- 平台
- YouTube
- 时长
- 4:47
- 片中语言
- 美国英语
- 播放量
- 407.2万
- 镜头间隔中位数
- 9.34s
- 人脸出镜
- 69%
- 暗部画面
- 0%
What happens when you talk to AI? 的文字记录
由语音模型根据影片自身的音轨转写，并非品牌撰写。点击任意一行即可跳到那一刻。
- 0:00When you send a message to an AI,
- 0:02there's a moment where it appears to be thinking,
- 0:04but what's actually happening?
- 0:06Is it reading the entire internet?
- 0:08Is it copying answers from a database?
- 0:10Is it just a fancier search engine?
- 0:13I'm Jane, and I work on user experience here at Anthropic,
- 0:16the company that makes Claude.
- 0:18Here's what's actually going on.
- 0:20AI models like Claude work by prediction.
- 0:23When you send a message, the model reads that message
- 0:26and draws on everything it learned during training
- 0:28to write back a response a little bit at a time.
- 0:31You'll see it appear word by word,
- 0:33and each one is chosen based on everything that came before it.
- 0:36Here's what surprises most people.
- 0:39The model writes one word at a time,
- 0:41but it doesn't think one word at a time.
- 0:44We'll come back to that in a moment.
- 0:46You've likely seen something like this before.
- 0:48When your phone's keyboard suggests Ben
- 0:50after you type, how have you, that's prediction too.
- 0:54The keyboard has learned which words tend to follow which.
- 0:56but there's a key distinction here.
- 0:58Your simple predictive keyboard
- 1:00is only looking at the last two or three words.
- 1:03It has no idea what you're trying to say
- 1:05or where the sentence is going.
- 1:07An AI model goes much further than this.
- 1:09It's been trained on an enormous amount of text
- 1:12and other kinds of information.
- 1:13This is known as training data.
- 1:15And when an AI model predicts,
- 1:17it isn't just looking at the last two or three words
- 1:20the way your keyboard does.
- 1:21What it's actually doing turns out to be much deeper
- 1:23and we'll see exactly how in a moment.
- 1:25Long before you ever talked to it, the model went through billions of rounds of the same exercise.
- 1:33See some text, guess the next word, see how close it got, adjust slightly, go again.
- 1:41Later, in a second stage called fine-tuning, the model's full answers are rated,
- 1:46sometimes by people, sometimes against a written set of guidelines,
- 1:50and the model is nudged toward answers that are more useful and less likely to mislead
- 1:55or cause harm.
- 1:57Both processes contribute to a model's
- 1:59generative capabilities.
- 2:01All of that takes place up to a certain date
- 2:04called the training cutoff,
- 2:06after which the model doesn't reliably know
- 2:08about facts or information without leaning on other tools
- 2:11such as internet search.
- 2:13While the model can search the internet
- 2:16to access information beyond the cutoff
- 2:18if the tool has search available,
- 2:20you can't assume the model has done a web search
- 2:22to craft its response.
- 2:24To be sure, you can ask it to, and when it does,
- 2:27it will typically share sources that you can review
- 2:30to confirm the output's accuracy.
- 2:32Now back to how the model thinks.
- 2:34Predict the next word sounds almost mechanical,
- 2:37like the model is just reacting one word,
- 2:39then the next, then the next.
- 2:40But to predict the next word well,
- 2:42you can't just look at the last few words.
- 2:44You have to work out where the sentence is going,
- 2:46what the paragraph is arguing,
- 2:48what a good answer would actually be.
- 2:49The model takes into account everything that came before,
- 2:52your uploaded documents, memory system prompt, your prompts, your conversation history,
- 2:56before outputting the next word or even syllable.
- 2:59And then that whole process repeats for the next one and so on until the response is done.
- 3:04So why should you care about any of this?
- 3:07Because when you understand that an AI model is a prediction system,
- 3:11you are able to work with it and interpret its outputs much better.
- 3:14The fact that it can write you something that's never existed before makes sense
- 3:18once you remember it's generating,
- 3:20not just pulling answers from somewhere.
- 3:22The fact that it sometimes states something false
- 3:25with total confidence, same reason.
- 3:27It's producing what a good answer would look like,
- 3:30and usually that lines up with reality.
- 3:33Occasionally it doesn't,
- 3:34which is why your judgment still matters.
- 3:37Knowing all that, a few habits
- 3:38that can get you better results.
- 3:40One, give it context to work with.
- 3:42Tell it who you are, what you're working on,
- 3:44and what a good result looks like.
- 3:46All of that becomes part of the pattern.
- 3:48Two, remember the cutoff.
- 3:50The model's knowledge stops at a certain date,
- 3:52so for anything recent like prices or the news,
- 3:55it could be out of date unless the tool tells you it searched the web.
- 3:58Three, ask for options.
- 4:00Because it's generating, not retrieving information,
- 4:03there's no single stored answer, so ask for a few versions of the draft
- 4:06or have it try again in a different tone.
- 4:08We encourage you to explore.
- 4:10Four, and above all, double-check the AI's outputs.
- 4:14A confident tone or a polished-looking result
- 4:16is just how the output comes out.
- 4:18It isn't proof that the output is accurate or done well.
- 4:21Consider the stakes of your question
- 4:23and double-check facts if wrong information would cause an issue.
- 4:26So that's the basic shape of how AI works.
- 4:29Every time you send a message,
- 4:30you're handing it the start of a pattern,
- 4:32and it completes that pattern
- 4:33based on everything it learned in training.
- 4:36Different models will likely complete that pattern in different ways,
- 4:39so it's worth experimenting to see what works for you.
- 4:42We'll keep sharing our research on this topic on Anthropik's blog.
112 行 · 901 字
More from Claude
All of Claude →The rest of their work we have taken apart.
3.5M viewsClaude
There’s hope in hard questions
1:3090 frames1.3s cut19% face
2.4M viewsClaude
What does AI actually know about you?
4:50156 frames6.1s cut83% face
Launch875K viewsClaude
Introducing Claude Design by Anthropic Labs
1:2240 frames3.3s cut0% face
403K viewsClaude
Getting started with Claude Cowork
4:16329 frames19.1s cut0% face
109K viewsClaude
Claude Cowork is now your Chrome side panel
0:4117 frames2.1s cut0% face
Cut at the same speed
Films whose median shot runs closest to 9.34s.
dennis müller
what if there was an app to create these videos in <5min?
0:4118 frames9.3s cut67% face
15 viewsTimestripe
Generate release notes from the commit history
0:3417 frames9.4s cut0% face
Launch1.2M viewsPerplexity
Introducing Voice Mode in Perplexity Computer. You can now just talk...
1:1833 frames9.5s cut9% face
7K viewsIcons8
Vector Creator. Create free personalized illustrations from 3000+ elements
1:0725 frames9.5s cut0% face
108K viewsFigma
No more slots gone wild → Set minimum and maximum layers → Only allo...
0:4219 frames9.6s cut21% face
LaunchShopify
Shopify Editions - Brand Film by Imaginary Forces
1:3037 frames9.1s cut3% face
223K viewsNanoleaf
Nanoleaf Canvas: Create Your Own Masterpiece
0:5927 frames9.0s cut0% face
113K viewsFigma
Figma Motion: design AND animate in the same file Live from Config 2026
17:311197 frames9.0s cut93% face
Launch54K viewsVaishak Kumar
Introducing text-to-motion. It’s Cursor, but for IRL code. Comment "v...
0:2512 frames9.7s cut0% face
This Week in AI
Thanks for having me on, @jason!
73:43640 frames9.0s cut98% face
148K viewsMagnific
Edit any image, layer by layer, in Magnific No more rebuilding a des...
0:2211 frames8.9s cut27% face
275K viewsPerplexity
We are rolling out Comet to all students worldwide. Ask Comet to man...
0:219 frames8.9s cut44% face
Lit and framed like this
Closest on how much of the running time has a face in it and how dark the frames are.
163K viewsPerplexity
1. Upload a photo. 2. Get turned into a Perplexity-style portrait. 3....
0:3010 frames70% face
26K viewsPetSmart
PetSmart Kitchen featuring GREENIES
1:0047 frames2.0s cut68% face
45K viewsLacoste
The Masterclass of Yoga featuring Novak Djokovic.
0:3025 frames1.5s cut72% face
Launch138M viewsBase44
Introducing Base44’s Big Game Ad Teaser “It’s App to You”
0:1515 frames1.5s cut67% face
Launch134K viewsKrea
announcing our $83M funding. s/o to our great investors @a16z @BainC...
0:103 frames67% face
58K viewsFigma
Config was a movie
0:093 frames67% face
29K viewsSketch Together
Open Peeps - A hand-drawn illustration library.
0:3430 frames0.9s cut67% face
Launch772K viewsSean Rich
This is an absolutely wild idea (and team!)
1:1027 frames12.9s cut70% face
2K viewsPinterest
Pinterest | Get Playful
0:3015 frames1.7s cut73% face
38K viewsheyclicky
i want to change
0:4936 frames1.9s cut67% face
2K viewsWispr Flow
How Clay drove 20% faster GTM execution with Wispr Flow
2:25162 frames1.9s cut67% face
330K viewsNike Football
France x Jacquemus
0:4743 frames1.2s cut74% face
More social
Same category, other makers, most watched first.
622M viewsBase44
Your ideas, your apps—Base44
0:5651 frames1.1s cut6% face
538M viewsBase44
Turn any idea into an app with Base44
0:5538 frames2.2s cut50% face
154M viewsBase44
Build Your Website with One Sentence | Base44
0:4835 frames0.9s cut31% face
81M viewsNike Football
Rip The Script | Nike Football
6:00310 frames1.1s cut60% face
75M viewsPetSmart
PetSmart | "Pickup in...STORE!"
0:1511 frames1.4s cut64% face
69M views1X
NEO The Home Robot Order Today
9:53271 frames2.8s cut25% face
69M viewsApple
Apple Education: Ready for every learning opportunity
1:4687 frames1.3s cut40% face
64M viewsBland
Phone calls suck. So, we bought a giant telephone car, raised $65,00...
1:4854 frames4.3s cut81% face
39M viewsFigma
Oh wait there’s more: we’re building 3D Transforms, here’s a preview ...
1:0423 frames4.0s cut78% face
36M viewshims
Getting Help is Simple, With Hims
0:3023 frames2.0s cut70% face
36M viewsFigma
Made in Figma (2025) Like this ❤️ we designed it for you
0:3872 frames0.5s cut7% face
35M viewsApple
Hello, MacBook Neo
3:50291 frames67.5s cut1% face