Language Models are Injective and Hence Invertible
Transformer components such as non-linear activations and normalization are inherently non-injective, suggesting that different inputs could map to the same output and prevent exact recovery of the in...
Researchers claim their tool can use generated text to discover the prompt that generated it: โlanguage models operate as maps on the *sequence* space rather than the embedding space โฆ all information about the input sequence is almost surely preserved end-to-end.โ arxiv.org/abs/2510.15511
01.11.2025 14:20 โ ๐ 16 ๐ 5 ๐ฌ 1 ๐ 1
So if the survival of Europe's industrial base wasn't enough of a motivation not to go for a Chinese EV, not having your car weaponized against you sounds like another solid one.
31.10.2025 09:41 โ ๐ 71 ๐ 37 ๐ฌ 5 ๐ 1
Definitely.
Also, a notification if someone youโre following is added to a blocklist *after* youโve subscribed to it + which blocklist it is, so you can choose to either ignore, report the inclusion, or unsubscribe from the blocklist.
01.11.2025 08:50 โ ๐ 1 ๐ 0 ๐ฌ 0 ๐ 0
a way that everyone can see I.e. if youโre thinking of using a given blocklist, but it contains dozens or hundreds of accounts that have been reported by other users + a count of reports, then maybe people would have a trust signal & wouldnโt be so quick to subscribe.
01.11.2025 08:22 โ ๐ 1 ๐ 0 ๐ฌ 1 ๐ 0
wouldnโt solve this unless it was overly broad & sensitive, based on wether youโd ever made a single slightly positive post on a topic.
It seems like blocklists need at least some degree of authority, together with a way to see whoโs been added, and importantly challenge/appeal/report blocks in
01.11.2025 08:22 โ ๐ 1 ๐ 0 ๐ฌ 1 ๐ 0
I donโt know that thereโs a good way to โfixโ blocklists.
At fist glance they seem like the opposite of starter packs BUT blocklists tend to mostly sweep up people who like/dislike a thing, even if they rarely post on it.
Theyโre a form of mass moralising/virtue signalling.
Topic filtering
01.11.2025 08:22 โ ๐ 1 ๐ 0 ๐ฌ 1 ๐ 0
The issue isnโt the people subscribing to blocklists, itโs the people operating the blocklists.
Itโs a degree of power with no clear accountability, process, or transparency - owners are free to add whomever they like a list, for any reason, regardless of the list name & subscribers are oblivious.
01.11.2025 07:55 โ ๐ 5 ๐ 0 ๐ฌ 2 ๐ 0
The Product Picnic
Writing on UX and product management to connect today's discourse with classic insights that the industry forgot.
Another casualty in the eternal war between LinkedIn and people who use LinkedIn: when people reply to their own post with a link (because links in the post are deboosted), the reply is hidden from the default comment view.
Meanwhile on Bluesky I can link to my silly little newsletter all day long.
30.10.2025 15:05 โ ๐ 30 ๐ 5 ๐ฌ 2 ๐ 1
The path from abusing foreigners to abusing the disabled is well-trodden. I I hope this nugget is bright enough to reflect on who else has walked it.
31.10.2025 07:40 โ ๐ 1676 ๐ 496 ๐ฌ 201 ๐ 20
Affinity just made pro editing and design free forever, and I think Adobe should be worried | Stuff
Affinity Photo, Affinity Designer and Affinity Publisher are now a single app that is free for everyone
If youโve been following the Affinity saga, hereโs whatโs happening. No, itโs not getting a sub. No, itโs not getting mashed into Canva. Instead, itโs going entirely free. Blimey. www.stuff.tv/news/affinit...
30.10.2025 18:48 โ ๐ 13 ๐ 6 ๐ฌ 6 ๐ 1
I very much doubt QT producers used Grokipedia to source this, but my guess would be that if, as many suspect, they move in right wing circles/groups, itโs probably the kind of thing they monitor and/or discuss.
30.10.2025 13:50 โ ๐ 0 ๐ 0 ๐ฌ 0 ๐ 0
Grokipedia
Grokipedia is an open source, comprehensive collection of all knowledge.
I hate to give it any coverage, but Grokipediaโs entry for Shaheen is interesting in comparison to Wikipediaโs, where it also describes her as an activist.
The likely relevant thing is the source stating that says she was recently appointed exec dir. of Tax Justice
grokipedia.com/page/Faiza_S...
30.10.2025 13:49 โ ๐ 0 ๐ 0 ๐ฌ 1 ๐ 0
Exactly right. Voters can only hold MPs to account if they understand what they're doing.
Yesterday's ECHR vote was a "ten-minute bill": a way of raising an issue in a short speech with almost zero chance of becoming law.
That even many journalists misunderstood this is bad for trust in politics.
30.10.2025 09:29 โ ๐ 217 ๐ 80 ๐ฌ 6 ๐ 4
So anyone named Al will be thought be AI, e.g. Al Gore?
30.10.2025 06:36 โ ๐ 278 ๐ 24 ๐ฌ 54 ๐ 10
I canโt stress this enough: you cannot understand politics without reading this paragraph about why thereโs no such thing as a 1/3rd pound hamburger
30.10.2025 02:26 โ ๐ 3518 ๐ 1087 ๐ฌ 85 ๐ 73
Generally agree - as long as companies are both upfront about it and open about the risks to consumers of doing it.
Because in this case, itโs almost identical to allowing a company to send a stranger into your house on an hourly/daily basis with just a โtrust me broโ ๐
29.10.2025 22:37 โ ๐ 1 ๐ 0 ๐ฌ 0 ๐ 0
Senators propose banning teens from using AI chatbots
Another bill promises to regulate AI.
Getting nervous the backlash against chatbots will put more wind in the age verification push's sails, because I cannot see a way you'd implement this without requiring age verification on Google, Microsoft, and any other major tech company experimenting with AI
29.10.2025 18:53 โ ๐ 31 ๐ 7 ๐ฌ 2 ๐ 1
Thatโs part of it, for sure, but not the complete picture .
Itโs also part of a long held common desire but loosely coordinated effort to offer alternative right-wing info sources to anything prominent they consider left wing or otherwise biased, to create a network of โtrustedโ citations sources.
29.10.2025 17:16 โ ๐ 11 ๐ 0 ๐ฌ 0 ๐ 0
I Tried the Robot Thatโs Coming to Live With You. Itโs Still Part Human.
1Xโs Neo wants to be your housekeeper. First, it needs to be controlled by a human in your home. Cool with you?
So the $20,000 humanoid robot housekeeper actually just webcams your house to a random person in a call center with an Xbox controller www.wsj.com/tech/persona... @joannastern.bsky.social
29.10.2025 01:49 โ ๐ 1383 ๐ 360 ๐ฌ 115 ๐ 269
However bad you think the police surveillance state has become, itโs worse. Here a tiny police force has blanketed a tiny town in a network of cameras to track every person and vehicle that goes through. Which leads a cop to this womanโs house to lie to her to coax a false confession.
28.10.2025 16:19 โ ๐ 805 ๐ 436 ๐ฌ 31 ๐ 36
CNBC headline: "Elon Muskโs AI-powered Wikipedia competitor goes live after a rocky launch". I've struck through "AI-powered" in red, and replaced it with "Wikipedia-powered"
Grokipedia attribution text: "The content is adapted from Wikipedia, licensed under Creative Commons Attribution-ShareAlike 4.0 License."
fixed that for you
28.10.2025 16:33 โ ๐ 4436 ๐ 1073 ๐ฌ 71 ๐ 51
Grokipedia literally copies entire Wikipedia entries verbatim where thereโs no right wing politics involved ๐คทโโ๏ธ๐
28.10.2025 14:30 โ ๐ 0 ๐ 0 ๐ฌ 0 ๐ 0
Lol at Musk, again, ๐
His plan to โrewrite the entire corpus of human knowledgeโ
Turned out to be โrewrite Wikipedia to be right wing using Grok AIโ
Exceptโฆ it actually turned out to be โrewrite just the bits of Wikipedia he dislikes or disagrees withโ
๐คก
28.10.2025 14:27 โ ๐ 0 ๐ 0 ๐ฌ 1 ๐ 0
A stat we don't talk about enough: Newspapers have lost 77% of their jobs over the last 20 year, more than any of the 532 other industries tracked by the Bureau of Labor Statistics.
27.10.2025 23:54 โ ๐ 2897 ๐ 1030 ๐ฌ 86 ๐ 88
Goodwin et al. misrepresenting C4 study leading their audiences to believe 51% of adverts are non-white (when in reality they featured a non-white person). They claim to want better integration but donโt like seeing it modelled in adverts.
28.10.2025 13:02 โ ๐ 39 ๐ 10 ๐ฌ 0 ๐ 0
Poppy pizzas, clearly not quiches before you pipe up.
Regurgitated from my Twitter account
We will fight them on the quiches
27.10.2025 19:36 โ ๐ 772 ๐ 184 ๐ฌ 43 ๐ 46
This is a sad but correct decision. Thank you, @python.org, for committing to your values when it matters.
I would love a list of organization that have made this kind of value-driven decision to donate to.
27.10.2025 16:49 โ ๐ 61 ๐ 19 ๐ฌ 2 ๐ 0
"Meanwhile, the regulations to bring in a requirement on providers to report CSAM content to the National Crime Agency have been revoked; they were due to come into force on 3 November and, at the time of writing, there are no further details available as to why they have been reversed." > Curious.
27.10.2025 13:10 โ ๐ 11 ๐ 7 ๐ฌ 1 ๐ 2
๐ฃTHREAD: Itโs surprising to me that so many people were surprised to learn that Signal runs partly on AWS (something we can do because we use encryption to make sure no one but youโnot AWS, not Signal, not anyoneโcan access your comms).
Itโs also concerning. 1/
27.10.2025 10:38 โ ๐ 2866 ๐ 1085 ๐ฌ 44 ๐ 185
CS PhD Student @ Northeastern, former ugrad @ UW, UWNLP --
https://millicentli.github.io/
Artist, model, photographer, intellectual (non-practicing).
Senior Editor, Tech & Policy, The Verge
I like utilitarianism, consciousness, AI, EA, space, kindness, liberalism, longtermism, progressive rock, economics, and most people. Substack: http://timfduffy.substack.com
Asst Prof at Cornell Info Sci and Cornell Tech. Responsible AI
https://angelina-wang.github.io/
Senior Technology Reporter at The Intercept
Got a tip? Signal: sambiddle.99 (Make sure you donโt contact me via a work-provided device)
https://sambiddle.com
The nonprofit organization behind the Python programming language. For help with Python code: http://python.org/about/help/
On Mastodon: @ThePSF@fosstodon.org
Mastodon: @nafnlaus@fosstodon.org
Twitter: @enn_nafnlaus
URL: https://softmaxdroptableartists.bandcamp.com/
#Energy #EVs #Ukraine #AI #Horticulture #Research
Principal Engineer/Architect for AI in Word at Microsoft.
Previous: ~14 years working on Windows, with a stint as small startup CTO in the middle. I also made Tweetium (may it rest in peace).
https://brandonpaddock.substack.com
๐ปI write about tech, media, politics, and consumer rights.๐บ๐๐คฆโโ๏ธ
Reporting research on human behavior and cognition. #PsychSciSky #Psychology #SciComm
PhD student at the University of Pennsylvania. Prev, intern at MSR, currently at Meta FAIR. Interested in reliable and replicable reinforcement learning, robotics and knowledge discovery: https://marcelhussing.github.io/
All posts are my own.
Illustrator & cartoonist | Guardian & Prospect (full scripts in alt text) | graphic novels | childrenโs books | WFC | My stuff: https://linktr.ee/stephencollinsart
Decision support, design, 30 years of data, architecture, augmenting human analysis, systems, cybernetics, safety, resilience, GOFAI, data cleaning & prep, engineering management, urban planning, botany, history, art, dada
Language and thought in brains and in machines. Assistant Prof @ Georgia Tech Psychology. Previously a postdoc @ MIT Quest for Intelligence, PhD @ MIT Brain and Cognitive Sciences. She/her
https://www.language-intelligence-thought.net
cybersecurity weather man. scanning the horizons for cloudy cyber. Expert at nothing except computer rubbish. Anti-ransomware since 2015.
Read Vulture Capitalism โฌ๏ธ
https://linktr.ee/Grace.blakeley
Software Development, AI, ML, Web, Blockchain. Opensource and Decentralization fan. Believer in lifelong learning. Speaking German too.
Drawing, painting, digital arts. Hockey fan. ZSC Lions Zรผrich, Nashville Predators. We don't deserve dogs.
we can only see what we think is possible.
Finance/Accounting/Tech. Interested in cognitive and social psychology, neuroscience, and policy.
Jamaican ๐ฏ๐ฒ | White Room Student | True Neutral