dojomouse

Yup, fair enough - having read the full overview from typesafe I agree they mean “won’t assign a probability to an outcome outside a defined set” rather than “won’t incorrectly assign a probability”.

LLM can be made deterministic too, they’re just generally not because the element of randomness is a feature.

There’s a big difference between deterministic and smooth though. Typical LLMs certainly aren’t reliably smooth, so the small prompt change might product a large and unpredictable output change. I’m not sure if that’s any better with the typesafe approach.

Hallucination is the result of a series of choices of options based on poor predictions of probability (or poor sampling from a probability distribution). As I understand the typesafe approach they’re allowing the set of options to be constrained, but the potential for a misrepresentative probability distribution is still there, and hence so is the potential for “hallucination”.

> The cloud deployment surface covered in our contract would not permit powering fully autonomous weapons, as this would require edge deployment.

… What?? Much of this seems duplicitous, but this isn’t even coherent. Is their implication that it’s not “autonomous” if it involves an api call to an external system? That mere definition would be extremely alarming.

I think it’s very much a magical fantasy computer :-) A thought experiment, rather than a claim to something that’s been built or exists.

The questions you pose are interesting ones, for the sake of the experiment I think at least:

- Programming

- Loading new software

- Adding new functionality

- Persisting user data

Could all at least in principle be achieved without changing the network architecture, but rather just providing the relevant data at the inputs (eg the full bit stream of the install file of the new software being installed) and having that lead to adjustments in activations (not weights or architecture) across the network which lead to any future inputs to the network resulting in the outputs that would be expected in the presence of the new software. Same deal for the other examples.

As to whether this is practical or achievable at present - not even remotely close in my view. But it’s still an interesting idea, even if just to think about why it wouldnt work and what that implies for future development direction of multimodal networks etc.

Perhaps. I was thinking along the lines of MarkBurns response - ML will allow us to efficiently look in those places we might otherwise only have searched by accident.

If ops point was rather that “accident”/“luck” are uniquely human… I don’t agree. Luck is when probability works out in your favour - and that can happen all the time with any sort of probabilistic search, which is rife in ML.

I think ML is likely to be material to us making many more such discoveries. So much of the current constraint is not in the knowledge to identify the interesting pattern, but the capacity to look for it at scale.

In New Zealand we have a pretty effective system that covers all fossil fuel use nationally, as well as several other greenhouse gas emission classes (notably not agricultural emissions though hopefully they’re included soon). The EU has something similar as do several other jurisdictions. It’s a solved problem at scale.

In NZ we also have an effective system for recognising and incentivising certain classes of forest carbon removals (which I think are a legitimate and important class of credits - unlike avoided emission credits which I agree are junk).

This is extremely cool! Nice work. What data does it rely on and do you have any plans to support other countries?

It wouldn’t help with kinetic energy harvesting from the raindrops as that would go into the funnel as heat.

It might provide a way to harvest the remaining gravitational potential energy of the rain (possible funnel being your roof and guttering) but the only upside is that you could concentrate the energy with something that’s already there (and hence harvest over a smaller area). The amount of energy (and hence value) available would be even lower - unless you had a really high roof.

This is also the reason I abandoned my high school scheme of hydro turbines at the bottom of downpipes.

As the comments below say - you need to be working at the scale of a few major geographic features as a funnel before it starts to get really interesting.

Love you for this! I had exactly the same “solar freaking roadways” thought, although at least that idea qualified by basic theoretical analysis of available energy and area for harvesting and conversion efficiency. It was an obviously terrible idea for other reasons :-) yet it still got a prototype…

I wasn’t sure about the droplet analysis so took your same numbers (25mm/h, 10m/s) and just worked out aggregate mass: 25mm over 1m^2 = 0.025m^3 = 25kg

0.5mv^2 => 1250J/h… so looks like we agree.

And to add a simple economic analysis of why this is such a dead-end idea:

Mawsynram, in India, is apparently the rainiest city in the world with roughly 10,000mm of annual rainfall - 10x the global average.

A given rain energy harvesting panel, deployed there, would generate 500,000J/yr… or 0.138kWh. That’s significantly less than what a typical rooftop 1m2 solar panel would generate in an hour on a sunny day. 0.138kwh is worth around 1.3cents at 10c/kWh.

A big roof might get you $1-$2/year. You couldn’t pay to clean your roof for that. You couldn’t even pay someone to answer an email enquiry about the install costs for your system for that. This solution would have to be VASTLY cheaper than paint to stand a chance of being viable.

There is a reason our existing systems to collect power from rainfall rely on vast existing landscapes and aggregation mechanisms (rivers) to concentrate the rainfall for us.

It is - in my view - a dead idea.

Correct, but what the parent here presents is a theoretical upper bound. A working product wouldn’t even get close. When the theoretical upper bound shows that something could never aspire to more than a vastly inferior alternative to existing proven technologies, the correct approach is to abandon it rather than invest in iterative improvements.

I agree we should keep an open mind regarding creative ways of collecting energy from the environment. But we should also abandon those which are quickly demonstrated to have no meaningful potential even if we were to perfect them.

It also ignores the fact that there’s no need for it to become “alive” or “conscious” to be a threat in the way he describes. It just needs to be an agent with an mis-specified, poorly specified, or maliciously specified goal. And there are already numerous examples of those. The only debate is around capability, and here he makes multiple references to “infinitely” capable. So the whole argument seems like wildly disingenuous strawman, consistent with his attempt to classify all those raising concerns as naive (or corrupt) cultists - not exactly the vibe from the likes of Geoff Hinton / Stuart Russell / Max Tegmark; all of whom generally act with far more integrity (it seems) than Marc Andreessen shows here.

Ironically I think the whole article is motivated by the thing he claims to condemn - namely: he’s a bootlegger, who has an interest in freedom of ai development.

Part 2 is much more interesting. Part 1 was very very weak.

> One reason is that there are some problems that a company will only encounter once it begins testing fully driverless operations. Waymo and Cruise’s problems with fire hoses and caution tape is a good example. A human driver would disengage FSD long before it got into situations like that, which means Tesla would be unlikely to have the training data necessary to train its cars to handle it properly.

The reasoning here seems flawed to me. I assume Tesla use periods when the human is driving for training data, so this is no constraint at all.