Uh, this is Hacker News. I thought people here didn’t give up so easily, even when they were "told" to. :)
1,651 karma · joined August 31, 2010
For me Google shows the .net site first the github one as second.
Asking chatgpt 5.2 (Auto mode) to search for the nanoclaw site, it says the same, first links the .net site and shows the github as an optional page. When I try to give it a hint by asking "are you sure?" it still even hallucinates that it's linked from the github:
"Yes — nanoclaw.net is the official documentation/site for the NanoClaw project, in the sense that it’s the project’s published homepage and is directly linked from its canonical open-source repository. It describes the project, features, installation steps, and links to the source code on GitHub, which is the authoritative source for the project’s codebase."
Chatgpt 5.2 (Thinking mode) and Claude gets it right the first try, they asnwer with the official .dev page first and claude shows the .net second as "another site covering the project".
I remember some papers about earlier models having around 15% prompt variability, and with different tool use sometimes there are even more significant jumps. And if I remember correctly the reasoning models improve some of these because lot of the early prompting tricks is included in them like "thinking step-by-step", "think carefully" and some other "magic" methods. Also another trick is to ask the models to rephrase the prompt with their own words because that may produce prompt that better align with their training prompts. For sure the big model developers are aware of these and constantly improving it, I just don't see too much discussion or numbers about it.
And based on that I could imagine with a combination of a camera and this method, you could train the model on data where both the camera and this method is seeing the individual and then continue to track them with the wifi sensing + the trained model even where the camera cannot see them anymore.
But yea real world is noisy, so it could be very challenging.
This approach relies solely on the "unencrypted parts of legitimate traffic". The attacker does not need to send any packets or "generate" their own traffic; they simply "listen" to the natural communication between an access point and its clients.
BFI is much more complex than simple signal strength. RSSI is an aggregation of information that the researchers describe as "not robust" for fine-grained tasks In contrast, BFI is a high-resolution, compressed representation of signal characteristics. This rich data allows the system to distinguish between 197 different individuals with 99.5% accuracy, something impossible with basic RSSI.
While older CSI methods often focused on walking directly between a specific transmitter and receiver (Line-of-Sight), BFI allows a single malicious node to capture "every perspective" between the router and all its legitimate clients.
Also CSI requires specialized hardware and custom firmware, this one isn't, just wifi module in monitor mode.
Another way he could benefit from this is when people want his skills to build them similar things, so it's basically already an advertisement for his skills.
"British reversible computing startup Vaire has demonstrated an adiabatic reversible computing system with net energy recovery"
https://www.eetimes.com/vaire-demos-energy-recovery-with-rev...
Short introduction video to reversible computing:
And at that point it will be a fight mostly between AI lawyers :-)
From that angle our artificial models seem very sample efficient, but it's all hard to quantify it without know what was "tried" by the universe to reach the current state. But it's all weird to think about because there is no intent in natures optimizations it's just happens because it can and there is enough energy and parallel randomness to eventually happen.
And the real mystery is not how evolution achieved this but that the laws of chemistry/universe allow self-replicating structures to appear at all. In an universe with different rules it couldn't happen even with infinite trial and error compute.
"Muon is an optimizer for the hidden weights of a neural network. Other parameters, such as embeddings, classifier heads, and hidden gains/biases should be optimized using standard AdamW."
And I just looked into this nanochat repo and it's also how it's used here.
https://github.com/karpathy/nanochat/blob/dd6ff9a1cc23b38ce6...
"that's because a next token predictor can't "forget" context. That's just not how it works."
An LSTM is also a next token predictor and literally have a forget gate, and there are many other context compressing models too which remember only the what it thinks is important and forgets the less important, like for example: state-space models or RWKV that work well as LLMs too. But even just a the basic GPT model forgets old context since it's gets truncated if it cannot fit, but that's not really the learned smart forgetting the other models do.
Yea, I know that was the case when I clicked on the thumbnails and couldn't close the image and had to reload the whole page. Good thing that you could just ask AI to fix this, but the bad thing is that you assumed it would produce fully working code in one shot and didn't test it properly.
https://youtu.be/lI1LCfTx2lI?t=525
There is also Kits.ai https://www.kits.ai/tools/ai-instruments
Videolabs was born from the VideoLAN community and started by maintaining the VLC ports on mobile. It is now the main contributor to VLC, hiring its historical developers, and building custom solutions around the VLC and FFmpeg ecosystems.
What does that mean? Was it too complex or costly to make it back then? Did that change now with your developments?
Better tech is nice but if it's to expensive or hard to mass produce then it could end up the same way as with Bose.