Even if you find a way to 100% watermark any text, couldn't you just use a non-watermark model? I have a hard time believing every AI company on heart would comply.

Hell, even if every AI company on earth decide to somehow apply watermarking to their next model, they would also need to apply it to all the previous version that they commercialize. Given that anybody could make a copy of an open source model right now and would be safe forever, this seems quite the lost cause to me.

By distilling, you are training a model via Reinforcement Learning (RL) to mimic the answer of a bigger model. To do that, you need all the steps that contributed to generating an answer.

To give you a better idea, imagine teaching a student how to solve math problem:

1. You give it the problem and the answer only (no thinking trace) 2. You give it the problem, the intermediary steps and the answer (full transcript)

I think you can agree that the second method is more likely to give a well-informed student.

The chapters on bats and eels are absolutely crazy.

https://www.tandfonline.com/doi/full/10.1080/03075079.2023.2...

"Our findings reveal that gradeless learning can increase motivation and reduce surface learning but has no significant effect on students’ later academic achievements."

I'm building agent to control machines at work and have to rely on small models, which are especially bad at remembering rules like this, so it is very apparent to me that soft rules like this are pretty much useless. Might be different for these huge models, but still it seems very shaky to me.

With these types of algorithms, massive amount of data usually perform better than any underlying model you could have.

I really don't understand that concern... So you have either a STT-LLM-TTS or fully native speech model that is instructed to answer calls. When does it generate an analysis? Are you giving the agent some tools to generate the analysis ? Maybe don't.

Also, in general I feel like training some classifier would be slightly better from a humanitary point of view, but I agree that in general reducing the work load of overworked public workers should be a priority of AI application.

A. If it's unrelated, it doesn't add anything to the post except aesthetics. B. On the aesthetics dimension: I don't think people are very impressed by the AI aesthetic, and tend to be annoyed by it. C. It seems that most people now assume that AI-generated image == AI-generated text.

I personally do not care about AI images or not, but it seems that a portion of people do.

A. You answer it from memory. B. Have to actually solve it, which from my experience requires a bit of time.

It just seems awkward to have somebody look at you while they wait for you to solve some riddle.

Glad to be able to put some numbers on it.

Is the goal of the "dopamine fracking" metaphor is to represent a "pure concentrated dopamine hit", I would argue it is a more apt representation of what hard stimulants do to your brain (i.e. a shortcut to massive amount of dopamine release).

The perverse effect of social media addiction is more akin to a slow I.V. drip where you are always in a slightly aroused state, which then becomes your baseline and any cut from it is a net negative in your life.

This is impressive work!

I come from sciences, where the problem of AI is more linked to AI-generated papers rather than coding. Again, the real underlying problem is the pressure to publish at all cost and the ranking grant system which rewards number of publication rather than quality.

I'm sure you could say the same of AI-generated images, sound or video used in advetising, or any field where AI generation offers a shortcut. The problem is always that you need to produce as much results as fast as possible and quality is not the objective. The problem is always the KPI.

However, with RLHF you don't need to have access to the original training set to steer a model towards a preferable behavior.

Sure, you can download the weights and run it yourself, but to run a 1T+ parameters you need roughly 2TB of VRAM (for full precision). That costs around $500k to $1M, not including electricity and rent.

In reality, they provide their models priced via API or subscription just like any other company. The "free weights" are only useful for other research labs. It seems that China understands that research should be shared and open, which is one of the reason that AI has been able to develop that fast in the last 10 years.