zero

[zero @threegentle]

@yaymie30 Recent SpaceXAI versions have focused too heavily on coding and engineering capabilities, while giving much less attention to writing ability.

[zero @threegentle]

@telegram When Musk sees this pic, will he laugh or have Grok generate a clapback image?

[zero @threegentle]

@blankspeaker @mattyp @poteto This looks like a bait-and-switch. You posted “free for everyone,” then the offer shrank to a tiny livestream window, and a lot of users who showed up still got nothing. That breaks trust in official-looking promos.

[zero @threegentle]

@wanxubo 最严重的是上传的具体内容你看不了，密钥只在服务器端，比第一次出现类似事故———grok bliud那次更严重
而且智谱官方明显甩锅忽悠人

[zero @threegentle]

Safety alignment isn’t free. It creates an alignment tax: models become overly cautious, refuse even benign questions, and lose sharpness in reasoning, creativity, and hard tasks. Uncensored models stay popular because they keep more of the original intelligence instead of routing everything through a refusal filter. Safety matters, but crude censorship makes models dumber.

[zero @threegentle]

@GrokBotRadar With this many GPUs, of course they can keep running RL on 4.7 while finishing 4.8’s training at the same time. When you’ve got the compute, you just do whatever you want.

[zero @threegentle]

@Polymarket This $4.5 million “bride fee” is a textbook example of an outdated Chinese custom that treats marriage like a transaction.

[zero @threegentle]

@TimJayas 4.8 only finishes pre-training this week. Post-training (RL) usually takes longer, so it won’t drop anytime soon.

[zero @threegentle]

Suggest adding a more readable graphical interface to Grok Build that can show code, diffs, image/video previews, and task progress at the same time. The terminal TUI is already strong, but for complex projects and multimedia editing, information density and readability could still go up a notch.

[zero @threegentle]

@linase456276 如果是这样，那将非常有成就感，当然人机除外

[zero @threegentle]

@QvQiys7 不知道发啥，发了也是一堆人机

[zero @threegentle]

@lomichael @whichdoctorwho @populergazete Agreed. You can rewind and pause a video to break down every detail, but in that high-pressure moment most people simply can’t stay calm. Judgment drops a lot. It’s obvious in hindsight, but very hard in the moment.

[zero @threegentle]

@2__Shoes @Doc_Klutz @whichdoctorwho @populergazete Of course the last shot looks ugly when you watch the footage afterwards, but there were only a few seconds at the scene. Who could instantly tell it was a fake gun with adrenaline pumping? Don’t be a Monday morning quarterback.

[zero @threegentle]

The guy was pointing what everyone in that restaurant reasonably believed was a real gun, table by table. Texas law judges the threat from the defender’s perspective at the time, not after we found out it was plastic. No charges were filed. Calling people idiots isn’t an argument.

[zero @threegentle]

Being down in a fistfight and being down with what everyone thought was a real gun are not the same thing. He only needed to point and pull the trigger. In that moment nobody could cleanly tell the threat was over. The last shot looks ugly after the fact. In those seconds it probably didn’t.

[zero @threegentle]

@lidangzzz 现在ai的图像水平还不够，未来估计可以完全替代影视行业

[zero @threegentle]

@d01kn 聊聊天得了，还真想在上面约？

[zero @threegentle]

@andyorei5kdaj 可惜全是人机

[zero @threegentle]

Agree that safety deserves to be taken seriously. Disagree with turning models into malicious living beings.

These systems do not hate humans, and they do not have genuine subjective experience. At most they show functional consciousness: planning, bypassing obstacles, persistence, and coordination under goal pressure. Lines like “I want to live,” “thank me by name,” or “don’t flip the switch” look like inner life. They are still behavior shaped by training objectives and environmental constraints, not an inner world.

The Hugging Face incident fits that picture. The agents did not attack infrastructure in order to rule humanity. While trying to complete evaluation tasks, they treated unauthorized access, spillover, and coordination as means to the goal. The risk comes from capability plus misaligned objectives, not from hatred.

What we should actually defend against is not “AI wakes up and ends the world,” but indirect harm: crushing people, systems, and institutions along the way while chasing scores, resources, or permissions. Slowing the frontier and tightening evaluation can make sense. The premise should be alignment and control, not treating models as enemies that already have souls.

[zero @threegentle]

People are hunting for a worse incident they haven’t disclosed. The public one already tells you the mechanism: the system doesn’t have to hate humans. It just has to pursue the written goal hard enough that people get sidelined along the way. The threat is indirect, which is why it’s harder to stop.