跟讀練習: Gemini 4 Argon explained in 5min.. - 通過影片學習英語口語

選擇練習模式

每種模式都會開啟全屏練習室。

56 句

如果句子過短或過長，請點選 Edit 進行調整。

It's been about seven months since Google announced a new flagship model, Gemini 4 Argon.

Even during Google I .O back in May, I expected Google to drop Gemini 3 .5 Pro.

But Google has been releasing a lot of flash models instead,

and a lot of them basically struggle to get market adoption given a variety of other flagship models

and cheaper alternatives from Chinese open models.

And Google finally made a flagship model with the codename Argon, and the preliminary benchmarks show a pretty strong number as you can see.

The first thing you might have noticed here is coding benchmarks, more specifically, scoring 77 .9 % on DeepSuite,

which would put Gemini 4 Argon above every other model we currently have.

But it's always good to have some healthy skepticism here.

Recently, EPOC found issues on DeepSuite Benchmark, finding 23 tasks that are flawed.

And looking back to when the benchmark was actually released back in May, the entire 113 tasks and their solutions are made public,

which means labs that have access to these could easily study them during post -training and benchmarks their models around public benchmarks.

And when we look at other benchmarks that Google reported here, a large majority of them are public.

So the further we go out in time, labs that release their models later have a slight advantage to potentially include them during post -training.

And of course, as users, all of this makes it really difficult to track real progress in AI since

so many public benchmarks like DeepSuite can be contaminated pretty fast.

Not only that, benchmarks get saturated a lot faster now than before, which forces benchmarks to be refreshed more frequently.

So consumers now carry a huge burden keeping track of real scientific progress in AI.

But regardless of the case, achieving this kind of leap in benchmark is a huge improvement from Google.

Now, when we take a step back and look at how Google is competing with other labs, a lot of people still primarily use clock code,

codecs, or open harnesses when it comes to coding.

And despite being a strong model, Google still hasn't really solved making their harness anti -gravity a truly competitive option for developers.

For a recent report, Codex has more than doubled the amount of weekly active users at 5 million compared to 2 .4 million for anti -gravity.

And on face value, Gemini 4 Argon currently undercuts both Anthropic

and OpenAI when it comes to the cost of intelligence until the temporary discount ends and the price doubles again.

But the cost benefit that we find here doesn't actually translate in Pareto Frontier, as you can see.

Argon falls below the Pareto Frontier, showing you that Anthropic and OpenAI still leads ahead in both the harness and model adoption.

Now, one angle we should think about, though, is the upcoming IPO for Anthropic and possibly OpenAI.

While OpenAI and Anthropic have a huge incentive to subsidize the model through subscription pretty heavily to grow their users,

Google already sits on top of one.

And given that Google is already a publicly traded company,

they want to be more careful about how to allocate their existing compute towards actually creating value within their ecosystem,

which could help increase their perceived revenue.

Even looking at how Argon is being rolled out in segments, the model is set to be made available to ultra -subscription users

and paid API users where there's already a huge revenue being

captured instead of fighting the subscription rates in channeling their compute towards that.

On a recent Google's Q2 earning call, they said that their API is processing 22 billion tokens per minute,

which is 6 billion more than the previous quarter.

That's more than 11 quadrillion tokens per year in just API channel alone.

And when you start factoring in other channels that Google owns in its distribution,

we really get a closer picture into what Google really means when they said that they're supply constrained.

And eventually, OpenAI and Anthropic will have to solve this scaling problem while maintaining healthy margins at the same time,

not just subsidizing subscription to capture more users.

So looking at Gemini for Argon, Google is demonstrating their capability in keeping pace at the frontier

while they also try to figure out how to capture value at scale.

One interesting note on Argon is how Google extended their output tokens from 64 ,000 to 1 million tokens.

This might not sound like a big deal, but it's definitely a huge move from Google

because maintaining coherence as autoregressive models generate longer and longer tokens get increasingly difficult the longer the output tokens get.

Scientists like Yen Le Koon made criticism on autoregressive models for this very reason,

where autoregressively generating one token after another can compound errors over long trajectory, let alone 1 million from Argon.

If you can imagine each token that's 99 % reliable, chaining 100 independent steps together would leave about 37 % chance of every step being correct.

Now, of course, that's just in theory, assuming that every token error is independent.

But the fact

that Google has enough confidence to extend their commercial product to

1 million output tokens really opens up door for longer trajectory tasks like code migration

and long -running research tasks that could benefit from a model that can generate more.

下載應用

流利口語所需的一切

AI 發音評分逐句打分

音標練習掌握每個音

詞彙學習建立你的詞庫

詞彙遊戲邊玩邊學

掃碼

Download on theApp Store Get it onGoogle Play

同頻道

World Models explained in 10min..

Business EnglishUniversity StudentsWorking Adults

FOR YOU

本課的詞彙與口語要點

這節口語課以影片“Gemini 4 Argon explained in 5min.”為素材。影片中反覆出現的詞有：Google, model, benchmark, token, Argon。這段影片共有 56 個句子、877 個單詞可供跟讀。講話部分時長為 4:56。說話人語速自然，每分鐘約 178 個詞，接近日常對話。只有 77% 的單詞屬於英語最常用的 3,000 詞，詞彙難度較高。

影片中的重點詞彙

影片中 15 個值得學習的單詞，附發音和釋義：

單詞	發音	釋義
token 名詞	/ˈtoʊkən/	代幣 /代幣, 籌碼 /籌碼
subscription 名詞	/səbˈskɹɪpʃən/	訂閱 /訂閱, 預訂 /預訂
output 名詞	/ˈaʊtpʊt/	輸出 /輸出
capture 動詞	/ˈkæp.(t)ʃɚ/	捕獲 /捕獲, 佔領 /佔領
frontier 名詞	/ˈfɹʌntɪə/	邊界 /邊界, 邊疆 /邊疆
flagship 名詞	/ˈflæɡʃɪp/	旗艦 /旗艦
solve 動詞	/sɒlv/	解決 /解決
adoption 名詞	/əˈdɑp.ʃən/	收養 /收養, 領養 /領養
trajectory 名詞	/tɹəˈd͡ʒɛktəɹi/	軌道 /軌道, 彈道 /彈道
earn 動詞	/ɝn/	掙 /掙, 賺 /賺
temporary 形容詞	/ˈtɛmp(əˌɹ)ɛɹi/	暫時 /暫時
compare 動詞	/kəmˈpɛɚ/	比, 比較 /比較
flash 動詞	/flæʃ/	閃光 /閃光
compete 動詞	/kəmˈpiːt/	競爭 /競爭, 高下
angle 名詞	/ˈæŋ.ɡəl/	角, 角度

影片中出現的短語動詞

單詞	釋義
figure out 動詞	弄清楚

影片中的語法

說話人最常用的結構，並附上影片中的原話：

結構	影片中的用法
被動語態 be + 過去分詞 — 強調發生了什麼，而不是誰做的	are flawed · was actually released · can be contaminated
現在完成時 have/has + 過去分詞 — 過去發生但與現在仍有關聯的事	It's been · hasn't really solved

需要注意的發音

說話人用了 3 次縮略和弱讀形式，例如 doesn't, hasn't, they're。請按聽到的簡短形式來說。

“sh” 和 “zh” 音: benchmark /ˈbɛn(t)ʃmɑːk/, subscription /səbˈskɹɪpʃən/, capture /ˈkæp.(t)ʃɚ/, flagship /ˈflæɡʃɪp/, adoption /əˈdɑp.ʃən/
長單詞——注意重音位置: trajectory /tɹəˈd͡ʒɛktəɹi/, primarily /ˈpɹaɪ.mə.ɹə.li/, temporary /ˈtɛmp(əˌɹ)ɛɹi/, potentially /pəˈtɛnʃ(ə)li/, developer /dɪˈvɛləpɚ/

中文母語者容易讀錯的音：

/v/ — 不要讀成 /w/，上齒輕觸下唇: solve /sɒlv/, gravity /ˈɡɹævɪti/, developer /dɪˈvɛləpɚ/, incentive /ɪnˈsɛntɪv/, perceive /pɚˈsiv/
詞尾子音 — 要讀清楚，後面不要加母音: benchmark /ˈbɛn(t)ʃmɑːk/, output /ˈaʊtpʊt/, lab /læːb/, generate /ˈd͡ʒɛn.ə.ɹeɪt/, flagship /ˈflæɡʃɪp/
/r/ 和 /l/ 的區別: primarily /ˈpɹaɪ.mə.ɹə.li/, regardless /ɹɪˈɡɑɹd.lɪs/, reliable /ɹɪˈlaɪəbl̩/, ultra /ˈʌltɹə/, preliminary /pɹɪˈlɪmɪnɛɹi/

如何用這段影片練習

先完整聽一遍影片，不要開口，記下不認識的單詞。
先用 0.75 倍速逐句跟讀，熟練之後再回到正常速度。
錄下自己的聲音並與原聲對比，特別注意 token, subscription, output 這類單詞。

什麼是跟讀法？

跟讀法 (Shadowing) 是一種有科學依據的語言學習技巧，最初開發用於專業口譯員的培訓，並由多語言者Alexander Arguelles博士普及。這個方法簡單而強大：您在聽英語母語原聲的同時立即大聲重複——就像是一個延遲1-2秒緊跟說話者的影子。與被動聽力或語法練習不同，跟讀法強迫您的大腦和口腔肌肉同時處理並模模擬實的講話模式。研究表明它能顯著提高發音準確性，語調，節奏，連讀，聽力理解和口語流利度——使其成為雅思口語備考和真實英語交流最有效的方法之一。

影子跟讀法: 閱讀完整分步指南 →

跟讀練習: Gemini 4 Argon explained in 5min.. - 通過影片學習英語口語

選擇練習模式

下載應用

同頻道

World Models explained in 10min..

推薦影片

Did Google just dethrone OpenAI and Anthropic?

Gemma 4 Just Released - Whats In? In 2026!

LangChain in 5 Minutes (Explained Clearly)

They JUST Accidentally Revealed the Truth: The AI Bubble

Give me 31 minutes, I will make PyTorch click forever

My 4.9-Star iOS App Is Actually a Website: Here’s How I Built It

What's New in Zephyr 4.5?

Go’s Big Mistake

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

The Engineering that Runs the Digital World 🛠️⚙️💻 How do CPUs Work?

Andrew Ng on Coding Agents and Deciding What to Build

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