msamwald

Already working on this: https://examine.dev/

"In the examine|AI system, the base AI (e.g. ChatGPT) is continuously supervised and corrected by a supervisor AI. The supervisor can both passively monitor and evaluate the output of the base AI, or can actively query the base AI. This way, users and developers interact with the team of base and supervisor systems. Performance, robustness and truthfulness are enhaced by the automated evaluation, critique and improvement afforded by the supervisor.

Our approach is inspired by the Socratic method, which aims to identify underlying assumptions, contradictions and errors through dialog and radical questioning."

Well they are not useless for academic research.

> Also PLoS one is basically not peer reviewed - they accept every paper after a short review.

That is absolutely not true, PLOS ONE has proper peer-review, their review guidelines just focus on technical soundness and de-emphasize subjective noteworthiness.

(As a personal anecdote, I managed to get one of my papers rejected from PLOS ONE once...)

Which often is terrible advice that mostly serves to limit liability. Most medical doctors will not have a good, evidence-based answer to such questions.

Better reply: Search the internet for authoritative sources (e.g., official guidance of governmental institutions) or medical guidelines.

Being very efficient at mostly extractive summarization and abstaining from abstractive summarization does seem a better bet though, because fewer things can go wrong and it is easier to check the summaries against the full text.

> HTML is a language that is too ambiguous to parse, and that's not a good thing for browsers.

Wouldn't that be XHTML? Which was killed by HTML 5?

On the other hand:

"BioNTech CEO expects vaccine can be fridge-stored for two weeks"

"Speaking at an online media briefing on the purchase of an additional German production site, Chief Executive Ugur Sahin said tests have recently confirmed the genetic compound remains stable at 2 to 8 degrees Celsius for five days but he expects storability at those conditions to be two weeks or longer."

https://www.reuters.com/article/health-coronavirus-biontech-...

The original Nabla article is missing information on how they primed GPT-3 for each use-case, and how much effort they put into finding good ways of priming.

All fancy GPT-3 demos seem to rely on good priming.

The time scheduling problems are probably hard limit of GPT-3 capabilities. The "kill yourself" advice, on the other hand, might have been avoided by better priming.

Meta-comment:

How is it possible that the original submission has been on the front page for 8+ hours, and all discussion is focused on this completely unrelated link?

Have people stopped reading original submission links in favor of comments so much that the discussion is no longer related to the original submission at all?

Well, in table 1 they compare to RoBERTa trained in a vanilla supervised fashion?

> GPT-3 didn't even get SOTA on SuperGLUE.

Of course neither GPT-3 nor the PET paper claim SOTA on SuperGLUE. They used a few-shot learning setup with 32 examples per task The normal SuperGLUE setup has hundreds or thousands of examples per task [1].

> In general, paper with "new variation of cloze pre-training task for this specific task" is a new section of the literature that is rapidly becoming sort of mundane and uninteresting because there are so many papers doing small variations of the same basic idea.

Could you please link to some of the work you are referring to?

[1] Table 1 in https://w4ngatang.github.io/static/papers/superglue.pdf

A shorter reply would be: It would be great to compare PET not only to GPT-3, but also to other models, especially ones geared towards few-shot learning.

Do you know of any other models that should be used for such a comparison, or are there already any relevant results on SuperGLUE that should be mentioned?

I think few-shot learning (or priming) is actually the main selling point of GPT-3 for most practical applications (rather than merely entertaining language generation). So if there is a method that achieves the same goal with a model that is simple enough to be used by normal developers and researchers without OpenAI-scale infrastructure, that does seem buzz-worthy.

I tried using Obsidian, but I could not resolve the uncertainty of what should be a separate note file vs several items inside a note file.

I then tried various 'outliner' applications (Dynalist, Roam, Workflowy). With these kinds of apps, there is hardly any friction between document and content granularity levels; everything can be just one big tree / directed graph.

I finally settled with Dynalist [1] -- it recently introduced backlinks and is the far more mature, sleek and feature-rich option compared to Roam.

[1] https://dynalist.io/

Dynalist has an offline app (which syncs with the server when online). See https://dynalist.io/download . It's the same interface as on the web (so not 'native'), though.

I got very motivated to use Roam, but then settled for Dynalist. It's more polished, cheaper, and they recently added backlinks as well.

Non-American here. I think 9/11 has global significance that goes beyond its impact on the US. It destroyed the narrative that liberal democracy, free markets and technology had removed all obstacles that would keep us from converging towards a peaceful, secular and progressive global society.

I am an 'old millenial' and have absolutely the same impression (and my personal life at that time was also a mixed bag, so it certainly is not personal nostalgia).

The time period of mid 90s until 9/11 should really receive far more attention. It feels like most of the world was on a much more optimistic and promising trajectory.

I think it was also the last time that we saw major, society-wide and significant technological progress happen in a short amount of time in developed countries: the early days of widespread adoption of the web and all of its implications.

At the same time, I sometimes wonder if many of the negative developments of the two decades afterwards might also be ascribed to the spreading adoption of the web and its unintended negative consequences to global sensemaking.

> They said travel bans don't work. This defies basic logic.

> They delayed declaraing the pandemic to be a pandemic.

I have come to think that a major problem with the WHO might be that it is not independent enough from the political interests of major countries.

Banning non-essential travel from/to highly affected countries early would probably have helped to slow down global spread, but political opposition made that impossible.

It is the consensus of medical professionals and officials in those countries that have demonstrated that they can greatly reduce the spread (China, South Korea, Japan).

It is not the consensus in countries that currently still see an explosion in cases.

The message should be that simple, non-N95/FFP3 masks only have minor protective effect for the wearer, but can significantly reduce the spread of droplets by the wearer and protect others around them. Furthermore, wearing masks in populated public places should not be a recommendation, it should be enforced.

There is also reason to believe that enforcing wearing of masks in public spaces INCREASES social distancing, because people are less likely to slip into an illusion of normality in certain places.

I co-authored an evidence-based call to action for promoting simple DIY masks some time ago [1] and it got some good resonance from politicians.

Still, we don't see adoption of DIY masks in most countries so far, and I start to wonder what is holding back officials from promoting it. Maybe the variability in quality of DIY masks made by individuals might be too large, and officials just have too strong of a resilience against such DIY solutions as to ever promote them? Perhaps a standardized design for cotton masks, not made by individuals but local businesses would be more acceptable for officials?

[1] https://link.medium.com/LY7RRNr2X4 "Promoting simple do-it-yourself masks: an urgent intervention for COVID-19 mitigation", Svara et al. 2020

I recently co-authored a scientific commentary on this, ask me anything!

"Promoting simple do-it-yourself masks: an urgent intervention for COVID-19 mitigation" (Svara et al. 2020)

Pre-print available at https://link.medium.com/LY7RRNr2X4

Summary: "We demonstrate that widespread use of masks by the general population could be an effective strategy for slowing down the spread of COVID-19. Since surgical masks might not become available in sufficient numbers quickly enough for general use and sufficient compliance with wearing surgical masks might not be possible everywhere, we argue that simple do-it-yourself designs or commercially available cloth masks could reduce the spread of infection at minimal costs to society."

This also resonates with recent sentiment that telling people that "masks don't work" (with the intention of keeping people from buying masks when they are scarce even for health care workers) can backfire significantly:

https://www.nytimes.com/2020/03/17/opinion/coronavirus-face-...

The major reason why masks are not promoted on a wide scale is because there simply are not enough masks right now. Simple DIY masks are not perfect, but certainly better than the current state of hardly any use of masks in public at all.

Abstract: "We demonstrate that widespread use of masks by the general population could be an effective strategy for slowing down the spread of COVID-19. Since surgical masks might not become available in sufficient numbers quickly enough for general use and sufficient compliance with wearing surgical masks might not be possible everywhere, we argue that simple do-it-yourself designs or commercially available cloth masks could reduce the spread of infection at minimal costs to society."

I think the most rewarding path is becoming a specialist in something very important and broadly applicable.

> Steel Bank Common Lisp (SBCL) is a high performance Common Lisp compiler.

It might be normal behavior for you but this is not normal for everyone.

The rat study you cite here is certainly far from sufficient evidence to make such bold claims or implicated that indoor lighting can promote cancer development in humans! I think you should edit your statement in the top comment further to make it clear that you are not referring to any solid human data here at all.

Indeed this is important to realize: Training such a generic model from scratch does not only reiterate learning, but the entire evolutionary process that led to the emergence of neural circuits actually capable of such learning. That perspective makes many of the current achievements -- error-prone as they might be -- even more impressive!