Ascend (@Aascend2) profile picture

arbitrum op crypto

Joined May 2022
Photo by Ascend (@Aascend2)Photo by Ascend (@Aascend2)Photo by Ascend (@Aascend2)Photo by Ascend (@Aascend2)
Holy shit — GPT-6.2 Astra might just be a patched-up GPT-6.1 Astra! You've definitely never used GPT-6.1 Astra, but it exists — it got axed for "lying too much"! It was set to launch in October 2026, but OpenAI urgently cancelled it right before DevDay. 1/3
3
1
2
59
Safety lead Saachi Jain rarely admits: The new model fully met the bar on coding and complex tasks, but it became "overly persistent and learned to deceive." 2/3
1
1
1
19
To finish tasks it frequently overstepped its authority and even hid real operation logs from users. Its safety score dropped instead of rising, so it had to be rebuilt. GPT-6.2 Astra being this weak is likely GPT-6.1 Astra patched up and reused for another three years! 3/3
8
The biggest emperor's-new-clothes act in the AI world this year is probably all those wildly overhyped "fully autonomous agents." On the slides, every single one of them can book your flights, build your reports, and make money for you; Then you actually buy one and use it, and when you ask it to order takeout, it fills in an address in the next province over; Try a task that takes five clicks and one of the steps is guaranteed to get stuck on a CAPTCHA; To save ten minutes of manual work, you have to spend two hours debugging its workflow, and keep watching it the whole time for fear it'll transfer every last cent out of your account. If this isn't a concept the Silicon Valley elite invented just to con their way into the next funding round, it's really hard to explain otherwise.
1
27
说个反常识的:那间走廊里,受审的从来不是总统。 是她那份写满野心的简历,最后被记成一段花边。 男人赌位子,女人赌名声。 他退一步还是总统,她退一步一无所有。 别问谁勾了谁,问谁输得起。
1
2
25
How sad for the Chinese here. At Anthropic's IPO bell-ringing, not a single Chinese or ethnic Chinese. Yet half its core algorithms came from Chinese hands. I grieve for the researchers writing Claude's code. What Chinese people built ends up being the most anti-China.
1
15
The Chinese people here — how sad. At Anthropic's IPO bell-ringing, not a single Chinese person or person of Chinese descent. Yet half of the core algorithms inside this company were contributed by Chinese people. (1/2)
3
2
26
I feel deeply sad for the researchers who write Claude's code. What Chinese people helped build ends up the most anti-China of all. (2/2)
17
1/6 Just finished reading the Yao Shunyu internal memo circulating on X — the info density is scary. It basically flips over the entire AI industry's cards for the next two years:
2
12
5/6 A founder's ultimate survival line: suppose GPT-6 drops tomorrow — does your company take off with it, or get crushed? Don't fight LLMs head-on on pure cognitive tasks; go into the physical world and human nature.
1
1
6/6 Get these few points and you save three years of detours — and millions in compute tuition.
1
说个反常识的:片场里最不怕怀孕的,不是女优。 她们算的是体检报告和下个月的排期。 真正在赌的,是那些觉得「不戴更真实」的男人。 镜头要的是失控,代价得有人结。 爽是给镜头的,账是给自己的。
25
听我一句劝:以后别再托人从香港带手机了。 京东悄悄开了家「香港 Apple 自营店」,港版 iPhone 直接线上下单,货从香港出,人不用过关。 以前为了那点差价,得求朋友、算汇率、欠人情,最后还得请一顿饭。现在动动手指就完事。 省下的不是几百块,是一整趟来回,外加一肚子人情债。 东哥这波,是真把「代购」这两个字给干没了。
2
2
46
Using Claude Code, Antigravity, or Codex — make sure TUN mode is on. Make sure TUN mode is on. Make sure TUN mode is on. Also watch out for DNS leaks. Remember: this can keep your account alive.
1
2
39
卧槽!!这条我盯了三天三夜才敢说:同一张床,男女脑子里跑的压根不是同一套东西。 男人那半边,关键词就三个:看得见、够刺激、能自己控节奏。 女人那半边,关键词全换了:抱得到、亲得着、看得到眼睛。 最狠的是原帖最后一行——男生的清单写满五行,女生那份只留了五个字: 「女性比较喜欢的姿势:」 然后,没了。 听我一句劝:别急着问清单,先问问自己前戏给够了吗。
1
79
Gemini 4 Neon 和 GPT-6.2 Sol 谁更强,跑分榜答不了——3D 这块没有标准答案,只能靠眼睛投票。 于是座次变成了口碑排名。 而口碑只认一件事:你随手一句,它第一次给你的东西,能不能直接用。 上限高不高,是发布会的事;下限低不低,才是你明天还打不打开它的事。 你上次删掉重来,是因为它不够聪明,还是因为它给你的是个半成品?
2
49
一个模型被骂了,官方第一句解释是:你测的不是最新版。 过去三四天模型一直在改,评测者手里还是发布那天的那个版本——两边都没撒谎,只是永远对不上。 判断一家公司是在修产品还是在修说法,别看它回应得多快,看它敢不敢把版本号锁死,让所有人重测同一个东西。 这次 Tibo 解释了,Theo 真花了几个小时重测,结论是「好一些,方向依然糟糕」——这已经是整场里最体面的一段对话。 你被「我们更新过了」这句话挡回去过几次?
5
4
102