HNHacker News
TopNewBestAskShowJobs

andreyk

5,023 karma · joined December 28, 2015

www.andreykurenkov.com
submissionscomments
andreyk··on How Twitter moderated the Covid debate
OK, so here's the key bit:

"Twitter executives did not fully capitulate to the Biden team’s wishes. An extensive review of internal communications at the company revealed that employees often debated moderation cases in great detail, and with more care for free speech than was shown by the government.

...

But Twitter did suppress views—and not just those of journalists like Berenson. Many medical and public health professionals who expressed perspectives or even cited findings from accredited academic journals that conflicted with official positions were also targeted. ... There were three serious problems with Twitter’s process [bots, contractors, and biased decision making by top level execs]".

This seems to be more a story on the complexities of content moderation than anything. There WAS a LOT of misinformation about covid, and a lot of it could cause real harm. At the scale Twitter operates at, false positives (incorrectly labeling things as misinformation) are impossible to avoid. How do you moderate at this scale without bots and contractors? Well, you can moderate less. But then, should Twitter have opted for more false negatives than false positives? I don't think so.

That is not to say this is article does not share useful information, but the headline is rather misleading. As before with the Twitter files, it shows that the government leaned on Twitter to do what it wanted but not able to mandate it, and that Twitter had internal debates regarding how to moderate these sorts of things. That seems reasonable?

andreyk··on AI chatbots are not a replacement for search engines
But that assumes search just involves document searching for retrieving information.

What about, I don't know, searching for a website with certain functionality/services?

What people really mean when they say LLMs will replace search is that it'll replace Quora/stack overflow/etc, which is clearly not the same thing!

andreyk··on “Nobody cares about your blog”
You are definitely allowed, it's just generally less common than citing other academic papers.
andreyk··on “Nobody cares about your blog”
Can be cited in various ways. See here: https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&q=a+br...

Google scholar tracks citations by itself in some mysterious way, I have not contributed to that at all.

andreyk··on “Nobody cares about your blog”
FYI, the actual title is "Nobody cares about your blog!" - pretty big difference from what is shown on HN

"The title of my blog today is a quotation from an email I received yesterday from Noemi De Santis of RECEPTIO, the Research Centre for European Philological Tradition."

It's all about how an academic plagiarized from the blog of this person, and then the academic's secretary completely dismissed it with statements such as - "I regret to inform you that blogs are not scientific texts, published by academic publishers, so their value is nil!"

Ironically, a long blog post I wrote in 2015 has over 50 citations (more than most of my papers as an academic, amusingly) according to Google Scholar, so clearly others don't feel the same.

andreyk··on Mastodon's federation model encourages specific instances with peculiar rules
Agreed - I run an AI-focused instance, but our FAQ explicitly states it's fine to post about non AI stuff. It's more about making a community that shares the interest rather than forcing all conversations to focus on that one interest.
andreyk··on I trained ChatGPT on my childhood journal entries to talk to my inner child
As others said, to be clear, no optimized-based training was involved ("To do this, I used OpenAI GPT-3 Playground. There's a maximum limit of how much content you can put in, so I selected snippets from around 40 entries from the ages of 7 to 18, but I focused on the ages between 10 and 14."). IE journal text bits were just put in the prompt as examples and then the author presumably told GPT-3 to talk to her as if she were her younger self.

To me it seems like entering a bunch of small snippets from a diary would result in a pretty generic impersonation of talking to your younger self, and that's kind of what the quoted bits read like. It's kind of disappointing, would be cool to see actually fine-tuned GPT-3 do this.

andreyk··on Enhance Speech from Adobe – Free AI filter for cleaning up spoken audio
I use Adobe Audition for podcast editing and they have a pretty nice feature built in for that already - not sure how this differs. I am also not a big fan of edit-by-transcription, since I do like to remove occasional long silences, weird sounds, etc. That being said the mic check looks pretty useful.
andreyk··on Twitter suspends pg's account [fixed]
"Should I step down as head of Twitter? I will abide by the results of this poll."

Wow. 50.9 opposed, 49.1 pro.

andreyk··on Twitter suspends pg's account [fixed]
I use Twitter, and most users are totally baffled, even more than they have been before with prior bans. Musk just keeps finding new ways to alienate users. It was one thing to ban links to Mastodon, but to Instagram and Facebook and link tree? That's going to piss off a lot of people...
andreyk··on Paul Graham is leaving Twitter for now
Related to this - what's the deal with the HN new pages having a bunch of "[dupe] Twitter bans promotion of other social networks"? Shouldn't posting the same link add to its upvote count?

Paul Graham is specifically leaving because of this, feels like a pretty major topic of interest to HN...

andreyk··on Twitter has banned Mastodon links in name and bio for being “malware”
As others have said, it is not allowed to post links with tweets to any Mastodon instance (see https://sigmoid.social/@thegradient/109521197175376332 for a screenshot)
andreyk··on Tesla FSD data is getting worse, according to beta tester self-reports
I've been testing FSD, and I WISH their system did more to limit its use in bad conditions. The perception (based on dashboard visualization) is much worse during rain, yes Tesla lets you keep using FSD even in heavy rain.
andreyk··on Perhaps it is a bad thing that the leading AI companies cannot control their AIs
I see how you made this connection, but it's not entirely fair. Guns are a tool with one use - hurting people (in self defense or not, that's the outcome). On the other hand, stuff like ChatGPT has many uses. So I don't think it's a great analogy. Something like 3D printers are closer - you can use them to make guns, but also lots of other stuff.
andreyk··on Perhaps it is a bad thing that the leading AI companies cannot control their AIs
AI like ChatGPT is a tool. Companies can impose some constraints on how the tool can be used, but ultimately they cannot control its users, and thus cannot control the tool. Yes, the tool has some undesired behaviors in response to some uses of it. But it's ultimately up to the user to know that and avoid negative consequences, and it's up to the company to make that as easy as possible.
andreyk··on The GPT Architecture, on a Napkin
I work in this field (PhD candidate), and what you say is true for smaller models, but not GPT-3 scale models. Training large scale models involved a lot more, as the OP said. It's not just learning rate schedulers, it's a whole bunch of stuff.

See this logbook from training the GPT-3 sized OPT model - https://github.com/facebookresearch/metaseq/blob/main/projec...

andreyk··on Memories: First exposure to computers
Feels like I am of an age (29) that is kind of a cut off point to first exposure to computers being special (my exposure being in 1998). Granted for many by then computers at home were not a big deal in the US, but growing up in Ukraine I still vaguely remember my first interactions with a computer when my dad brought it home from the bank where he worked as a programmer -- we did not get personal computers til later. I don't have precise memories, but remember the general feeling of it being this cool piece of tech and being able to use it as being special, and Starcraft being so awesome. Exposure to video game consoles was also a big deal even beyond this age; it was very rare for people in those parts to have them.

Of course, in other parts of the world it is probably still be a special experience some people don't get to have as a normal part of growing up.

andreyk··on The political orientation of the ChatGPT AI system
Not very meaningful - given different prompts it could easily answer from a different political perspective. Still, could be informative as to the composition of the training dataset...
andreyk··on The Twitter Files
They did not ask to handle it, the person from Twitter replied "handled" (to another person at Twitter, for the record). It's in the screenshot of the email with the request.
andreyk··on The Twitter Files
What was asked was "More to review from the Biden team". I suppose having some sort of access to request reviews of specific tweets is concerning, but this alone does not demonstrate the degree of influence they had over Twitter's moderation decisions to me. But, it's telling that more damning emails weren't shared...

So ok, there was some new information (from which it's hard to confidently extrapolate much, I would say). But mostly it's a rehash of a known story, right?

andreyk··on The Twitter Files
It does not seem like this really exposed any new information - Twitter publicly justified its actions as being about its hacked data policy back when this all happened (https://www.nationalreview.com/news/twitter-cites-hacked-mat...). All this shows is some internal disagreement as to whether that was the right call. By the way, Facebook also suppressed the news story citing its misinformation policy (https://www.politifact.com/article/2020/oct/15/look-behind-n...).

It was already widely criticized at the time(https://variety.com/2022/digital/news/twitter-files-blocked-...): " Twitter subsequently reversed the decision, saying that it had updated its hacked-materials policy and would not retroactively apply it to the New York Post. Other news outlets, including the New York Times, have since reported that the laptop did in fact belong to Hunter Biden and the documents on it were authentic. Predictably, Twitter’s blocking of the Post became a rallying point for Republican politicians accusing the social network of censoring conservative viewpoints. "

So... doesn't seem like a big deal, this is just confirming what was already known?

andreyk··on Comparing Google and ChatGPT
I seriously don't get this argument. Google can implement this themselves! It's not like they can't train a large language model akin to GPT-3 (they already have) or deploy it. And as others pointed out, language models are seriously not reliable right now in terms of producing true information.
andreyk··on Twitter 2.0: Our continued commitment to the public conversation
Oh I agree, this post won't change anything at this point. But I guess that's why it was released (and as others noted, also for regulators / to save face after weeks of chaos).
andreyk··on Twitter 2.0: Our continued commitment to the public conversation
TLDR: this post is entirely to calm advertisers, it can be boiled down to "don't worry, we still prioritize brand safety like before".

It does have this gem in it:

"What has changed, however, is our approach to experimentation. As you’ve seen over the past several weeks, Twitter is embracing public testing. We believe that this open and transparent approach to innovation is healthy, as it enables us to move faster and gather user feedback in real-time. We believe that a service of this importance will benefit from feedback at scale, and that there is value in being open about our experiments and what we are learning. We do all of this work with one goal in mind: to improve Twitter for our customers, partners, and the people who use it across the world."

What a weird thing to say... A/B tests are a thing, does anyone buy that experimenting with new things by rolling out new features to all users at once is a good strategy?

andreyk··on Artist feeds childhood diary into GPT-3 to have a chat with herself
For some reason this article does not have the actual chat (just a short excerpt)... You can see more of it here https://twitter.com/michellehuang42/status/15970060064036208... and here https://twitter.com/michellehuang42/status/15970060108202106...

It's fun, but not so different from a conversation you could have with a psychic who is good at cold reading. It'd be more interesting to see what would have happened if GPT-3 was actually fine-tuned with all the journal entries, I think.

I do agree a lot with this conclusion: "This is the stuff I think that has the most interesting ramifications: more broadly, more immersive human / computer interface loops, from conversation with virtual therapists to in-game interactions for virtual worlds, given there is user input, AI could be used to train highly customizable responses or generate unique storylines per use."

andreyk··on Intentionally making close friends
I think the "this is a skill" just refers to the various things needed to form and maintain friendships (being good at conversations, taking social initiative, following up, etc.) More of a set of skills. And to be fair, my interpretation is that it is a set of skills both people in the process of friendship possess and need to practice to make things work (it is possible after all to be a bad friend or a bad partner due to your own lack of effort or skill with eg communication or vulnerability).
andreyk··on Intentionally making close friends
Very brief summary: a personal anecdote about forming good friends followed up by advice for how to have better conversions (vulnerability, curiosity) and how to meet more people who will become friends (meet people but filter out most of them, take initiative with following up), and lastly how to deepen friendships.

As someone who has a fair number of close friends, I think this is a good post with lots of solid advice. And it's written well! Personally I've been lucky to just naturally meet and form my best friends over time, but I definitely think taking initiative and following up are important to maintaining a close friendship.

A general point in this post I also like is that making and keeping close friends may require work and energy, rather than being something that life just throws your way. I intentionally try to periodically message friends, come out to places they live, just generally keep in touch; as one gets older people move about and it's harder to maintain your closest friendships, but it is possible!

Last comment: this is a very pragmatic and analytical post with a lot of discussions of things you can do, and not much discussion of how you should feel. I'd add this - just CARE. Appreciate the people in your life and let that appreciation guide you.

andreyk··on CICERO: An AI agent that negotiates, persuades, and cooperates with people
ah, good catch; I was going off of the blog post
andreyk··on CICERO: An AI agent that negotiates, persuades, and cooperates with people
This is super cool, but it's worth pointing that it still relies on supervised learning from a large dataset of human players, and that it's not a general learning algorithm (there is a whole bunch of problem-specific aspects to the model). It not having learning via RL (trial and error, as humans do) or self-play is kind of disappointing.

Don't get me wrong, it's an impressive advance, but just as with AlphaGo it's important not to overgeneralize what this means. I would not be surprised if a lot of people jump to talking about what this means for AGI, but with this learning paradigm it's still pretty limited in applicability.

andreyk··on CICERO: An AI agent that negotiates, persuades, and cooperates with people
Agreed. This is using an AI to play a game, IRB seems like overkill. I guess the only potential problem is if it went off the rails and spouting toxic language, but that presumably was not a real possibility.
← PreviousPage 3 of 21Next →