谷歌Gemini 4 Carbon首批实测曝光！手感直逼Opus 5.5
ASI启示录
ASI启示录
新智元
在小说阅读器读本章
去阅读
在公众号小说中沉浸阅读
新智元报道
Argon全面开放才
8天
，
谷歌的下一张王牌，就杀进竞技场了！
就在刚刚，
谷歌最强的Carbon被曝开测
。
合作伙伴已经抢先拿到，LMArena上也开了盲测，
发布就在眼前
。
消息一出，测试者全扑进了Arena。
一天之内，Carbon的实测在X上炸成一片，
有人拿它一口气做完了
整关超级玛丽
。
上一代旗舰的热乎劲还没过，下一代就在门外排队了
。
谷歌联创布林说，谷歌开始踏上
「RSI的跑步机」
，
模型正帮着造下一代模型
。
超级玛丽一次通关
Carbon的3D强到离谱
Arena上的Carbon，名字还挂着「gemini-4-argon」。
测试者认出它的办法说来好笑——就看慢不慢
。
网友Harshith发现，挂这个名字的其实有两个模型。一个又快又不错，另一个慢得出奇，作品却真的惊艳，
后者才是Carbon
。
测评账号Lumina用超级玛丽的提示词，
让Carbon做出了一整关游戏
。
主角一路顶砖块、吃金币、钻水管，下到地下关再爬出来，最后跳上旗杆、走进城堡，
屏幕弹出「COURSE CLEAR」
，金币数和评级都结算了出来。
Lumina还把同一道体素宝塔题，
交给Argon、Carbon、Opus 5.5和Sonnet 5.5同场比
。
在Lumina看来，Carbon的质量明显要高过Argon，只是视任务不同，
要慢2到4倍。
网友Synopsis的体素山谷，把画面又往上推了一层。
樱花沿着河岸开，红色拱桥架在水上，木屋亮着灯，
一场雨说下就下，整片山谷跟着暗了下来。
网友Salio做了一座3D城堡，
光照能在黄昏、蓝调时刻、月光和薄雾晨光之间切换
，城下的石桥一路亮着灯。
知名爆料人chetaslua把Gemini的吉祥物交给Carbon，让它做一段讲碳为什么稳定的3D动画。
画面一开始，吉祥物盯着一颗碳-12原子核。随后6个电子按1s²2s²2p²排好，轨道杂化成4个一模一样的sp³瓣，再和邻居共享电子凑满8个，
最后搭成一整块金刚石晶格
。
chetaslua说，
这道题对之前的Gemini模型来说非常难
。
拿Carbon讲碳，谷歌这次的命名，算是被网友玩明白了。
从碳-12原子核讲到金刚石晶格
开发者Pankaj Kumar则让它做了一个能交互的3D水晶柑橘。
然而，到了SVG这一项，评价就开始分化了。
网友Binkg让Carbon画了一个可乐瓶的SVG，瓶身挂满水珠，标签上的小字和瓶盖都画了出来，旁边还能调霜冻、气泡和背光。
但Haleemah看了一整天的测试后说，
Carbon最强的还是3D
，好的SVG她还没见到。
SVG之外，前端也还是老问题。Lumina说，Carbon的前端设计和审美依旧平平，缺了Anthropic模型的那股味道。
同一个模型，3D能猛涨，前端却原地踏步，是因为两类题的考法不同。
3D场景里，几何、光照、碰撞全由代码决定，算错一点画面就崩，
对不对一眼就能验出来
。
前端考的是留白多大、字号几号、按钮放哪，
没有标准答案，只有人的偏好
。
所以3D能靠算力和
强化学习
一路往上推，
前端
却只能靠大量人的反馈慢慢磨。
更慢但不偷懒，Carbon专治Argon的懒病
测试者晒作品时，几乎都会补一句慢，可很少有人抱怨，
因为它不偷懒了
。
偷懒，偏偏是Argon上线以来被骂得最狠的毛病
。
它的跑分很漂亮，可开发者上手不到一周，「懒」的名声就传开了。
科技媒体Every测完Argon，给出了5条使用建议，第3条是让大家务必复核它的活。
在38次基准运行里，
有17次，Argon把代码审查这一步打开又关上
，
中间只隔了30毫秒
。
同一份建议里还藏着一组数字。把Argon从低推理档拉到高档，
用时只变成原来的1.7倍
；Opus 5.5拉到高档，
要花18倍的时间
。
也就是说，
叫Argon多想一想，它也没多想多少
。
到了Every CEO Dan Shipper的资深工程师基准上，Opus 5.5和Sonnet 5.5都拿了85%，GPT-6 Astra拿了81%，
Argon只有21%，在8个模型里垫底
。
而到了Carbon这里，情况直接反转。
Business Insider此前曝出，谷歌员工在内部试用Carbon后说，它写代码
「感觉像极了Opus 5.5」
。
Argon刚铺开就出Carbon
谷歌踏上RSI跑步机
从Argon到Carbon，谷歌这一轮的节奏快得惊人。
9月30日官宣Argon，先只交给Fairwind计划里的网络安全防御方；10月13日，Argon向开发者和订阅用户全面开放；
又过了8天，Carbon就进了合作伙伴的测试
。
谷歌出牌越来越快，布林给出的答案便是——
RSI（递归自我改进）
。
在最近一段圆桌对谈里，他曾表示，
现在研究员已经和模型手拉手，一起做下一版模型了。
Argon的安全报告显示，今年9月，DeepMind
超过九成的代码合并请求里，
AI
写的部分都在75%以上
。
完全交给AI、人类插话不到3轮的合并请求，也
从4月的7.5%涨到了11.9%
。
DeepMind每月的代码合并请求，颜色越深，AI写的比例越高
当模型开始帮着造模型，Argon和Carbon挨得这么近，也就不难理解了。
参考资料：
https://x.com/Lentils80/status/2112770669549068684
https://x.com/steph_palazzolo/status/2112183963452059734
https://x.com/steph_palazzolo/status/2112205820440326307
编辑：摩西
秒追ASI
⭐
点赞、转发、在看一键三连
⭐
点亮星标，锁定新智元极速推送！
预览时标签不可点
不喜欢
Scan to Follow
Got It
Scan with Weixin to
use this Mini Program
Cancel
Allow
Cancel
Allow
Cancel
Allow
×
分析
微信扫一扫可打开此内容，
使用完整服务
:
，
，
，
，
，
，
，
，
，
，
，
，
.
Video
Mini Program
Like
，轻点两下取消赞
Wow
，轻点两下取消在看
Share
Comment
Favorite
听过
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Google Gemini 4 carbon's first physical exposure!
ASI Apocalypse
ASI Apocalypse
Shinjimoto.
Read this chapter in a novel reader.
Read.
Reading in WeChat Official Account novels
New Ji-won.
Argon's full open.
8 days
，
Google's next ace, he's in the arena!
Just now,
Google's strongest carbon is exposed.
。
The partners got it first and the LMArena made a blind look.
publish is right here.
。
As soon as word came out, all the testers jumped into Arena.
In one day, Carbon's measurements were blown to pieces on X.
Someone finished it with a breath.
Turn on Super Marie.
。
The last generation of flagships is still hot. The next generation is waiting outside the door.
。
Google Zemblin said Google started to set foot.
RSI's Runner.
，
Models are helping to build the next generation.
。
Super Marie's customs clearance.
Carbon's 3D is so strong.
Carbon on Arena also has the name "gemini-4-argon".
It's funny how the tester recognizes it -- it's slow.
。
Netizen Harshith found out that there are two models of this name. One's fast and good, the other's slow and amazing.
The latter is Carbon.
。
The evaluation account number Lumina uses Super Marie's prompt.
Make a whole game of carbon.
。
The main character came up with bricks, ate gold coins, drilled water pipes, went down to the underground gate and crawled out, jumped on the flagpole and entered the castle.
"COURSE CLEAR"
The gold coins and ratings were settled.
Lumina also has the same bodyware pail.
Give it to Argon, Carbon, Opus 5.5 and Sonnet 5.5. That's right.
。
According to Lumina, the quality of carbon is clearly higher than that of Argon, depending on the mission.
Two to four times slower.
Netizen Synopsis's body Valley pushed the picture up one level.
The cherry blossoms are running along the river bank, the red arch bridge is mounted on the water, the cabins are lit,
In a rainstorm, the whole valley went dark.
Netizen Salio made a 3D castle.
Lights can be exchanged between dusk, blues, moonlight and thin morning light.
The stone bridge beneath the city was lit.
Chetaslua, a well-known comedian, gave Gemini ' s mascot to Carbon to make an animation of 3D on why carbon is stable.
At the beginning, the mascot was staring at a carbon-12 atomic core. Then six electrons were lined up by 1s22s22p2, the orbits were melted into four identical Sp3 petals, and eight were shared with neighbours.
Finally, a whole diamond crystal.
。
Chetaslua says:
This is a very difficult question for the old Gemini model.
。
Carbon talking about carbon, Google's name this time is sort of being played by netizens.
From the carbon-12 atom core to the diamond crystal.
The developer Pankaj Kumar made it an interactive 3D crystal citrus.
However, at the time of the SVG, the evaluation began to become fragmented.
Netizen Binkg has asked carbon to draw a Coke bottle of SVG, full of beads, small words on labels and caps, with frosting, bubbles and back light.
But Haleemah, after a full day of testing, said,
Carbon's best is 3D.
Good SVG she hasn't seen yet.
Beyond SVG, the front end is still a problem. Lumina states that the front-end design and aesthetics of Carbon remain flat, missing the taste of the Anthropic model.
The same model, with a 3D surge and a forward end, is due to the fact that the two subjects are considered differently.
In the 3D scene, geometry, light and collision are all code-described, and the miscalculation goes down.
Right?
。
The front end is how big it is, how many numbers, where the buttons are.
There's no standard answer. There's only one.
。
So 3D can be calculated and...
Enhanced learning
Pushing all the way up,
Frontend
But it can only be slow with a large number of feedback.
Slower than that, but not lazy. Carbon's treating Argon's lazy.
When the tester suns his work, he almost makes it slow, but few complain.
Because he's not lazy anymore.
。
Lazyness. Argon's one of the toughest hairs ever. Diseases
。
It runs pretty well, and the developer does it in less than a week, and the reputation of "slazy" spread.
The technical media, Every, had completed the Argon survey and had made five recommendations for its use, and article 3 made it imperative to review its work.
In 38 baseline operations,
Seventeen times, Argon opened and closed the code review.
，
Only 30 milliseconds between.
。
The same recommendation contained a set of figures. I'm not sure what I'm talking about.
It'll only be 1.7 times the old one.
;Opus 5.5 to the top,
It'll take 18 times.
。
In other words,
Tell Argon to think about it. It doesn't think much.
。
Up to the Every CEO Dan Shipper's senior engineer benchmark, Opus 5.5 and Sonnet 5.5 took 85%, GPT-6 Astra took 81%,
Argon's only 21%. Bottom
。
And here's carbon, it's a direct reverse.
Business Insider was exposed before Google employees tried Carbon internally and said it wrote the code.
"I feel like Opus 5.5."
。
Argon's just out of carbon.
Google on the RSI.
From Argon to Carbon, Google's round beats are amazing.
On 30 September, the Minister announced that Argon would be handed over only to the cyber-security defense of the Fairwind plan; on 13 October, Argon was fully open to developers and subscribers;
Eight days later, carbon was tested by partners.
。
Google is getting faster, and the answer from Bryn is:
RSI (renewal self-improvement)
。
In a recent round-table conversation, he said,
Now the fellows have joined hands with the modeler to make the next version.
Argon's security report indicates that in September this year, DeepMind
More than 90 percent of the requests for code consolidation were made by the government.
AI
It's written over 75%.
。
It's a three-round merger request for full AI.
From 7.5% in April to 11.9%.
。
Deepmind's monthly code merge request, the darker the color, the higher the AI writing ratio.
It's not hard to understand how Argon and carbon get so close when the model begins to make.
References:
https://x.com/Lentils80/status/2112770669549068684
https://x.com/steph_palazzolo/status/2112183963452059734
https://x.com/steph_palazzolo/status/2112205820440326307
Editor: Moses
Second after ASI.
⭐
It's good, it's good, it's good.
⭐
Light up the sign and lock the new schizophrenics to the extreme!
Non-pointable tabs for preview
I don't like it.
Scan to Follow
Got It
Scan with Weixin to
use this Mini Program
Cancel
Allow
Cancel
Allow
Cancel
Allow
×
Analysis
A WeChat sweep opens this up.
Use full service
:
，
，
，
，
，
，
，
，
，
，
，
，
.
Video
Mini Program
Like
Take it easy.
Wow
Quietly cancel and watch.
Share
Comment
Favorite
Yes.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.