• 1 Post
  • 20 Comments
Joined 1 year ago
cake
Cake day: August 2nd, 2023

help-circle


  • The whole “it’s just autocomplete” is just a comforting mantra. A sufficiently advanced autocomplete is indistinguishable from intelligence. LLMs provably have a world model, just like humans do. They build that model by experiencing the universe via the medium of human-generated text, which is much more limited than human sensory input, but has allowed for some very surprising behavior already.

    We’re not seeing diminishing returns yet, and in fact we’re going to see some interesting stuff happen as we start hooking up sensors and cameras as direct input, instead of these models building their world model indirectly through purely text. Let’s see what happens in 5 years or so before saying that there’s any diminishing returns.


  • Gary Marcus should be disregarded because he’s emotionally invested in The Bitter Lesson being wrong. He really wants LLMs to not be as good as they already are. He’ll find some interesting research about “here’s a limitation that we found” and turn that into “LLMS BTFO IT’S SO OVER”.

    The research is interesting for helping improve LLMs, but that’s the extent of it. I would not be worried about the limitations the paper found for a number of reasons:

    • There doesn’t seem to be any reason to believe that there’s a ceiling on scaling up
    • LLM’s reasoning abilities improve with scale (notice that the example they use for kiwis they included the answers from o1-mini and llama3-8B, which are much smaller models with much more limited capabilities. GPT-4o got the problem correct when I tested it, without any special prompting techniques or anything)
    • Techniques such as RAG and Chain of Thought help immensely on many problems
    • Basic prompting techniques help, like “Make sure you evaluate the question to ignore extraneous information, and make sure it’s not a trick question”
    • LLMs are smart enough to use tools. They can go “Hey, this looks like a math problem, I’ll use a calculator”, just like a human would
    • There’s a lot of research happening very quickly here. For example, LLMs improve at math when you use a different tokenization method, because it changes how the model “sees” the problem

    Until we hit a wall and really can’t find a way around it for several years, this sort of research falls into the “huh, interesting” territory for anybody that isn’t a researcher.




  • I read that and was prepared to have my mind blown. Not really impressed, though. That article says this:

    And you can literally say anything about North Korea, the most absurd thing you could imagine, and people would believe it.

    That links to this article, which says:

    The country has been in the news of late, as ongoing negotiations between the Trump and Kim Jong-un administrations appear to have soured. The chief casualty of this diplomatic failure, the New York Times (5/31/19) breathlessly reported, was Kim Jong-un’s negotiating team, with the vice chair of the North Korean Workers’ Party, Kim Yong-chol, being sent to a forced labor camp in “the latest example of how a senior North Korean official’s political fortune is made or broken at the whims of Kim Jong-un.”

    The linked NYT article says this:

    Now, he has suddenly become the latest example of how a senior North Korean official’s political fortune is made or broken at the whims of Kim Jong-un. This week, leading South Korean newspapers reported Kim Yong-chol’s fall from grace. One of them, the conservative daily Chosun Ilbo, went so far as to report that Mr. Kim had been banished to forced labor, with many of his negotiating team members either executed or sent to prison camps.

    South Korean officials and analysts cautioned that it was too early to say with precision what was happening inside Kim Jong-un’s opaque regime. South Korean news media offered differing conjectures, including whether Kim Hyok-chol, the North’s special nuclear envoy to the United States, had been executed by firing squad in March, as the Chosun Ilbo reported, or was still under interrogation.

    But they all agree on one thing: Kim Yong-chol and his negotiating team, which had driven Kim Jong-un’s diplomatic outreach toward Washington, have been sidelined, as the North Korean leader sought a scapegoat to blame for his disastrous second summit meeting with Mr. Trump, held in Hanoi, Vietnam, in February.

    That seems pretty reasonable? It says that the official has found disfavor, says what one other paper reported with language of “went so far as to report”, and also notes that it’s hard to say for sure because North Korea is very opaque.

    The FAIR article then says:

    There was one problem: Kim Yong-chol appeared only a few days later at a high profile art performance alongside Kim Jong-un.

    Yeah, that’s hard evidence he wasn’t executed, but that’s about it. Situations like this can change on a whim in a dictatorship. Maybe Kim Jong-un had a good breakfast and decided that the official’s forced labor could be done.

    FAIR also says this in that article:

    North Korea is also a favorite location for wacky and easily disprovable stories. The BBC (3/28/14) originally reported that all men were required to wear their hair like Kim Jong-un, with other haircuts banned.

    The BBC article has a correction that it’s university students and not all men (which is missing from the FAIR article), so is that true? And it’s weird to say that stuff like that is wacky when stuff like this apparently happens:

    A second, and unprecedented, TV series this winter showed hidden-camera style video of “long-haired” men in various locations throughout Pyongyang.

    In a break with North Korean TV’s usual approach, the programme gave their names and addresses, and challenged the fashion victims directly over their appearance.

    That looks legit, with footage on youtube. Is there any reason to think that’s fake? That certainly confirms my mental model of North Korea as a wacky dictatorship if it’s true.

    EDIT: FAIR’s other statements in that article are dunking on the worst possible interpretations of what people say, which just makes FAIR seem like it has a chip on its shoulder about North Korea for some reason. I’d take what they say about North Korea with a grain of salt.



  • This is interesting, but the post is very inaccurate. The first picture is Portrait of a Moor by Jan Mostaert, and there’s no indication that it’s a portrait of Reasonable Blackman:

    https://en.wikipedia.org/wiki/Portrait_of_an_African_Man

    The second picture was drawn by a modern person, and isn’t even supposed to be Blackman, that’s what the artist thinks Edward Swarthye might’ve looked like:

    https://www.historyextra.com/membership/black-faces-of-tudor-england/

    All that aside, here’s what the book Black Tudors has to say about him:

    A surname alone cannot confirm a person’s ethnicity. Although Reasonable’s surname would seem to indicate the colour of his skin, it is in fact an old English surname, derived from the Old English Blaec mann, as are ‘Black’, ‘Blackmore’, ‘Moor/More’ and ‘Morris’. It could also be spelt Blakeman, Blakman, Blackmon or Blackmun. A John Blakman was living in England in 1206 and the name was fairly common until the thirteenth century. By the Tudor period, the name was found in Eynsham, Oxfordshire, Fowey, Cornwall, and Berkhampstead, Hertfordshire. Henry VI had a chaplain named John Blacman, a fellow of Merton College, Oxford. A different John Blackeman was buried at Grey Friars Church, London, in July 1511. A third man of the same name was a benefactor of St John’s Hospital, Coventry. None of these men was African.

    ‘Blackman’ may have originated in reference to a dark complexion, but by the sixteenth century it cannot be assumed to signify African ethnicity. As William Camden noted in 1586, ‘surnames began to be taken up … in England about the time of the Conquest, or else a very little before’. Theoretically, a man called More in 1566 could have had a Moorish ancestor from five hundred years before, but it is a rather remote possibility. We cannot even assume that ‘Blackman’, or names like ‘Moor’ or ‘Niger’, were originally assigned to men of African origin. Wilfred Niger was nicknamed Niger or ‘the Black’ in around 1080, after he painted his face with charcoal to go unrecognised amongst his enemies at night. The names could also refer to dark hair (Black), or to someone who came from a place called Moore (in Cheshire), More (in Shropshire), Blackmore (Essex), Blackmoor (Hampshire, Somerset) or Blakemere (Herefordshire), or even to someone who lived on or near a moor. In Scotland, the surnames ‘Muir, Mure, Moor, Moore, More’ referred to ancient ‘residence beside a moor or heath’.

    It is only because Reasonable Blackman was also described as ‘blackmor’ and ‘a blackmore’ that we know he was African. ‘Blackamoor’ or its variants was the most popular term Englishmen used to describe Africans, appearing in some 40% of references to individuals in the archives, and in literature from at least 1525.



  • There’s at least one example you can look at, the Jenkins CI project had code like that (if (name.startsWith("windows 9")) {):

    https://issues.jenkins.io/secure/attachment/18777/PlatformDetail

    Microsoft, for all their faults, do (or at least did) take backwards compatibility very seriously, and the option of “just make devs fix it” would never fly. Here’s a story about how they added special code to Windows 95 to make SimCity’s broken code work on it:

    Windows 95? No problem. Nice new 32 bit API, but it still ran old 16 bit software perfectly. Microsoft obsessed about this, spending a big chunk of change testing every old program they could find with Windows 95. Jon Ross, who wrote the original version of SimCity for Windows 3.x, told me that he accidentally left a bug in SimCity where he read memory that he had just freed. Yep. It worked fine on Windows 3.x, because the memory never went anywhere. Here’s the amazing part: On beta versions of Windows 95, SimCity wasn’t working in testing. Microsoft tracked down the bug and added specific code to Windows 95 that looks for SimCity. If it finds SimCity running, it runs the memory allocator in a special mode that doesn’t free memory right away. That’s the kind of obsession with backward compatibility that made people willing to upgrade to Windows 95.



  • If you’re writing code that generic, why wouldn’t you want str to be passed in? For example, Counter('hello') is perfectly valid and useful. OTOH, average_length('hello') would always be 1 and not be useful. OTOOH, maybe there’s a valid reason for someone to do that. If I’ve got a list of items of various types and want to find the highest average length, I’d want to do max(map(average_length, items)) and not have that blow up just because there’s a string in there that I know will have an average length of 1.

    So this all depends on the specifics of the function you’re writing at the time. If you’re really sure that someone shouldn’t be passing in a str, I’d probably raise a ValueError or a warning, but only if you’re really sure. For the most part, I’d just use appropriate type hints and embrace the phrase “we’re all consenting adults here”.