The learning rate in modern optimisers is so large that the piecewise-linear loss landscape really looks indistinguishable from a smooth function. The lr you'd need to use to ensure that the next step lands in the same linear patch is ridiculously small, so in practice the true "felt" landscape is something like a smoothed average of the exact landscape.

More insightful than what is conserved under the scaling symmetry of ReLU networks is what is not conserved: the gradient. Scaling $w_1$ by $\alpha$ scales $\partial E/\partial w_1$ by $1/\alpha$ and $\partial E/\partial w_2$ by $\alpha$, which means that we can obtain arbitrarily large gradient norms by simply choosing small enough $\alpha$. And in general bad initializations can induce large imbalances in how quickly the parameters on either side of the neuron learn. 

Some time ago I tried training some networks while setting these symmetries to the values that would minimize the total gradient norm, effectively trying to distribute the gradient norm as equally as possible throughout the network. This significantly accelerated learning, and allowed extremely deep (100+ layers) networks to be trained without residual layers. This isn't that useful for modern networks because batchnorm/layernorm seems to effectively do the same thing, and isn't dependent on having ReLU as the activation function. 

Thus, the γ value is a “conserved quantity” under gradient descent associated with the symmetry. If the symmetry only holds for a particular solution in some region of the loss landscape rather than being globally baked into the architecture, the γ value will still be conserved under gradient descent so long as we're inside that region.

Minor detail, but this is false in practice because we are doing gradient descent with a non-zero learning rate, so there will be some diffusion between different hyperbolas in weight space as we take gradient steps of finite size.

I suspect the expert judges would need to resort to known jailbreaking techniques to distinguish LLMs. A fair interesting test might be against expert-but-not-in-ML judges.

Sorry to be blunt, but any distraction filter that can be disabled through the chrome extension menu is essentially worthless. Speaking from experience, for most people this will work for exactly 3 days until they find a website they really want to visit and just "temporarily" disable the extension in order to see it.

For #5, I think the answer would be to make the AI produce the AI safety ideas which not only solve alignment, but also yield some aspect of capabilities growth along an axis that the big players care about, and in a way where the capabilities are not easily separable from the alignment. I can imagine this being the case if the AI safety idea somehow makes the AI much better at instruction-following using the spirit of the instruction (which is after all what we care about). The big players do care about having instruction-following AIs, and if the way to do that is to use the AI safety book, they will use it. 

Do you expect Lecun to have been assuming that the entire field of RL stops existing in order to focus on his specific vision?

Very many things wrong with all of that:

1.  RL algorithms don't minimize costs, but maximize expected reward, which can well be unbounded, so it's wrong to say that the ML field only minimizes cost. 
2. LLMs minimize expected log probability of correct token, which is indeed bounded at zero from below, but achieving zero in that case means perfectly predicting every single token on the internet. 
3. The boundedness of the thing you're minimizing is totally irrelevant, since maximizing $f\left(x\right)$ is exactly the same as maximizing $g\left(f\right(x\left)\right)$ where g is a monotonic function. You can trivially turn a bounded function into an unbounded one without changing anything to the solution sets.
4. Even if utility is bounded between 0 and 1, an agent maximizing the expected utility will still never stop, because you can always decrease the probability you were wrong. Quadruple-check every single step and turn the universe into computronium to make sure you didn't make any errors.

This is very dumb, Lecun should know better, and I'm sure he *would* know better if he spent 5 minutes thinking about any of this.

The word "privilege" has been so tainted by its association with guilt that it's almost an infohazard to think you've got privilege at this point, it makes you lower your head in shame at having more than others, and brings about a self-flagellation sort of attitude. It elicits an instinct to lower yourself rather than bring others up. The proper reactions to all these things you've listed is gratitude to your circumstances and compassion towards those who don't have them. And certainly everyone should be very careful towards any instinct they have at publicly "acknowledging their privilege"... it's probably your status-raising instincts having found a good opportunity to boast about your intelligence, appearance and good looks while appearing like you're being modest. 

Weird side effect to beware for retinoids: they make dry eyes worse, and in my experience this can significantly decrease your quality of life, especially if it prevents you from sleeping well.

Basically, this shows that every term in a standard Bayesian inference, including the prior ratio, can be re-cast as a likelihood term in a setting where you start off unsure about what words mean, and have a flat prior over which set of words is true.

If the possible meanings of your words are a continuous one-dimensional variable x, a flat prior over x will not be a flat prior if you change variables to y = f(y) for an arbitrary bijection f, and the construction would be sneaking in a specific choice of function f.

Say the words are utterances about the probability of a coin falling heads, why should the flat prior be over the probability p, instead of over the log-odds log(p/(1-p)) ?