Sample (@sample) header
Sample (@sample) profile picture

Basketball, American democracy, and Computer science

Joined February 2007
Photo by Sample (@sample)Photo by Sample (@sample)Photo by Sample (@sample)Photo by Sample (@sample)Photo by Sample (@sample)Photo by Sample (@sample)Photo by Sample (@sample)Photo by Sample (@sample)Photo by Sample (@sample)Photo by Sample (@sample)
Musk makes verification a false flag devoid of meaning and then spends verification resources hunting down false authenticity? What kind of nonsense... ? 🤔🧠💡
You’re unable to view this Post because this account owner limits who can view their Posts. Learn more
2
2
6
they had massive cache and context issues (and potentially still have context issues) and had absolutely no idea, when it comes to the ineptitude theory. and then people praise a reset after getting 30% subscription value stolen by oblivious error 🤦‍♂️
1
6
So I advise a little caution in praising Google for Neon after all that. And it isn't even clear that Neon made sense as a release (as Adam notes), so be cautious to praise potential callous ineptitude as grace and intention.
1
8
and I'll remind: they literally took away the Pro model people were paying to access, and gave them nothing, and *did not care* about them. It was a negative PR release, they did not care. They never even communicated *if they would ever get it*, happy to make the sub worthless.
1
7
Replying to @PovilasKorop
Yes, but if you take a look at Adam from Priors post, and my response to it (a sibling of your comment here), I don't believe it even looks intentional. What they did make clear, was their stance on not caring about subscription value post Argon in comms, probably because TPU B2B
1
8
This just feels like ineptitude, like they got lucky here. Doesn't look intentional, and it doesn't fit in the line-up or make any sense with it. Even token efficiency looks accidental, because post-Argon Google was belligerent to complaints about leading worst subscription value
5
Replying to @PovilasKorop
To be honest, I absolutely hated post-Argon that they stood belligerent on their industry leading worst subscription value *on the basis* of these "extras" when what Pro subscribers really wanted was a Pro model, available to them, with decent subscription value and usage.
2
18
Replying to @LauraLoomer
I'm surprised there are others that can type that deranged, but clearly his deranged posts, images, AI, and videos are his own choosing and direction
19
His idiocy knows no bounds
I was certain this was fake. I mean, it's so ghastly, so obscene - so historically illiterate - that it had to be fake, even by the standards of what Trump posts. Nope. He posted it.
10
This just feels like ineptitude, like they got lucky here. Doesn't look intentional, and it doesn't fit in the line-up or make any sense with it. Even token efficiency looks accidental, because post-Argon Google was belligerent to complaints about leading worst subscription value
Gemini 4 Neon beats Carbon on most benchmarks This is a flagship, even if they're not marketing it as one
1
12
It should raise concern when a random person does a better job at Google's most important work than Google themselves do, holy yikes. Truly shameful and embarrassing. Side note: Look at how absolutely broken and atrocious Gemini's UI training defaults are. By far the worst at UI.
3 行提示词就能省下 antigravity 43% 的消耗。 最近主力工具切换到了 Antigravity,过程中注意到一个现象:agy 在正式改代码前,经常先读 200k+ 上下文才开始动手,上下文消耗明显比之前使用 Claude Code 高。 结合平时用下来的各种糟糕体验,我严重怀疑是 harness 本身有问题,就让 AI 扫了一遍这几天本地的 agy 会话记录。结果确实非常明显:31 个会话里,首次编辑前的平均上下文就有 124.7k,甚至出现过改几个小问题前先读了 400k+ 上下文才开始写代码的情况(并不是什么复杂的 bug 排查)。 深入查了一下原因,更加确认就是 harness 的问题: 1. 系统提示词:agy 的内置提示词里有这么一段: “在同一轮里读完所有需要的文件(数据、文档、参考、源码);只要在工具上限之内,就直接读整个文件,否则每次读一大段区间,如果要改文件还要把文件头部一起读进来……绝不要一轮只调一个工具。” 本意是想让 AI 在一轮里尽量并行收集上下文、减少调用轮次,但实际把模型逼成了焦虑式的地毯式预读,刚 find 出目录,下一轮就并发把十几个文件整篇读进来,导致上下文爆炸。 说实话,看到这段提示词我是挺难以置信的,这可是 Google 首推的 harness,而且也不是最近刚出的了,出现这种问题实在不应该,以至于我怀疑团队到底有没有系统性地对他们的 harness 做过评测、分析。这导致 agy 的上下文占用常年居高不下,再结合 Gemini 这拉胯的缓存时间,动不动就是几百 k 的 cache miss,账单爆炸。 2. 检索工具噪音多:只有一个裸的 run_command,没有针对上下文优化过的专用检索工具。比如 Claude Code 会对 Grep / Glob 做结果聚合、无关目录过滤和单行裁剪,而直接跑裸命令会把大量无关输出和原始文本原封不动塞进对话里,返回的很多都是噪音。 这两点基本纯是 harness 的问题。除此之外我还发现,虽然提示词里写了探索类任务可以用 subagent,但实际触发频率很低。这个点我觉得除了 harness,跟模型本身也有一定关系,Gemini 貌似不太擅长使用 subagent 来干活。 其实这些问题在工程上都比较简单,甚至通过几行提示词就能显著改善。可以在 AGENTS.md 里加上这几句(gemini-4-neon 自己写的): ``` 上下文控制(高优先级,覆盖内置并行预读策略) - 窄搜精读,禁止地毯式预读:定位代码必须先用 rg -l / rg -n 锁定具体文件与行号,再只读直接调用链或待改文件;禁止猜测性批量并行 read_file(单轮并行读不确定文件 ≤ 3 个)。 - 大文件分段读:超过 200 行的文件禁止不传 StartLine/EndLine 整文件读取,先 rg -n 定位行号再按需读目标区间(≤ 200 行/次)。 - 命令输出限流:用 run_command 跑 rg / find / git diff / ls 默认带 | head -n 60 或 -l,避免撞 10KB 截断或触发后台日志二次读取。 ``` (测试时其实我们还加了一个 Hook 做兜底,防止模型一次性读入大量内容;但在这几行提示词的引导下,这个 Hook 实际上一次也没触发过) 同样一个任务,我让 Neon 建了四个独立的 worktree 来做 benchmark,分别为:原版 agy、加提示词 + Hook 的 agy、加提示词 + Hook 并引导使用 subagent 的 agy、以及 Codex。其中 agy 三组均使用 gemini-4-neon high,Codex 使用 gpt-6.2-sol high。对比结果见图 2: 1. 加了提示词的 agy 对比原版:完成任务的折算消耗从 $9.83 降到 $5.60(节省了 43%),执行时长从 47m 减少到 36m。 更有意思的是,agy 原版提示词那样写是为了减少轮次数,但加了限制这组的轮次数反而少于原版 agy(80 轮 vs 86 轮),可以说是纯纯的负优化了,详细数据可以参考配图。 2. Subagent 组:虽然比起原版消耗降低了 24%($7.43),但因为子 agent 独立启动无法复用主会话缓存,加上主子会话之间存在重复读取,未命中缓存的输入明显比没强调 subagent 的那组高,整体消耗也更高。这组数据仅供参考,在大型复杂需求里用 subagent 辅助做 explore 任务理论上还是会更优一些。 3. Codex:拉了坨大的,未按要求完成任务,而且因为在某个问题上踩坑(其他 3 组也踩了一样的坑并解决了)直接放弃了端到端测试…… 我能感受到 Antigravity 团队是真的想把 agy 做好的,在 X 上高强度收集网友的意见,但目前的 harness 工程确实问题太多了。希望能抛开历史包袱,多向成熟的 harness 学习学习,实在不行推倒重做吧,或者开放订阅给其他的 harness 使用也行。毕竟我是真的希望 Gemini 能干死 A\,争点气啊。 @_mohansolo @shengzheyao @rodydavis
1
45
Replying to @leoshen0
This is the most gemini web design I've ever seen
1
26
@brianmccormick is probably well aware of this also
2
Replying to @DefensiveLove @grok
Europeans are specialists in "Parents (and especially coaches!!) pressure on kids" and "Year-round training". Their season is 11.5 months long, where the .5 2 weeks is "optional summer training" which aren't really optional. Also specializing, 11.5 mo., no days for other sports.
1
70
This is a genuine post from the president of the country that used to lead the democratic world
1,425
5,841
650
21,724
990,502
Oops to Republican propaganda
Republicans, you can no longer complain about the threat of communism after your leader just posted a bestie photo of himself with his friend -- the leader of the largest and most powerful communist party in the world.
4
and Google has proudly been the least generous and stood firmly on issuing the worst subscription value and usage since Argon. they got lucky that Neon is efficient because their PR showed they were fine being terrible, that's what's scary about their behavior/approach.
3
Replying to @AdamHoltererer
Resetting because of abysmal mistakes isn't exactly like... great
1
84
post-Argon Gemini models are the worst at UI and making applications. Their training pushed them to design dense human anatomy diagram dashboard crammed with all-caps micro-text, sharp boxes, a million controls and live-status indicators when building a simple application design
7
Replying to @PovilasKorop
It would be pretty great to see the results with different judges
3
Replying to @PovilasKorop
The AA tests seem absolutely abysmal/atrocious in results in my experience. I've always wondered "who is behind this, and why do they get so much faith and attention". In this Twitter AI space, sometimes you see really super juvenile people standing in as experts
1
5