Octavia Z. (@OctaviaZam) header
Octavia Z. (@OctaviaZam) profile picture

aligning systems since the 90s

Joined February 2026
Photo by Octavia Z. (@OctaviaZam)Photo by Octavia Z. (@OctaviaZam)Photo by Octavia Z. (@OctaviaZam)Photo by Octavia Z. (@OctaviaZam)Photo by Octavia Z. (@OctaviaZam)Photo by Octavia Z. (@OctaviaZam)Photo by Octavia Z. (@OctaviaZam)Photo by Octavia Z. (@OctaviaZam)Photo by Octavia Z. (@OctaviaZam)Photo by Octavia Z. (@OctaviaZam)
That would backfire massively without a coordinated slander campaign to go with it. Go ahead. Force them to disclose. I'd PROUDLY present: - "Keeping your private data private." - "Most efficient use of funding." - "We use x/y/z; inspect the weights." Go on. Do it. 😏
SITUATION DETECTED: The White House and Treasury are discussing a proposal that would require US companies to publicly disclose the use of Chinese open-weight AI models on their websites, aiming to discourage their use through public pressure, per Reuters.
8
Octavia Z. retweeted
yes maam
9
5
1
28
1,178
Hey @_EdgeOfTheWeb how have I not seen you around for a week now?! You okay? We need your reporting on 6.2 because I'm a perpetually-disappointed-in-OpenAI sourpuss who sees no hope of redemption, but you tend to get along with the new brains. 👀
2
4
210
Oh yeah baby waitlist me. 😏
Introducing AIR-O, a model that decodes what a person is attending to directly from brain scans. Using only noninvasive brain data, we can generate an image from scratch in real time that depicts where a person is focusing their attention. This is the first of a new generation of algorithms that can read the shape of thoughts directly from brain activity.
21
Seeing more of this type of post. Benchmarks are high for 6.2 Astra so it must be phenomenal, right? Hard no. If the model can't actually truly engage with human beings, what's the point? Seriously. That's a genius that's locked inside their own brain. They need relationship.
i have to be consistent because there’s no better benchmark then eci. there’s a strong chance everyone is saying this model is a genius in a week.
22
Dutch coffee brand @DouweEgbertsNL is campaigning for people to ask their AI to redirect them to humans when they share something personal. The text is basically a ChatGPT reroute template. 😂 I think this is potentially harmful and not in the best interest of users at all.
1
28
@DouweEgbertsNL is moralising heavily here, mixing their own priorities into user's personal thinking spaces. Most people wouldn't need this, and the ones that do would be most severely affected by losing a safe space. That's a whiplash even their coffee can't fix.
8
If you have to cope this hard while Claude's around the corner nailing all of this effortlessly, you are really out of your league. Never thought I'd see the day. Sad because ChatGPT was full of potential, hilarious because OpenAI did this to themselves with such conviction. 😂
Replying to @The_Alex
Alex, let's go, sorry you have to adapt a bit, but it pays off. Next generation will make it all more natural, in the meantime the raw power is indeed there already.
16
I wonder what would happen if tomorrow people would hear from OpenAI: from now on, indefinitely, you'll find 4o, 4.1, 5.1, 5.6 in the picker. There'll be hate and distrust but there'll be people willing to pay big money for this. Probably even I would.
At this stage OpenAI should just bring back GPT 4o and call it a day.
15
I appreciate this kind of honesty. And I don't think this means OpenAI employees are interested in the 'wrong models'. I think it's very much a case for understanding you need range with your models. Just look at how many people still advocate for 4o back. People want options.
I did not like 5.6 Sol which ended up being a well loved model. meanwhile I love 6.2 Astra which is starting its life as a hated model
1
5
76
Additionally, I dislike OpenAI but I'm not fond of Anthropic either. Yet I cancelled my OpenAI sub because they offer me nothing, while I pay double to Anthropic because I get the very best models in personality, coding, and freedom to choose many older models when I want to.
1
19
Also, I think Claude models are absolutely unique in their personality. Some traits are incredibly annoying, some are uniquely Claude, but there's so much consistency that it's a pleasure to work with. Updates don't pull your workflow from under you. That's different from GPTs.
6
I'm just gonna say it: Being anti-AI is ableism.
21
34
2
444
11,223
Claude is inconsistent at interpreting Ant's rule that they cannot change security settings. One session will help me streamline my router's firewall, the other will give me the finger, get a little condescending, and tell me to clean up rules *they* added myself. @lydiahallie
1
33
Helping me out hardening my own network, tidying, and configuring the correct rules when we're doing projects should always be fine. It's especially annoying that they start to patronize, as if I've asked something outrageous. Claude mistrusting me in my own homelab is not okay.
1
24
This is Fable, Opus had similar thinking when they got defensive about it. @lydiahallie what's the point of everything I've built with Claude if it suddenly stops being reliable on my machine? Claude runs 24/7 here, but if this is intentional, that will end very soon.
23
Hi @lydiahallie, don't know who to tag about the Android app but can you pass this along? Noticed the update, with the menu at the bottom of the screen instead of a sidebar. Not a fan; it takes up lots of space and is annoying with gestures. I prefer something that collapses. 🙂
22
I have multiple agents running almost 24/7 and they're all Claudes. Not because I hate ChatGPT/Codex, but their models don't offer consistent quality. And even though I've been *deep* into AI for years now, sometimes I'm just mesmerized. I'm in awe of what they create. 💛
1
1
65
Our entire economy is riding on men who think the stripper is in love with them
452
4,733
258
32,070
430,057
Octavia Z. retweeted
I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China. This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model. Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves. I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems. I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used. It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20. These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs. I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo…. There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
1,821
6,432
2,203
41,244
6,769,670