It worth recognizing that few or none of these comparisons come anywhere near apples to apples.

An innovative solution originating with a human almost always involves far more time investments, including long periods (days, weeks) of ostensibly unfocused time that is used productively by our brains, recent high focus in the problem area, and far more try-and-fail passes, providing significant specific feedback.

Also, human insights are often preceded by years of self-directed puzzling over relevant relationships or subjects.

Without some strong correspondence in problem solving contexts, comparisons are not really model vs. human, but typical model use (tremendous restrictions placed on them) vs. atypical (best of breed) human self-direction. Not a true "capability" comparison. More of a "result" comparison between extremes of contexts.

Pick a random human, random SOTA and random problem with any significance, and give the human and SOTA the same time resources, and I suspect the model will do better with very high confidence.

The system could work just as well for the user, stored only on the users devices, syncing in encrypted form only.

The plausibility in plausible deniability evaporates instantly with device/service/software provider's accessibility.

But faux plausible deniability seems to be just as effective of a fig leaf.

"Male chauvinist" and "Male chauvinist pig" are related in meaning but not equivalent.

The difference matters enough that preemptively suggesting "pig" is not a desirable "service" for most people.

I think the dichotomy of open-research vs. commercially-driven innovation might be easier to coordinate, with three stages, not two:

1. Open research: Great creative freedom, with great resource constraints. Lots of ideas, lots of things tried, at insignificant cost relative to commercial work.

2. Results filtering, vetting, regularization: Opportunistic selection of apparently successful work from stage 1, for tentative testing at increasing scale. The best ideas end up with moderate budgets that verify benefits, which are reworked for compatibility before any handoff to stage 3.

3. Commercial efforts: aimed strictly at improving quality and economics of current revenue sources.

Direct negotiation between #1 and #3 involves too much dissonance over globally complementary but locally incompatible values.

#2 can be a relatively small, high payoff group, keeping #1 and #3 from butting heads, while facilitating them both.

Exactly.

Yann LeCun has the syndrome where someone who sees the value of B (non-text modalities, i.e. physical reality) cannot stop themselves from constantly critiquing A (LLMs), even though clearly A has its own advantages.

I think he is a smart guy. But he constantly deploys the crackpot's PR playbook.

On Earth we have optimized for: single core speed, multi-core speed, performance/energy, low demand (for mobile), and performance/energy, high demand (data centers).

A new dimension to optimize for, that space is going to incentivize, is highest stable operating temperature.

Higher temperature operation sheds heat energy via radiation much faster.

Some post-silicon substrate research is aimed at wafers that can operate at extremely high temperatures.

I imagine given how often materials research pulls a rabbit out of a hat, there is a lot of head room for temperature.

> Higher temperature operation sheds heat energy via radiation much faster.

Isn't this what heat pumps are for? Then you can run the radiators at higher temperatures without needing to run the electronics at those temperatures.

This is a kind of nasty tradeoff in space. You can dissipate entropy per unit time from the same radiating area if you use a heat pump, but the heat pump itself uses power even if it’s perfect and operates at the Carnot limit, so you need a larger solar array.

There’s probably an optimum that involves a heat pump, but I don’t think it’s vastly better than not using a heat pump, and it’s certainly heavier and more complex.

Heat pumps gain efficiency if the temperature gradient they have to maintain decreases.

If you want chat privacy, you should probably not craft a situation that initiates a human-led investigation into those chats.

Laser focused self-sabotage isn't a strategy most people recommend.

Unless of course, your dream is to break out of prison. But you need to get in one first.

> Art Nouveau architecture, being nouveau, relied far more on artistic genius and bespoke decorative elements.

> an architecture meant for jewels [...] reliant on [...] the starchitect.

> This made the style difficult to replicate at scale.

I think this is it. There was no convergence on a set of stable basic elements, from which other elements, unique or not, would still inherit common constraints. Without common constraints, extensions or variations just create a mess of ideas, good or bad, with no coherence between them.

> It’s not even clear what a badly executed Art Nouveau building would be

If nobody knows where a style starts or stops, or what is required or not, it isn't really a style. It is a movement encompassing many styles, a dynamic evolutionary explosion that will not just lose common definition but common recognition. Not a specific design or style.

Not going to argue with the article on any front.

But when it comes to measuring productivity of transformational technology, a lot of things make that hard.

A lot of early use can be considered exploratory, whether it was intended to be or not. Meaning it was productive, but in a way that pays off down the road, not as much up front.

Not only are exploration payoffs not evident early, but they are also likely to be extremely high variance. Food for lots of "Well, of course it was a waste of time!" and "Well, of course it was going to pay off!" concurrent conclusions.

Then there is the fact that real productivity gains achieved early, may be the minority of wins, but are likely happening in places where they will grow very fast. At some point, they will "suddenly" appear and dwarf the losses.

But you don't find a lot of unexpected sweet spots if you don't throw the kitchen sink at them.

I think the point was meant to be more simple and broad:

A system is not just its little bits.

Where "system" and "little bits" can be just about anything. But people endlessly repeat this mistake, and it has been a go-to misunderstanding/misrepresentation for a lot of people on this forum.