jprete

Goddamn. I looked them up and there's clearly an ingroup meaning that went over my head. I was hoping for a Meetup competitor that wasn't enshittified on day one.

I'm not in the club or rave scene - practically the opposite - but it astounds me that the FT thought they could draw useful conclusions about an underground scene by analyzing publicly-posted events on a site named Resident Advisor.

This is a good characterization. I'm precision-driven and know what I need to do at any low level. It's the high-level definition that is uncertain. So it doesn't really help to produce a dozen prototypes of an idea and pick one, nor does it help to fill in function definitions.

The Bitter Lesson claimed that the best approach was to go with more and more data to make the model more and more generally capable, rather than adding human-comprehensible structure to the model. But a lot of LLM applications seem to add missing domain structure until the LLM does what is wanted.

My post is limited to pointing out what I think is a flaw in the reasoning. I am not taking a stance on the larger question.

Substack is not too far from the blog format.

> ...they would need to be feeding those snippets in almost real time into a system which forwards them onto advertising partners who then feed that information into targeting networks...

I think this is totally mistaken. An ad seller which also wants to respect privacy keeps this data in-house and does the ad targeting themselves. The advertiser never needs to see personal information for this kind of market to give people ads related to overheard conversations.

I was thinking something very similar. Posting about a problem adds information back to the system, and every company selling model time for money has a vested interest in patching publicly visible holes.

This could even be automated; LLMs can sentiment-analyze social media posts to surface ones that are critical of LLM outputs, then automatically extract features of the post to change things about the running model to improve similar results with no intervention.

Any problem set that can make news is implicitly an LLM benchmark.

It's not that kind of banished-word list.

I don't see any reason to assume they removed it unless they're very explicit about it. Model publishers have an extremely strong vested interest in beating benchmarks and I expect them to teach to the test if they can get away with it.

I think these are the key findings (from the abstract): "...both human and mouse adipose tissues retain cellular transcriptional changes after appreciable weight loss. Furthermore, we find persistent obesity-induced alterations in the epigenome of mouse adipocytes that negatively affect their function and response to metabolic stimuli. Mice carrying this obesogenic memory show accelerated rebound weight gain, and the epigenetic memory can explain future transcriptional deregulation in adipocytes in response to further high-fat diet feeding."

I.e. human and mouse fat cells epigenetically change in response to weight gain*, they don't change back in response to weight loss, and in the mouse case they demonstrated that the mice with the changes gained the weight back faster.

I would (naively) expect this to also be true in humans, which would explain why rebound weight gain is incredibly common.

I personally see this as significant because it's evidence that obese people and formerly obese people are directly going to have more trouble getting thin, at the level of cell biology, than people who were always thin.

*The abstract implies this about human cells but doesn't explicitly say so. I presume it's a known fact about human fat cells.

The main article is paywalled and appears to be a layman's translation. The first reference (visible in the preview) looks like the biologists' version. https://www.nature.com/articles/s41586-024-08165-7

If you mean user-agent-wise, I think real users vary too much to do that.

That could also be a user login, maybe, with per-user rate limits. I expect that bot runners could find a way to break that, but at least it's extra engineering effort on their part, and they may not bother until enough sites force the issue.

I hope this is working out for you; the original article indicates that at least some of these crawlers move to innocuous user agent strings and change IPs if they get blocked or rate-limited.

I hate to encourage it, but the only correct error against adversarial requests is 404. Anything else gives them information that they'll try to use against you.

I think the GP is referring to people who either don't care whether AGI takes over from humanity, or who actively prefer that outcome.

The economy isn't meant to serve us. It's an emergent system that evolves based on a complex incentive structure and its own contingent history.

Autocomplete distracts me enough that it really needs to be close to 100% correct before it's useful. Otherwise it's just wrecking my flow and slowing me down.

Giving them an LLM to help with syntax errors, at this stage of the tech, is deeply unhelpful to their development.

The foundation of a computer science education is a rigorous understanding of what the steps of an algorithm mean. If the students don't develop that, then I don't think they're doing computer science anymore.

It's plausible that, without Musk's less-than-a-third, nobody else would have put in any serious money.

I think it's entirely a legal dodge to pretend that they aren't gutting the non-profit mission.

What exactly does AGI mean in this hypothetical? The term means different things to different people.

TFA?

The problem is that the "we" who are busy trying to make humans irrelevant seem to be completely unconcerned with the effects on the "we" who will be superfluous afterwards.

I think successful groups almost always have a core of devoted people who will show up on time, rain or shine, enough for the event to succeed. If you were alone for hours waiting for other people, that was a really bad sign IMO.

Book clubs and other discussion-focused groups that are open to new attendees often put up a "flag" so the new people can find the group in a public place. It's just not a literal flag, but some other kind of sign.

If the group's activity is easily visible and distinct from what everyone else is doing, it's not necessary.

I think you're both right and wrong. You're right that capitalism has become a paperclip machine, but capitalism also wants AI so it can cheaply and at scale replace the human components of the machine with something that has more work capacity for fewer demands.

AI models have historically found lots of ways to game systems. My favorite example is exploiting bugs in simulator physics to "cheat" at games of computer tag. Another is a model for radiology tasks finding biases in diagnostic results using dates on the images. And of course whenever people discuss a benchmark publicly it leaks the benchmark into the training set, so the benchmark becomes a worse measure.

If it's gaming the system, then it's much less likely to reliably come up with novel proofs or useful new theoretical ideas.