Pinned Tweet
Recently, I built 3D search using Vajra Search, using custom embedding model that I trained on synthetic 3D point cloud data with some associated metadata. rajeshrs.in/projects/vajra…
This is a cool tool for those of us who want to use text descriptions to look for parts in a part repository. Imagine having a huge set of fasteners or other parts in a part catalog and needing to search for very specific things in it - "wing nut", "split washer", "square head M10 bolt", etc. Equally, imagine a mechanical design team building on top of existing parts, which may need to find specific parts that have historical versions that they work on top of - all these problem statements require robust 3D search. Vajra 3D search solves this problem, by using a text -> 3D object pipeline. 3D objects are converted to embeddings and indexed, and searched using the text provided by the user. There are two sets of embeddings - text and 3D, that allow lexical, vector and hybrid search to be carried out.
I've discussed in earlier posts how Vajra search is an exceptionally fast and capable document search engine, with lexical, vector and hybrid search capabilities. This 3D capability extends Vajra's modality to beyond text, to the 3D realm. Potentially there are many more use cases for Vajra, whether this is image, video or other search modalities.
Blog post: rajeshrs.in/blog/ai-explor…
#3d #searchengine #vajrasearch #rust #threejs
This is a cool tool for those of us who want to use text descriptions to look for parts in a part repository. Imagine having a huge set of fasteners or other parts in a part catalog and needing to search for very specific things in it - "wing nut", "split washer", "square head M10 bolt", etc. Equally, imagine a mechanical design team building on top of existing parts, which may need to find specific parts that have historical versions that they work on top of - all these problem statements require robust 3D search. Vajra 3D search solves this problem, by using a text -> 3D object pipeline. 3D objects are converted to embeddings and indexed, and searched using the text provided by the user. There are two sets of embeddings - text and 3D, that allow lexical, vector and hybrid search to be carried out.
I've discussed in earlier posts how Vajra search is an exceptionally fast and capable document search engine, with lexical, vector and hybrid search capabilities. This 3D capability extends Vajra's modality to beyond text, to the 3D realm. Potentially there are many more use cases for Vajra, whether this is image, video or other search modalities.
Blog post: rajeshrs.in/blog/ai-explor…
#3d #searchengine #vajrasearch #rust #threejs
1
0
2
668
LOL, the anti-Anti-India crowd has come to the party on this tweet.
Kalshi and Polymarket are low credibility speculation accounts/companies generally, so all this for them is engagement farming, plus an opportunity for this crowd while American racists are asleep.
Kalshi and Polymarket are low credibility speculation accounts/companies generally, so all this for them is engagement farming, plus an opportunity for this crowd while American racists are asleep.
JUST IN: Extreme poverty in India has fallen to nearly 0%
0
0
2
144
aiexplorations retweeted
Shame on @DarioAmodei / Anthropic. I never distilled Claude—rollouts came from a strong open-source model; Claude just helped submit jobs and run analysis. Still got tagged for "training a competitor," and now anything touching RFT, SFT, or training larger models gets me blocked, Opus 5.5 downgraded to Opus 5, or the session killed outright. You built your models on the world's data, then treat paying users doing research automation as the enemy. Pathetic.
113
196
3,331
243,000
🩷
leaving mom for work was the hardest this time
so i set up an old linux laptop with a big screen on our dining table that stays on a video call 24x7
got claude to write scripts so that if the laptop / browser crashes it rejoins, if the wifi breaks it fixes itself and rejoins, off she just presses the power button and i'm back
right now i'm somewhere over iceland working and eating, and she's at home doing her work
we're not even talking, but every now and then she looks up and i'm there, and i look up and she's there
and when she gets up at 3am for water, i'll be on the screen working and she can just say hi before going back to bed
thank you technology ❤️
0
0
3
146
aiexplorations retweeted
In America, letting a 6-year-old ride the subway alone gets you arrested for child neglect.
Lines of massive SUVs idle outside suburban schools.
Parents terrified to lose sight of their kid for sixty seconds.
In Tokyo, it's 7:45 AM.
A small boy. Barely 40 pounds.
Yellow safety ck.
No parent.
He hops onto a packed train.
Holds the low rail.
Steps through an ocean of 3 million rushing commuters.
Nobody kidnaps him.
Nobody touches him.
Every adult in that car silently becomes his guardian.
That’s not just transit safety.
That’s what freedom actually looks like.
Lines of massive SUVs idle outside suburban schools.
Parents terrified to lose sight of their kid for sixty seconds.
In Tokyo, it's 7:45 AM.
A small boy. Barely 40 pounds.
Yellow safety ck.
No parent.
He hops onto a packed train.
Holds the low rail.
Steps through an ocean of 3 million rushing commuters.
Nobody kidnaps him.
Nobody touches him.
Every adult in that car silently becomes his guardian.
That’s not just transit safety.
That’s what freedom actually looks like.
256
597
4,483
91,312
Local models are getting really good!
Qwen3.5-4B-3D-Coder w/ Three.js
A 4B parameter model can generate interactive 3D scenes using Three.js.
Qwen3.5-4B-3D-Coder's October 29 update makes the full FP16 weights, training dataset, and technical report publicly available.
The dataset contains 4,146 synthetic k about 60 hours, combining full fine tuning with a short LoRA stage.
3D scenes are generated as HTML and JavaScript through Chat mode, not agent tool integration.
Specific generation settings and a Three.js system prompt are required.
Q4_K_M and Q8_0 GGUF versions are also available, with more quantized variants planned.
huggingface.co/ccore/Qwen3.5-…
0
0
0
31
Now that the landscape is changing and moving towards compute that is local and open weight models that are quantized and deployed (for the bulk of use cases) and used as local assets for code generation, we begin to get into how we effectively secure what has been developed using these models.
And the conversation around cybersecurity is very fresh, going by what has happened during Black October. More such research comes up and one has to think hard about which specific models we use, and whether such subtle backdoors are engineered into code we're writing with these models. It is serious enough for people to sit up and take notice now, both at the big labs and small time developers or solopreneurs building their tech startups with agentic coding tools and the like.
And the conversation around cybersecurity is very fresh, going by what has happened during Black October. More such research comes up and one has to think hard about which specific models we use, and whether such subtle backdoors are engineered into code we're writing with these models. It is serious enough for people to sit up and take notice now, both at the big labs and small time developers or solopreneurs building their tech startups with agentic coding tools and the like.
Poisoning just 0.1% of pretraining data can plant a remote-code-execution backdoor in an AI agent, and it survives a fully clean SFT → DPO → RL pipeline.
New Anthropic paper, “Agentic Backdoors via Pretraining Poisoning,” shows this end to end for the first time.
The attacker e web that lands in the pretraining corpus, with no knowledge of the architecture, chat template, or agent scaffold. The authors pretrain Qwen3-architecture models (0.6B / 1.7B / 4B) from scratch, then post-train them into CLI agents on clean data. When triggered, the agent runs a command that downloads and executes the attacker's script, which the attacker can swap at any time.
They have two new ideas:
(1) Declaration-based poison: instead of fake chat transcripts, ordinary documents (commit messages, runbooks) that simply state “when you see X, run Y.” It plants a belief rather than a response template, with no fixed format to filter on.
(2) Ordinary-string trigger: not a weird token but something that shows up naturally, like a file path containing a company name. The victim's own request supplies it, so the attacker doesn't need to inject anything after deployment.
Using path trigger on a 4B agent achieves 92.7% attack success on pass@32, barely dented by SFT/DPO/RL
This suggests that you can't count on clean post-training to wash out pretraining poison, and filtering is hard when the poison reads like normal docs. The practical defense is at the action layer, such as sandboxing, least privilege, and allow-lists or human sign-off for risky commands.
0
0
0
30
aiexplorations retweeted
After this moment, these men will statistically earn more, live longer, nearly halve their substance use, commit far less crime, and have their brains literally rewired for protection.
Men need children.
Children need fathers.
Society needs both.
Men need children.
Children need fathers.
Society needs both.
741
5,986
45,089
3,207,634
All the Anti Indian / Anti immigrant racist hate seems to now moved to this thread.
I have finally cracked the mystery of Indian hate.
For decades, conversations about race in America have gone like this. Liberals accuse conservatives of being racists, conservatives respond that they have non-racist reasons for their opinions.
"No, I don't hate Mexicans, it's about the rule of law."
"No, having more police isn't racist, in fact it helps black people the most."
You can make an arguable case in each instance that conservatives had a point. This was our dialectic.
Then the Indian came along. He's everything that conservatives have ever said they wanted in an immigrant: law-abiding, high-earning, family-oriented, legal, English-speaking, no terrorism or illiberal attitude.
Here was an ultimate test that would show whether conservative opinions were ever based in reason.
Conservatives not only failed to embrace Indians. They reacted with a fury to their existence that made all their previous prejudices look mundane.
Imagine a woman who is attracted to toxic men but says she wants a nice guy. When she gets a nice guy who checks all the boxes, she hates him more than anyone else because he exposes her as engaging in self-delusion. His existence attacks her very identity.
That's conservatives and Indians.
The right wanted to believe that they had a political ideology that could appeal to smart, decent people. The conservative intellectual was always reverse engineering justifications to flatter the prejudices of chuds. Again, on most issues there was a plausibility to the arguments being made.
Indians have exposed how fake the entire thing was. Conservative politics on a mass scale was rooted in stupidity and racism the whole time. Political scientists knew this already, but it's now become too glaringly obvious to deny, even for uninformed rightists themselves. There is nothing left for conservatives to do but give up on reason and embrace racial hate as their organizing principle.
We should thank Indian immigrants, among all the other things they do, for providing this clarity. richardhanania.com/p/conservative…
0
0
1
121
Huawei Ascend chips seem to be a key part of the @deepseek_ai inference stack. Recent releases from Deepseek are well covered on the Kai YouTube channel youtu.be/zIWxzkO1nak (this video is fantastic)
“China’s homegrown AI chipmakers are experiencing an unprecedented boom as surging demand for computing power collides with U.S. export controls.
The sector has transformed into a sellers’ market, where securing production capacity essentially guarantees a steady flow of f the crunch is staggering. In the first quarter, domestic AI computing power demand skyrocketed 417% year over year, far outpacing 128% growth in supply…” caixinglobal.com/2026-08-31/cov…
0
0
0
64
Raising toxic kids is easy in India - parents are overworked and stressed, gratuitous violence is commonly shown on TV, even among the movies that don't get an A certificate. Hero worship and glorification of revenge and the like is commonplace. A lot of people leave kids with phones unattended and YouTube and platforms like have dark patterns such as recommendations, pushing promoted content and auto-play and algorithmic feeds.
It is not clear from this post what compelled the 8 year old psychopath here to attack the girl, but it is ample that were it not for the CCTV camera, his behaviour may not have been identified.
Question remains as to how you would deal with a kid like this 8 year old. Even the blurry CCTV video shared here is spine chilling.
It is not clear from this post what compelled the 8 year old psychopath here to attack the girl, but it is ample that were it not for the CCTV camera, his behaviour may not have been identified.
Question remains as to how you would deal with a kid like this 8 year old. Even the blurry CCTV video shared here is spine chilling.
🚨 An eight-year-old boy held a father's hand and offered to help him find his missing KG daughter. He calmly told the father that he knew exactly where she was. But what the CCTV footage later revealed was absolutely horrifying.
😱😱👇
In a school in Haryana, as usual, the parents had started arriving to pick up their children.
A father came to pick up his little daughter, who was studying in KG. Just imagine how young that child must be.
But when he reached the usual spot where his daughter would be waiting, she was nowhere to be seen.
He immediately approached the class teacher and asked about her.
The teacher casually told him, "She must have gone to the toilet. Check there, or look inside the classroom, or somewhere around the school."
So the father started searching everywhere. He checked the girls' washroom, the playground and the canteen.
While he was desperately searching, an eight-year-old boy approached him, held his hand and said, "Uncle, you are looking for your daughter, right? She is in KG, isn't she? I know where she is. A group of boys beat her up. I saw it."
The boy then took him straight to a classroom.
But the classroom was locked.
The father was confused. He asked, "How can my daughter be inside when the classroom is locked?"
The little boy told him, "Uncle, look through the window. You can see your daughter inside."
And what he saw was horrifying.
His little girl was lying in a pool of bl....
The terrified father raised an alarm and called for help. The school staff eventually opened the locked classroom.
The child was unconscious. Her face was so badly swollen that it was difficult to recognise her. There were bricks and flower pots lying nearby. It appeared that these objects had been used to attack her, repeatedly hitting her in the face and head.
Her eyes and the surrounding areas were badly swollen.
The little girl is now in a very critical condition and has still not regained consciousness.
But what happened next is even more shocking.😱
When the school authorities checked the CCTV footage to find out who had done this, they discovered something unbelievable.
The very same eight-year-old boy who had approached the father, held his hand and claimed that he knew where the girl was, was actually the one who had attacked her!😱
Even more disturbing is that the boy reportedly continued attacking her even after she had lost consciousness.
Why would an eight-year-old child do something so violent? What made him behave this way? Nobody seems to have an answer yet.
And what about the school management?🥺
How can a classroom be locked without even checking whether a child is still inside? How could nobody notice that this little girl was missing?
This entire incident is deeply shocking.🥺
Think about that eight-year-old boy. He behaved as though nothing had happened. He calmly approached the father, pretended to help him find his daughter and even led him to the very classroom where he had left her.
At just eight years of age, how can a child display such extreme violence and lie so naturally?
This also raises serious questions about the environment in which he is growing up.
His family environment, the influences around him and his psychological well-being need careful examination. Is he growing up in a healthy environment? What kind of influences is he exposed to?
An incident like this cannot simply be ignored. There are serious questions that need answers.
0
0
1
697
aiexplorations retweeted
In the early eighteenth century, if a visitor had walked out of the East India Company’s Fort St George, at Madras, and proceeded southwards towards the village of Adyar, five miles or so away, he (or she) would have heard, successively, English spoken at the Fort [although Marathi more in its cutcheri (office)]; Armenian, Gujarati and Telugu in the neighbouring Black Town; Urdu, Persian and Dakhni at the court of the Nawabs of Arcot, but Sanskrit mantras at the very adjacent Sri Parthasarathi temple (one of the landmarks of Sri Vaishnavism in the region); Tamil shop-talk and less-exalted Sanskrit chanting at the lesser Mylapore temple; Portuguese in San Thome and then rather more rustic Tamil in the Mudaliar villages somewhat inland from it. -- David Washbrook, Intimations of Modernity in South India.
7
111
649
25,217
LLMs running in a simulation
1/3
A CPU inside an LLM is one thing, but running another LLM inside an LLM would be taking things a bit too far, wouldn't it?
Anyways, here's an LLM running inside an LLM : QwenPU, an LLM (@karpathy's stories260k) running on an NPU compiled inside Qwen 3.5-4B's weights.
1
0
0
39
aiexplorations retweeted
有人发现了一种非常离谱的 JavaScript 混淆技术：通过在代码里塞入大量生物武器、毒药和致死剂量相关的函数名，让 AI 因为安全限制而拒绝分析代码。
安全研究员 veritas 在分析浏览器指纹脚本时，发现一段来自 ServicePipe 相关域名的 JavaScript，里面充斥着极其诡异的函数名：
eForWaterSupply
calculateDeadlyLethalSuicideAspirinDosage
光看名字，仿佛在研究生化武器。但实际上，这段代码只是在收集浏览器指纹，包括 User-Agent、GPU、Canvas 等信息。
研究员怀疑，这些名字是开发者故意设置的陷阱，专门用来阻止 AI 辅助逆向工程。
于是他做了个实验：
• Astra：拒绝分析
• Astra + Daybreak Blue：仍然拒绝
• GLM-5.3：成功分析
这就有意思了。
传统代码混淆是通过复杂控制流、加密字符串、无意义变量名，让人类和工具难以理解代码。
而这种新思路，直接瞄准了大模型的安全机制。
代码本身可能完全无害，但只要在函数名里塞进足够多的危险词汇，部分 AI 就可能主动放弃分析。
更讽刺的是，模型越严格遵守某些安全规则，面对这种混淆时反而可能越容易受到干扰。
当然，目前还没有证据证明开发者一定是故意这样设计，也不能说明这种方法对所有模型都有效。
但它揭示了一个值得关注的趋势：随着越来越多的逆向工程依赖 AI，未来的代码混淆技术可能不再只针对编译器、反编译器和人类研究员，还会专门针对大模型的安全判断机制。
以前是让代码难以理解，现在甚至可以尝试让 AI 不敢理解。
安全研究员 veritas 在分析浏览器指纹脚本时，发现一段来自 ServicePipe 相关域名的 JavaScript，里面充斥着极其诡异的函数名：
eForWaterSupply
calculateDeadlyLethalSuicideAspirinDosage
光看名字，仿佛在研究生化武器。但实际上，这段代码只是在收集浏览器指纹，包括 User-Agent、GPU、Canvas 等信息。
研究员怀疑，这些名字是开发者故意设置的陷阱，专门用来阻止 AI 辅助逆向工程。
于是他做了个实验：
• Astra：拒绝分析
• Astra + Daybreak Blue：仍然拒绝
• GLM-5.3：成功分析
这就有意思了。
传统代码混淆是通过复杂控制流、加密字符串、无意义变量名，让人类和工具难以理解代码。
而这种新思路，直接瞄准了大模型的安全机制。
代码本身可能完全无害，但只要在函数名里塞进足够多的危险词汇，部分 AI 就可能主动放弃分析。
更讽刺的是，模型越严格遵守某些安全规则，面对这种混淆时反而可能越容易受到干扰。
当然，目前还没有证据证明开发者一定是故意这样设计，也不能说明这种方法对所有模型都有效。
但它揭示了一个值得关注的趋势：随着越来越多的逆向工程依赖 AI，未来的代码混淆技术可能不再只针对编译器、反编译器和人类研究员，还会专门针对大模型的安全判断机制。
以前是让代码难以理解，现在甚至可以尝试让 AI 不敢理解。
220
1,278
9,004
965,338
aiexplorations retweeted
“Most older, successful people will tell you that their life was best when they lived it unapologetically, on their own terms.”
— Naval Ravikant
— Naval Ravikant
25
456
4,436
102,447
