hatthew

I first noticed this a year ago when I read a gmail AI summary, then glanced at the main text and saw it was AI generated. I feel like there's good fodder in there for a dystopian sci-fi story about a future where nobody communicates directly with one another, it's all AIs translating, but the AIs slowly start to drift.

My understanding is that to unambiguously quantify information, you need to have a known model within which that information fits. In this context, talking about LLMs adding value (or not) via inserting new semantic information, we're assuming that public knowledge on the internet is not new semantic information, and the quantity of information of contained in text talking about public knowledge is equal to the ~32 bits needed to point to that knowledge.

But then the LLMs aren't adding new information, they're just reading your code and translating that information from e.g. python to english. Rather than sending someone an LLM-generated doc to someone, send the code and your own personal thoughts on the code. If your reader doesn't want to read and understand your code, they can ask an LLM to analyze it, within the context of their specific use case and your personal thoughts if any.

LLMs are wonderful at adding noise and okay at removing noise. My point is that they're not very useful at adding signal. If you wrote 10000 bits of noise and 1000 bits of signal, I would rather receive those 11000 bits from you, and if necessary ask an LLM to remove noise based on what I consider noise. If you can point out to the LLM what it should consider signal vs noise regardless of context, it should have been easy to not write that noise in the first place.

That's kind of my point. If the 1000 bits are well known, then their inclusion isn't new semantic information. By pasting an LLM's output, you're deciding for your reader that they don't already know that information, and deciding that your LLM prompt is better than whatever they would to to obtain that information if they lack it. IMO, it's much better to give your readers the 300-bit key, and let them decide for themselves if they want/need to get more information, and if so, how.

Sure, but then that's less "making an LLM write for you" and more "making an LLM research for you" which I don't think is what TFA is talking about

Sure, but that seems orthogonal to what I'm saying, no? We could continue clarifying the definitions of information in this context, but my fundamental argument is LLMs don't add information to my writing that couldn't be added by the reader themself.

This of course has the potential to change with personal LLMs that can have shared private context with me. However, that isn't a defense for sending people AI slop, it just turns it from "LLMs don't add value" to "LLMs may add value when used judiciously."

That was my assumption. The alternative is that the author is a blatant hypocrite, which seems less likely given their blog's history.

As I have been saying for years:

Writing is fundamentally the transfer of information from your brain to my brain. If you have 1000 bits of semantic information you want to transfer, you can't give 300 bits of semantic information to an LLM and have it fill in the remaining 700, because it doesn't know what those 700 bits are. If it's able to guess those 700 bits correctly, then they aren't true semantic information, and you really only have 300 bits you want to transfer. You might as well transfer those bits to me directly, rather than having the LLM add on an extra superfluous 700 bits that I then have to filter out.

Update: The name tilcayo has been associated with oncilla by a bolivian government survey of fauna: https://repositorio.uasb.edu.bo/items/0e30ad14-971b-4d99-9aa...

Yeah that's what I'm curious about. "Known to locals" is very vague, and there are a lot of situations that could fit that description. The situation you mention is probably the most reasonable concrete interpretation, but my guess is that there was a game of telephone from locals -> scientists -> journalists -> wiki, and any nuance we could read into has been lost.

> And they said, ‘I don’t know! My grandpa called it tilcayo, so I call it tilcayo.

I'm curious if the locals knew it was distinct from other similar species, e.g. Oncilla/L. tigrinus, or if tilcayo is the locals' name for a broader set of 2+ visually similar species?

Yes.

Note: Generally "dark" (as opposed to "bright") means emphasis on the lower frequencies, which is a good description for brown noise, but they aren't formally the same. Formally, brown noise refers to a power distribution that drops ~6 dB per octave, and I have no idea how strictly apple's "dark noise" matches the technical specifications of that.

I'm not familiar with the case, but to me it sounds like the bicycle manufacturer is encouraging people to use bikes in the drive through, maybe even providing specific tools explicitly for that purpose?

Ok?

Maybe the subset of japanese content I'm exposed to is biased, but personally I encounter doubled vowels (obaasan) more often than macron vowels (obāsan). The latter is explicitly "modified hepburn romanization". Even if the latter is more common in general, it's worth noting that the former is the original. The article seems to say that the former is weird and nonstandard, which afaik is completely false.

Also, the particles は を へ aren't always romanized the way they're pronounced; I can think of several places I've seen を written as wo despite being used as a particle.

My understanding of the result that was found is that the blowup doesn't happen in the real world, and only happens in an NS simulation. The bottom line is that NS is insufficient to model the real world, because in this case the real world is more stable than the model. [Take this with a grain of salt, I barely knew of NS before a couple days ago]

Huh? Nobody's talking about that because it's old news. We already talked about it the first few times that AI made notable progress on a difficult math problem. Now, most people who care about the intersection of AI and math just assume that Lean was involved.

I feel like this whole thing hinges on one point. OAI says[0]:

> However, our proofs differ significantly and even the precise results proved are different in the Euler case (forced vs unforced).

How true is this?

If the implied statement is true (i.e. OAI couldn't have stolen results because the results were so different anyways), then I feel like it's pretty clear that OAI solved the problem on their own, and offering some amount of credit to Tristan and Levent is generous. Though on the other hand it's dirty to have even attempted a scoop in the first place.

If the implied statement is false, and Tristan/Levent's results are a substantial portion of the solution to the millennium problem, then it's probably unintentional but clear plagiarism. I think it's plausible to assume that OAI has trained their models on Tristan's codex conversations, and so regardless of legal ownership, the academic ownership definitely includes Tristan and Levent.

The rest of the drama (individual statements and wordings, e.g. by Sebastien) seems like a bit of a red herring. Worth noting, but not worth basing conclusive judgements on regarding academic misconduct. Regardless, OAI does not seem like the good guys.

[0]: https://openai.com/index/navier-stokes-solution/

I like the part where the value is actually still 8

Investments need a return on investment. If "the thing you wanted to exist now does" is the only return, that's not enough to support anywhere near the current amount of digital product creation happening today.

I find json annoying to write a lot of the time, but I do enjoy that it's a lot more difficult to get wrong than yaml. The json spec is pretty clear, and can be learned and memorized by any programmer in 5 minutes. I'd much rather write in a format that's slightly annoying than a format that's slightly ambiguous.

The "File" menu is about CRUD operations on file-level objects.

I think eventually (~decades) the concept of the menu bar will go away, but until it does, File is a fine name for the first menu on the bar. I don't know what it will get replaced with, but it will go the same way as the Save icon—a relic that was a good design at some point, but no longer has much basis other than inertia, and will slowly get forgotten as it becomes obsolete.

If you're like me and struggled to parse the title, my understanding is, "To obtain certain source code from Google, you could previously reference git tags, but now you have to fill out a form and wait for a human to give you a google drive link."

This thought occurred to me too, but then I realized even 37% is very high. In a reasonable society, most individuals' earnings should go up all the time. The downward pressure that should exist is high earners retiring and low earners just starting their career. A mildly idealized society should probably have 3% go from unemployed to employed, 3% go from employed to (voluntarily) unemployed, and the remaining 94% increase their earnings.

My experience is that current models are pretty good at making functional changes without too many more bugs than a human would make, but are still bad at making good high quality changes. I have a degree in software engineering specifically, so perhaps I am overly sensitive to design issues.

I feel like those examples are considered difficult because they're niche topics, but aren't actually all that difficult in a general sense. What I consider truly difficult are things like taking a ticket and implementing it in a preexisting codebase, using a clean and reasonable design that fits the existing style and makes sense to a human, and avoids the footguns I learned by working with the codebase for over a day.

A little bit of hand waving, but I don't mind it much. A cloud of steam rising from a boiling pot doesn't have a well defined border or consistent shape, but a human looking at a nearby object can usually make a clear and reasonable decision about whether or not that object is "inside" the steam cloud.

I have a hypothesis that that majority of ad clicks are misclicks where someone closes the new page within <1 second.

One of the ways I think about this is to treat AI as if it were another person. If I want to give information to Bob but I think Alice might do a better job of writing something, I'm not going to ask Alice and then copy-paste her response to Bob. I'd either tell Bob to ask Alice, or loop Alice into the conversation. Me being the middleman between Alice and Bob can cause several issues including delayed communication, miscommunication, and obscured provenance.

Alice being AI or a real person doesn't change those factors much.