nsagent

You left off an important caveat from the next sentence:

> But of course I have been very lucky, there.

You do have to be lucky enough to be relatively secure in life, but nearly everyone in tech can achieve an adequate living. Not everyone will be a top earner and there are of course compromises you have to make, but should actually consider if the trade-offs for potentially earning less yet being more fulfilled at work are ones they are happy with.

In some sense I'm the starving artist in my family. I eschewed money and moving up the tech management hierarchy to pursue my passion. In pursuit of my goals, I've previously been unemployed and in recent years paid 10x less than I was in big tech.

I'm totally happy with that trade-off.

Agree wholeheartedly. I still have a big box shrink wrapped copy of Planescape: Torment on my bookcase. I'd add Arcanum to that list as well.

Those games are the reason I just finished a PhD focused on NLP. I've long been interested in giving players even more freedom to roleplay within the confines of narrative driven games.

The research platform I created based on Disco Elysium [1] has enabled some cool new research I just finished up for my dissertation that I hope to publish soon.

[1]: https://pl.aiwright.dev

That's not an outlandish hypothesis. I'm from a family of medical doctors and entrepreneurs. They are all very conservative, while I'm more of a classical liberal.

I gave up working in the tech industry to pursue a PhD with the hopes of becoming a professor. Considering I research NLP, I could be making top dollar in industry, but have opted for the academia route. My family mostly don't understand and view the PhD as a diversion. Despite securing a postdoc at Cornell, they still ask me to reconsider working in industry so I can become rich and powerful, which they equate with fulfillment in life.

There are frequently better lenses for examining these issues, like culture and socioeconomics. These explain a larger portion of the variance than race in many circumstances. For example, black immigrants from Africa have a very different outlook than American descendants of slavery [1], with black immigrants having higher educational attainment than any other demographic [2] in the US. A similar finding has been observed in the UK [3]. It's true that systemic racism has disadvantaged African Americans even after the Civil Rights movement, but interestingly the income gap is wider for black men than it is for black women [4] indicating there are more factors at play than racism alone since black women face similar amounts of discrimination as black men [5].

That said, I don't discount the impact that both implicit and explicit racism can have. I just think it's important to take a more holistic view rather than falling back on identity as the main causal factor.

[1]: https://www.nytimes.com/2019/11/08/us/slavery-black-immigran...

[2]: https://www.chicagotribune.com/2007/03/18/black-immigrants-c...

[3]: https://www.lambeth.gov.uk/sites/default/files/2021-10/the_a...

[4]: https://www.brookings.edu/articles/the-inheritance-of-black-...

[5]: https://www.apa.org/news/press/releases/2020/07/black-women-...

I similarly stopped listening to NPR around 2018, because they really leaned into news commentary with a very one-sided bias. I could no longer stand the preachiness of their newer programming, and it seemed that the more established folks like Robert Siegel and Steve Inskeep got roped into toeing the line as well. I think the Trump era really broke a lot of news sources I previously relied on.

Looks like a slight modification of FActScore [1], but instead of using Wikipedia as a verification source, they use Google Search. They also claim to include a wider range of topics. That said, FActScore allows you to use whatever knowledge source and topics you want [2].

[1]: https://arxiv.org/abs/2305.14251

[2]: https://github.com/shmsw25/FActScore?tab=readme-ov-file#to-u...

I also recently ditched the streaming services I was paying for (Netflix, Youtube, and Max). I'd love a recommendation for an external Blu-ray drive for ripping media onto a streaming server.

That said, I recently discovered Tubi, which is completely free, has a crazy amount of shows and movies, and ad blockers seem to work fine. Re-watching Andromeda while I finish up my dissertation. It's been great fun.

This paper addresses that issue and allows fully independent training of the experts:

https://arxiv.org/abs/2303.14177

You're right that parameter updates typically require full state, but during my PhD I've explored some possibilities to address this limitation (unfortunately, my ideas didn't pan out in the time I had). That said, there is research that has explored this topic and made some progress, such as this paper:

https://arxiv.org/abs/2303.14177

Good point, they are quite different objectives.

Their approach is one that works for simple directives: "Go to ship" or "Pick up iron ore" which lends itself well to sandbox-like games (which seems to be a major focus looking at Deepmind's tech report). Similar research has been done in Minecraft [1].

These instruction following agents are more an RL achievement than a language understanding achievement. On the other hand, Disco Elysium has over a million words of dialogue, and solving the quests requires an agent to understand and reason about language much more extensively. People have looked at text-based game agents, like Microsoft's TextWorld [2], but these are much smaller in scope and not easily adapted for humans-in-the-loop.

My work bridges that gap, focusing on the language aspect, rather than navigating a 3D world. Again, they are definitely different objectives, but as a sole researcher there's no way I can compete with Deepmind's budget and manpower anyway. Just look at the extensive author list in the tech report. So it doesn't make sense to necessarily focus on outcompeting them in producing a better generalized RL agent (in fact I merely use GPT-4). Instead, I made a publicly available experimentation platform that allows others to be able to build upon this work, which is valuable for the community at large.

At least, that's my take.

[1]: https://sites.google.com/view/steve-1

[2]: https://www.microsoft.com/en-us/research/project/textworld/

Sometimes I wish I had the reach of Google Deepmind. I created a sandbox environment for the text-heavy RPG 'Disco Elysium' [1]. The current research I'm focused on is having an agent use a natural language interface (via text generation) to solve quests in the game.

The project required lots of reverse engineering on my part to make a web-based facsimile of the game such that it's possible to conduct controlled experiments on the language capabilities of current agents.

Hopefully what I've created will be useful for others, because unlike big tech, I've released all my code under the AGPL [2].

[1]: https://pl.aiwright.dev [2]: https://git.sr.ht/~dojoteef/pl.aiwright

Yes, but the majority of the funding goes to the increasingly bloated institutional overhead. NYU takes 61% of research grants [1], while Columbia takes 64.5% [2]. That doesn't include other fees that PIs might pay in addition. These percentages keep going up year-over-year and are even into the 70% range at some institutions.

[1]: https://www.nyu.edu/content/dam/nyu/research/documents/OSP/N... [2]: https://www.finance.columbia.edu/sites/default/files/content...

Tbh, I'm surprised there aren't any federated dating sites. Not sure that will magically fix all the issues with current dating platforms, but removing the profit motive while making hosting sustainable seems like a win.

I think this also partly explains the LLM hype — people can be as confidently incorrect as LLMs, or maybe LLMs are as confidently incorrect as humans since they are trained on text from social media.

The article does state this though:

  According to the American Property Casualty Insurance Association (apcia), a trade association, last year insurers paid out $1.08 in claims for every $1 in premiums they took in.

Can't speak to Pkl, but for Jsonnet I made it possible to fully define then load neural network model architectures directly from Jsonnet config files [1], rather than relying on Python's unsafe pickle module.

Since neural networks often have many repeating features, using a traditional configuration language requires repeating the same structures a lot, whereas using Jsonnet you can use `std.repeat` instead. You can see some examples of this in the readme of my package.

[1]: https://pypi.org/project/cresset/

The problem is just about any foreign particulates at a high enough concentration will cause adverse effects when inhaled [1].

If the machine had a large residue of metal dust from manufacturing, or a grinding in the motor which made it into the airstream, there are severe consequences [2].

A similar issue could if there was some sort of wood dust [3] that made it into the airstream (e.g if they had a wood trim for aesthetics, and it was shaved down by the motor and made it into the air stream). If instead just the dampening of the wood caused mold spores to grow that made it into the air stream [4], that could also cause major issues.

So yes, in this case it's foam, but most foreign particulates are going to be a problem.

[1]: https://www.ccohs.ca/oshanswers/chemicals/lungs_dust.html#se...

[2]: https://lunginstitute.com/can-breathing-metal-dust-hurt-you/

[3]: https://www.osha.gov/wood-dust

[4]: https://www.mayoclinic.org/diseases-conditions/aspergillosis...

It's not just cartels, but also self-citations. For example, the original preprint of this paper [1] has 5 self-citations, with the version accepted to EMNLP increasing that to 10. This kind of behavior is sadly pretty common now and when you mention it to conference organizers they just shrug.

[1]: https://arxiv.org/abs/2304.02210

My mom worked as an ER physician in Pennington Gap, VA during the early years of opioid abuse there (late nineties through through the early oughts) [1]. She would constantly complain about drug seekers there and refused to give out pain meds. According to her, she was the only doctor who would refuse to give out pain meds.

Fast forward to 2022, she needed heart surgery and was in the ICU for three weeks afterward. She never asked for pain meds and regularly refused them when asked. I think she had became so accustomed to denying people pain meds, that she even denied herself, though it was clear she was in pain.

[1]: https://www.nytimes.com/2019/08/18/health/opioids-purdue-pen...

Not sure how mrmincent did it, but when I searched DDG, it didn't come up with anything. I then switched to Brave Search and it was in the middle of the first page, just past the fold. I specifically searched for:

  john carmack time double

I posted a related paper a couple of days ago [1] titled "Promotion of Scientific Publications on ArXiv and X Is on the Rise and Impacts Citations"

[1]: https://news.ycombinator.com/item?id=39124309

I'm pretty sure that this is intentional. The site's called Daft Social for a reason. There isn't even a link back to the homepage from a given user page and there's no way to find users unless they post their profile link somewhere. Maybe that philosophy will change, but personally I really like just how brutally minimalist the site is. The site was created by Pascal Deville, who's also behind https://brutalistwebsites.com/

I don't think they take security any less seriously than most social networks. If you're worried the sending/receiving servers have access to the email address, then it seems you would have to be equally concerned account recovery via email considering most email is not encrypted.

  we find that in the first 5 years from an initial publication peer-reviewed conference papers submitted early to ArXiv gain on average 21.1+/-17.4 more citations, revised on ArXiv gain 18.4 +/- 17.6 more citations, and promoted on X gain 44.4 +/- 8 more citations. Our results show that promoting one's work on ArXiv or X has a large impact on the number of citations, as well as the number of influential citations computed by Semantic Scholar, and thereby on the career of researchers.

Really makes me rethink deleting my Twitter account. It's clear network effects exist, but an average treatment effect of 44 more citations is pretty huge.

No joke, I was just reviewing a paper for NAACL a couple of weeks ago and the authors made up a brand new task, then stated that they had achieved the state-of-the-art on it. In my review I mentioned that was a meaningless statement for a new task, but I guess some researchers are incentivized to claim that in their papers.

Academics with surprising results are rewarded with more fame, prestige, funding opportunities, etc. It's not a big surprise that people naturally cheat, in big or small ways.

The general public often think that academia is merit-based with the smartest being rewarded, but as you know it's a more complicated picture than that. You're not alone in your thinking; I'm pretty sure everyone in academia recognizes the problems. It's just that enough people benefit from the current incentive structures that the occasional scandal isn't enough for academics to reassess the predominant paradigm.

That's what I ended up doing recently [1]. I'm doubtful I would even know if someone did end up breaking the license agreement though.

[1]: https://pl.aiwright.dev/about/license/

I agree, it's not a radio protocol and I wasn't trying to imply that it was. The person I was responding to posited it might be one and I pointed them to the protocol that's used without explicitly correcting them on that point.

It looks like they implement something similar to Bonjour [1] using multicast DNS [2]. Not sure if it's basically a reimplementation like Avahi [3] or if it's a slightly different protocol (only skimmed the code [4] briefly).

[1]: https://en.m.wikipedia.org/wiki/Bonjour_(software)

[2]: https://en.m.wikipedia.org/wiki/Multicast_DNS

[3]: https://en.m.wikipedia.org/wiki/Avahi_(software)

[4]: https://github.com/localsend/localsend/tree/main/app/lib/pro...

It uses Bonjour [1], which has an LGPL reimplementation named Avahi [2]

[1]: https://en.m.wikipedia.org/wiki/Bonjour_(software)

[2]: https://en.m.wikipedia.org/wiki/Avahi_(software)