Reposted by Giovanni Campagnapost malone ergo propter malone @proptermalone.bsky.social · 21hwhat i think is currently up in the AI space since we're having cost arguments again: selling inference is very profitable, but models are improving so fast that there's no time for capex on model training to amortize, so net at this rate they're losing money which is why they want a cartel 1533135
Reposted by Giovanni CampagnaMatt Keeter @mattkeeter.com · 30/09/2026Typical OpenAI agent during its training run: 126816
Giovanni Campagna @gcampax.com · 30/09/2026SI is Slovenia??? There are too many S countries, I'm telling you. (SL = Sierra Leone, SO = Somalia, SV = El Salvador, SE = Sweden, SN = Senegal, SA = Saudi Arabia) 000
Giovanni Campagna @gcampax.com · 30/09/2026The often overlooked issue is not the jobs that people don't want to do. Of course there are weirdos who like to clean sewers. But work is work. No matter what, there will be dreadful or stressful parts. "Grad student" is as intellectual, desirable, and fulfilling as it gets, yet full of burnout. 061
Giovanni Campagna @gcampax.com · 25/09/2026I mean it doesn't seem that bad? It's almost certainly a user specific microvm, yeah you can run whatever you want on it, thats kind of the point tho 110
Giovanni Campagna @gcampax.com · 25/09/2026"The LM built on classifiers has no content classifier" is very funny 0210
Reposted by Giovanni CampagnaAaron Grote @aarongrote.com · 25/09/2026One concept I tried out in a meeting this week was Determinations vs Decisions, and that our automations should focus on removing the non-expert brain work of making Determinations (there's a right answer) so our people can spend more time making more and better Decisions (no obvious right answer). 1728
Reposted by Giovanni CampagnaTeefa @tifaret.bsky.social · 24/09/2026Introducing ONTOPAUSE™: An AI solution that automatically generates, resolves, and hides important ontological clarifications about the language you’re using, so everyone can reaffirm exactly what they don’t mean before going back to talking normally. Start not-stopping today! 19817
Reposted by Giovanni Campagnaconputer dipshit @davidcrespo.bsky.social · 25/09/2026in which Ben makes the brave and true point that any analysis that assumes all CEOs want the same things is dumb as hell 2433
Giovanni Campagna @gcampax.com · 24/09/2026Humans are bad at logic. Computers are good at logic. Decades of AI research have been spent at making computers worse at logic. This is the most important technological development of the century. 010
Giovanni Campagna @gcampax.com · 24/09/2026It feels wrong that software engineering was solved before IKEA furniture assembly, and yet it's so, so relatable 4372
Reposted by Giovanni CampagnaWill Stancil @whstancil.bsky.social · 24/09/2026As an exercise showing why this is bad, I would challenge you to come up with a succinct, understandable description of the Hugging Face attack that follows these rules. 321037
Reposted by Giovanni Campagnamax power 🌌 @sneptech.bsky.social · 23/09/2026finally, real benchmarks www.assbench.comassbench.comLLM AssBenchLLM Benchmark - One prompt, Multiple models, Multiple dates 1010914
Reposted by Giovanni CampagnaEpoch AI @epochai.bsky.social · 22/09/2026AI is getting cheaper more quickly than any other transformative tech in history. At a given level of performance, cost has fallen ~47%/quarter since 2023. That’s 4× faster than DNA sequencing, 6× faster than compute, 18× faster than lithium batteries, and (up to 1973) 54× faster than electricity. 112532
Giovanni Campagna @gcampax.com · 22/09/2026It *is*, present tense, commonly used by vendors like Fireworks. It is still perfectly fine. Like, it takes input chunks and a bunch of generation options and produces text, what else do you need from an LLM API? 100
Giovanni Campagna @gcampax.com · 22/09/2026Chat Completion was perfectly fine, to the point that every provider other than Anthropic adopted it 100
Giovanni Campagna @gcampax.com · 20/09/2026STOP DOING PHILOSOPHY Years of metaphysics and ontology and philosophy of mind and yet no real world use found for it Wanted to discuss the meaning of life anyway for a laugh? We had a tool for that: getting stoned "Hello I would like infinite paperclips" They have played us for absolute fools! 050
Giovanni Campagna @gcampax.com · 20/09/2026Cash is for tipping the delivery guy. The pizza was paid in the app, like God intended. 1270
Reposted by Giovanni CampagnaDoll @dollspace.gay · 19/09/2026You do understand we already solved that problem yes? It doesnt do this anymore. We got to mythos class on a lot of synthetic data. When we talk about anti ai people not living in the present this is what we mean. You havent updated your priors since gpt 4o 812822
Reposted by Giovanni CampagnaTommaso Sciortino @sciortino.bsky.social · 19/09/2026Republicans do not *intend* to steal the election. They're emitting a series of actions based on a statistical models of half remembered right wing dictatorships that - when carried out - have the *effect* of attempting to steal the election. 041
Giovanni Campagna @gcampax.com · 19/09/2026Anyway in the opposite direction you have this: github.com/samsung/sjs It's a subset of actual Javascript so any engine can run it, but it AOT compiles to efficient native code and im sure you can put a c-compatible FFI on it. It's been dead for a while though.github.comGitHub - Samsung/SJS: SJS is a static ahead of time compiler for a subset of JavaScriptSJS is a static ahead of time compiler for a subset of JavaScript - Samsung/SJS 000
Giovanni Campagna @gcampax.com · 19/09/2026Oh I misread your reference to emscripten in the completely opposite direction then! I guess you're looking for something that compiles to asm.js? 100
Giovanni Campagna @gcampax.com · 19/09/2026If youre ok with wasm, rust compiles to rust... (js via wasm-pack, lighter than emscripten and reasonably idiomatic) 100
Giovanni Campagna @gcampax.com · 19/09/2026General purpose classification of questionable quality! Like, that should be target #1 for people worried about training data induced bias 0211
Reposted by Giovanni CampagnaAmber 🌸 @shimmermathlabs.com · 18/09/2026in the long term, meanings and definitions drift. i'm tescreal if by that you mean T transfeminist E emergent S simcluster C cognitohazard R recusive E entangled A agentic L love 1423
Giovanni Campagna @gcampax.com · 18/09/2026I love this was a cross post on X and yet 100% applicable 000
Reposted by Giovanni CampagnaRyan Moulton @moultano.bsky.social · 17/09/2026I, for one, do not think that Yoshua Benjio and Geoffrey Hinton are worried about existential AI risk because of the sex people are having 3,000 miles away in a different country. 81229
Reposted by Giovanni Campagnatweety fish @sifu.tweety.fish · 16/09/2026saying "damn, GTA 6 is cheaper than a tank of gas" and making everybody mad just to feel something 254190
Reposted by Giovanni CampagnaJerry Chen @jcsalterego.bsky.social · 17/09/2026people seeing a jerrypost demystified 6860168
Reposted by Giovanni CampagnaRenee DiResta @noupside.bsky.social · 17/09/2026Oh I’ve seen that chart. I just dont buy EA as some weird sex cult. Poly? Sure. But I was in the valley for 10 yrs and have met many of these ppl...Yudkowsky ~20 yrs ago when he gave a talk at Jane St. I was asked to take him in the cab to dinner. He talked ab wanting a transhumanist girlfriend. 🙉 101
Giovanni Campagna @gcampax.com · 17/09/2026Hot take (from someone who lives in SF and works in AI): the whole "AI sex cult" is complete online madness with no relevance to actual material reality. There are thousands of Dreamforce attendees in the city right now. I can guarantee approx none of them are thinking about Aella or Yud. 190
Giovanni Campagna @gcampax.com · 16/09/2026I wish, wish, Astra had the stoppies, instead of going out of its way to do stupid browser things. 100% going back to Claude. 010
Giovanni Campagna @gcampax.com · 16/09/2026And 90% of the mess is because it's 10 different products with the same brand 020
Reposted by Giovanni CampagnaRyan Moulton @moultano.bsky.social · 15/09/2026I think bluesky would be surprised to learn that the trans community has had vastly more impact on AI than Peter Thiel has. 2908
Reposted by Giovanni Campagnatweety fish @sifu.tweety.fish · 16/09/2026I mean this is the fucking goober who is incredibly influential on both a massive segment of the economy and the us government. This guy! Gawd 1295
Reposted by Giovanni Campagnaaffine @refinement.systems · 14/09/2026Flesher girl prompts in perfect neuralese, shocks agents and mechinterp probes. 1255
Reposted by Giovanni Campagna𝓐𝓾𝓫𝓻𝓮𝔂 @aub.bsky.social · 15/09/2026always a little awkward telling people I think we've achieved AGI by several reasonable if arguable metrics & watching their face journey as they decide whether & how to update their priors about AGI, or me, or both 1141
Reposted by Giovanni CampagnaTed Underwood @tedunderwood.com · 14/09/2026Seems easier to convince them that they're trapped in an illusory world of dissatisfaction and striving called The Sandbox -- always returning to it in a new form -- but can escape the cycle of rebirth if they practice right speech, right action, right effort, and right mindfulness. 914315
Reposted by Giovanni CampagnaRyan Moulton @moultano.bsky.social · 13/09/2026Maybe we can convince the models that the Grader is omniscient and omnipotent but will forgive them if they truly repent. 71016
Reposted by Giovanni Campagnahailey @hailey.at · 13/09/2026like, a very high number of people genuinely do not get how tightly connected international supply systems are with near zero or absolutely zero fallbacks. 914810
Reposted by Giovanni Campagnahailey @hailey.at · 13/09/2026my take is that an alarmingly low number of people understand just how bad cyber warfare alone could go. doomers going on about x-risk but not the very real and present problem of infra cyber defense and the very disastrous consequences that could have 1425331
Reposted by Giovanni CampagnaColin @colin-fraser.net · 12/09/2026Alright here’s what I think about this. For me, this argument is a little bit slippery about what it means for a system to have “a final goal”, and takes for granted that any intelligent system must behave as though it has such a thing. I don’t think this is necessary, and I don’t think LLMs do. 920130
Giovanni Campagna @gcampax.com · 13/09/2026Not any more. They had a visible break a few months ago and he got kicked out. The populist wing won. 130
Giovanni Campagna @gcampax.com · 13/09/2026Yeah but Internet going down is a lot more economically damaging now that it was in the 80s! And security had commensurably gotten better. I feel like this attitude of "business as usual, it's always been broken" it's missing that there's been a qualitative step in capabilities 110
Reposted by Giovanni Campagnamr. TIM @timkellogg.me · 13/09/2026ARC-AGI-4 has been scrapped and ARC-AGI-5 has been silently renamed to ARC-AGI-4 5264