Imnimo

In my book a great school would likely have a low-ish acceptance rate. And so they could (even though they may not be happy about it) absorb some amount of declining applications by adjusting their acceptance bar. WKU's acceptance rate is like 95%. They're already taking ~everyone who applies. I question whether school quality truly plays no role at an institution which struggles to find a student they wouldn't admit.

I only know the story from word-of-mouth. My understanding is that the submitted paper wasn't written/presented well, and reviewers had trouble understanding the significance of what was being proposed. But take that story with a big grain of salt.

An interesting bit of trivia about the Burrows-Wheeler transform is that it was rejected when submitted to a conference for publication, and so citations just point to a DEC technical report, rather than a journal or conference article.

>In a world where there’s enough AI capability to process the entire web and rewrite every page to remove something, the cost of “changing history” is much reduced, so we can expect more of it.

I gotta be honest, this scenario is not a concern that impacts my choice of email provider.

I'm just not sure that I would trust that the view/description of the item ChatGPT shows me and the thing I'm actually agreeing to buy are the same thing.

We can look at what he did when he had the opportunity to help reduce it, for one.

It's very hard for me to envision something I would use this for. None of the examples in the post seem like something a real person would do.

Sorry, I realized I didn't quite write what I meant to. I didn't intend to say that LLMs are non-Markovian from a theoretical standpoint. I meant to say that the language generation task is non-Markovian from a theoretical standpoint, because the next word can depend on arbitrarily distant history.

A common distinction is that whatever you take to be your state space should be fixed-sized. So an LLM has a fixed context window, which might be extremely large, but with a long enough conversation, the oldest tokens will fall out of the context window, and your process will be non-Markovian. But with a large context size, most conversations will be small enough that they fit entirely within the context, and then the process is Markovian.

So, people can correctly call LLMs Markovian in practice, and also non-Markovian from a theoretical standpoint.

I think of it as conceptually little bit like the distinction between a formal Turing machine which requires an infinite tape, and a practical computer with a finite amount of memory. Your computer acts as a Turing machine for the real computations you use it for, but there exist some computations that would require more memory than you have. From a theoretical standpoint, your compute is merely a finite state automaton.

I looked at the example for computer science basics for a 7th grader interested in food. Explanations include:

"a list can be used for a recipe"

"a set can be used to list all the unique ingredients you need to buy for a week's meals"

"a map can be used for a cookbook"

"a priority queue can be used to manage orders in a busy restaurant kitchen"

"a food-pairing graph can show which ingredients taste good together"

Maybe I'm over-estimating the taste of 7th graders, but I feel like I would get sick of this really quickly.

My understanding is that the way they do this is have some number of model instances generating solution proposals, and then another model which chooses which candidates to submit.

I haven't been able to find information on how many proposals were generated before a solution was chosen to submit. I'm curious to know whether this is "you can get ICPC gold medal performance with a handful of GPT-5 instances" or "you will drown yourself in API credit debt if you try this".

Still extremely impressive either way.

I notice none of the pieces of advice are "don't keep a hundred thousand dollars in a Coinbase account".

Just be careful you aren't doing the classic, "my linear regression works way better when I independently sort the inputs and targets"!

If you are pre-idea today, does OpenAI believe your startup will still be relevant in the face of the AGI progress they forecast to make in the time it takes you to ship?

I make this mistake every time I see their name.

I won't even consider a role where I'd have to occasionally be on call, let alone 996.

I think SAM and DINO are the two off-the-shelf image models I've gotten the most mileage out of.

It's also interesting to look at what happened in Luanda. During the civil war, a large portion of the population fled to the capital for safety. The city went from half a million in 1970 to ~9 million today. Now downtown Luanda is one of the most expensive places in the world to live, and the city is surrounded by endless stretches of slums that are visible on satellite maps.

They are making progress on building new housing, but it's a very uphill battle.

I was foolishly expecting clarification given the title.

I dunno, "with tools" means different things for different models. It depends on what tools you give it access to. HLE demands a lot of specialized stuff. Like an interpreter for the esoteric programming language Piet for two questions. If you're not standardizing the set of tools, these aren't apples-to-apples numbers.

The way I see it, her job had two parts - reign in Elon, and then run the show. But she couldn't (or wasn't interested in) doing the first part, and so her tenure was a failure. Gwynne Shotwell at SpaceX does a great job at both, by contrast.

I can at least understand "I am going to a different AGI company because I think they are on a better track" but I cannot grap " I am leaving this AGI company to work on some narrow AI application but I still totally believe AGI is right around the corner"

But the students would quickly learn not to try to cheat in this way. It would be a non-issue.

>Generative AI's main impact on higher education has been to cause panic about students cheating, a panic that diverts attention from the already immiserated experience of marketised studenthood. It's also caused increasing alarm about staff cheating, via AI marking and feedback, which again diverts attention from their experience of relentless and ongoing precaritisation.

>The hegemonic narrative calls for universities to embrace these tools as a way to revitalise pedagogy, and because students will need AI skills in the world of work. A major flaw with this story is that the tools don't actually work, or at least not as claimed.

>AI summarisation doesn't summarise; it simulates a summary based on the learned parameters of its model. AI research tools don't research; they shove a lot of searched-up docs into the chatbot context in the hope that will trigger relevancy. For their part, so-called reasoning models ramp up inference costs while confabulating a chain of thought to cover up their glaring limitations.

If AI tools do not actually work, how are students able to cheat with them? It seems like that would be a problem that would solve itself - a student would attempt to use AI to cheat, it would fail to complete the assignment, and the student would get a bad grade.

I just don't get why Scale and/or Alexandr Wang are so important to Meta. Like sure, data is good and all, but does Scale really bring something so unique and valuable to the table? What vision or insight does Wang offer that's worth so much?

>Its results suggested that certain common indoor plants, like pothos, ivy and ficus, may absorb carbon dioxide and release oxygen through photosynthesis, and provide a natural way to remove toxic volatile organic compounds (VOCs) such as benzene, formaldehyde and trichloroethylene from the air.

I don't understand why there's a "may" in front of "absorb carbon dioxide and release oxygen through photosynthesis"

This is an interesting way of squeezing extra feedback from raw text, but I'm a little skeptical that it's the best way to spend training flops. It feels like most "next tokens" are pretty low information (even after filtering for entropy like they do). Does it make sense to spend a bunch of compute on a reasoning trace on them? Maybe if you're harshly data limited, but not compute limited?

So if you're a business that sends sensitive data through ChatGPT via the API and were relying on the representation that API inputs and outputs were not retained, OpenAI will just flip a switch to start retaining your data? Were notifications sent out, or did other companies just have to learn about this from the press?

In the linked Biden press release I see an announcement of tariff changes under section 301 and section 232. From today's article:

>Other tariffs imposed under different powers, like so-called Section 232 and Section 301 levies, are unaffected, and include the tariffs on steel, aluminum and automobiles.

It seems that Biden did not, at least in the instance you link to, do the thing that the courts have ruled against today.

I think the study is saying that if you just read the diagnostic criteria, you tend to over-report symptoms and can be detected as likely faking, and that happens less with the ChatGPT-generated guide (although the effect seems to be of questionable magnitude).

>Our findings demonstrated that the AI-coached simulation group consistently moderated their symptom overreporting and cognitive underperformance compared to the symptom-coached group, as evidenced by group effects in mostly small to medium size (though nonsignificant in underpowered Bonferroni-corrected pairwise comparisons). This effect is also reflected in lower sensitivity rates for detecting individuals in the AI-coached simulation group compared to the symptom-coached group.

Here "symptom-coached group" is the group that was just given a handout of the diagnostic criteria.