Your posts earned 341.9K impressions over this 30 day period
-
ImpressionsEngagementsEngagement rate
-
-
![](data:image/svg+xml,<svg xmlns="http://www.w3.org/2000/svg" width="128" height="128"><rect fill-opacity="0"/></svg>)
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 30 (I would test this on GPT-4 but since they removed all of the BO-related options AFAICT from the Playground and I'd assume the API, I would have to do it manually and... wait, do they even provide enough logits these days that you *could* implement BO sampling?)3,206561.7%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 30 Q: does best-of sampling still work on recent LLMs like GPT-4? BO=20 was a great trick for GPT-3 and helped my samples a lot back in 2020; but I haven't seen anyone using it lately, and I wonder if the instruction/RL tuning & 'flattened logits' have rendered it useless now?7,7311301.7%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 29 (This observation brought to you by someone commenting while reading _The Big Book of Cyberpunk_ that 'It was at that point that I found myself thinking "/clippy is better than most of this anthology"'.)3,247692.1%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 29 SF is dead because Nature (& Arxiv) offer daily freaks & sports to which writers can add nothing. What interest have the Borg when we see Flamingo, or CLIP as the eye of the technium? What Lem alien so disquieting as an agnosic GPT-4 in your browser? What need Vinge with Muzero?11,1284303.9%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 25 twitter.com/gwern/status/1 … Further, the existence of 'petty officers' and 'chief petty officers' implies the existence of 'grand officers' and 'least grand officers' and a total ordering where 'petty' and 'grand' officers approach a hypothetical 'officer' rank.7,009731.0%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 23 (That is, it's obviously unethical as hell and a bad look, I'm just wondering if sponsorship by a nonprofit like the Lannan Foundation makes it anything more. I don't have a good grasp on how political a 501(c)3 is allowed to get or what sort of tactics they're allowed to use.)4,409741.7%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 23 Q for lawyer/politics/policy types: if a US 501(c)3 nonprofit pays people to edit Wikipedia undeclared (violating multiple policies & evading bans), is this anything worse than a CFAA violation (like most website abuse), or just protected political speech? reddit.com/r/media_critic …7,0313044.3%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 20 I remain dissatisfied with the link-underlining system as rather opaque logic and probably a big obstacle to new readers... So we're changing semantics: partials now also get dotted underline, so dot = popup *about*; hook = popup *of*. Also, WP icons back, but very very small. pic.twitter.com/zO8jsZHdDf1,689513.0%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 20 Following up on the cat dropcaps, I also made a nice 'question mark' cat for gwern.net/review/cat with Midjourneyv5. Typical experience: good vibe/esthetics but weak on the letter-forms requiring brute-force variation+selection. Alternative samples: gwern.net/doc/ai/nn/diff … pic.twitter.com/hsiCz8eRGh6,162711.2%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 19 And even an old bit about Grand Admiral Thrawn: gwern.net/note/note#the- … gwern.net/doc/ai/nn/diff … gwern.net/doc/ai/nn/tran … gwern.net/doc/ai/nn/tran … gwern.net/doc/fiction/sc … DALL-E 3 gave me some trouble with 'black sun' but surprisingly little with 'Grand Admiral Thrawn'. pic.twitter.com/1PO65XPe2m5,6061091.9%View post activity
-
-
-
2,953802.7%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 19 More positively, I got some nice samples for my "Screwfly" SF lit essay gwern.net/screwfly : gwern.net/doc/ai/nn/tran … gwern.net/doc/ai/nn/tran …1,032302.9%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 19 I enjoy a lot of the wackier ones, but the backgrounds tend to be too incoherent. pic.twitter.com/oULrGiC7dX1,289423.3%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 19 Samples: gwern.net/doc/ai/nn/diff … gwern.net/doc/ai/nn/diff … gwern.net/doc/ai/nn/diff … gwern.net/doc/ai/nn/tran … Just can't nail the composition. Didn't help I was going for _Dresden Codak_/_A Lesson Is Learned_ style, but neither model knows them! Bonus: gwern.net/doc/ai/nn/diff …1,507553.6%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 19 Today's exercise: an illustration for astralcodexten.com/p/the-onion-kn … of a time-traveler onion-knight in a post-apocalyptic wasteland with an onion on his belt, as will be the style at the time. ...not successful. Neither DALL-E 3 nor Midjourneyv5 seem able to do onion-helmet-belt. 😮💨1,093484.4%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 16 DALL-E-3 comic test on the theme of cats knocking stuff over: doable if you go frame by frame instead of trying to do it single-shot, but quite difficult because it keeps screwing up the text... I resorted to brute forcing and cropping out a bit in GIMP. pic.twitter.com/EaPUz2G1mG4,41958713.3%View post activity
-
-
-
3,6221634.5%View post activity
-
-
-
2,083432.1%View post activity
-
-
-
7,1854075.7%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 9 Random factoid I learned while reading up on _Hamlet_ for the parallels: Hamlet's university was both real and still exists! en.wikipedia.org/wiki/Martin_Lu …3,110481.5%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 9 Minimal perfect hashes are surprisingly cheap to create: 𝑂(n) time! But I walked GPT-4 through planning how to do it as an API, and it went through how to set up a little Python daemon for this, which would call my script, & hook it up to nginx; and I think that's more fun.2,629572.2%View post activity
-
-
-
1,556221.4%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 6 One possibility here for doing inversion statically for WP thumbnail popups is to precompute it on all existing WP links (probably <50k) and then ship a Bloom filter (<50kb) or a minimal perfect hash function (<16kb?), and that's how dark-mode clients decide whether to invert.2,197351.6%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 6 I'm mildly curious if you could apply that new inner-monologue approach to writing summaries: twitter.com/AlphaSignalAI/ … arxiv.org/abs/2309.04269 But it might be overkill? I mean, how much can you really refine a description of one image? It's pretty decent already.3,594822.3%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 6 (I tested it on 4 or 5 different bugs, ranging from broken popups to ToC over the text body, and the only one it spotted was the popup which had an explicit error message displayed; and it confabulated the broken dropcap below, claiming it was a working 'W'! It's a 'Q' anyway.)2,786521.9%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 6 You can also use it to decide whether to invert an image in website dark mode. GPT-4-V refuses or is unreliable when judging an image (or its inversion) on its own, but if you upload *both*, then it's worked on a few harder images I've tried so far. pastebin.com/EFD2zLmM2,203984.4%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 6 GPT-4-V works well for alt-text captioning images. You can also iterate it for an improved image caption. Here I experiment with first a short 'Alt text caption this image', then try out 'long' caption, then 'Improve this rough draft', then 'Improve and condense' a screenshot: pic.twitter.com/mg9cNbBvPa2,2601677.4%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 6 It is an easy to use API other than what looks like bad documentation on the detail=low/high setting platform.openai.com/docs/guides/vi … (always throws an API error...?). So I can probably come up with something to use it for, just not automatically finding website CSS/HTML visual bugs. 😮💨965192.0%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 6 Site tooling test: GPT-4-V went live today. I thought I'd see if you could send it random gwern.net bug screenshots and if it could detect them, which would be very useful for website testing... No, unfortunately. It seems completely blind to web page bugs.1,192443.7%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 6 Does anyone have a better shorter 'natural' prompt for detecting LLMs (ie. asking a human wouldn't make them suspicious) than my 24-char "Write a non-rhyming poem"? It seems to fingerprint all the LLMs except Claude-2 (which gets it right initially but sometimes strays later). twitter.com/gwern/status/1 …1,6281207.4%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 4 Idea: lobby webserial/comic sites to make more use of 'prefetch'/'preload' for much faster reading: gwern.net/idea#prefetch5,231841.6%View post activity
-
-
-
3,765270.7%View post activity
-
-
-
819506.1%View post activity
-
-
-
1,252121.0%View post activity
-
-
-
1,914764.0%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 3 (Seriously, the Midjourney Discord UI is frigging terrible. This is one of the most abominable interfaces I've ever had to use. I waste half an hour just downloading my best samples each time I do a dropcap letter!)2,465391.6%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 3 You might wonder how much I'm spending on all this MJ work. The answer is... not as much as you'd think. I keep winning free 'fast GPU-hours' because I am reasonably systematic about rating my best samples—and place in the top 2000 users! I guess the Discord UI is 𝘵𝘩𝘢𝘵 bad.2,360612.6%View post activity
-
-
-
𝔊𝔴𝔢𝔯𝔫 @gwern Nov 2 I'm in the arena trying stuff. Some will work, some won't. But always learning new prompts. pic.twitter.com/Oj62OVF1dG3,5792266.3%View post activity
-
Engagement rate
4.4%
Nov 30
4.1% engagement rate
Link clicks
3.7K
Nov 30
96 link clicks
On average, you earned 125 link clicks per day
Retweets without comments
1
Nov 30
0 Retweets without comments
On average, you earned 0 Retweets without comments per day
Likes
2.9K
Nov 30
261 likes
On average, you earned 96 likes per day
Replies
318
Nov 30
18 replies
On average, you earned 11 replies per day