For example, it could be truly true that a developer is roughly as good as a 1.5b model. It could really be true, in which case we’re not valuing these models for their true simulation power (yet). Might be the best interview test, you must generate hand written code that’s better than a small model (or show better judgement).
For the presidency, the current benchmark to beat is GPT2 it seems.
I wouldn't want to torture the author by force-exposing him more to Trump's inanities. That's already 5x longer than I can stand before the urge to tear my ears off becomes unmanageable.
You're generous. I would've gone with Eliza, but on second thought that's an insult to Joseph Weizenbaum.
Actually this type of applications should be the killer application of AI.
The extreme analogy is that robot instead of human should explore the nuclear incident of Chernobyl and Fukushima.
Now we're trying to build automated analysis of ECG interpretation using AI. In order to interpret standard 12-lead ECG waveforms signal of 5 minutes for Afib detection that's equivalent of 60 minutes duration. For long duration multi-lead Holter ECG it's at least 24-hour, but I think you get the idea.
You can hire cardiologist that interpret the ECG but you probably need sub-specialist cardiologists specializing in ECG interpretation for the best results but only handful of them exist in the world with ratio of 1:100K or 1:1M for experts:populations, depending on where you're living. These expert cardiologist would rather spent the time doing more meaningful exercises for them like pace maker & IPC surgeries, teaching future cardiologists, and having holidays once in a blue moon.Even they can mistake due to human errors and other limitations. Thus this exercise is tailor made for AI.
I understand it's made for personal use but if it's posted on the public web at least a disclaimer would've been nice.
The difference between posting an image url without and with a random string in the url.
Blue line is successful requests (people viewing the image I posted), green are unsuccessful requests (people trying to find other files).
Second blue bump is the screenshot with a randomized „hard“ to guess url. First bump the default iOS screenshot name in the url.
There are unusual things about Trump's communication: extreme informality, seemingly not running things past advisors or editors (for both speech and writing), and probably making things a lot more personal than other presidents have done (talking more about himself; talking more about his experiences, opinions, and emotions; talking more about individual friends and enemies). Maybe also a far greater willingness to make digressions, whether personal or not.
Still, I think the prior commenter is right that we can be surprised by how disfluent almost everyone's extemporaneous speech appears when precisely transcribed. This is often hidden because published interviews or speeches in the press are usually extensively edited, so you won't see something like
"And I, um, I said, well you know, I said that we-- we're going to fix this pro--problem, uh, this problem that's been affecting a lot, a lot of people."
Instead, you'll see something like
"And I said that we're going to fix this problem that's been affecting a lot of people."
There was a big controversy about Dan Quayle as vice president and about George W. Bush as president because they both seemed to make a lot of speech errors and misstatements. I think I remember some linguists complaining that the oddities of both Quayle's and Bush's were probably being exaggerated merely because people were paying such close attention but were unaccustomed to seeing carefully transcribed speech. And I think the same is true for Trump's speech, even though he is also flouting various communication norms in both his writing and his speech, sometimes consciously.
Earlier presidents also had more patience for delivering prepared written speeches, from paper or from a teleprompter. For many of their speeches, they were extemporizing less, and when they did extemporize, they were comparatively better at keeping on-topic.
Humans can indeed make sense. Not to be too Swiftian, but in some countries, children even go to school for it.
For semantic analysis, however, git is just not the right tool. It's a chronological graph that affords diffs. For code.
We need Python NLP and spacy here. But even the best tooling won't get far. A compiler would abort nonsequential logic and unsatisfied contracts and grammar.
An important business presentation would have structure and facts. Inside the theoretical classroom, a public speech is different from casual, random remarks. Unless the speech is intended for entertainment (e.g. comedy, theatre) or some dark usage, such as propaganda.
From TFA:
> the cyclical pattern of his speeches, little snippets of “the best words” and talking points assembled like a ransom note cut from a magazine
That's gold!
It could take 20 minutes until we reach the conclusion, at which point he finally explains what the purpose of the final formula is and why we want it.
I got the habit of reading his book in reverse before the lectures; reading section by section in reverse order from the end. This way the mathematical calculations had a clear goal and were faar easier to follow.
... brilliant maths, but he was fully and utterly incompetent at teaching it. And he had a bit of an ego about how many students fail each year because "they're lazy".
Me and a few friends did deep recaps to de-tangle the explanations using his book, rephrasing it in a easier and shorter format; and he accused us of cheating because our scores deviated from the normal distribution.
All this to say; sometimes clever doesn't correlate well to great with words.... though dont take that as a endorsement of trump.
It's certainly not what is drawing attention to this thread, but Trump's ideas. However we should be aware that transcribing someone's speech with high fidelity 1°) will inevitably make them look dumb 2°) but only because we're not used to do this task, in fact our brains constantly do the opposite, fixing and filtering what we perceive in order to make out what is being communicated. You need some training to do that in fact!
Nonetheless, the author decided to included repair initiators ("uh") but not pauses, because I think he wanted to underline how dumb Trump sounds (pauses do not carry this connotation). This argument was used during the 2012 French presidential election against candidate Eva Joly (a prosecutor) in an attempt to tackle critics about her Norwegian accent on more objective grounds. Moving beyond right-wing gut-driven glottophobia, some people on the left thought they had found reasons to validate this criticism by examining improvised declaration she made during interviews through the filter of written transcription. I can accept this argument only if the author's article is a transcription of a spontaneous monologue (which is impossible of course).
Anyway I stress my point: we underestimate how messed up our actual linguistic production are. We're all too focused on the ideal of written language.
https://assets.cambridge.org/97811084/17211/frontmatter/9781...
>SPONTANEOUS SPOKEN ENGLISH
>A new, thought-provoking book on the theory of grammar and language processing, Spontaneous Spoken English is based on the analysis of authentic speech produced in real time. Drawing on insights from cognitive psychology, neurology, and conversation analysis, the author offers a fascinating, easy-to-follow account of why spoken English is structured the way it is. The traditional product-based approach to grammar is given up in favor of a dynamic, speaker-based perspective that integrates language-structural, neurocognitive, and dialogic aspects of speech production. Based on fresh empirical research, Haselow argues that grammatical knowledge rests upon two cognitive principles of linearization called microgrammar and macrogrammar, which are shown to interact in various ways. The book discusses a broad range of speech phenomena under an integrated framework, such as the omnipresence of “unintegrated” constituents (e.g. discourse markers), ellipses, or the allegedly “fragmented” character of syntax, and explains the mechanisms of processing efficiency that guide syntactic planning.
There are people who take these transcripts and re-enact what he's actually saying, but in their voice.
Doing that makes it immediately obvious that he's a rambling moron.
I feel like you have a pet discovery about "spontaneous spoken English" and now you are seeing it everywhere even when it doesn't apply. Like one of those If you have a hammer everything looks like a nail kind of things.
Trying to search that post.
Edit: in the discussion there was a link to a do a YouTube video where to movie characters were playing word badminton with each other.
Edit: this clip https://youtu.be/swqfFHLck1o
https://wiki.improvresourcecenter.com/index.php?title=Harold
I think what bugs you in Trump's "speech" is that it looks like a one-sided conversation, and as such, features the kind of structure you'd in fact expect to find in a conversation:
┌ A: "Are you coming tonight?"
│ ┌ B: "Can I bring someone with me?"
│ │ ┌ A: "Boy or girl?"
│ │ │ ┌ B: "What difference does it make?"
│ │ │ └ A: "A question of balance."
│ │ └ B: "A girl."
│ └ A: "That works."
└ B: "Alright, I'll come."
So the git structure is not what is problematic with Trump's speech, it is in fact in line with the findings of conversational linguistics. But there is something beyond the mere inadequacy between the format and the role of president.
The conversational yet monologic nature of his speech, the total lack of transition markers, the constant violation of relevance continuity, the parasitic asides for his ennemies, down to the explicit:
>"You know, I do the weave. You know what the weave is? I'll talk about, like, nine different things, and they all come back brilliantly together."
indicate that the needle that ties everything together is himself. Here's the conduct of topical continuity in his speech, the sole coherence point. Illness or strategy ?
https://www.unige.ch/clf/fichiers/pdf/10-Berthoud_nclf17.pdf
Topic management, enunciative processes, and conversational sequences
[...]
Indeed, these processes do not concern only the speaking subject: the topic is expressed for a listener, and its formulation is oriented toward them. In speaking, the speaker immediately positions a recipient in front of them, who becomes a co-speaker (Culioli 1985) and thus directly participates in this construction. From this perspective, the development of the topic depends on the reciprocity of enunciative perspectives, which manifests and materializes in interaction, where the topic does not belong to a single speaker but is collectively constructed, thereby constituting an activity of co-enunciation.
[...]
The participants are sensitive to the question, “Why this, now, addressed to me?” which notably arises when transitions, introductions, or topic recycling cannot be described and analyzed as such. The first analysts of the topic and conversation are therefore the speakers themselves:
1 Roger: I'm gonna blackmail you
2 Al: fuck you
3 Ken: hhh
4 Roger: better not I become pregnant easy hch // hch heh hhhhh
5 Louise: heh hch hehhhh take birth control pills
6 Roger: hehh heh
7 Al: hey Isaw// saw airai neat//joke
8 Ken: the Utile green pills?
9 Al: I went down to the Ports O'Call Village, not to be changing the subject but she brought it up
10 Roger: Not to be change- "I wouldn't change the subject"
11 Al: but where was a birth con- they had a joke shop with a birth control pill and it was made out of styrofoam. put it between you lee- legs'n press very hard
12 Louise: hch hhhhh
13 Al: heh heh heh
(Sacks, 1992, 1:539)
I can’t learn this way. Great description of some of my instructors. I gravitate towards design and engineering. Goals are first. This sounds like play: “…and this happened, and this happened…” nope. Not me.
>So, what does the neurotic want to know? I'm slowing my delivery here so you can hear clearly, as every single word is important. He wants to know what is real in that which is his passion – in other words, what is real in the effect of the signifier.
And I think this is exactly what Trump is on about. He wants to see for himself the effect of his own name as he utters it (hence the 3rd person). I don't think he came with that remark about "the weave" by accident. I think it was planned and "discussed" internally as the explicit, self-aware display of madness that would allow him to string this nonsense together, and by uttering it, absolve himself from the toxcicity of his own thought process and make it our problem – thus achieving the roaring effects he sought to find in the very name of Trump in the first place.
He really isn't anywhere near as coherent as the average person, and I'm not American, so I don't even have a dog in this fight. I've only heard him talk, and I usually can't make out any point at all, which I personally have very rarely encountered with the average person (for the sober ones, at least).
I understand that speech and conversation branches. That's normal, that's not an issue. What makes it notable is how random the jumps seem to be, the seeming lack of logical connections between them, and the apparent randomness of when they rejoin the main branch, if they ever do.
The examples you show display very predictable stack-like behaviour which is very different to what is shown in the article.
The idea helps when talking with autistic people who might have quite extreme versions of weaving. A conversation between two weavers is kind of an exchange of blunt facts about their world views, which does sound a lot like disparate unrelated tidbits. In contrast to storytellers, who have a discrete path and story to tell, with a start and a finish.
It failed badly. While it's easy to construct sentences that way, it puts immense cognitive load on the receiver. Probably because you don't have any structure in which to put all these terms thrown at you, we reasoned, so we reversed the order to get structure (connectives) first, followed by the terms.
It wasn't better. Turns out, receiving a structure with a bunch of holes to be filled later also results in high cognitive load.
Why? My guess is that the cognitive load mainly comes from the number of unfinished structural connections. To minimize that, you need to transmit a tree in such a way that the terms come as close as possible to the connective. In other words, not bottom or top first, but "side first".
I believe this is why infix notation is so popular. While you parse "A and B" or "X + Y" you never have more than one open connection. When you parse "(+ X Y)" you have two open connections after reading the "+". Five levels deep that begins to matter a lot.
I like the purist lispy idea of operation-first expressions, but I struggle to make my mind actually work like that. If you like clojure-type threading macros, consider that they do something similar to infix notation: they reduce the number of open structural connections during parsing.
I feel it too, it's a higher context language and I agree that it is probably the fact that you are holding onto more unresolved threads at a time. But perhaps that's just because I didn't grow up with it? I would love to find out.
An interesting observation related to this is that on top of the sentence order differences, things are generally spoken about from the largest concept to the smallest which is different to English as well.
So where we would say "I ate lunch at the park today", in Japanese you might say Today, I at the park ate lunch.
In the second sentence it feels like there is a cliffhanger until we get to the end, the smallest details are often the point of a sentence, and so it's like waiting for the punchline. My brain is on hold until we get there, but in English I must admit I can tune out of a sentence early on and usually get the gist anyway.