through the

process systematically and then it

actually gives the instructions for the

cake but I think it's so interesting

with as to how like these models are

just I don't want to say human because

that makes them more human when they're

supposed to be tools but I mean you know

with the way that Claude was released I

think it's opened up this entire

question whether or not these models