Zhaoxin (Chinese market)

Maybe some are like that, but have you seen Marques Brownlee's review of the Xiami SU7 ? High end, high quality. Priced at equivalent of $40-45K, and in Marques opinion competitive with anything here at $70-75K point.

Tesla's aren't exactly flawless in this respect either - plenty of reports of fires and recalls like Cybertruck glued-on body panels falling off.

> First they claim this is because of tariffs...but aren't vehicles with final assembly in the U.S. free from most tariffs?

Does that apply any more? It used to be that "made in USA" cars might actually have less actually "made in USA" content than a "foreign" car due to these bizarre cross-border rules, but Trump's tariffs are/were anyways ADDITIONAL on top of pre-existing tariffs, and generally so high that they dominate.

SCOTUS recently declared Trump's tariffs illegal, but Trump responded same day that he would just impose tariffs under a new excuse. Maybe in a years time SCOTUS will shoot that down too, and in the meantime nobody can guess what is going to happen.

> Then, they claim their product is not competitive with Chinese competitors...with a vehicle they were going to assemble in the U.S.? That also makes no sense. Why wouldn't the manufacture it in Asia?

China's level of factory automation is insane, and they have the whole supply chain there which is a huge factor. It may well not be possible for another country to compete unless they can match these advantages, even if based elsewhere in Asia.

You would imagine Honda has been keeping close tabs on Chinese EV developments, but it seems they are making very progress, and I have no idea how long the planning cycle would be for Honda to create whole new lines. Presumably whatever sunk costs they have made sense at the time.

It's a very rational response, and certainly impossible for a business to plan when the rules (e.g. tariffs) are changing on a daily basis

Certainly China seems far ahead of everyone, US as well as Japan, in being able to build low-cost desirable EVs.

Of course the tariff playing field is far from level, and Trump is likely to flip-flop on what he's imposing on which country based on what he had for lunch, what presents they've sent him, etc.

It may be that even though China is the low-cost producer than Japan still has an advantage in the US market because Trump is choosing to penalize them less than China, but that can change in a heartbeat and therefore isn't something that it makes sense for a manufacturer to plan around. Even if Trump made some "commitment" to Japan over tariffs, his words are meaningless, as I'm sure the whole world is well aware.

In the meantime, the official goal of tariffs is supposedly to encourage domestic manufacture, but Trump's buddy Musk doesn't seem to be stepping up to the challenge, no anyone else. The car manufacturers have all learnt that it's more profitable to sell fewer expensive SUVs than more entry level cars, and new cars in the US are increasingly becoming a luxury item than many can no longer afford.

What I've often wished for, when driving, is to be able to talk to the map, to update what it's doing - I wonder if this will let me do that, and if so will it be available via CarPlay ?

The sorts of things I'd like to say/ask are:
- please let's make a stop at X
- please go via the highway (H)
- how far to the next gas station/services/etc

> Personally I would be surprised if world models offered anything of value given they operate at such a low level.

You're thinking of the animal approach in the wrong way. Forget all the "world model" type, and just think of it as a predictive model, a near cousin of an LLM, that learns to predict next perceptual input(s) rather than next token form an historically gathered training set.

Let's also note that the input to an LLM really isn't text or symbolic sub-word tokens - it's really the high dimensional embeddings that are created at the input layer...

Now, contrast, or rather compare, this to the animal (let's say human) visually scanning lines of text in book, or street signs or whatever. The input will also be high dimensional embeddings, just ones that originated as visual input, and what-follows-what is exactly the same whether you are learning in real-time or learning from a frozen dataset. Obviously you will learn exactly the same as an LLM would have learnt given the same data frozen as a training set.

So, the animal can do (and we do!) exactly the same as an LLM, but it can also do a lot more, so it's capability is a superset of what an LLM can do.

Finally you should remember that continual learning and human/animal intelligence (AGI) are the two holy grails of AI research, and there are no easy answers, so if you think there are then you should realize you are misunderstanding something. If LoRA was the answer to continual learning, then they'd be using it. If looping the output of a model back into itself (re: your first reply in this thread) was all that was needed for animal intelligence, then we'd already have AGI.

LoRA is just an efficient way to fine tune a single model. It's not about merging different models.

Merging models is not even well-defined. What would it mean? What would be a principled criteria for deciding how to merge them when there are conflicting weight updates needed?

How do you address the privacy concerns of merging models? Are you really proposing to merge proprietary/private data from multiple companies and/or individuals then redistribute the merged changes to everyone? Sounds like a non-starter to me !

Sure the industry is fine tuning models woith LoRA, but they are NOT then sharing their private updates with each other !!

> I don't understand this line of argument. What makes world models any different in these regards?

Let's call it animal intelligence, not "world models" (which means nothing). The difference to LLMs is that LLMs by definition are pre-trained. The animal intelligence model by definition requires continual real-time learning (NOT pre-training) - it's not a nice to have "extra" to avoid groundhog day with your LLM intern, but rather core to the approach. My point was just that we don't currently have such an algorithm.

> ICL is "realtime"

Sure, but it's not really learning. As soon as the context changes (or gets compacted) then all your internal activations (which is the only place this "learning" manifested itself) change, and your "learning" is lost. How does this help address the need to be able to learn continually all day (and also for this to be real learning - weight updates)?

> Training up something like a LoRAs is not that hard. People can do this shit frorm their own workstations.

LoRA doesn't help with what we're discussing. It's a way to make fine tuning more efficient, but it's fundamentally not an incremental learning algorithm, and if you try to do multiple consecutive LoRA fine tunes, then you are liable to experience catastrophic forgetting of your earlier fine tune.

Also while LoRA is more efficient then tuning without it, it is still far from realtime.

Not really - continual learning from real-world inputs completely disrupts the whole "pre-train then serve to everybody" LLM approach. Instead you've now got every model instance running and experiencing different things and needing real-time learning.

Not only do you need a billion or so instances of that real-time learning algorithm running in parallel vs the "build a datacenter, train once" approach, but you need to invent that so-far elusive incremental training algorithm in the first place.

You could shoot for a poor man's version where you try to batch learning up to end of the day rather than do in in realtime, but think what that means for trying to pick up a new skill that requires dozens, hundreds, or more of rapid experiments. Things that should take hours to learn would now take months, and it's not even clear how that would work. After last night's learning update, does the model now need to go back and put itself (and the external world!) back into the exact state where it was when it was about to learn something so that it can now do the next step?!

The difference "predict the real-word, not the training set" SOUNDS minimal, but the consequences are far from it.

No - the opposite, it's the Milky Way (not the earth) that is almost as old as the universe.

I'm not a big fan of LeCun - his level of recognition seems far in excess of his actual accomplishments, and his main claim to fame seems to a somewhat questionable claim to have invented CNNs, a long time ago.

That said, I do think LeCun is correct (but hardly alone) is saying that LLMs won't get us to AGI, and that we need a different approach, more akin to animal intelligence.

While LeCun does talk about animal intelligence, there is also this focus on "world models" and physical grounding, and it's not clear what his actual research/development direction will be. It seems he was basically pushed out of Meta (put in a position where he was bound to quit), and this new venture is a reaction to that ... if he is just trying to create a new cozy home for himself where he can slow-walk his JEPA line of reseach, while talking up "world models", then this is not going to work!

To give LeCun the benefit of the doubt (which he may not deserve), the real focus may be on animal intelligence rather than "world models" and JEPA, in which case having a large well funded lab pursuing this approach is significant.

While the animal intelligence approach would result in the (artificial) animal learning a behavioral world model (not at all like what some confusingly like to also call a "world model" that LLMs have internally), of how the world evolves and reacts, that seems an odd thing to focus on.

The real difference between the animal intelligence approach and an LLM is that while an LLM predicts training sample continuations, and stops learning once it is trained, an animal predicts the real world (via it's perceptual inputs), including how the world reacts to it's own actions, and learns continually.

The LLM is a copying machine. The animal is a learning machine. Since the animal is learning about the unknown, and how to achieve outcomes (how it's actions affect the world), it has potential to be creative and make new discoveries.

OK, so possibly complex cells evolved more than once, but apparently still a rare enough event (and/or one dependent on a lot of other things happening first) that it took a long time (2B years) to happen. If eukaryotic cells, and multicellular life, had appeared sooner it might suggest some inevitability to this progression, but it seems that at least one reading of the slow emergence is that this was a very low-probability event.

No :) I was thinking of Milky Way age.

I expect that life itself is not rare - that the universe is teaming with life, but maybe only at simple prokaryotic cell type of level. The emergence of "life" (self encoding self-replicators) from non-life seems somewhat inevitable when a few conditions are in place.

The case for advanced civilizations being rare is from looking at the earth (a sample of one, but still ...). Earth is almost as old as the universe. Simple life here arose almost immediately about 4B years ago, but it took another 2B years for eukaryotic cells to emerge, the Cambiran Explosion of diversity of life only happened 500M years ago, humans only emerged a few million years ago, and our ability to transmit radio signals happened yesterday.

Earth seems like a goldilocks planet, so why to expect a much different timeline on other planets? Some of these transitions such as from simple replicator cells to eukaryotic ones with the complexity necessary for multi-cellular life are apparently far from a slam-dunk (having taken 2B years to happen here, and by all estimates only ever having happened once). If life here on earth, almost as old as the universe itself, only became capable of transmitting a radio signal 100 years ago, then why to expect that life on other planets is so much further than ahead of us? Maybe there is another radio-transmitting civilization on the other side of our galaxy, but if they only just started transmitting 100 years ago, then the radio signals are still in transit and will take another 100,000 years to reach us ... assuming they were transmitting at whatever absurd power levels would be necessary for us to be able to detect it.

I'd be interested to see some analysis of how strong of a signal an alien civilization would need to be transmitting for us to have any chance of detecting it with our networks of radio telescopes.

Sure we can hear Voyager's weak signal, which is impressive, but in the galactic scale of things it is right beside us, only just having left our solar system.

Any potential aliens are much, much further away ... On a scale where our sun is a grain of sand, the closest star is another grain of sand 600 miles away, with radio signal strength weakening according to an inverse square law.

Of course it's almost certain that the closest alien civilization (assuming one exists) capable of radio transmission isn't so conveniently close by, and if it was on the other side of our galaxy (100,000 light years away, not just 4), then what sort of transmitter power would they need to be using? The inverse square law is brutal.

What if the nearest civilization if not even in our own galaxy?

Rather than injecting chemicals into your dick, which sounds a bit extreme, couldn't they just do a "Spinal Tap" and shove a foil-wrapped pickle down there?

Do Olympic officials actually inspect your junk before/after ski jumps ?