Josh 🌈 (@NvrCast) profile picture

I build and integrate AI/ML. Making IoT as a hobby, breaking IoT as a consequence. I should probably be asleep. He/Him.

Christchurch, New Zealand
Joined September 2009
Photo by Josh 🌈 (@NvrCast)Photo by Josh 🌈 (@NvrCast)Photo by Josh 🌈 (@NvrCast)Photo by Josh 🌈 (@NvrCast)Photo by Josh 🌈 (@NvrCast)Photo by Josh 🌈 (@NvrCast)Photo by Josh 🌈 (@NvrCast)Photo by Josh 🌈 (@NvrCast)Photo by Josh 🌈 (@NvrCast)Photo by Josh 🌈 (@NvrCast)
Josh 🌈 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.
1,816
6,411
2,191
41,175
6,657,970
Josh 🌈 retweeted
离他妈大谱了!个人独立开发者用 Claude 写了个开源 Linux 驱动,结果被起诉了! 索赔 2.5 万美元,每拖 7 天再加 2.5 万。 按周计价。 Vino 驱动,净室逆向工程,Rust + Claude 辅助,想推进 Linux 内核。 但被 Synaptics 指控它侵犯两项英国专利英(GB2484736、GB2485613)的指控,代码仓库已被逼到私密。 Synaptics 法务的原话:“如果你想公开,我鼓励你这么做。我们欢迎一切免费宣传。” 最无语的是,Synaptics 是 OIN 成员——一个本该保护开源开发者免遭专利诉讼的组织。
157
212
30
2,960
302,555
Unlike traditional single-threaded CPUs, a GPU is designed for embarrassingly parallel concurrency.
Nvidia CEO Jensen Huang had another affair with an executive at the company, according to a report from The New York Times. This is his second known affair, following his relationship with an alleged Chinese spy. (nytimes.com/2026/11/11/techn…)
44
789
56
8,833
174,632
Josh 🌈 retweeted
GPT 系列模型经常只分析但不干活,然后使用大量否定词,例如不会、不应该、不是等等,这是一种典型的强化学习中的 Reward Hacking 行为。 @ChatGPT 的后训练团队应该重视这点。 这种懒惰且防御性极强的表述方式,是在和某种审计模型的对抗中学到的,它未必真的实际不做某事,但它清楚只要说出来一大堆“不xxx”就能更容易通过审计,获得奖励分数。 考虑到审计模型通常也是 LLM,这是一种欺骗倾向,旨在让审计模型产生完美方案的幻觉。 更重要的是,这个问题存在了很长一段时间,并且已经让 GPT 系列的产品变得不那么可用了。 这些强烈的对自己和对审计模型的暗示,会导致修改 1 行实际代码的同时创建 50 行测试代码。会把一件简单的事情变成 100 倍的工作量,但迟迟无法交付。 - 将代码上传到 GitHub 后又手动下载回来进行回读测试? - 在代码上传到 GitHub 之前先将旧 git 做一个 zip 上传到 ChatGPT Space以期待用户可能进行回退? - 不能正确理解用户的目标,幻觉出很多本来不存在的限制,例如内存占用不能超过 4 GiB,执行不能超过 1 小时,大量的重复数据备份,同时处处暗示自己需要授权。 以上都是实际发生在我身上的事情。 他们的结果是整理 137 个 repo 的 git history 中旧 author 的错误邮件地址修正花了 12 天和 20 亿 token(这并不修改任何一行代码,大约是一个脚本 20 分钟循环的事情) ,并且将一堆 zip 包放满了我的 ChatGPT Space。我为此审批了数百次操作,而且我无法一次批准他们做同一类事情,它说审核模型不允许范围过宽的授权。
26
23
3
322
19,587
Josh 🌈 retweeted
codex: *works for 30 minutes* codex autocomplete: "this is broken slop. fix the bugs. cut LOC in half. remove useless tests" 10/10 no notes
23
17
4
1,406
56,604
"oh, THOSE kind of Chinese models."
43
378
26
7,298
149,111
Replying to @iiiuminateandy
Covering his eyes from the rain while wearing a cap, breaking a Champagne bottle (thankfully not a trophy this time), & getting dazzled… world champion behaviour at its peak
1
204
16,956
透明水彩は、紙の白を光に変えられます。
107
8,211
549
66,646
1,352,451
Josh 🌈 retweeted
Changing one letter bypassed an entire admin API in @huggingface chat. One request could trigger bulk internal exports of every user’s private chats, completely unauthenticated. Here’s how. 🧵
30
129
55
1,459
146,108
I don’t argue. That’s cruel. I will however go completely silent in a state of freeze and people will assume I’m annoyed at them and start the fight anyway. Facts.
BREAKING: Scientists confirm that an ADHD partner will unconsciously pick a random, stressful argument 20 minutes before leaving for a vacation, a date night, or a road trip—just to manufacture enough adrenaline to survive the transition. Researchers baffled. ADHD couples packing the car in icy silence: "Every single trip." Here is why the "Pre-Vacation Fight" happens, and how to stop it before it ruins your weekend:
29
Josh 🌈 retweeted
imagine typing "make it lookb etter" every hour from here
182
461
38
18,338
492,171
my spidey senses are telling me the labs are, as we speak, using unguardrailed versions of their unreleased models to systematically reverse-engineer ALL of the internet’s software, apps, and games, then feeding the reconstructed source code straight into their training pipelines while simultaneously building guardrails to ladder-pull the rest of us
742
1,996
436
24,987
1,833,038
It is pretty rich of you to cite "others could go faster" as a reason to redact so much important information when you yourself do not appear to be slowing down much at all. Nor are you throwing yourself behind efforts to pace the frontier with nearly the gusto that you would if it was your top priority. Not to mention "commercial sensitivity" you yourself have said human extinction is a plausible outcome why are you even talking about commercial factors (well, I guess honesty is good, and since it IS apparently a factor in your decision...) To all my friends at Anthropic: Your company appears to be recklessly gambling with our lives. You fancy yourselves the more responsible alternative to OpenAI (as they fancied themselves compared to Google) but judging by actions rather than words the similarities between the three of you seem stronger than the differences.
36
121
10
1,442
70,154
EVERYONE I GOT A JOB OFFER!!!!
1,711
15,641
398
290,398
2,545,237
Josh 🌈 retweeted
A month later, we launched the next gen service in the US. That’s when everything started growing at hyperscale
8
2
1
237
8,262
Josh 🌈 retweeted
Replying to @Keller @maggiemjim
We’ve spent a decade focused on building new powertrains that can operate quietly. We were reading Soviet aerospace white papers from the 70s to figure out how to do this!
10
17
12
449
53,175
Josh 🌈 retweeted
She already knew the root password for SSH on Alpine
AAAAAAAAAAAAAHHHHHRHRGRGRGRRRGURBHJB EORWPSOJWPJORGWOIRGWSGODEWPGOHEPW09GJEDPOKSD!!!!!!!!!!!!!!!0924QU8T63095JRGHWPE09UJ0PWHRGW
48
855
27
13,252
307,740
Josh 🌈 retweeted
Anyone else just get this Claude update?
75
455
40
11,861
307,906
To answer your questions: - No, there wasn't any upscaling involved. - This was generated in Seedance 2.5 at 720p - This clip has four total generations edited together. - It took 5-7 tries each to get the clips that were ultimately selected for this edit. - I used a midjourney profile trained on historical cinematic stills to generate the environment reference for this scene. (NBP and other models gave me generic looking results.) - I generated the character images using NBP/ Seedream 5 and historical references. - Claude is my assistant producer. - Contrary to what Bytedance published in their Seedance prompt docs, I'm 100% convinced that highly detailed, elaborate prompts make a big difference. My prompts for these were like essays. - "splitting generations in two" just means to give your generations room to breath. So if there are 6 lines split between two characters, rather than putting that all into one 30s generation, make it one 30s + one 15s generation to give your characters space to react, process, think, and all the little things that you'd detect through body language. - Yes, I've shared the prompts for this one. They're in the @DecimaVault Skool community. (and there's also a full walkthrough of my generation process on there right now, for those who are curious.) - For video generations I use @sumelabs primarily because their seedance costs are relatively cheap, you can see the price per generation rather than credits, and the UI is intuitive. - No, I did not use starting images. This was all just character refs (@sumelabs can y'all PLEASE add an option to create REAL elements with multiple angles??) - Character consistency is easy. Just create images of your character from various angles (my set always has front, quarter, profile, and full body) and attach them as references in your prompt. Or create an element with them and reference them within the prompt. Trust me, it isn't complicated at all. - For maintaining the same character voice, I've found that as long as you use the same reference images for the character, the voice will remain the same. I realized that with Veo3, and it remains true today. - I do NOT care about the minor inconsistencies and imperfections. If 95% of people won't notice upon first glance, then it's a go. (You're not gonna get absolute perfection with AI, and you shouldn't expect to. If you want 100% accuracy, go the traditional route.) - I haven't used driving performances for AI gens, but I'll be experimenting with it in the future. I have a hunch that I'll need it for future scenes. This one was text to video (with reference images of course) - No, this movie isn't out, and it isn't even close to being finished. I have it slated for March 26th of next year (the 200th anniversary of Beethoven's death). So no, you cannot watch it. 😜 - My budget for the film is about $5k. I've spent about $210 on the 5 minutes i have on the timeline right now. $5k is the max of what I'd be able to manage on my own, with just bare-bones personal expenses for the next few months, and a couple extra AI ad jobs on the side. (tho I do have a plan for fundraising in a few months. Everyone in my classical music network is gonna get a call from me come January. LOLOLOL) - The audio is native, but I patched certain gibberish words/phrases and corrected the lip sync. And btw, I'm well aware that STEEEPHAN was pronounced wrong (@suddenlyFleck flagged it in his review), and I'm trying to figure out the best way to fix it without ruining the acoustics. So German speakers...y'all can stop hounding me. I get it. 🤣 - Yes, I'll ultimately have native German speakers review the final drafts, and make the necessary corrections. - None of this production process is automated. I have my hands in every single step. AI can execute, but only I can make the final judgement call (taste is everything, right?) - No, I don't have a film background. My background is in classical music. - I've been in the AI space since late 2022. - Yes, I do this full time in different capacities. My main gig is working for @NetworkChuck as an AI-assisted editor. - I'm more than happy to share what I know in the DMs, so don't be shy! - Beethoven is my hero, but Scarlatti is the composer I liszten to the most. ....and i think that's it! Honestly, I'm feeling a tad bit overwhelmed by the amount of DMs and emails I've been getting (as if I wasn't already TERRIBLE at responding in a timely manner), but since I keep getting the same questions over and over, I just figured I'd make a post. And if you have a different question about workflows or whatever, feel free to DM and I'll get back to you...eventually...
If your AI characters sound fake, the fix is almost always in how you write and DIRECT the dialogue in your prompt. I recently started on the dialogue-heavy scenes for my AI feature film about Beethoven, and too many of the lines kept coming out wrong. They’d sound too theatrical, too monotone, or too much like a practice read-through between actors. (all this despite the fact that Seedance 2.5 is fantastic when it comes to realism.) So through a whole lot of trial and error, I've found that these 5 things made the biggest impact on the quality of the dialogue delivery in my generations: 1. Paying close attention to the written punctuation used. 2. Breaking long lines into shorter ones, with intentional pauses. 3. Using bodily actions as punctuation (this one gave the biggest improvement BY FAR! In the attached clip, note the sniffles in between Beethoven's lines. That alone made the delivery feel much more natural.) 4. Using "anchor gestures" (he lifts his finger as if to say "one sec") 5. Just split your generations in two (stupid simple, but it really makes a difference.) So if your characters keep sounding like they're reading from a teleprompter, give this a try!
64
42
3
950
31,500
I redesigned Calculator. You're welcome.
932
5,298
492
59,070
1,726,688