Shadowing-Übung: Pick Problems That Survive When AI Solves Coding - Ara Khan, Cline - Englisch Sprechen Lernen mit Video

Übungsmodus wählen

Jeder Modus öffnet einen Übungsraum im Vollbild.

419 Sätze
Wenn Sätze zu kurz oder zu lang sind, klicke auf Edit, um sie anzupassen.
1
Hey guys, this is Era.
2
I'm a founding engineer at Klein, and today I'm going to be talking about recursively self -improving coding agents
3
so this is a very interesting conversation in the sense
4
that everyone wants to understand what rsi is how we accomplish it how we go through it
5
and how it affects our career
6
so i'm going to start with giving you just three goals of this entire conversation just three simple goals
7
and that would help you understand like what we get out of it
8
and what the future uh for this looks like
9
so the goal number first is that understanding the intent of the Frontier Labs, right?
10
Everyone who's outside of the Frontier Labs is always trying to speculate, like, hey, what is the topic in OpenAI?
11
Or what is DeepMind going to come out with?
12
Or what are they trying to make?
13
And what's the future going to look like?
14
Who has cracked the RSI?
15
So, like, this kind of helps you get a rough outline of what their roadmap looks like.
16
And the second thing is that, like, if the models are improving so fast,
17
like, how do you think about your career over the next few years, right?
18
That is really what we want to understand.
19
We want to understand, like, okay, things are changing so fast like the ground underneath us is shifting
20
so quickly right now like it's like there used to be a time
21
when you could like demonstrate your side projects on your personal website
22
or github and now it's just like you could literally one shot them using any frontier coding vision and with
23
that it's like it's like oh well how do we demonstrate proof of work how do we demonstrate you know
24
that we know what we're doing
25
and how do we think of our career for the next coming years right
26
so it's very interesting times that we're living in
27
and then the third and the most important goal of the conversation is escaping department underclass right
28
so i know this seems like a joke
29
but this really is a very valid possibility
30
that people are contemplating everyone's like oh well if everything
31
that we want to be doing can potentially easily done by
32
a frontier model like what are we what are we good for
33
and like where does software go where where is the future of technology going to go
34
and what are we going to do as as things evolved
35
so fast right so i think that there's like a lot of like fear mongering
36
but also there's like some very valid merge where people are concerned
37
that like if all the revenue of the planet is going to go to um a few hardware companies
38
that make a frontier gpus and then a few few uh ai companies
39
that that make the best models like how are we how are the rest of us going to survive
40
and what are we going to do here
41
so i think i'll walk you through a few examples and a few ideas that
42
Like really, really help you reconsider the career choices and the life choices you're making and what your problems you're picking up.
43
And then that would help you like realign what it is
44
that we should be working on and what it is
45
that you could work on that would stand the test of time
46
and help you both just not just like on a psychological level but like financial and other levels as well.
47
So to give you a very precise example, let's start with something very precise that I had done.
48
So a couple of months ago, I came up with recursive improvements on IA coding harness.
49
So this was the first public example of an automatic eval improvement.
50
So to give you an example, we used to do this thing where we would measure the quality of a coding agent using evaluation scores
51
so on terminal bench 2 .1 kimmy k3 with klein with
52
score um usually like you know um 60 70 percent
53
and then depending on like how good the model is like you get better scores on terminal bench 2 .1
54
so we noticed that when kimmy k3 came out our score was 77 .5
55
which is a good score
56
but we were like oh is this is this truly is this truly the best score
57
that this model can get on this benchmark
58
and also like what does the model release paper claim about what their best score is
59
so the model's release paper said that they could score up to 85
60
or even more so we were just like okay okay
61
so we are not getting the sort of score
62
and often what we do with model releases incline is
63
that we try to like optimize your score
64
and try to get the best out of the model so
65
when we started hill climbing on the on the model like we we realized this we realized that like,
66
you know, there's a lot of steps that we do, says that we do it gracefully,
67
says that we pursue the right kind of problems and do it the right way.
68
And we started thinking like, okay, there's a lot of groundwork here.
69
How do we automate it?
70
And this is a demonstration of like how we automated it.
71
So I'll just leave that as a placeholder for now.
72
And we'll get back to it later where we talk about in detail of like, how does this automatic improvement loop of a client harness work?
73
So this basically means that like you have a coding engine harness
74
that could just like basically improve itself and become better on the same set of problems.
75
So let's start with the ground terms.
76
What is recursive self -improvement?
77
Recursive self -improvement is the idea that an AI system that helps build a more
78
capable successor
79
which repeats the process you know until you hit our our artificial general intelligence right
80
you have better systems you have better improvements you have a better system
81
and then you know you have singularity
82
so a class example is like a self -reporting coding agent right
83
it's after brewing coding agent is just like a coding agent
84
that can improve itself
85
so you know this this is all like a pie in
86
the sky idea like it's like it's like you know yeah sure this looks cool
87
but how did this even come out to be like where how did we even get to here so
88
So if you notice in Topic Roadmap in 2021, in 2021's roadmap, they very specifically said
89
that they want to develop a general AI system that can interact conversationally with a user to create, edit, and explain software.
90
It would display all the capabilities of a software engineer colleague.
91
It would write functions.
92
It would identify bugs.
93
It would design an API, discuss style, and best practices, right?
94
So we don't expect to reach this goal right away, but it's in the direction in which we're striving.
95
It's interesting how this was more than five years ago.
96
More than five years ago, people had predicted
97
that there would exist a day where you will have like a full -blown general assistant
98
that can basically act like a colleague, which we do have right now.
99
Then in 2022, ChatGPT got released.
100
We all know how that happened.
101
It was a fun moment.
102
It was like the first time things got sparked for us.
103
Then in 2023, 2024, there was the first time
104
Chat's models could call functions.
105
So these functions would be simple functions like, hey, what's the weather or, you know, what's the what's,
106
you know, what's this competition or what are these competitions result in Python?
107
And then finally, in 2024, when Sonic 3 .4 .5 was released,
108
that was when the agents could like truly call like actual tools the way we call like think of tools.
109
So it's like reading a file, editing a file, you know, reading, searching something on web, listing, listing a folder, searching and grabbing through,
110
you know, a file system, things like that.
111
So, and in 2024, we had the intuition that, like, chat box, what if chat box could actually talk to each other, you know?
112
So then that was the first time when we had, like, you know, multi -agent conversations.
113
Now, at the time, these things were very janky, right?
114
Like, they were just, like, taped together in this, like, weird, weird, bizarre ways where you have just, like, these classes and these things and then,
115
like, these hierarchies of classes.
116
And it was, like, it was all, like, kind of, like, a very janky attempt to, like, make us think that,
117
like, These models could have these pipeline of flows or responses that can all communicate with each other.
118
But a lot of it was mostly deterministic, and it was not as free -flowing as it is right now.
119
And we started getting the intuition.
120
So this was the autogen project.
121
But only in 2025, we started getting the decisions, like, what if the agents just kept getting better and faster and cheaper?
122
What if that just kept happening?
123
And then in 2026, we had the intuition that, like, what if the agent could improve itself, which is, like, the final,
124
you know, the final part of the whole process.
125
So let me give an example.
126
Like, what does it mean for an agent to improve itself, right?
127
So, like, it's, like, a lot of times when we think of,
128
like, evaluating an agent, we initially used to think of evaluating an agent in the form of, like, just a single -term conversation.
129
So we would tell agents that ask a simple question, like how many toes does a dash animal has?
130
It could be, let's say, how many toes does a cat have?
131
And then the element would go through and the element would prepare a response and then we would grade the response.
132
So like a cat has maybe 10 toes, 18 toes, whatever many, I actually don't know how many toes a cat has. But anyway,
133
yeah, so then we grade the response as like a yes or no. But like when you work with an agent,
134
there's a ton of things that you're doing.
135
So let me give you an example.
136
Let me give you an example.
137
So, for example, you're going to ask an agent, like, hey, set up this MCP for me.
138
So now it's not just like, you know, a trivial one -shotable thing.
139
Like, it will, like, search online, like, oh, what's the MCP?
140
What is the config?
141
So I'll, like, edit and download some soft files.
142
Like, maybe I need to database update the MCP tool.
143
Maybe I need to read and edit some files.
144
Maybe I need to run some bash commands, right?
145
So you do all of that, and then the agent responds, like, hey, I did it.
146
And then you grade that with a bunch of tasks to gauge,
147
like, hey, did the agent actually accomplish the thing that it said it would do?
148
So to do that thing, the second part, which is, like, more intricate, where you're actually, like, truly, you know, testing agents,
149
like, to their best extent, like, it's, like, you know, we use this thing called Terminal Bench.
150
Terminal Bench is a benchmark for, you know, AI agents in terminal environments.
151
And that is something that we have a lot of, like, experience with at Client.
152
And it's like it's like one of the best ways to
153
like really gauge like is your coding agent accomplishing in general the task
154
that you're doing that you want to be doing with a certain model.
155
So there's a couple there's many different parts of like evaluation suit.
156
So ideally, you want that to run in, you know, you want that to run in individual sandboxes.
157
Then you have like independent graders that like check everything like.
158
How many turns it took, how many tokens it took, how much latency it took, how many tool calls did it make,
159
what was the trajectory looking like, and then grading all the results for things like, hey, did you not reward a hack?
160
Did you not cheat through the solution?
161
Did you not just look up the solution online?
162
Things like that.
163
So basically, a lot of these data sets of evaluation problems have a set of tasks.
164
So in case of terminal bench 2 .1, they have 89 tasks.
165
So what we do is like we spin up 89 parallel containers where they all run online
166
and they run in parallel and they're sandboxed and then they each try to solve a problem.
167
So then they're all in parallel and independent of each other.
168
And then at the end, we get a score.
169
So our score could like give us measures like is the model performing well and what kind of failures it has.
170
Right.
171
So like you run the evaluation, you get a score.
172
So you get an aggregate score.
173
And the aggregate score could reveal a set of failures.
174
Failures could be of any kind.
175
Failures could be, hey, the refile tool is not working.
176
Sometimes the inference is so slow that, like, you know, the agent's working, it's doing, it's making progress.
177
Maybe it's running a circle, but it's so slow that it literally, like, runs out of time.
178
It runs out of, like, 45 minutes or whatever.
179
Many that were given for the task and just, like, stopped working.
180
So once you do that, once you understand, hey, this is what we ran, this is what the failures we got,
181
we bucket the failures in different categories.
182
So, like, 10 failures of, you know, this kind of tool called 10, 20 failures of, like, say, inference not working or something else, right?
183
And then once we bucket the failures into different categories, we're, like, we try to make an improvement to the harness.
184
Because, like, what if there's, like, something wrong with the harness that we could change just that the model score would improve, right?
185
And on and on and on.
186
You do this iteratively.
187
It takes a while.
188
It's just, like, you know, hill climbing is, like, a very groundwork -y kind of process
189
and you know usually usually we we took we used to take a lot of time trying to do this
190
but it was a great way for us to like learn how to improve models
191
so one way to help climb on public evals is that like
192
and you know we could have a thing where you know
193
we started doing this thing where we could just like what
194
if we could have like a long -running agent
195
that runs for a while and automatically improves the model score
196
and that's what we did so we ran gpd 5 .6 soul
197
that led the work and that was like the main agent
198
and then uh kimi k3 was just running evals
199
so what it was doing is
200
that like dbd 5 .6 was like hey run these evals
201
get these scores let me look at the failures let me look at where it went wrong
202
and then i'll make changes to client's harness
203
and then i'll run these evals again and then i'll keep trying
204
so again you can see how like very quickly this can like go out the rails
205
because this is supposed to be like the running evaluation itself takes like an hour
206
then you have to read the feedback then you have to like see
207
if you're properly interpreting the feedback to make reasonable changes to the harness
208
that are like not reward hacking they're just like actual genuine improvements
209
and then run the uvl again so we ran
210
that for 17 hours and when we ran
211
that for 17 hours there was like a legitimate hell climbing
212
and what we found that was very interesting this happened with a very detailed prompt
213
that accomplished this is that like we had a baseline score of 207 percent and then
214
And, you know, we found issues like loop detection.
215
Loop detection is when a harness just runs in circles around the problem, right?
216
We found, like, background process grinding, where background processes would run too long.
217
We found some file index issues and other things as well.
218
I mean, when I say we, what I've actually made is GPT -5 .6 solve, right?
219
So that found those issues, that found and bucketed those issues.
220
It's like, hey, how about I make those changes in the harness and I run the eval again?
221
And it iteratively built the hill climbing process completely autonomously on its own without my interference.
222
And 17 hours later, when I come back to it, it has improved how to harness its performance in so many ways.
223
So see what does that mean?
224
What does that mean?
225
It just basically means that we are at a point where the experience could make the harness better.
226
That there's a two -level learning loop where the harness does some action,
227
interacts with the environment, understands whether or not it did the right thing or the wrong thing and gets some reward.
228
And like almost, and then has, there's like a meta agent that's observing the main harness
229
and finding the changes to it and just like incorporating those changes and improving.
230
Now, I think in my personal opinion, it's much harder to do in real time because if you do it in real time,
231
you over -index on solving certain problems or you're just like too curtailed to a specific thing.
232
But I do think that this is something
233
that works like on a broader scale of like working on a wider range of problems.
234
And you want to be able to justify the changes you're making on a broader applicability
235
so that the changes you make here will apply to a broader range of problems as well.
236
So what does that mean?
237
What am I trying to get at here?
238
The point I'm trying to get at is that the models are
239
so good right now that if they can improve the harness, they can improve a lot of other things.
240
So then what else can they improve?
241
So this is a post.
242
primarily from OpenAI, where GPU to 5 .6 SOL autonomously rewrote and optimized its own production kernels,
243
and those production kernel improvements removed the end -to -end serving costs by 20%.
244
It also had 15 % higher token generation efficiency, and this was completely done by 5 .6 SOL on itself.
245
So what that means is that inference is becoming much cheaper, and that means that if for a company like OpenAI,
246
If you're having 20 % lower serving cost for a frontier model, that could be billions of dollars of improvement in price.
247
And you could do that as well for your own models.
248
Another interesting thing is that when people have kernel writing competitions, so Core Auto is a company which has kernel writing competitions.
249
And you'll notice that they mentioned that the person who won the kernel writing competition,
250
the XLCC said that I've never handwritten a GP operator it's all ALM and agent driven.
251
So notice that like how like the best inference improvements, kernel improvements are at this point completely exclusively done by a swarm of agents
252
or generally just like agents doing some magic to you know magically improve like the kernel code and have faster inference, cheaper inference, yeah.
253
So basically like the improvement surface that you think of is not just like the harness improvement, it's literally the whole surface stack right
254
so the whole stack from like evaluation to harness improvements to
255
training data to post training to world interaction to inference like
256
literally the entire part of the pipeline like all the way
257
from like electricity going through your machine to like making an error request going through the model models doing inference
258
and making a response
259
and getting back to you like every single part of the pipeline we can optimize it
260
so the whole rsi loop has like um there's a couple different parts.
261
So it's like pre -training, Nematron examples of like, you know, AI curated pre -training data that was good.
262
There's like post -training examples, which are, you know, where there's like constitutional AI examples of like post -training improvements.
263
Then inference, I gave you the example of GPT -5 .6.
264
So for evals and eval generation, I give you our own example of how we improved ourselves on our own evals.
265
In general, what you want to think of is
266
that like anytime a model is trying to do something
267
and it's like not able to do it
268
or not able to accomplish it
269
so for example like imagine you're having a long conversation where you ask the model something obvious
270
so like it's like or something that you think is not necessarily obvious
271
but it's like it's hard but you would assume
272
that a model could potentially do it
273
so to give an example let's say you want to make add a new column to a database
274
and then fill that column right and then put a cache in front of it
275
so what happens is that that's the whole thing
276
but for some reason the way
277
it pushes the cache in front of the whole database like it kind of doesn't do it right
278
so at that point you were like oh no set this cache differently set this as the ttl
279
and then you gave like a correction signal and then
280
that kind of like fixes your problem
281
and you kind of go about your day
282
but the way i want to think of it is
283
that i want to think of like every single time like
284
a human correction to the digital landscape is a part of a delta that needs to be covered
285
And what that means is
286
that every single time you're finding the edge of model capabilities of how far it can push for infants, how much it can push for this,
287
how much it can push for solving your problems, you want to find the delta and we want to push that frontier a little bit further,
288
a little bit further.
289
And that's what post -training does, right?
290
So post -training is in that post -training.
291
We find these really interesting and really hard long horizon problems for coding agents or for other kinds of problems.
292
And then just like slowly he'll climb the models on those kind of problems.
293
So anything that has a verifiable loop can be optimized.
294
But this is not just limited to software.
295
Like literally like the physical world itself could have a verifiable loop.
296
So to give an example, like let me give you a very funny kind of bizarre example where like it's like you look at this
297
and you look at this like robot and it's it's typing something
298
and it's like really really like interesting what the robot is typing
299
so notice how like notice how the robot is itself typing like how it wants to like solve robotics
300
and imagine if the robot itself could like interpret the results of of like the answers
301
and just like keep improving then that means that like this is you know basically like an ASI robot.
302
Another example that I could give you from my personal life is
303
that like yeah I have this chip that has the ability to see
304
and has a speaker and has a microphone inside of it
305
so I was trying to set up this with my machine and it was
306
GIFs that I had for my girlfriend
307
and what I find very interesting about this is as I was trying to set
308
it up it's very slow it was slow to respond it had higher latency it had an agent in
309
and of itself right so what I did was I connected this chip to my computer
310
and I let I let coding agents run on this hardware
311
and basically run on this hardware automatically make tests to it so like
312
my computer would talk to my hardware give it some tests like hey do this do
313
that and just like what do you see right now
314
so then it would notice the end -to -end latency of
315
like the inference response of like how fast my hardware is responding
316
and then what I found is
317
that like it was able to do the breakdown of like every single case and scenario where there was like higher latency,
318
slower response, or just the algorithm wasn't right.
319
And I could then set up a recursive loop where both the firmware could be updated, the tool calls, the inference model used,
320
the speech -retext model used, all of those could be improved automatically with like an open router key, which I'd given to a coding agent.
321
And then overnight with self -running experiments with the hardware, my this tiny little piece of beautiful hardware
322
that I have was like deeply optimized
323
so it was this also like even hardware could be a part of the RSI loop
324
so what happens
325
when you connect these loops right
326
that's the next question like you could have cheaper inference
327
you could have like more experiments budget you have like you know better data
328
and better models so is this singularity it's a right question to ask
329
and I think
330
that the part of the parts of the loop are already right like we're we we are really
331
we really are very close to parts of the rsi loop right
332
and what do we do like what do we do about this
333
and it's like how do we think of
334
that in terms of like our career and how do you think of
335
that in terms of you know our lives
336
and what do we want to do next
337
so here's my advice to everyone who's watching
338
this like please please please please move closer to the rsi loop right
339
so it's like right now like people are working on all these like different kinds of software
340
which is like you know agents for you know construction agents for this agents for
341
that and it's like i think
342
that you're you know well the way i see those things is
343
that like i think of those as like you're missing the forest for the for the trees you really are
344
and it's like you have this like very interesting god -like technology
345
that can change the course of humanity
346
and then we're not ambitious enough to pursue the best problems we really are not
347
and we're just like i think that it's not just like the problem isn't just
348
that we're not ambitious
349
and we're also like what we're pursuing a lot of those
350
things are just like one model resell away from like being completely like wiped out right
351
so like it's also like it's also like it may not be
352
that intellectually simulating because like all you're essentially all we're essentially doing is like sitting all day
353
and it's like hey cloud do this for me a cloud do this for me
354
and i don't know if that's like i don't know
355
if that's like the most intellectual simulating or like uh
356
or like um thoughtful use of a person's time like maybe maybe you know we're supposed to be solving harder
357
and more uh more frontier problems and i think
358
that like it's like the ceiling is like our ambition
359
so instead of like working on these tiny things my suggestion to people is to like you know do things
360
that of course like goes without saying
361
but you should do things you believe in
362
and you know i'm just a guy on the internet
363
but do things to help you move closer to the rsi loop right
364
so like anything that's just like the end -to -end pipeline of like
365
hey this is where like the models are lacking like it's like whether that's inference whether that's post -training, whether that's literally like physical interaction,
366
whether that could be, you know, sending end -to -end pipelines of truly any kind, you know, whether that's hardware problems,
367
whether that's networking problems, whether that's building a sense of community, things that just, you know,
368
things that just like help you as an engineer get closer to the benefits of the improvements in the model.
369
So this is a very interesting slide from a topic.
370
So they mentioned that like you want to have like this, you're in this space where it's like,
371
you know, you only you only have like one to five years left to contribute to luxury to the world.
372
What do you do?
373
Right.
374
So it's like they gave great suggestions where like prefer ambitious, targeted and fast collaborations.
375
Right.
376
And, you know, in general, like the timescale, like all the timelines that you give to yourself,
377
like they should be just completely squeezed out you completely squeeze out
378
because it's like a lot of times and you also have to think through
379
that like if if the model if the problem
380
that you're working on is too ambitious like it's like is
381
it probably is it probably better to wait for just like the next next frontier model to come out
382
or not like that's that's something you also have to wait through right
383
and uh you you ideally want to work on problems
384
that benefit from scale and you know you also at the same time want to avoid projects
385
that can be solved by save alone because then you know
386
if even if you like it's like if working on a problem
387
and then the next model release comes out it solves the problem
388
so that's not that's probably not a good thing to do
389
so it's very like you have to like think philosophically of Like, what is the endeavor that benefits uh greatly from other improvements
390
but like isn't also completely wiped out from it
391
and then ideally like it's like if the
392
if the endeavor could also uh be a part of the the even like the improvement pipeline uh
393
that could be amazing
394
so you see a lot of companies raising on say um making data pipelines
395
and there's a lot of money there there's a
396
a lot of companies working on physical hardware and stuff
397
so i think if you're a software engineer watching these it's like I think
398
that if you're if you feel like you're like here the delta of your intellect
399
and ability in your day -to -day
400
it's like not contributing as in like you have the intellect you can make a difference
401
but like if cloud is just like one -shotting all those problems like in your day -to -day
402
and you're just like kind of like becoming a meat proxy
403
you should work the thing through take a stop back
404
and think through it's like okay where are the problems where you're actually truly uh contributing
405
and that they will stand the test of time and then
406
And that kind of reflection and realignment right now would,
407
I think, like have like very interesting consequences for you in the long term. So, yeah.
408
So thank you so much.
409
You know, thank you so much for your time.
410
Like, I think that I think that if you're watching this and if you've gone this far,
411
I think that you are definitely contemplating like one of the most interesting improvements in the course of humanity.
412
And, you know, there's many philosophical arguments here.
413
Some people might say, hey, you know, I think I maybe I'm off on the timelines
414
or maybe maybe I'm just being too pessimistic or maybe I'm being too optimistic.
415
But I would love to hear all those thoughts.
416
So this is my QR code.
417
Please feel free to scan it and let me know all your thoughts.
418
Thank you so much.
419
Very kind of you to give me your time.

App herunterladen

Alles, was du zum fließenden Sprechen brauchst

KI-AussprachebewertungJeder Satz bewertet
IPA-TrainingJeden Laut meistern
WortschatzBaue deinen Wortschatz auf
VokabelspielSpielend lernen

Gleicher Kanal

FOR YOU

Empfohlenes Video

Wortschatz und Sprechhinweise zu dieser Lektion

Diese Sprechübung basiert auf dem Video „Pick Problems That Survive When AI Solves Coding - Ara Khan, Cline“. Dieses Video enthält 419 Sätze und 5370 Wörter zum Nachsprechen. Der gesprochene Teil dauert 26:30. Der Sprecher spricht schnell, etwa 203 Wörter pro Minute – rechne mit verbundenen und verkürzten Lauten. 86 % der Wörter gehören zu den 3.000 häufigsten im Englischen; den Rest solltest du dir vor dem Üben ansehen.

Wichtiger Wortschatz in diesem Video

15 lernenswerte Wörter aus dem Video, mit Aussprache und Bedeutung:

WortAusspracheBedeutung
harness Verb/ˈhɑɹ.nəs/vorspannen, aufzäumen
loop Substantiv/luːp/Schlaufe, Schlinge
hardware Substantiv/ˈhɑɹdˌwɛɹ/Eisenwaren
frontier Substantiv/ˈfɹʌntɪə/Grenzland, Grenzgebiet
solve Verb/sɒlv/lösen
pipeline Substantiv/ˈpaɪpˌlaɪn/Pipeline, Rohrleitung
terminal Substantiv/ˈtɚmɪnəl/Terminal
climb Verb/ˈklaɪ̯m/klettern, steigen
evaluation Substantiv/ɪˌvæljuˈeɪʃn̩/Beurteilung
kernel Substantiv/ˈkɝ.nəl/Kern
bench Substantiv/ˈbɛnt͡ʃ/Bank, Sitzbank
robot Substantiv/ˈɹoʊ.bɑt/Roboter
toe Substantiv/təʊ/Zeh, Zehe
edit Verb/ˈɛdɪt/bearbeiten, redigieren
accomplish Verb/əˈkɑm.plɪʃ/vollenden vollziehen, ausführen

Phrasal Verbs, die du hören wirst

WortAusspracheBedeutung
come out Verb/ˌkʌm ˈaʊt/herauskommen
set up Verb/ˌsɛt ˈʌp/aufbauen, für den Einsatz vorbereiten
get back to Verbsich melden
think through Verbdurchdenken
come up with Verbausdenken
go out Verbhinausgehen, herausgehen
look up Verb/ˌlʊk ˈʌp/bergauf gehen, aufwärtsgehen
pick up Verbaufheben, aufnehmen

Sätze zum Wiederholen

Kurze, vollständige Sätze aus dem Video, die du im Alltag wiederverwenden kannst:

  • How do we automate it?
  • It could be, let's say, how many toes does a cat have?
  • Did you not cheat through the solution?
  • Did you not just look up the solution online?
  • What am I trying to get at here?

Grammatik in diesem Video

Die Strukturen, die der Sprecher am häufigsten verwendet, mit den genauen Worten aus dem Video:

StrukturIm Video
„Used to“ used to + Verb – eine frühere Gewohnheit oder ein Zustand, der nicht mehr giltused to be · used to do · used to think
Passiv be + Partizip Perfekt – wichtig ist, was geschieht, nicht wer es tutwas released · were given · was completely done
Relativsätze who / which + Satz – eine Zusatzinformation über eine Person oder Sacheitself, which is · part, which is · company which has
Present Perfect have/has + Partizip Perfekt – eine vergangene Handlung, die jetzt noch wichtig isthas cracked · I've actually made · has improved

Aussprache, auf die du achten solltest

Der Sprecher verwendet 67 Kurzformen und abgeschwächte Formen, zum Beispiel you're, we're, I'm. Sprich sie in der kurzen Form, so wie du sie hörst.

  • Die „th“-Laute: underneath /ˌʌndɚˈniθ/, algorithm /ˈælɡəɹɪðm̩/, thoughtful /ˈθɔːtfəl/
  • Die Laute „sh“ und „zh“: evaluation /ɪˌvæljuˈeɪʃn̩/, accomplish /əˈkɑm.plɪʃ/, ambitious /æmˈbɪʃ.əs/, cache /kæʃ/, intuition /ˌɪn.tjuːˈɪʃ.ən/
  • Lange Wörter – auf die Betonung achten: evaluation /ɪˌvæljuˈeɪʃn̩/, ideally /aɪˈdi.ə.li/, intuition /ˌɪn.tjuːˈɪʃ.ən/, essentially /ɪˈsɛnʃəli/, potentially /pəˈtɛnʃ(ə)li/

Laute, die Deutschsprachigen schwerfallen:

  • /w/ — mit runden Lippen, nicht wie das deutsche „w“ (/v/): hardware /ˈhɑɹdˌwɛɹ/, reward /ɹɪˈwɔɹd/, wipe /waɪp/, consequence /ˈkɑn.sɪ.kwəns/, squeeze /skwiːz/
  • Stimmhafter Auslaut — /b/, /d/, /g/, /z/, /v/ am Wortende nicht verhärten: solve /sɒlv/, cloud /ˈklaʊ̯d/, reward /ɹɪˈwɔɹd/, valid /ˈvælɪd/, gauge /ˈɡeɪd͡ʒ/
  • /æ/ — offener als „ä“: evaluation /ɪˌvæljuˈeɪʃn̩/, ambitious /æmˈbɪʃ.əs/, parallel /ˈpæ.ɹəˌlɛl/, cache /kæʃ/, valid /ˈvælɪd/

So übst du mit diesem Video

  1. Höre dir das ganze Video einmal an, ohne zu sprechen, und notiere die Wörter, die du nicht kennst.
  2. Beginne mit 0,75-facher Geschwindigkeit, sprich Satz für Satz nach und wechsle zur normalen Geschwindigkeit, sobald es leichtfällt.
  3. Nimm dich auf und vergleiche mit dem Original; achte dabei auf Wörter wie harness, loop, hardware.

Was ist die Shadowing-Technik?

Shadowing ist eine wissenschaftlich fundierte Sprachlerntechnik, die ursprünglich für die professionelle Dolmetscherausbildung entwickelt und durch den Polyglotten Dr. Alexander Arguelles populär gemacht wurde. Die Methode ist einfach aber wirkungsvoll: Du hörst englisches Audio von Muttersprachlern und wiederholst es sofort laut — wie ein Schatten, der dem Sprecher mit nur 1–2 Sekunden Verzögerung folgt. Anders als passives Hören oder Grammatikübungen zwingt Shadowing dein Gehirn und deine Mundmuskulatur, gleichzeitig echte Sprachmuster zu verarbeiten und zu reproduzieren. Studien zeigen, dass es Aussprachegenauigkeit, Intonation, Rhythmus, verbundene Sprache, Hörverständnis und Sprechflüssigkeit signifikant verbessert — was es zu einer der effektivsten Methoden für die IELTS Speaking-Vorbereitung und reale englische Kommunikation macht.

Shadowing-Technik: die vollständige Schritt-für-Schritt-Anleitung lesen →