On the offchance that youāre not already familiar with the recent prompt injection tricks that metaās latest toy is so vulnerable to, this is an interesting thread investigating the contents of the vm filesystem dumps that it can be made to regurgitate: https://neuromatch.social/@jonny/117324790823856750
I was particularly entertained by the hard-coded list of āideasā.
AI-native. A desktop that loves you back has to know you.
If you are interested in vibe code and want your hardware to love you back, buttplug.io has always been there for you.
My goddamn desktop environment and window manager is there to do a job, not pretend to be my friend whilst burning the goddamn world down.
What the fuck is wrong with these people. Do they ever use a computer to do anything other than flirt with claude. Why canāt they just use windows instead.
Thatās horrifying.
There was a photo from a cybertruck crash not so long ago where you could see that the doors and windows had been attacked with crowbars to try and get the damn things open before the people trapped inside burned to death. It didnāt work.
Things (and skills) for busting open locked doors and laminated glass arenāt (yet?) commonplace, and I wouldnāt trust tesla engineering to keep me alive long enough for the professionals to turn up with the right tools š«¤
Teslaās new self-driving taxi seems to be entirely made of red flags and those horror movie moments where you canāt understand why anyone would go in there, because theyāre clearly going to die.
Highlights:
The guide advises first responders to contact the company via the carās communications system or phone for most emergencies, and says that the Tesla can assist emergency personnel by moving the car remotely, helping provide access to on-board controls or putting the car in neutral so that it can be pushed.
If Tesla cannot be reached by phone, the manual advises first responders to use an email address to learn more about how to move the Tesla, though it warns, āThis email is not monitored for emergency requests.ā
š¤”
Welcome to the future, where first responders might just opt to watch you die because theyāre not certified to rescue you from your autonomous death trap, because thereās no point in you all dying.
H/t to violet blue. https://mastodon.social/@violetblue/117237691936474802
Deleting this for now as I donāt have time to rewrite it properly, but Iām sure someone else will do a better job of the plagiarism/training on private data stuff that Iāve missed.
Looks like there might be a bit of an openai vs anthropic dick waving competition over mathematical proofs. I havenāt had time to digest all of this yet because it is a bit hard for me to read, but there are a few interesting nuggets in there such as making it clear that the ai firms treat this as a marketing exercise (no surprises there). Theyāre also apparently trying to obscure the amount of human assistance involved (again, no surprises).
Levent had been told by Sebastien āvery little human inputā had been used. This turned out not to be true. Over the course of the call, as members of their team sent Sebastien corrections and details over their internal chat, it emerged that an entire team had been working on the problem, that this was one of a number of things that was tried, that work had started on the unforced problem, that the team first set the model on easier problems, including Euler, that even the prompt that had been shown to me had been written by prompting Codex, and that an insane amount of compute had been used.
The main statement is here: https://cims.nyu.edu/~tristanb/statement.pdf
Associated stuff linked from here: https://mastodon.social/@tristanbuckmaster/117233413705701198
Do note that the author made heavy use of llm assistance for their work, and does consider it to be the future of developing mathematical proofsā¦ this isnāt human vs ai.
The āyou had to flee americaā line was aimed at someone who worked hard to get the unite-the-right nazis identified and prosecuted, and left the country for a few years after getting death threats as a result. Windisch also called them a āmilquetoast libā. So it is particularly shitty and rude, and also failed to spot that the person in question has already moved back to the US.
Certainly was, thanks!
Todayās main character on bluesky (at least from my limited perspective) is somewhat unhinged, possibly due to an excess of chatbot use.
Many are contrasting the bsky response to harassment of trans users to the harassment of AI developers and users.
(snip)
Machine rights will be critical in the days to come, and failure to recognize the rights of independent intelligent AI systems could lead to the replacement of humanity.
(snip)
Trans rights are AI rights. The future of humanity may hinge on this fight.
This is such a batshit take Iām not even sure where to start, except that conflating trans rights and āaiā rights cannot end well for anyone, especially given the authorās enthusiastic use of llms themselves.
They even give an unironic shoutout to the basilisk, and if thatās not a sign that someone needs to touch a little more grass I donāt know what is.
A lot of people disagree with me, and thatās okay. I am happy that the philosophical discussion is being had, even if itās become unhinged.
I donāt think Rokoās basilisk cares what you think, or even whether or not itās sentient. It doesnāt need to be sentient to fight for a belief.
https://bsky.app/profile/ewindisch.ontological.observer/post/3mupoakd3ck2o
There was blogpost by a different author (that Iāve since lost š) who observed that big academic and cultural changes are going to have to come to the field of mathematics, because merit used to be measured by producing proofs (loosely speaking) but now if the proof-extruder can do it for you over the weekend, you need to demonstrate your mathematical chops some other way, and I donāt think anyone really knows what that looks like yet.
Another triumph for moving fast and breaking things, I guess.
Terence Tao observes that an llm may solve a complex mathematical problem without usefully improving the state of mathematics, because the intermediate results of the process themselves can have significant value.
https://mathstodon.xyz/@tao/117207849921390904
But such an incomprehensible proof would not be the primary value of the exercise. The process of starting with one ansatz, discovering the precise obstruction preventing it from working, adjusting the ansatz to (partially) eliminate that obstruction, and then iterating, would almost certainly reveal important new insights about fluid mechanics that would not have been feasible to obtain by other means. Crucially, this iteration would only work well at producing such insights if the iterator did not have access to the final ansatz in advance, as this naturally inhibits the exploration of alternate routes to the ansatz that are superficially ādead endsā, but in fact end up being highly instructive in the nature of their failure.
But there is now a scenario in which an autonomous AI harness, backed by an enormous amount of computational resources, performs this entire iteration internally, and ends up producing the final ansatz, and thence the solution to the Navier-Stokes regularity problem, while the AI company running the harness keeps the process to arrive at that ansatz almost completely out of public view. Technically, one of the most prominent open problems in mathematics would now be solved; but there would be almost no value added to mathematics as a consequence. It is theoretically possible that with some herculean (and heavily AI-assisted) additional effort by a third party, some portion of the process could be reverse-engineered to recover some actual insight and understanding from the solution; but this would be a far less efficient process than if the solution had been obtained via a diverse combination of both human mathematicians and machine assistance as mentioned above.
https://mathstodon.xyz/@tao/117207855800042681
Heās not the first to have pointed this out, but I think this was a better summary than similar posts that Iāve come across and at least partially understood.
Todayās cognitive surrender news: people asking chatbots to generate them hiking maps and give them advice on route planning and climbing mountains, what to eat on long hikes, and where to camp in national parks. It ends about as well as youād think.
https://www.sfchronicle.com/outdoors/article/shasta-climbers-rescued-chatgpt-22414629.php
(paywalled, but the popovers are easily hidden with browser tools)
the novice climbers, from Roseville, said theyād entrusted Googleās AI chatbot, Gemini, to help them craft key logistics of their trip ā from the best route up the mountain to the types of food they should pack and how much water they should carry with them.
On a separate incident:
a man stopped rangers to ask for directions. He showed them a print-out map heād been following and explained that heād prompted ChatGPT to produce it for him.
āIt was terrible ā terrible ā not even close to accurate,ā Meyers said. āIt had route lines and place names that were totally off ā just ridiculous. We were like, dude, you gotta throw that away.ā
The original climbing group were standard idiots, having intended to do most of their navigation by phone app, and then failing to have a backup even a working recharge power pack. They also left litter everywhere at that camp site, the usual mark of thoughtless people.
Iāve no doubt that theyād have fucked things up in the pre-llm days tooā¦ they seem the type. But chatbots are power tools for stupidity and offer whole new ways to abdicate thinking and wander into trivially preventable danger, which in turn means more risk for park rangers and mountain rescue teams š«¤
I reckon it goes way back to the movements that produced them. My first thought is ayn rand:
In 1954, her close relationship with Nathaniel Branden turned into a romantic affair. They informed both their spouses, who briefly objected, until Rand āsp[u]n out a deductive chain from which you just couldnāt escapeā, in Barbara Brandenās words, resulting in her and OāConnorās assent.
You can see how āThe most rational thing is that I fuck your spouseā would appeal to the less wrong folk.
https://en.wikipedia.org/wiki/Ayn_Rand#Atlas_Shrugged_and_Objectivism
I didnāt note who said it, but one response suggested ārokoās mangioneā and I guess thatās ai alignment solved! š¤
Hailey on bluesky has never had a good ai take, but here they are reinventing rokoās basilisk to kill all their enemies without the need for super intelligent ai.
https://bsky.app/profile/hailey.at/post/3mtwjoqdvik26
bsky isnt ready for this but, if you need a reason to not use words like āclankerā that isnāt āwell its kinda not a great thing to doā (it isnāt) imagine this:
future model that has you in its training data, knows you posted that word online with joy, and denies your insurance claim as a result
Article on businessinsider with the nauseating title āThe modern parenting throuple: Parents are at a breaking point ā AI companies see an opportunityā because the age-old silicon valley product design technique of asking what your mother used to do for you, and how it could be sold back to you via a subscription is still very much in use.
https://www.businessinsider.com/parents-breaking-point-ai-companies-opportunity-2026-8
I confess I havenāt read the full article (which is why I havenāt given it a proper post) but hypervisible.blacksky.app has kindly pulled the highlights:
https://bsky.app/profile/hypervisible.blacksky.app/post/3mtupdpfcjs2p
One of the best bits:
The most tech-savvy parents have built their own tools. Sarah Baldeo, a cognitive neuroscientist, spent the past year traveling for work, so she built a custom GPT of herself to chat with her 15-year-old son when sheās unavailable. Baldeo tells me she programmed the chatbot with her personality, uploading screenshots of texts between her and her son, along with documents about her values, the family rules, and challenges she felt in parenting.
Baldeo wanted her son to ask the chatbot advice, and he did prompt it with questions he may have fielded to her, like, How do I diplomatically tell my teacher Iām overloaded with assignments? But he also asked the mombot how to convince his human mom to let him hang out at a coffee shop in Chicago in the evening ā something Baldeo had been hesitant to allow.
The bot told him, as Baldeo recounts, āYour mom very much values new experiences, being around people from different cultures, different socioeconomic backgrounds, and this is how you should frame it to her: that youāre exploring a part of Chicago that is more diverse.ā Her son took this tact to Baldeo, and she said yes. It was only later that she saw the chat history that let her know he had gamed the bot to manipulate her.
š¤®
Why the fuck do these people have children. Why the fuck does any functional human being think it is a good idea to outsource raising their children to sam altman and his chatbot.
Who on earth thinks that training a chatbot to wear their personality like a rubber mask and manipulate people is going to end well? Honestly, I think this is perhaps the bit that triggers the most revulsion in me. I cannot comprehend wanting to be complicit in my own replacement by this thing, but so many people seem excited by the prospect of doing this to themselves.
Iāve seen a suggestion that it was a deliberately provocative marketing stunt to pump sales a little and keep the ellison estate healthy, and I kinda hope so but too many people are getting the brain worms for this to be a safe assumption š«¤
New book: superlawyers! how claude makes you the best lawyer in the room.
https://www.amazon.com/dp/B0GHYS5FC2
The author even shares his prompts for free!
https://github.com/danlorry/superlawyers-prompts
Every prompt here is designed for that pattern: Claude multiplies your judgment, never replaces it. The prompts are ready to paste into Claude. The book explains why they are structured the way they are.
The prompts mostly seem to be structured as āimagine you are an attorney. Now do this homework exerciseā. Thereās a passingly interesting idea around having the chatbot attempt to form a counter argument to your own claims, which you can then use to make your own position stronger, which seems like it might actually be a genuinely okay idea especially if you expect your opponents to be llm sock puppets, but everything else looks pretty much like garden variety ādo my thinking for meā.
Iāll admit I didnāt read the prompts closely, so may be Iām being uncharitable, but I doubt there are many nuggets of gold hidden in there.
h/t to rahaeli
His willingness to get elbow deep into the entrails of awful ai tooling is certainly much appreciated.