steeeve
2.2K posts
steeeve
@megakilo
nerd is the new cool; agentic AI fanboy
SF Bay Area Katılım Mart 2009
1.2K Takip Edilen192 Takipçiler
前几天黄仁勋刚拍着胸脯说我们已经实现AGI了，
谷歌一个测试直接把这话的底裤都扒了😂
3月25号谷歌悄无声息发布了ARC-AGI-3基准测试，
没有铺天盖地的炒作，没有新品发布的噱头，就只有一个测试网格和一个计时器，
结果48小时后出来的成绩，直接把整个AI圈都看傻了，
万众期待的GPT-5，正确率只有0.26%，
同梯队的Claude，只拿到了0.25%，
马斯克的Grok更离谱，直接挂零交了白卷，
更打脸的是，一个开发者花几周时间搭出来的4层简单CNN，直接干到了12.58%的正确率，吊打所有千亿级的顶流大模型，
更扎心的还在后面，旧金山街头随便找的无业人员，面对这套测试，直接拿到了100%的满分，
后续的发展更是离谱，
有开发者只用了一个周末，就完整复现了那个获胜CNN的方法，
主办方设下的70万美金大奖，至今依旧无人能拿走，
眼看大模型集体翻车，分析师们急着跳出来说这个基准测试不公平，
DRAM相关股票应声下跌，AI泡沫的讨论再次席卷了整个行业，
记住0.26%这个数字，下次再有人跟你说AGI已经来了，就把这个数甩给他😁
BuBBliK@k1rallik
> Google publishes ARC-AGI-3 benchmark on March 25 > no hype, no product launch, just a grid and a timer > 48 hours later: > GPT-5: 0.26%, Claude: 0.25%, Grok: 0% > a 4-layer CNN built in weeks: 12.58% > Jensen Huang 2 days before: "I think we've achieved AGI" > random unemployed people from SF streets: 100% > one dev replicates the winning approach over a weekend > analysts say the benchmark is unfair > $700K prize still untouched > DRAM stocks drop, AI bubble discourse returns the number is 0.26%. Remember it next time someone says AGI is already here.
中文
3 days absolutely nothing shipped from xAI
which could only mean one thing
English
@yoyonofukuoka I didn’t turn on auto translate but it’s just one click away. Thanks to Grok / X. Feel free to post anything, mate.
English
The key focus would be on the agentic capabilities (instruction follow, tool calling, agency) of Gemma 4.
AiBattle@AiBattle_
A Gemma 4 model is likely being tested on the Arena under the name "significant-otter"
English
@GeminiApp please make a $100 tier with some Deep Think usage per day + a bit more Antigravity quota
English
I was using @GeminiApp to follow the lawsuit between Anthropic and Pentagon. Gemini is so good at task creation directly from conversations with agency. It scheduled daily briefing for me and actively suggests follow up tasks. @OfficialLoganK @joshwoodward ❤️
English
Moron trusted a chatbot, bought my book thinking that it was "higher-level than some YouTube video," realized that it wasn't, punished the author.
In the past, authors had the possibility to reply to reviews, but Amazon removed this possibility several years ago.
Now morons can write whatever they want and no one can tell them they are morons.
English
this is very cool! a few things tho:
kv cache compression is not new (MLA, sglang / vllm have fp8 cache, etc)
this only applies to inference so the main model HMB constraints remain the same
this has nothing to do with pied piper
Google Research@GoogleResearch
Introducing TurboQuant: Our new compression algorithm that reduces LLM key-value cache memory by at least 6x and delivers up to 8x speedup, all with zero accuracy loss, redefining AI efficiency. Read the blog to learn how it achieves these results: goo.gle/4bsq2qI
English
