3,484 karma · joined January 15, 2014
I think this claim is acceptable in context, but taken alone, this needs to be qualified. Some aspects of language are universal, like abstract information structure and other pragmatics, but no model is going to speak rural Khmer dialects well anytime soon because not enough of it has been digitized, fundamentally speaking, hence qualifying what it means to "be good at all languages".
(Note: I'm speaking vaguley, on the same vague level as your comment. I can be more precise as to what my indications indicate)
A lot of simulation we only have exponential-time algorithms for. Motion planning, protein folding, etc. For a lot of these today, the SOTA is to use an NN model to learn the heuristics from data. OP's claim only rings true if one can only think of just the algorithms that undergrad CS now studies.
1. what makes you confident that the rules of particle physics alone can lead to a perfect simulation of macroscopic phenomena? (as you call it "emergent behavior")
2. what makes you confident that our universe is not "turtles all the way down", that as soon as we develop more precise and powerful instruments we will discover new "fundamental phenomena" that requires new theory? This has happened before where neat Newtonian physics was found to fail at edge cases detectable by more precise instruments.
3. considering that we have no feasible way of snapshotting a 1 cubic inch of the world, and even if we have done that, no feasible way of simulating its dynamics at the qft level for one second, and yet we have "approximate" physical theories with less invasive instruments that can do that (for specific purposes), is the ideology of calling macroscopic phenomena an emergent behavior of particle physics still relevant? In a similar bent, there exists a Starbucks in certain places in the world, and there does not exist a Starbucks in other places. This is a fundamental truth of the present world, and many peoples' lives are oriented around that and affected by it. Nevertheless, it cannot be precisely predicted by mathematics where there exist Starbucks and where there does not exist Starbucks one year into the future.
4. I hope you can see that just as Newtonian physics was useful for describing an approximation to the world but still fails at giving knowledge about sufficiently complex systems and sufficiently small scales, you now see that likewise particle physics is just an "effective lens" for a particular genre of human activity.
But you and I have ignored all that and are blind to all these things that are part of reality that you still don't have knowledge of, or abide by mathematics. If we simply put our attention to the parts of experience that have order that can be described by maths and blind ourselves to everything else, then sure, maths seems very powerful.
> Claude repeats man twice like Fagles, despite that not being in the Greek.
This style is not unique to Fagles. In Greek poetry modifying words (polytropon, for preceding andra ...) can be in apparently insensible locations not found in prose and speech, but accepted in poetry for the sake of meter and poetic style. In Greek, this heritage makes it acceptable, and the marked cases provide redundancy that allows the very flexible poetic word order. Here, there is a choice to capture the sense putting "man" before Muse as the focus of the sentence as in the Greek, and then repeat "the man" again before "of many turnings" because it is otherwise more ambiguous to just say "of many turnings" after "Muse".
> Using driven and wandering is a weird repetition
I don't think it's negatively weird, it's a choice that tries to capture the range of senses that a Greek might associate with it.
> It also drops the "very" for no apparent reason (also not present in Fagles or Lattimore)
Lattimore doesn't really capture this. Fagles captures this by using a more expressive "time and again" for "far"/"many".
Maybe one can argue that Claude choses to reflect extremity in this by giving both "driven" and "wandering".
> once
This is a very understandable gloss that I had chosen to supply "after" instead. It is not an innovation by Fagles, but a choice Fagles from a small range of valid common choices.
> It also uses citadel rather than city
I think citadel is altogether an appropriate general translation as well. It is described as "holy", and the Greek is a Homeric "ptoliethron" that everyday Attic/Ionic speakers would consider an archaic word, where the everyday word is the "polis" that you and I are familiar with.
If your opinion is that Claude does not have a consistent editorial voice that is preferable for translation, sure I agree. If your opinion is that it translates while being aware of and perhaps taking cues from prior translations, sure I agree, with the caveat that it also sees dictionary definitions, the Homeric Greek itself, common ways Homeric is translated, and more literal glosses, and it seems to me in fact to take more influence from these (at least, just by looking at this first line). But it is difficult for me to agree that it just mashes up two prior translations.
As for the criticism of editorial voice I wonder if it can be solved by a matter of prompting with expertise.
> ἄνδρα μοι ἔννεπε, μοῦσα, πολύτροπον, ὃς μάλα πολλὰ πλάγχθη, ἐπεὶ Τροίης ἱερὸν πτολίεθρον ἔπερσεν·
man(acc.) me(dat.) tell(pres. imp.), Muse(voc.), many-turn(acc.), who(nom.) very(adv.) far(acc.) wandered(aor. 3sg. m.p., but note active sense of "to drive off"), after(adv.) Troy(acc.) holy(acc.) city(acc.) (he) sack(aor. 3sg.)
Glosses my own, text from https://scaife.perseus.org/reader/urn:cts:greekLit:tlg0012.t...
I can't evaluate how true or sensationalist this story is, but this bearish article suggests to me that Tesla robotaxis today isn't yet the success you are painting https://electrek.co/2026/07/03/tesla-robotaxi-miami-service-...
I don't think I was clear and explicit, it being tedious to write, and I apologize for that. I also apologize for shifting the goalposts as I had not written out my own position, which is not exactly in the "grandparent commenter"'s position (that I had not previously given enough attention understanding), but it is also not in agreement with yours. I don't mean to say that we are not presently in the "bitter lesson" (your idea of what the bitter lesson says) regime. I definitely think that a lot of progress can be done right now by emphasizing the humanoid robotics platform as a foundation. What I mean to say is that I don't know if that platform with the hardware we have today is sufficient for parity with human housekeeping tasks in the domains that we wish it to have parity. The bitter lesson itself (not your understanding of it) is in fact silent on this as it is in relation to feature engineering, where it is a clear point, but you seem to be adapting it uncritically wholesale to mean something more than what it is written about. My position is that, it is unclear whether today's sensor platform is sufficient for parity. It is less strong than the blog post author's "Why Today’s Humanoids Won’t Learn", it is a "We can't say whether or not today's humanoids will learn", but it is something that also contradicts a "the bitter lesson means today's humanoids will learn" thesis.
The self-driving car supports my claim, because after so much investment in capital and time, we ended up with a car with comparatively expensive LIDAR sensors as our preferred platform.
It is not reporting on an opinion of a representative or proxy of the European Commission.
LeCun thinks that LLMs are a bad fit for AI that understands the physical, dynamical systems that we inhabit, and that understanding this is necessary for AGI/ASI.
I don't know that Hutter is bearish on LLMs, but Hutter is interested in AI that can reason exceptionally well given infinite compute, and approximations of such a reasoning AI. I think he is open to the idea that LLMs can be such an approximation.
This is the narrative for us in developed nations, but the majority of users today are people who were in developing countries and got a mid-tier smartphone to chat with friends and do banking with the same values as Apple users.
1. Older ML models encoded in their architecture and lack of expressivity a bias to simplicity; which aided interpolation.
2. Overparameterized models instead use regularization to nudge parameters to simpler and more robust representations, while still memorizing the noise. In this manner, we still achieve generalization performance OOD. Moreover, the softer nudging and fundamental architectural expressivity allows for "data-specific" generalizations and representations that may be impossible to represent in small models. 3. At the critical point between the two regimes, the model is expressive enough to memorize; but not expressive enough to simultaneously both do that and encode general patterns.
I wonder how this understanding translates to these researchers' models of deep learning.
Similarly, a call center worker may not care about having their accent being heard, but wants to get their numbers up, without struggling with a customer that isn't familiar with their accent, and enjoys the ease of speaking in their own accent than having to use one that distant customers are accustomed to. Likewise a customer probably just wants their problem fixed, without the effort of getting accustomed to an accent that they rarely encounter. This meets your definition of deplorable, but analogous to the former scenario, perhaps not deplorably so.
This isn't true in the way you are thinking of. An accent can pronounce words the same way that another accent distinguishes. An accent can pronounce word x that another accent pronounces word y. What comes to mind immediately: in Indian English accents, RP/GA fricative "th" is pronounced as the aspirate, while the RP/GA aspirated "t" is pronounced retroflex, so naively, "three" can be misheard as "tree".
The working-class accent that I use where I'm from (not India) is syllable-timed (stress does not lengthen the duration of a syllable), and uses pitch lexical stress, rather than intensity/loudness for it, and stress itself is frequently very differently located compared to RP or GA. For "th" as well, we collapse it into t/d.
All in all, for someone who has heard it for the first time or rarely, it can be extremely disorienting to listen to a very distant foreign accent.
- Is it dehumanizing to Filipinos that Filipinos probably now do their job more efficiently without having to learn an accent that they are not exposed to?
- Is it dehumanizing to Filipinos that they no longer enjoy having their accent heard as a externality of a counterfactual arrangement?
- Is it dehumanizing to the customers that the company does not expect their customers to be cosmopolitan enough to understand a foreign accent with ease?
- Is it dehumanizing to the customers that the customers are now more sensorily shielded from a current-day reality regarding globalized providers of service?
- Is it dehumanizing, not due to this decision itself; but the globalized arrangement, to Canadians that they cannot expect to hold such a job and get by in Canada? Or perhaps to Filipinos, that such a job might be low-paying in their own country (or in respect to non-domestic goods that need to be purchased from outside their polity)?
- Is it dehumanizing, regarding not this decision, but the offshoring decision, that such decisions can be made without consent by employees and contractors?