Shadowing Practice: Why AI Tokens are so Expensive - Computerphile - Learn English Speaking with Video

בחר מצב תרגול

כל מצב פותח חדר תרגול במסך מלא.

407 משפטים

אם המשפטים קצרים או ארוכים מדי, לחץ על ערוך כדי להתאים אותם.

I thought, given how much it's been in the news about costs of AI and, you know, agentic coding and how much you pay per month and all this stuff,

I thought we could look at what is a token and why is it so unbelievably expensive?

This is happening everywhere.

So, for example, Amplopic have more caps on.

For their premium users, you might hit a cap where you have to go

and wait for a while sooner than you used to.

for the famous recent one was the GitHub copilot change,

which changed from a request-based monthly cost to a AI token credit monthly cost,

which has substantially reduced the amount you can do per month.

And we can talk about why that is and why essentially it was inevitable, given how expensive it is to run those agentic coding models.

Well, you could call this video, What is a Token?

But we're going to answer that straight away right the token is either a word

or a piece of a word that is it it's not an interesting question right

so if I write the cat sat on the mat you might intuitively

if you're a person split this up per word

but we don't incorporate spaces kind of in our head we

just use them to find the words large language models and modern AI

will use the spaces as well full stops will be tokens curly brackets will be tokens

if you have chinese characters they might be one

or two tokens right because they're more complicated

so you might have you know this might be a token

and then this might be a token with the space included

and you know you might have

that i'm just making this up right you know something like

that the exact split will depend on who's running the model

and the tokenizer they're using they're based around this idea of frequency

so for example because the is so common on its own it would make sense for it to be a token.

A modern language model will have access to let's say a hundred thousand of these tokens

which will include standard British things but also other languages

and code symbols and special Unicode characters and things like this right

so most of the tokens will be in a small subset of a few thousand

that get used most of the time but there'll be many more

that it could produce if the if the need arose

but they haven't been decided they

if they emerged no they've been decided they've been decided the tokenizer is essentially a string parser

that goes in front and at the end of a large language model what happens

when these go through a network is these get turned into a numerical representation called an embedding

which means that this one will have some high dimensional vector of numbers

that represents what the word the means right not a lot

and cat will be different and app will be different and

so on that part is learned so

that part will be bespoke for whichever large language models being used

but they probably use whichever token either they've always used

so we did a video a

while back with lewis where we did i don't know mike

in a nike on a hike mike on a bike yeah

a bike on a mic mic at a strike yeah yeah different tokens

but they're different tokens

but they would be in some senses in similar similar words

will be in similar space probably right it will come out of a training process it makes sense

that they would be um the in some sense the tokenization

part is is almost a sort of system development issue one thing

that is interesting is you would think well why would you why would you have a hundred thousand tokens

when you're writing let's say javascript

because javascript only uses a subset of the you know lots of them aren't valid tokens after a an

if statement there's only a certain number of tokens you can

you can predict otherwise it doesn't make sense that's true

but part of the power of these models is their huge size

and their massive pre-training and

so it actually isn't particularly cost-effective necessarily to have a model

that's only trained on code what you tend to do is refine a model that's trained on everything else

so you will just have all the tokens included

that could do javascript or they could write english poems

and then later you will refine that model into a specific domain by giving it extra examples.

That presumably helps with people prompting it as well and asking it to do things for them.

Yes and actually you can ask these models to do things in base64

and they will pretty much do it right because to it it's just input tokens.

This isn't in some sense this isn't the interesting part.

The interesting part is

is when you talk to a model how much does it cost

because that's what we all want to know

and the answer is right the answer is we're not sure but actually we can come up with some guesses.

these models work in something called an auto-aggressive way and what

that means is

that they put they take in this input they look at

the whole input as a big list they make a lot of decisions

and then they output a single next token so

if I gave it the input the cat sat on the

right these tokens here it would with high likelihood return as its next prediction this token here

and that's true of code and all other kinds of things

if I write half an if statement and I ask it to complete it for me, it will use the rest of the context and its pre-training to produce likely output tokens.

The issue is, and the thing that perhaps a lot of people haven't quite, certainly, you know, in the media hasn't grasped and in sort of standard audience hasn't grasped,

is how unbelievably inefficient that is, right?

So when you ask me a question and I answer it, what I don't do is quickly reread the whole transcript of this previous conversation in order to answer it.

I suppose I might if I've completely forgotten what I'm saying.

But in general, I can hold some things in my head and then I can answer the question.

Large language models don't work that way.

Everything goes in every single time.

So when you want to put the next statement in and the next statement you're putting in this, then you're putting in this, then you're putting in this, and it gets bigger and bigger and bigger.

And that comes at a cost.

It comes at a particular cost for coding agents where the size of its input goes up hugely

and that's I think where it gets quite interesting and the cost becomes a little bit confusing.

Let's imagine we're just on a website talking to a large language model.

In some sense it's not a very difficult question, I ask it 500 tokens worth of question,

it produces some thousand tokens worth of thought process which is the bit it writes before it answers, but may or may not be seen by the user.

And then maybe it outputs a 500 token response.

So the actual whole token cost and input tokens is my input.

I mean, let's look at a concrete example.

You might have a system prompt, which let's say is 1,000 tokens.

100

Conservatively, they'd be much bigger than this.

101

And that's things like you are an AI agent.

102

Yeah, don't tell them what the system prompt is.

103

Don't answer rude questions, this kind of stuff.

104

And the current time and the operating system are on, and maybe some actual user-specific stuff.

105

You then have the user query, which let's say is 100 tokens quite a short question.

106

Then you have the thought, which is the output tokens, but it produced in response to these two things added together.

107

So the actual input to the model was 1100 tokens.

108

And then for the first thought word, the output then goes back in.

109

So now it's system plus user query plus the first word of the thought.

110

So it's 1101 tokens, and then 1102.

111

And this loops over and over again.

112

Oh, so like you said, when it's predicting the word mat on the cat sat on the, after mat, you have to have the cat sat on the mat again to work out what comes next.

113

Yes.

114

Yes.

115

You don't just say you said mat, what's the next word?

116

You have to say, don't forget all this other stuff that came in first.

117

Right.

118

Now, we'll talk in a minute about caching because that's a big thing for speed ups and efficiency.

119

But that's basically what happens.

120

So if this thought process is 10,000 tokens, I asked a complicated but brief question right

121

and then its response is let's say 1 000 tokens there's a lot going on

122

because the actual output was what was 10 000

123

and another thousands 11 000 tokens right

124

and the input was symbolically 1 100 tokens

125

but that's not how it works it works on a loop

126

and so actually it ran a forward pass through a gpu 11 000 times right

127

and each time the context in the gpu got slightly bigger right with with the previous thing

128

which is why it costs so much electricity to run these models right

129

because you don't just say here's your output sentence you say

130

here's the first word of your output sentence let's go again now it's not quite as inefficient as this

131

because I'm being a bit unfair right

132

because actually what they do is something called kv caching

133

which I think probably will leave to another video

134

but ultimately what happens is if you're adding the word map to this sentence

135

and then you add another word let's say the word to right new sentence

136

so this is a new token

137

when you actually put this through the network to predict the

138

next word we don't have to calculate the relationship between cat

139

and at and cat and on because we already did

140

that the previous go so we cache those values

141

so a lot of the intermediate representations in the network are stored as you go

142

so that when you add another one you don't have to completely do everything from scratch

143

which makes a huge difference for very very big context windows

144

for this sentence it makes almost no difference at all but

145

if you've got a million documents you're putting in you're trying to find the best bit of each one

146

and your context window is full up

147

and it's all gonna this will this will help of course

148

there's an implementation question of how long do you keep this cache live for

149

if i ask a long question of a large language model it caches all of these values

150

and then i go off for a coffee

151

and come back later it's got to have dropped

152

that from the gpu because it's got other people to talk to right

153

and so in actual fact it will then have to put all those back in again and That's called pre-filling.

154

It's almost a system design issue, right?

155

The more aggressively you can cache some of these things the

156

better your performance will be To the point where it becomes it's not helpful, right?

157

Or the user is too slow

158

Asking questions and you've had to get rid of the cache

159

because they have a very short time to life time to live, right?

160

So you obviously cats sit on the mat all the time.

161

Yeah, not a huge amount of

162

tokens in use in that example for the kind of

163

And you're imagining kind of a chat bot

164

or something doesn't seem to be using a lot of tokens Where does this you know, what's the deal with all this?

165

Yeah, everyone's very cross about the cost, but they're mostly cross about the cost of coding agents, not them going online and chatting to a chatbot with simple questions like which what did the cat sit on?

166

And the issue is that this scales up very, very quickly.

167

And we're going to talk about that with a very simple example.

168

But the thing to remember is whether you're caching it or not caching it, every time you go forward through a network, everything is in there.

169

And if you then wait a few minutes and you go in again, right, and you do or don't get cash hits, this gets longer.

170

Suppose I ask a follow up question to this.

171

Right.

172

Well, in order to answer that, it has to be able to read the original query, the thoughts and the response, which is some 11,100 tokens.

173

And my new query, which is another 200 tokens.

174

So that input is bigger than the one before.

175

The input to the one before was 1,100 tokens and a bunch of output.

176

Now it's nearly 11,000 tokens.

177

And then I ask another question and now it's 20,000 tokens.

178

And I ask another question and it's 30,000 tokens.

179

Or it goes and reads a file which itself is 50,000 tokens.

180

And now every time I ask a question, it has to have this file in the context as well, which adds 50,000 to every single iteration.

181

That's when the price gets very, very big, very, very quickly.

182

And agents, coding agents, are particularly good at scaling this cost up very, very quickly.

183

And so this is what's been

184

so interesting in the last few months where the cost of these has kind of become more known.

185

And some companies are switching to what I might sort of call a slightly more realistic costing model.

186

But let's see.

187

Let's think about the difference between a chatbot and a coding agent.

188

A coding agent has more autonomy.

189

So what happens is it has access.

190

I can talk to it, but it also has access to my files and it can make what we call tool calls.

191

a tool call is just instead of outputting directly to me it outputs to the system

192

and says I would like to do this thing I would

193

like to read this file I would like to read the

194

specific lines of this file I would like to delete this file the system then goes off

195

and does that thing maybe it asks for permission first maybe it doesn't

196

and then and then the model continues to output maybe to me directly

197

or maybe a code change and so the reason we call it an agent is

198

because it's going off

199

and kind of doing this thing I'm not responsible for here are five files

200

that I've pasted in the chat please can you give me

201

this bug fix it is there is a bug can you go off

202

and find it in these five files and

203

and action the response right that's the different thing

204

which as a programmer would be a lot easier right you

205

don't want to have to paste everything in the context window to begin with

206

so the problem the problem is that it's the same sense of scaling up as we had before.

207

Most of the cost of these models are measured in input tokens and output tokens.

208

So most of the modern providers will charge you somewhere around two to three dollars for a million input tokens

209

and something like $15 for a million output tokens.

210

They are what have you sent into the model and what streaming text did it send out.

211

So if I put in the cat sat in the and it sends out Matt, I've only got I've only been charged for one output token which is kind of cheap unfortunately

212

coding models don't only write the word mat so let's imagine

213

that I fire up a coding agent so there's loads of them right

214

but I I um I was recently I will show you in a moment

215

but I was having a play around with github copilot on the command line and the I asked it a question

216

hang on a minute I've gone into you know word and I've produced a bunch of boxes I basically

217

I don't know if you remember back to our Dijkstra video where we had little cards.

218

I thought that might be quite a nice way to visualize this.

219

So I've got myself a system prompt here of some unspecified number of tokens.

220

The system prompt is, you know, the instructions that the manufacturer's given it, right?

221

You are a GitHub copilot coding agent and you must solve Mike's problems because he doesn't want to do it for himself.

222

So we put in the system prompt always.

223

And then I ask it a question.

224

So my question might be, or a programmer's question might be, I've got a bug in my authentication code when someone types a certain thing it crashes

225

please go and find this bug and fix it for me right so the system prompt is let's say

226

4,000 tokens I never see them but they are there

227

and my my my query is let's say a nice modest 200 tokens

228

so the input cost to the model in terms of input tokens is 4,200 tokens right

229

which I'm going to write let's go with another color right

230

so 4200 tokens now there's a very good chance I'm going to do my maths wrong

231

but we'll keep going now what does what's the first thing

232

that a coding agent will do is it will have a

233

good think to itself right it will produce some thought text

234

so it outputs it outputs its initial thoughts the user has

235

asked this I've reminded myself I'm a coding agent maybe I should do this

236

or perhaps I could do this and it churns away for a

237

while and maybe produces let's say 2,000 output tokens

238

so they cost my output but I've only paid for 4,200 input tokens

239

so so far so good right this has not cost me a huge amount of money yet

240

brace yourself right now in response to thought number one it decides it's going to make a tool call

241

if it's going to actually answer the question it needs to look at the file that's the problem

242

and so what it does is it makes a tool call

243

and it's going to say I'd like to read file one

244

so let's say index.html or login.php or something.javascript right and

245

that is a very short output

246

so that's maybe 100 tokens it's just a small bit of structured json between some special characters

247

that tells the system read that file for me it then the system does

248

that and dumps the file contents into the chat right like this you won't see

249

that file contents because it will be hidden from the user

250

but it's in there and that file is is not small so that file is maybe 5,000 tokens

251

right so now we have a problem

252

because now the large language model has to read this file

253

and decide where the bug is and what it has to do

254

so this is now an input of four nine eleven 11,300 tokens, right?

255

11,300 tokens.

256

Not too bad, right?

257

But that's in addition to the 4,200 I've already paid for, right?

258

Which was just its initial thinking and a tool call.

259

Okay, let's keep going.

260

So it thinks for a while longer, right?

261

Like this.

262

And that's going to be another 2,000 tokens.

263

Now, one of the interesting things about large language models is that the thoughts can be very, very short, or they can be incredibly long and there are different models

264

that do different things it's difficult to know going in they might get stuck in a loop

265

and think for ages and achieve nothing

266

or they might be very very succinct right one of the lovely unknowns

267

so in response to tool called thought number two it's going

268

to make another file read of something else dot php

269

or database dot pi

270

or whatever i'm literally making up file names um why they're all programmed in different languages is someone else's

271

because i wrote it it the system will return the file content

272

So the tool call was maybe another 100 tokens and the file contents is perhaps another 4,000 tokens.

273

Now this goes back into the model

274

and it has to make a decision on what to do in terms of fixing the code.

275

So this is now 11,300 plus another 2,000.

276

So 13,300, 17,400 tokens.

277

right 17,400 input tokens I've not even got my answer yet right um okay it's getting a bit grim

278

but all right let's keep going

279

so now it thinks to itself again well why wouldn't it

280

right I'm running out of page I'm gonna go like

281

that look at that um so this thought is another 2,000 tokens

282

it's just a just a guess it doesn't really matter for

283

this example it finally comes up with a solution it's read

284

the files it knows what the code fix is it knows what my query was

285

and it's done it in it's thought through the logic of it in the thinkings uh thinkings right

286

and so it finally makes a tool call to actually code patch the code

287

which is done kind of using a diff it says take

288

out these lines of code add these lines of code right

289

and maybe that's 1500 tokens right maybe I'm being a bit harsh

290

but let's say it's a big big code change my code is not good let's just recap

291

so the initial query

292

that went to the model was these two right like this the next query

293

that goes to the model includes that information

294

because it has to that's the original query you can't just read the file

295

and then go well now i've forgotten what i was meant to do with it so so

296

that is this bit it takes this it does some thinking it asks for a tool call those output tokens

297

and then it gets 5 000 tokens back

298

so these add up to 11 300 in addition to the 4 200 you already paid for.

299

Then it thinks again.

300

It makes another tool call and it outputs some file contents.

301

And the next query is 17,400, right?

302

In response to which it finally starts coming up with some answers.

303

It says, okay, here's Fort 3.

304

That's 2,000 tokens.

305

Here's the code patch, 1,500 tokens.

306

And that returns a result that's a success of, let's say, 50 tokens that goes,

307

yet we've successfully successfully done that right and that means that the final

308

the final prompt that goes through this model is this one

309

which is 17 400 plus another 2000

310

so 19 400 20 900 plus another 50 is 20 950 tokens right input

311

and some thousands of output tokens in the various thoughts, tool calls.

312

And of course, what it also then does is produce a final response that says, I've done a really good job.

313

That's also, why not another 500 tokens?

314

I've run out of paper.

315

There we are.

316

So we asked it to do a bug fix.

317

It read two files.

318

It thought to itself a few times in between these steps, and then it made a patch, right?

319

And it fixed my code, let's say.

320

And in doing

321

so, it's cost me 20,000 plus nearly 20,000 is 40,000 plus

322

another 10 plus another five is somewhere around 55 to 60,000 tokens and a few thousand tokens of output as well.

323

And this is a very, very simple request where it's only read two files and just answered the question, right?

324

Now, suppose this doesn't work.

325

I say, now you broke it again.

326

And I put another user query down here this gets shipped in

327

and this is going to start getting big very very quickly

328

so let's look at a real example of a fun bit of code

329

that I wrote using github copilot to show how big this

330

can get you know what I've always wanted to code

331

and could never be bothered was was was a classic windows 3.11 starfield screensaver

332

so I went on to copilot

333

and I've had it produced this starfield screensaver it's quite fun you know I can go to warp

334

and it does little effects for me.

335

This is completely coded by Copilot and essentially Sonnet and I haven't done anything.

336

I just wanted to see how much it would, how many tokens it would use.

337

So I asked about five or six prompts to get to this point.

338

The first one was things like, can you add a little bit of color to the stars?

339

Can you make them go faster and some slower?

340

Things like this.

341

At one time it introduced a bug, I had to get to fix his own bug.

342

And if I type, so this is the

343

conversation I've been having it's over a few prompts

344

and each time it outputs some text

345

that it thinks to itself then it outputs some code changes

346

and then it outputs its response to me to say what it's done I can see

347

that my current token uses for this session

348

which was I think six prompts in total one file initially now three files is two million input tokens

349

and 47,000 output tokens right it's a lot a lot of tokens it's a lot more than I would do

350

if I was just chatting on the web interface there are some companies

351

that are measuring whether they're moving to AI by how many

352

tokens people are using I mean this is this is ludicrous

353

if this is like it's like measuring my quality as a driver by how quickly I wear through my tires

354

that is an unsustainable practice there's no doubt it was obvious

355

it was unsustainable I think to those of us who understand what a token is before you know

356

if somebody was measuring my use of a broom how many times i used it yeah just swept it backwards

357

and forwards yeah and it turns out you've just created some pulley system

358

but just automatically does it like that episode of the symptoms where he's got

359

that bird doing the s to keep the thing the

360

when you have an incentive like

361

that of course people are going to ask really long-winded questions they're going to ask things

362

that cause the ai to get stuck in a loop where

363

its thought process is much longer they're going to use the more expensive models they're going to use models

364

that think for longer periods of time and have to look over more files

365

and in a surprise to absolutely no one

366

that is completely unsustainable in terms of cost you cannot afford

367

to do it right there will be some companies maybe

368

that will give this a go but for a normal company

369

that has normal paid developers right

370

so not the big tech companies just normal companies who are producing regular products to be able to spend

371

that amount of money on tokens you have to be able to see immediate product return

372

so you know we've developed five new features

373

that are making us this much more money

374

or we have a new product that exists but six months ago didn't

375

and it's good right that has not been answered yet yes

376

that might happen i remain suspicious

377

that the quality of code is good enough for it to do

378

that across all companies right um

379

and there are there are other issues like the fact

380

that you have to kind of know what you're doing in

381

order to you know this is not a good example my my example is not a good example of coding

382

because I haven't read it right so it could be doing all kinds of things

383

I certainly wouldn't inflict it on someone else's computer a more

384

reasonable way to use these models would be small succinct questions very quick fixes

385

or the kind of code completion thing we already see

386

so you're in something like codex

387

and you say I've written half a for loop can you

388

just finish this off for me that's quite a quick you

389

only need to read a small amount of context that's quite a

390

quick thing the agentic stuff i think will have its place but i think it is worryingly costly and

391

we've been seeing a lot of subsidization right github copilot people are only annoyed about

392

that because it was being subsidized before

393

when you pay a flat fee per month

394

and you can have an agent go off for hours

395

and hours doing all these millions of tokens of course you're going to think that's absolutely fantastic

396

but of course when it gets to the when it gets

397

when the rug gets pulled that and they say okay you're paying per token now

398

and you realize that that

399

that file you just read in is 10 000 tokens being put in over

400

and over again suddenly you're kind of thinking well actually could i have avoided

401

that could i have done something else that meant it didn't cost quite

402

so much money that's what we're going to have to answer over the next year and see how that goes.

403

So I can see, for example,

404

that split 171 is something about exchange of information between internet identities, right?

405

I don't, you know, who knows.

406

The next thing for me to do is to embed every single one of these chunks into the embedded space

407

and stick them in a database.

הורד אפליקציה

כל מה שאתה צריך כדי לדבר בשטף

הגיית AIציון לכל משפט

תרגול IPAשלוט בכל צליל

אוצר מיליםבנה את מאגר המילים שלך

משחק אוצר מיליםלמד תוך כדי משחק

סרוק QR

Download on theApp Store Get it onGoogle Play

אותו ערוץ

Turing & The Halting Problem - Computerphile

Business EnglishUniversity StudentsWorking Adults

FOR YOU

המלץ על סרטון

137 משפטים

Expensive RTX 5090 for LLMs? NO. Use This Instead. (Tesla V100 SXM2 + Z8 G4, #RACERRRZ)

Daily ConversationUniversity StudentsWorking Adults

171 משפטים

Why the Arab World never united

Daily ConversationUniversity StudentsWorking Adults

74 משפטים

LangChain in 5 Minutes (Explained Clearly)

Daily ConversationUniversity StudentsWorking Adults

182 משפטים

We Let $200 GPT-Pro Think for 7 Days on Unsolved Math

Daily ConversationUniversity StudentsWorking Adults

23 משפטים

What is an AI Token? | LLM Tokens explained in 2 minutes!

Daily ConversationWorking Adults

41 משפטים

No, Really — Is A.I. a Bubble? | The Ezra Klein Show

Daily ConversationUniversity StudentsWorking Adults

55 משפטים

When AIs act emotional

Daily ConversationUniversity StudentsWorking Adults

198 משפטים

Jack Dorsey’s New Dessa AI Voice Sounds Scary Human

Daily ConversationUniversity StudentsWorking Adults

13 משפטים

What is an API?

Daily ConversationUniversity StudentsWorking Adults

אוצר מילים והערות הגייה לשיעור הזה

שיעור הדיבור הזה ברמה C1 מבוסס על הסרטון “Why AI Tokens are so Expensive - Computerphile”. המילים שחוזרות הכי הרבה: token, model, file, output, cost. בסרטון הזה יש 407 משפטים ו-4967 מילים לתרגול Shadowing. הדיבור נמשך 25:21. הדובר מדבר מהר, כ-196 מילים בדקה, אז צפו לצלילים מחוברים ומקוצרים. 85% מהמילים נמצאות בין 3,000 המילים הנפוצות ביותר באנגלית; כדאי ללמוד את השאר לפני שמתחילים.

אוצר מילים מרכזי בסרטון הזה

15 המילים המתקדמות ביותר בסרטון, עם ההגייה והמשמעות שלהן:

מילה	הגייה	משמעות
token שם עצם	/ˈtoʊkən/	אֲסִימוֹן, זִ׳יטוֹן
cache שם עצם	/kæʃ/	מִצְבּוֹר, מַחְבּוֹא
embed פועל	/ɪmˈbɛd/	לְשַׁבֵּץ
subset שם עצם	/ˈsʌbˌsɛt/	תת קבוצה
ludicrous שם תואר	/ˈluː.dɪ.kɹəs/	מגוחך
parser שם עצם	/ˈpɑɹ.sɚ/	מפרק, נתח תחבירי
pulley שם עצם	/ˈpʊli/	גלגלת
subsidize פועל	/ˈsʌbsɪdaɪz/	סִבְסֵד, סִבְּסֵד
programmer שם עצם	/ˈpɹoʊɡɹæmɚ/	מְתַכְנֵת, תׇּכְנִיתָן
curly שם תואר	/ˈkɝli/	מתולתל, מתולתלת
symptom שם עצם	/ˈsɪm(p)təm/	תַּסְמִין, סִימְפּטוֹם
numerical שם תואר	/n(j)uːˈmɛɹ.ɪ.kəl/	מִסְפָּרִי
confuse פועל	/kənˈfjuːz/	בלבל
grim שם תואר	/ɡɹɪm/	עגום
autonomy שם עצם	/ɔˈtɑ.nə.mi/	אוֹטוֹנוֹמְיָה

Phrasal verbs שתשמעו

מילה	משמעות
come back פועל	שָׁב
go up פועל	עָלָה

משפטים שכדאי לחזור עליהם

משפטים קצרים ושלמים מהסרטון שאפשר להשתמש בהם בשיחה יומיומית:

I mean, let's look at a concrete example.
You don't just say you said mat, what's the next word?
I've run out of paper.
Can you make them go faster and some slower?
I don't, you know, who knows.

דקדוק בסרטון הזה

המבנים שהדובר משתמש בהם הכי הרבה, עם המילים המדויקות מהסרטון:

מבנה	בסרטון
הווה מושלם ממושך (Present perfect continuous) have/has been + -ing — פעולה שהתחילה בעבר ועדיין נמשכת עכשיו	I've been having · we've been seeing
הווה מושלם (Present perfect) have/has + past participle — פעולה שקרתה בעבר ועדיין יש לה חשיבות או השפעה עכשיו	it's been · haven't been decided · they've been decided
סביל (Passive voice) be + past participle — הדגש הוא על מה שקורה, לא על מי שעושה את הפעולה	been decided · they've been decided · is learned
משפטי זיקה (Relative clauses) who / which + clause — מידע נוסף על אדם או דבר	process which is · thought, which is · response, which is

הגייה ששווה לשים לב אליה

הדובר משתמש ב-114 קיצורים וצורות מקוצרות, כמו I've, don't, you're. אמרו אותם בצורה הקצרה, כפי שאתם שומעים אותם.

הצלילים של “sh” ו-“zh”: cache /kæʃ/, inefficient /ˌɪn.ɪˈfɪʃ.ənt/, authentication /ɔːˌθɛntɪˈkeɪʃn̩/, iteration /ˌɪt.əˈɹeɪ.ʃən/, visualize /ˈvɪʒuəˌlaɪz/
מילים ארוכות — שימו לב להטעמה: inefficient /ˌɪn.ɪˈfɪʃ.ənt/, authentication /ɔːˌθɛntɪˈkeɪʃn̩/, iteration /ˌɪt.əˈɹeɪ.ʃən/, symbolically /ˌsɪmˈbɑːlɪklɪ/, unspecified /ʌnˈspɛ.sɪ.faɪd/

צלילים שדוברי עברית מתקשים בהם:

/w/ — שפתיים מעוגלות, לא /v/: query /ˈkwɪɹ.i/, whichever /wɪt͡ʃˈɛvɚ/, warp /ˈwɔːp/, backward /ˈbækwɚd/, sweep /swiːp/
/æ/ — פתוח יותר מה-“אֶ” העברית, לא כמו “a” ב-“cup”: cache /kæʃ/, chatbot /ˈtʃætbɑt/, grasp /ɡɹæsp/, programmer /ˈpɹoʊɡɹæmɚ/, transcript /ˈtɹæn.skɹɪpt/
/r/ — הלשון לא נוגעת בשום מקום, לא ה-“ר” הגרונית: query /ˈkwɪɹ.i/, prompt /pɹɑmpt/, predict /pɹɪˈdɪkt/, refine /ɹɪˈfaɪn/, screensaver /ˈskɹiːnˌseɪ.və(ɹ)/

איך לתרגל עם הסרטון הזה

הקשיבו לכל הסרטון פעם אחת בלי לדבר, ורשמו את המילים שאתם לא מכירים.
התחילו במהירות 0.75×, תרגלו Shadowing משפט אחר משפט, וחזרו למהירות רגילה כשזה מרגיש קל.
הקליטו את עצמכם והשוו למקור, תוך תשומת לב למילים כמו token, cache, embed.

מהי טכניקת השאדואינג?

שאדואינג היא טכניקת למידת שפה מבוססת מדע שפותחה במקור להכשרת מתורגמנים מקצועיים ופופולרית על ידי הפוליגלוט ד"ר אלכסנדר ארגואלס. השיטה פשוטה אך עוצמתית: אתה מקשיב לאודיו באנגלית ילידית וחוזר עליו בקול רם מיד — כמו צל העוקב אחרי הדובר עם עיכוב של 1–2 שניות בלבד. בניגוד להאזנה פסיבית או תרגילי דקדוק, שאדואינג מכריח את המוח והשרירים בפה לעבד ולשחזר דפוסי דיבור אמיתיים בו זמנית. מחקרים מראים שזה משפר משמעותית את דיוק ההגייה, האינטונציה, הקצב, הדיבור המחובר, הבנת השמיעה ושטף הדיבור — מה שהופך אותו לאחת השיטות היעילות ביותר להכנה לIELTS Speaking ולתקשורת באנגלית בעולם האמיתי.

טכניקת Shadowing: קרא את המדריך המלא שלב אחר שלב →

Shadowing Practice: Why AI Tokens are so Expensive - Computerphile - Learn English Speaking with Video

בחר מצב תרגול

הורד אפליקציה

Turing & The Halting Problem - Computerphile

JUST $1 For Billions of Tokens?! (DeepSeek v4 Pro)

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

Crypto Coin vs Token (Differences + Examples)

Expensive RTX 5090 for LLMs? NO. Use This Instead. (Tesla V100 SXM2 + Z8 G4, #RACERRRZ)

Why the Arab World never united

LangChain in 5 Minutes (Explained Clearly)

We Let $200 GPT-Pro Think for 7 Days on Unsolved Math

What is an AI Token? | LLM Tokens explained in 2 minutes!

No, Really — Is A.I. a Bubble? | The Ezra Klein Show

When AIs act emotional

Jack Dorsey’s New Dessa AI Voice Sounds Scary Human

What is an API?

אוצר מילים והערות הגייה לשיעור הזה

אוצר מילים מרכזי בסרטון הזה

Phrasal verbs שתשמעו

משפטים שכדאי לחזור עליהם

דקדוק בסרטון הזה

הגייה ששווה לשים לב אליה

איך לתרגל עם הסרטון הזה

מהי טכניקת השאדואינג?