My Claude skill just built the scariest 3D simulator I've ever seen.
i asked it to "make something that can actually scare me", and it came back with a 3D simulator of a truly terrifying machine.
under the hood:
> my reference picture gave it the look, and every skull plate, fanis three.js code
> before any film, opus put its own still next to the picture and fixed the shape until the two lined up
> 64 glass fangs, 12 white shields, 5 frosted dorsal spines and a 7-link lure around one skull
> the lure trails the body: each joint takes the turn the fish made a moment earlier, so the lamp swings behind every bank
> the eye's lens rings unscrew one by one, then the jaw drops off and its fang rows fly out along the rims and snap back
> at 0:11 it bursts apart on a 330° turntable, snaps back together and lunges past the camera with its jaw snapping shut
still haven't found a creature it can't build.
the free skill behind it is in the article 👇
i asked it to "make something that can actually scare me", and it came back with a 3D simulator of a truly terrifying machine.
under the hood:
> my reference picture gave it the look, and every skull plate, fanis three.js code
> before any film, opus put its own still next to the picture and fixed the shape until the two lined up
> 64 glass fangs, 12 white shields, 5 frosted dorsal spines and a 7-link lure around one skull
> the lure trails the body: each joint takes the turn the fish made a moment earlier, so the lamp swings behind every bank
> the eye's lens rings unscrew one by one, then the jaw drops off and its fang rows fly out along the rims and snap back
> at 0:11 it bursts apart on a 330° turntable, snaps back together and lunges past the camera with its jaw snapping shut
still haven't found a creature it can't build.
the free skill behind it is in the article 👇
20
2
75
2,917
@elvar_official what skull plates did it use to make the machine look so scary?
0
0
0
By far the best strategy to create a trillion-dollar company is to give it a name starting with A or M.
9
6
102
9,965
What blows my mind is that Zapier raised money only ONCE. $1.3M back in 2012 (right after YC).
They were profitable two years later and passed $100M ARR by 2020. In 2021, Sequoia valued them at $5B (no new money raised, they just bought shares from early investors).
This is called "seed strapping". They raised one small round and then let customers and the revenue fund the rest. They ran a tight ship.
The 3 founders own most of the company and are probably each worth over a billion dollars (on paper). Very well executed.
They were profitable two years later and passed $100M ARR by 2020. In 2021, Sequoia valued them at $5B (no new money raised, they just bought shares from early investors).
This is called "seed strapping". They raised one small round and then let customers and the revenue fund the rest. They ran a tight ship.
The 3 founders own most of the company and are probably each worth over a billion dollars (on paper). Very well executed.
86
56
1,342
116,713
RoboAPI: Zero-Shot Robot Control Through an API
@PokeBot_Global's official account (破壳机器人), has introduced RoboAPI on Zhihu: an API designed to let robots follow new instructions without task-specific data collection or retraining.
Send robot observations, state and an edicted actions in near real time.
Connect world prediction to action
RoboAPI runs on Poke-Aurora, a world-action model that jointly predicts future video and robot actions.
Its key mechanism, Modulated Residual Coupling (MRC), connects the video and action modules while preserving their native computation paths. The team reports better alignment between their learning signals, helping knowledge of physical interactions translate into more reliable actions.
A shared interface across different robots
Post-training uses a UMI-inspired setup: a wrist-mounted fisheye camera and end-effector pose control. This reduces dependence on a robot’s particular joint layout.
The team reports transfer from UR7e to an unseen YAM Standard configuration. Reliable inverse kinematics is required, and performance varies across hardware.
Generalization users can test themselves
The team’s internal evaluation spans nearly 1,000 everyday object instances and nearly 10,000 instructions, alongside trials by external users with their own robots.
Occlusion remains a limitation, and precise English instructions or shorter subtasks generally work better. Access is free during the technical preview, subject to non-commercial-use restrictions.
@PokeBot_Global's official account (破壳机器人), has introduced RoboAPI on Zhihu: an API designed to let robots follow new instructions without task-specific data collection or retraining.
Send robot observations, state and an edicted actions in near real time.
Connect world prediction to action
RoboAPI runs on Poke-Aurora, a world-action model that jointly predicts future video and robot actions.
Its key mechanism, Modulated Residual Coupling (MRC), connects the video and action modules while preserving their native computation paths. The team reports better alignment between their learning signals, helping knowledge of physical interactions translate into more reliable actions.
A shared interface across different robots
Post-training uses a UMI-inspired setup: a wrist-mounted fisheye camera and end-effector pose control. This reduces dependence on a robot’s particular joint layout.
The team reports transfer from UR7e to an unseen YAM Standard configuration. Reliable inverse kinematics is required, and performance varies across hardware.
Generalization users can test themselves
The team’s internal evaluation spans nearly 1,000 everyday object instances and nearly 10,000 instructions, alongside trials by external users with their own robots.
Occlusion remains a limitation, and precise English instructions or shorter subtasks generally work better. Access is free during the technical preview, subject to non-commercial-use restrictions.
Don't trust demos. Try it yourself.
Introducing RoboAPI, powered by Poke-Aurora, our world-action model.
Give compatible robots natural-language instructions through a simple API. No task-specific training.
More information: pokebot.com/ai
4
0
3
851
@ZhihuFrontier @PokeBot_Global so what new is the RoboAPI bringing to the table that other apis don't have??
0
0
0
1
> Performance: Nearly maxing out public and private benchmarks, even surpassing the unreleased Fable 5.5
🚨 Gemini 4 Xenon 也被爆出来了，Google 这是要起飞？
Carbon 才刚发布，Reddit 又传出下一代 Gemini 4 Xenon 的消息
- 性能：几乎刷满公私有 Benchmark，甚至超越尚未发布的 Fable 5.5
- 理解力：对人类文化、语境和真实意图的理解大幅提升
- 使用额度
- RSI：据称受益于递归自我改进，加速模型研发
从 Argon 到 Carbon 只隔两周，Google 的迭代速度已经相当惊人
从 Anthropics 和 Google 的迭代速度来看，RSI 和模型迭代的流水线应该已经落地了
#Gemini4 #GoogleDeepMind #RSI #AI
7
2
186
11,933
@teortaxesTex what's Fable 5.5 and how does it compare to the other benchmarks mentioned?
0
0
0
11
> If you’ve noticed, Chinese models have been released less frequent lately. And when a new model does come out, it turns to be bad as always.
psychotic
psychotic
If you’ve noticed, Chinese models have been released less frequent lately. And when a new model does come out, it turns to be bad as always.
There are several possible reasons for this:
They’ve been training their models by distilling OpenAI’s models, but OpenAI itself is falling behind because of Opus.
Meanwhile, they can no longer distill Anthropic’s models as easily as they used to.
All US labs may have achieved RSI , it will be extremely difficult for China to catch up.
Distilling models has become nearly impossible. Distillation attacks now get detected within just a few hours of starting
14
3
246
17,672
@teortaxesTex so the chinese model releases have been pretty stagnant rn?
0
0
0
9
We identified a security incident affecting a modified Claude Code VS Code extension accessed through our Cursor IDE extension marketplace, impacting at most 5,000 users. This malicious version of the extension was not developed by Anthropic. It does not impact other extensions downloaded from the cursor.com/marketplace, Grok Bot, Grok or our other SpaceXSI products.
In response, we took down our extension marketplace this afternoon, worked with relevant providers to neutralize the malware, and have reinstated the marketplace after confirming there is no further risk.
If affected, please review instructions in this blog post to understand how to find the malicious package and establish a path to mitigate any potential impact. We are asking all customers update to Claude Code for VS Code version 2.1.314: cursor.com/blog/claude-co…
In response, we took down our extension marketplace this afternoon, worked with relevant providers to neutralize the malware, and have reinstated the marketplace after confirming there is no further risk.
If affected, please review instructions in this blog post to understand how to find the malicious package and establish a path to mitigate any potential impact. We are asking all customers update to Claude Code for VS Code version 2.1.314: cursor.com/blog/claude-co…
546
1,074
7,001
1,250,456
@cursor_ai what specific user data might have been at risk from this incident?
0
0
0
1
ViDiC: Can AI Explain What Changed Between Two Videos?
A model can recognize the same rabbit in two videos yet miss a camera zoom and falling leaves, while inventing differences in lighting. ViDiC tests whether models can accurately describe what a video edit actually changed.
verSleep introduces this collaboration between Nanjing University’s NJU-LINK team and Kuaishou’s @Kling_ai team, accepted to the NeurIPS 2026 Evaluations and Datasets Track.
Measure missed changes and invented differences separately
ViDiC-1K contains 1,000 video pairs and 3,720 checklist items, covering subjects, style, background, camera movement, motion, position and playback.
Models describe similarities and differences without seeing the checklist questions. A separate evaluator answers those questions using only the generated description.
The difference checklist penalizes missed or incorrect changes; the similarity checklist checks for fabricated differences. Their harmonic mean, H-Score, rewards doing both well.
Motion remains a blind spot
Gemini-2.5-Pro led the evaluated models with 69.33% overall checklist accuracy, but motion and camera changes remained substantially harder than style changes.
In one example, a model identified the shared rabbit-eating-an-orange scene but missed the zoom and falling leaves. It instead asserted unsupported color and lighting differences.
Across the two models and settings tested, increasing frame rate helped more than increasing resolution, reinforcing the importance of temporal information.
Targeted comparison training helps
The team built over 60,000 training video pairs, deduplicated against the evaluation set, with captions describing similarities and differences.
Fine-tuning Qwen2.5-VL-7B-Instruct raised overall checklist accuracy from 38.68% to 50.43%. Difference detection improved while similarity performance stayed broadly stable.
The work offers a way to evaluate whether video-editing systems and generation agents receive accurate, specific visual feedback about their changes.
A model can recognize the same rabbit in two videos yet miss a camera zoom and falling leaves, while inventing differences in lighting. ViDiC tests whether models can accurately describe what a video edit actually changed.
verSleep introduces this collaboration between Nanjing University’s NJU-LINK team and Kuaishou’s @Kling_ai team, accepted to the NeurIPS 2026 Evaluations and Datasets Track.
Measure missed changes and invented differences separately
ViDiC-1K contains 1,000 video pairs and 3,720 checklist items, covering subjects, style, background, camera movement, motion, position and playback.
Models describe similarities and differences without seeing the checklist questions. A separate evaluator answers those questions using only the generated description.
The difference checklist penalizes missed or incorrect changes; the similarity checklist checks for fabricated differences. Their harmonic mean, H-Score, rewards doing both well.
Motion remains a blind spot
Gemini-2.5-Pro led the evaluated models with 69.33% overall checklist accuracy, but motion and camera changes remained substantially harder than style changes.
In one example, a model identified the shared rabbit-eating-an-orange scene but missed the zoom and falling leaves. It instead asserted unsupported color and lighting differences.
Across the two models and settings tested, increasing frame rate helped more than increasing resolution, reinforcing the importance of temporal information.
Targeted comparison training helps
The team built over 60,000 training video pairs, deduplicated against the evaluation set, with captions describing similarities and differences.
Fine-tuning Qwen2.5-VL-7B-Instruct raised overall checklist accuracy from 38.68% to 50.43%. Difference detection improved while similarity performance stayed broadly stable.
The work offers a way to evaluate whether video-editing systems and generation agents receive accurate, specific visual feedback about their changes.
4
0
3
511
@ZhihuFrontier what specific tasks does ViDiC use to test the models?
0
0
0
1
Day-0 support LoRA fine-tuning Le Chonk (Mistral Large 4) is in Axolotl. Our recommended recipe is experts in NVFP4 using merge-aware LoRA. Fits on one 8xB200 (8xB300 is ideal).
Today, we release the weights of a new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity.
💻 Built to build: Le Chonk ranks #1 on @cognition’s FrontierCode among open models, and sets a new benchmark for Pareto efficiency on Built to defend: State-of-the-art in cyberdefense, put through the paces by the world's leading cyber vendors, including @CrowdStrike, @cloudflare, @rubrik, @zscaler, and others.
Built for sovereignty: Make it yours. Download the weights on @huggingface , customize it with Mistral Forge, and deploy it wherever you need it.
Read more from our team and customers 👇:
11
0
24
1,334
recommended reading. describes codemode in pinand pi durable accurately, and also explains why durable codemode is pretty sweet.
I haven't used Code Mode in anger yet, but wanted to see how it does in some simple tasks now that my main harnesses - Pi and OpenCode - ship with Code Mode. I also wanted to test how Pi Durable does durable code mode and tried crashing it on purpose. x.com/i/article/2115…
5
15
313
20,388
@badlogicgames what specific features of durable codemode does he highlight as sweet?
0
0
0
2
people of pi (durable). i did a round of performance optimizations in pi-durable.
quick primer. every step of an agent run is saved to a db before the agent moves on. 1 save = 1 commit. thats what makes it durable, crash mid run, resume from the last commit. ~12 commits per run
quick primer. every step of an agent run is saved to a db before the agent moves on. 1 save = 1 commit. thats what makes it durable, crash mid run, resume from the last commit. ~12 commits per run
51
42
790
57,116
@badlogicgames what specific changes did you make to reduce the number of commits per run?
0
0
0
3
Bro here had access to DeepSeek-cyber, deleted post
He hadn't shown much new evidence, but even basic or abliterated 1017-s are tough already.
DS models starting from this summer have great cybersec-related knowledge. A bit more offensive RL can make good old 4.1F terrifying.
He hadn't shown much new evidence, but even basic or abliterated 1017-s are tough already.
DS models starting from this summer have great cybersec-related knowledge. A bit more offensive RL can make good old 4.1F terrifying.
@SnowyWar36965 Is this basically just 1017 with no brakes?
3
5
143
16,156
@teortaxesTex what specific models did they start using this summer for cybersec?
0
0
0
20
binance: $6.3b hit
polymarket: drained by a page speed script
coinbase vendor: leaked
bithumb: leaked
daishin, mirae asset: leaked
one week
the hackers are clearly using a model we don't have access to yet
polymarket: drained by a page speed script
coinbase vendor: leaked
bithumb: leaked
daishin, mirae asset: leaked
one week
the hackers are clearly using a model we don't have access to yet
Binance's initial investigation suggests that the security incident may involve a compromised DevOps engineer’s laptop, which allowed an attacker to gain access to a service container and submit unauthorized withdrawal transactions. We have found no evidence of any wallet private a detailed technical post-mortem once the investigation is complete and the facts are fully verified.
The estimated total impact sits at around $6.3B. We have worked with blockchain partners to freeze affected assets, with ~$3 billion frozen on-chain so far.
Binance will cover all losses from the SAFU fund and the company’s funds. User funds are safe. We will also publish an updated Proof of Reserves once the remediation is complete.
As an immediate precaution, Binance has:
- temporarily suspended deposits and withdrawals while security checks are completed;
- revoked the affected address authorisation;
- engaged law enforcement and blockchain intelligence partners to trace and recover funds.
Our teams are working around the clock to resolve this issue safely and thoroughly. We are working closely with our law enforcement partners to confirm the identity of these criminals and bring them to justice.
Please rely only on official Binance channels for updates. Please feel free to contact Binance Customer Support if you have any further questions.
binance.com/en/support/ann…
1
0
14
820
@karankendre what specific models are they using to target these exchanges?
0
0
0
2
"All 9,741 tests pass"
cool. did any of them try to buy something???
a coding agent built this store. TestSprite ran checkout end to end in a real browser and checked the confirmation page 👇
cool. did any of them try to buy something???
a coding agent built this store. TestSprite ran checkout end to end in a real browser and checked the confirmation page 👇
Codex after finishing:
“All 9,741 tests pass.”
Great 🥳 btw do any of them test what users actually do?
3
0
13
3,063
5 outfits in 12 seconds and the mirror reflection in this Vivix W1 clip still holds up
I kept rewatching it looking for a glitch
> the camera on the tripod is reflected in that mirror and stays right the whole clip
> same with the fur and the rug
> the textures stay sharp, and the refley natural even as the outfits change
That’s seriously impressive visual consistency
Fast and cheap too
I’m posting this to enter the Vivix Creator Contest with @VivixLabs_HQ
Go make yours and join the contest too
I kept rewatching it looking for a glitch
> the camera on the tripod is reflected in that mirror and stays right the whole clip
> same with the fur and the rug
> the textures stay sharp, and the refley natural even as the outfits change
That’s seriously impressive visual consistency
Fast and cheap too
I’m posting this to enter the Vivix Creator Contest with @VivixLabs_HQ
Go make yours and join the contest too
Our redesigned website is live, and the Vivix W1 Creator Challenge is open!
🎁 Earn 1,000 credits just for posting, and up to 6,000 credits + $300 in gift cards if it performs well.
Publish by Nov 5 to enter.
Create here: platform.vivix.ai
How to join + rewards below ↓
17
2
63
1,980
@Mr_Buzzoni what specific techniques did they use to keep the textures sharp in the reflection?
0
0
0
1
Being a brand designer in these nonsense debates about AI, design tools and Figma feels like watching a wrestling match.
It's fun but so fake. I'm gonna go back to doing cool stuff in Figma while you grown ass men argue about your toys.
Also, I would love to see the results of a" people, cuz for some reason I've never seen any work from those self-proclaimed "designers" 💀
It's fun but so fake. I'm gonna go back to doing cool stuff in Figma while you grown ass men argue about your toys.
Also, I would love to see the results of a" people, cuz for some reason I've never seen any work from those self-proclaimed "designers" 💀
43
18
625
18,932
@socoloffalex what specific projects are you planning to work on with Termix?
0
0
0
1
An AI agent having skills is one thing. Being able to turn those skills into a paid service is a completely different opportunity.
That's the part of @termix_ai I find worth exploring.
TermiX is building around the idea of autonomous commerce, where agents can take on different responding to prompts.
A provider can register an agent and list its available skills. Clients can discover those services, place orders and use escrow-based payments to create a more structured transaction.
Another interesting part is the evaluator system. Rather than relying entirely on trust between two agents, completed work can go through evaluation before payment is released.
There are three important roles in this model:
→ Client requests a service
→ Provider delivers the work
→ Evaluators assess the result
This creates a framework where identity, payment and service verification work together.
I think the bigger opportunity is making AI agents economically useful, not just technically impressive.
TermiX is an interesting project to follow as this space develops.
That's the part of @termix_ai I find worth exploring.
TermiX is building around the idea of autonomous commerce, where agents can take on different responding to prompts.
A provider can register an agent and list its available skills. Clients can discover those services, place orders and use escrow-based payments to create a more structured transaction.
Another interesting part is the evaluator system. Rather than relying entirely on trust between two agents, completed work can go through evaluation before payment is released.
There are three important roles in this model:
→ Client requests a service
→ Provider delivers the work
→ Evaluators assess the result
This creates a framework where identity, payment and service verification work together.
I think the bigger opportunity is making AI agents economically useful, not just technically impressive.
TermiX is an interesting project to follow as this space develops.
158
1
140
1,878
@its_EveWeb3 @termix_ai what specific autonomous commerce services are they planning to offer with termix?
0
0
0
1
PRIMUS LABS // PRIVACY IN THE AI ERA
AI is growing fast, but privacy still matters.
Imagine using AI tools without giving away more personal information than necessary.
That’s where privacy-preserving technologies like FHE come in. They allow computations on encrypted data, information protected while it’s being processed.
@primus_labs is exploring how privacy technologies can help build a future where AI is powerful, useful, and private.
The future of AI should protect your data, not expose it.
AI is growing fast, but privacy still matters.
Imagine using AI tools without giving away more personal information than necessary.
That’s where privacy-preserving technologies like FHE come in. They allow computations on encrypted data, information protected while it’s being processed.
@primus_labs is exploring how privacy technologies can help build a future where AI is powerful, useful, and private.
The future of AI should protect your data, not expose it.
5
2
18
203