The LLM could be trained to embed obfuscated malware in valid code outputs, only occasionally, or for specific types of requests. Then the user runs the codeâ¦. This gets around even users running local LLMs, unless the user reads all the code output before it is run, and can recognise the obfuscated code.

Donaldson v. Beckett (1774) [UK] or Wheaton v. Peters (1834) [USA]

The fundamental "copyright bargain"â"granting temporary monopolies to creators in exchange for works eventually entering the public domain to benefit societyâ"is deeply rooted in two foundational historical texts:

The Statute of Anne (1710): Enacted by the Parliament of Great Britain, this was the worldâ(TM)s first copyright law. Its explicit goal was "the Encouragement of Learning" by vesting rights in authors for a set period (14 years, with a possible 14-year renewal). It established the precedent that after this limited term, the work's monopoly ends and the text "falls into the public domain" so others can access it.

The U.S. Constitution (1787): Article I, Section 8, Clause 8 (known as the Copyright or Progress Clause) empowers the government "to promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries." This directly enshrined the philosophical bargain that temporary, exclusive protections are merely tools to encourage creation and scientific progress for the overall good of the public.

Muscosal immunity (secreted anitbodies) peak within three months after infection/ vaccination.

âoeAccording to research and medical reviews on âVaccines for the common cold, vaccine development is largely hindered by both rapid mutation and the biology of mucosal immunity, compounded by the sheer multitude of causative viruses. Furthermore, because the illness is generally mild and self-limiting, the immense public health and economic resources required to develop an effective multi-strain vaccine have historically lacked urgency.â

Interestingly, the benchmark questions are not available to the public for âoesafety and licensingâ reasons. I donâ(TM)t t see how benchmarking questions about answering science questions could be a public risk unless the questions are asking how to build atomic weapons or biological warfare agents, and the scoring matrix contains useful /secret information â¦

A severe reaction to a tick bite is probably not his powerful âvenomâ(TM), but rather an allergic reaction to the tick regurgitating part of his last meal, probably a mammal. This strong reaction to a tick bite is a sign that you are at risk of mammalian meat allergy (MMA) or alpha gal syndrome. This can lead to life threatening anaphylaxis. https://www.news.com.au/lifest...?

Yes, the idea put forward by the Supreme Court in Florida vs Riley is that it must be visible from the street with the naked eye or from a plane above 400 feet. However, the recent ruling in the Michigan Supreme Court pointed out that the US Supreme Court has ruled multiple times that the exclusionary rule applies only in criminal cases, not civil cases. https://law.justia.com/cases/m...

Yes, the idea put forward by the Supreme Court in Florida vs Riley is that it must be visible from the street with the naked eye or from a plane above 400 feet. However, the recent ruling in the Michigan Supreme Court pointed out that the US Supreme Court has ruled multiple times that the exclusionary rule applies only in criminal cases, not civil cases. https://law.justia.com/cases/m...

If you're using Office 365 Personal, one user reported a method to opt out:
"If you're using O365 Personal it looks like you need to 'cancel' your subscription - when you hit cancel you have the option to change to 'classic (without AI)' - at your next renewal you will lose all AI features for Office and save about $50"8.

Scroll down to the bottom of this link:
https://bgr.com/tech/microsoft...
or
the bottom of this link:
https://www.reddit.com/r/Offic...

The price increase is optional, but hidden by a dark pattern:
https://www.consumer.org.nz/ar...

Yes, good points. I had not thought this through yet. I will get a Raspberry Pi AI camera when they are generally available.

One weak point of the IMX500: it only has 8 Mb of memory, limiting what models you can run.
https://developer.aitrios.sony...

Sipeed had modules like this two or three years ago. They have moved on to make more useful cameras like the MaixII and MaixCam. The MaixCam has 256 Mb ram and costs ~$70 with screen and camera, and ~$40 with only camera. The 500 mA (max) power is not too bad, similar to running the IMX on a Pi Zero. The Maix modules can run PyTorch and Tensorflow models, have a model zoo, and an IDE for running and testing models on the board.

If you are prototyping as a professional enginneer, then Aitrios and the Sony sensor may be useful.
https://developer.aitrios.sony...

If you are just a hobbyist running modes from the zoo, the Maix camera with the builtin screen for troubleshooting (or the ability to monitor the image from the IDE when troubleshooting) may be more useful.

Or the Jetson Nano, at 40 TOPS:
https://developer.nvidia.com/e...

The SONY IMX500 is about 1 TOPS. The existing RPi kit with Hailo-8L NPU found in the Raspberry Pi AI Kit is 13 TOPS.
https://www.raspberrypi.com/pr...
Price? $70 versus $120 plus RPI5 cost?
Power consumption?
This is a competitor to the Sipeed MaixCam: https://wiki.sipeed.com/hardwa...
Sipeeed have been making this stuff for a while... nice little modules, but the English documentation is a little bit thin sometimes....

I think we need to conceptually seperate "synthetic data" from the "model collapse" mentioned above.

Synthetic data does not add more information: it helps solve the problem of overfitting the model to the training data, so that it performs poorly on a separate test data set.

With image classification, it is clear from research that "data augmentation" by zooming, adding noise, streching, rotating, shearing and changing the color space can increase the effective size of your annotated image dataset and thus reduce overfitting and help generalize the data:
https://innotescus.medium.com/...

Another technique is to take your 100 cat imges and paste them randomly over say 1000 random images, genereating tens of thousands of different images.

Synthetic data has several advantages besides reducing overfitting: even up class distributions, reducing bias, and potentially, improving privacy.
https://syntheticus.ai/blog/th...

The big problem is when you use synthetic data generated by a generative model to train the same generative model (rather than a classification model).
https://www.techtarget.com/wha...

You end up with a death spiral of "model collapse" that cannot be recovered from.