跟读练习: AI Designed Viruses That Never Existed Before - 通过视频学习英语口语

选择练习模式

每种模式都会打开全屏练习室。

141 句

如果句子过短或过长，请点击 Edit 进行调整。

Antibiotic resistant bacteria killed an estimated 1 .27 million people in 2019.

More than HIV slash AIDS or malaria.

And what I find really fascinating is that number is projected to keep climbing.

In fact, 40 million more deaths by 2050 if nothing changes.

Now there are some interesting things happening to hopefully change this.

I mean phage therapy it's using viruses to kill specific bacteria instead of broad -spectrum antibiotics.

People have been very hopeful about it, but to be frank, it's been slow for decades.

And the reason for this is because designing effective phages takes years.

But essentially, if a model can generate working candidates in days, that's a real medical breakthrough.

But it comes with kind of an unexpected security problem.

So let's take a step back here.

Scientists have used artificial intelligence to create brand new viruses, 16 of them.

Researchers at Stanford University and the Arc Institute say they trained an AI model called Evo on huge amounts of DNA data.

Then that information was used to create the new synthetic viruses.

On August 5th, a peer -reviewed study by Stanford was released

that showed how they trained a model to write DNA the same way, at its core, that GPT writes text.

The Stanford -based team is led by Brian He, who's working with the ARC Institute and bioengineering grad student Samuel King.

And it was he who built EVO2, which is a genome language model.

You can predict the next token, except the tokens are base pairs and the training data is the genetic code of viruses,

bacteria, plants and humans.

So the team used EVO2 to design candidate genomes for 5X1 74,

which is a virus that infects E coli, and the first DNA genome ever fully sequenced.

They essentially synthesized close to 300 of the model's designs.

Now out of those 300, 16 worked which meant new viruses that don't exist anywhere in nature.

I mean, some killing E coli more effectively than the original.

Now, the virus itself that was created is harmless.

It really only infects E coli.

But the paper, the Stanford paper that came out of it,

proving this came out to what I found very fascinating and a lot of others did.

It was from biosecurity researchers at John Hopkins and they basically call the findings a source of urgent biosafety and biosecurity questions.

Yeah, here is the actual technical problem.

DNA screening systems catch sequences by matching them against a database of known threats, things that already exist.

But this sequence has no match because nothing like it existed before it was created.

And to be fair, there's an industry -led screening system already.

I mean, the International Gene Synthesis Consortium has more than 40 gene synthesis companies that agreed to a shared screening protocol.

And even more than 175 developers of AI biodesign tools have

even signed commitments to only use the synthesis providers who follow those standards, which I think is pretty exciting.

People are seeing that this could be a real threat and taking action, hopefully ahead of time.

But all of that infrastructure really answers one question.

Does this sequence look like something we already know is dangerous?

Adoption is essentially voluntary and uneven, especially among smaller providers.

And none of it was built to catch a sequence that resembles nothing in any database.

I mean it's making something that doesn't exist yet.

And that's the real hole of Evo 2 exposed.

It's a mismatch between the rules that were set in place to catch these things and what something can do,

what a generative model can do and now produce something that hasn't existed before.

How do we put guardrails for that?

Running Evo 2, which by the way is open source right now, at full scale takes some serious hardware easily.

It's open source but it probably takes a quarter million dollars or more in compute.

It's really out of reach for someone working alone

but well within reach for a university lab or a startup

or a state program and

that barrier tends to drop over time

now one caveat worth knowing evo2 in its raw form couldn't reliably design a specific phage like phi x1 74.

it really doesn't roll off the tongue easy

it couldn't design one of those on its own

it has to take an extra step supervised to narrow the model's focus on that phage family.

So this was a deliberate targeted process by people who really understood what exactly the model needed to be doing.

Not a case of typing a prompt in and a virus comes out.

Much more complicated than that.

But it is still a real shift.

I mean, a predictive model tells you about biology that already exists, whereas a generative model invents biology that never did.

How does something like EVO2 actually work?

And why is this different from a normal biology AI tool?

I mean, they've been around for a while now.

Most AI biology tools are narrow.

AlphaFold predicts how an existing protein folds, which, yes, is extremely useful, but it's answering a question about something that already exists,

whereas EVO2 is generative.

For scale, to put this in perspective for you,

9 trillion nucleotides is enough genetic material to cover the entire human genome about 3 ,000 times over.

And EVO2 gets there with just 4 billion parameters, which is smaller than GPT -3s, going back a while now,

$170 billion, which should really tell you the power here came from what it was trained on, not from the size of the model itself.

Three years ago, four years ago, an LLM writing usable code was cool.

It's a cool party trick.

Here's how I can develop some code.

But now, I mean, it produces full production ready software.

Capability curves in AI don't move straight.

They move in jumps, in leaps.

And we've watched this happen again and again, especially when it comes to language and code.

And we're watching the opening move of the same jump when it comes to genome design in biotech.

Okay, I want to go back to the hit rate because I feel like we skimmed over that earlier.

16 out of 300.

It's not high for the 16 out of 300 viruses that actually took.

I mean, that puts this closer to the GPT -2 era than the GPT -4 era of the technology.

It's on the same curve though.

Every generative model improves on more data, better outputs, better architecture.

But let's zoom out to say two to five years.

What's going to happen with this?

I really don't think anyone's saying something like EVO2 is designing a human genome next year

but every capability curve

we've watched has moved toward the harder version of the problem

not away from it now bacteriophage run a few thousand base

pairs bacteria runs into the millions humans into the billions the gap between 16 working phage gmones

and a model that reliably designs genomes for more complex organisms is an engineering gap

and i really think it's not a a fundamental one

that will come with time all right

but this isn't the first time biology has hit a moment

like this back in february 1975 around 140 scientists actually gathered

to figure out what to do about a brand new capability

100

which at the time was recombinant dna

101

which is engineering genes across the species for a first time

102

now paul berg who was the organizer of it looked for an existing framework to follow and found nothing.

103

So instead the scientists wrote their own voluntary safety guidelines before any regulations forced them to.

104

And those guidelines really shaped biotech policy for the next 50 years.

105

How was 70s by the way 50 years ago?

106

It's wild.

107

But 50 years later scientists reconvened in the same convention center for a 50th anniversary meeting

108

and stated the reason was for the same kind of moment was currently happening

109

that happened in the 70s but this time with AI driven biotech.

110

Basically a call for the people building this to set guardrails before anyone makes them.

111

Evo2's own paper does exactly that.

112

It argues for oversight.

113

The open question I think on everyone's minds is whether

114

that instinct spreads to the next lab

115

that picks up an open source genome model

116

or what guidelines will they follow whether it will stay limited

117

to the one team careful enough in this case thank goodness to publish

118

it now i want to be clear here none of this necessarily makes ai designed medicine dangerous by default

119

i think there are so many incredible breakthroughs that we are seeing

120

because of this phage therapy against resistant infections could save lives

121

and that's the outcome his team is actually working towards

122

but it also the flip side you have to talk about

123

that because a technology

124

that can do enormous good can also do enormous harm from the same capability

125

and that is something

126

that we need to really have conversations around how to put guardrails for how to govern it before it scales not after.

127

A lot of coverage jumped straight to the scariest version of this story.

128

Researchers in the US have for the first time used artificial intelligence to design new viruses.

129

What could possibly go wrong?

130

There's also concern that the technology could create new and yet unheard of bioweapons.

131

Let's be honest here.

132

If somebody actually wanted to cause harm with a pathogen today, training a genome model from scratch is definitely not the easy path.

133

The big deal here isn't that phage is dangerous.

134

It's that the thing standing between a generative model

135

and a dangerous sequence just got proven not to work while the regulatory conversation is still stuck on chatbots.

136

Now, I'm not asking if this specific phage we spoke about is scary.

137

I'm asking whether you're comfortable with what it proves is now possible.

138

Leave in the comments your thoughts.

139

Do you like topics like this around biotech?

140

I'm kind of obsessed lately with biotech, so let me know.

141

All right, I'll see you in the next video.

下载应用

流利口语所需的一切

AI 发音评分逐句打分

音标练习掌握每个音

词汇学习建立你的词库

词汇游戏边玩边学

扫码

Download on theApp Store Get it onGoogle Play

FOR YOU

本课的词汇与口语要点

这节口语课以视频“AI Designed Viruses That Never Existed Before”为素材。视频中反复出现的词有：model, Evo, genome, viruses, design。这段视频共有 6 个句子、1694 个单词可供跟读。讲话部分时长为 10:54。说话人语速自然，每分钟约 155 个词，接近日常对话。只有 78% 的单词属于英语最常用的 3,000 词，词汇难度较高。

视频中的重点词汇

视频中 15 个值得学习的单词，附发音和释义：

单词	发音	释义
virus 名词	/ˈvaɪ.ɹəs/	病毒
genome 名词	/ˈd͡ʒiː.noʊm/	基因組 /基因组, 基因體 /基因体
sequence 名词	/ˈsiː.kwəns/	序列, 順序 /顺序
biology 名词	/baɪˈɑ.lə.d͡ʒi/	生物學 /生物学
gene 名词	/d͡ʒiːn/	基因
scientist 名词	/ˈsaɪəntəst/	科學家 /科学家
synthesis 名词	/ˈsɪnθəsɪs/	綜合 /综合
essentially 副词	/ɪˈsɛnʃəli/	本質上 /本质上, 基本上
rail 名词	/ɹeɪ(ə)l/	軌 /轨, 鐵軌 /铁轨
curve 名词	/kɚv/	彎子 /弯子
researcher 名词	/ˈɹiˌsɝ.t͡ʃɚ/	研究員 /研究员, 研究者
harm 名词	/hɑːm/	損害 /损害
whereas	/(h)wɛɹˈæz/	然而
narrow 形容词	/ˈnæɹəʊ/	窄
scary 形容词	/ˈskɛə.ɹi/	恐怖

视频中出现的短语动词

单词	发音	释义
come out 动词	/ˌkʌm ˈaʊt/	出, 出來 /出来
figure out 动词		弄清楚
go back 动词		回去, 返回
pick up 动词		拾取, 撿 /捡

值得反复跟读的句子

视频中简短而完整的句子，可以直接用在日常对话里：

Does this sequence look like something we already know is dangerous?
I mean, it's making something that doesn't exist yet.
How do we put guard rails for that?
I mean, they've been around for a while now.
Do you like topics like this around biotech?

视频中的语法

说话人最常用的结构，并附上视频中的原话：

结构	视频中的用法
被动语态 be + 过去分词 — 强调发生了什么，而不是谁做的	is projected · was released · is led
现在完成时 have/has + 过去分词 — 过去发生但与现在仍有关联的事	have been · have used · hasn't existed
定语从句 who / which + 从句 — 补充说明人或事物	people who really · parameters, which is

需要注意的发音

说话人用了 19 次缩略和弱读形式，例如 don't, I'm, couldn't。请按听到的简短形式来说。

“th” 音: synthesis /ˈsɪnθəsɪs/, breakthrough /ˈbɹeɪk.θɹuː/, gather /ˈɡæðɚ/, synthetic /sɪnˈθɛtɪk/, pathogen /ˈpæθəd͡ʒn̩/
“sh” 和 “zh” 音: essentially /ɪˈsɛnʃəli/, artificial /ˌɑː.tɪˈfɪʃ.əl/, regulation /ˌɹɛɡjʊˈleɪʃən/, infection /ɪnˈfɛkʃən/, publish /ˈpʌblɪʃ/
长单词——注意重音位置: biology /baɪˈɑ.lə.d͡ʒi/, capability /ˌkeɪ.pəˈbɪl.ɪ.ti/, essentially /ɪˈsɛnʃəli/, artificial /ˌɑː.tɪˈfɪʃ.əl/, voluntary /ˈvɑ.lənˌt(ɛ)ɹi/

中文母语者容易读错的音：

/v/ — 不要读成 /w/，上齿轻触下唇: virus /ˈvaɪ.ɹəs/, curve /kɚv/, provider /pɹəˈvaɪ.də(ɹ)/, voluntary /ˈvɑ.lənˌt(ɛ)ɹi/, developer /dɪˈvɛləpɚ/
词尾辅音 — 要读清楚，后面不要加元音: virus /ˈvaɪ.ɹəs/, sequence /ˈsiː.kwəns/, scientist /ˈsaɪəntəst/, synthesis /ˈsɪnθəsɪs/, curve /kɚv/
/r/ 和 /l/ 的区别: rail /ɹeɪ(ə)l/, voluntary /ˈvɑ.lənˌt(ɛ)ɹi/, necessarily /ˌnɛs.əˈsɛɹ.ə.li/, regulation /ˌɹɛɡjʊˈleɪʃən/, regulatory /ˈɹɛɡ.jʊ.lə.t(ɔː.)ɹi/

如何用这段视频练习

先完整听一遍视频，不要开口，记下不认识的单词。
先用 0.75 倍速逐句跟读，熟练之后再回到正常速度。
录下自己的声音并与原声对比，特别注意 virus, genome, sequence 这类单词。

什么是跟读法？

跟读法 (Shadowing) 是一种有科学依据的语言学习技巧，最初开发用于专业口译员的培训，并由多语言者Alexander Arguelles博士普及。这个方法简单而强大：您在听英语母语原声的同时立即大声重复——就像是一个延迟1-2秒紧跟说话者的影子。与被动听力或语法练习不同，跟读法强迫您的大脑和口腔肌肉同时处理并模仿真实的讲话模式。研究表明它能显着提高发音准确性，语调，节奏，连读，听力理解和口语流利度——使其成为雅思口语备考和真实英语交流最有效的方法之一。

影子跟读法: 阅读完整分步指南 →

跟读练习: AI Designed Viruses That Never Existed Before - 通过视频学习英语口语

选择练习模式

下载应用

推荐视频

Viruses: Components & Shapes - Microbiology Acellular Pre-Nursing, Pre-Med | @LevelUpRN

Did Google just dethrone OpenAI and Anthropic?

What if quantum physics could eradicate illness? | Jim Al-Khalili for Big Think

Explaining how AI is killing the PhD in less than 10 minutes

Why Ebola Keeps Coming Back

Session #1: Life Sciences in the 21st Century – A Pathway to Scientific Discovery - Rotem Sorek

This Woman Cured Her Cancer with an Insane Method

How AI is helping brands identify consumer trends faster

Editing Your DNA, with Nobel Prize Winner Dr. Jennifer Doudna

Inside the secret race to build artificial intelligence | The Gray Area

Explanation of Vaccine Adjuvants

We Can Now Simulate a Human Brain, Scientists Show