Post B7tf96hhkenc4EIlUm by Sassinake@mastodon.social
 (DIR) More posts by Sassinake@mastodon.social
 (DIR) Post #B7tJhqmmjDuFqGq7U0 by fribbledom@mastodon.social
       1 likes, 1 repeats
       
       It's amazing AI companies can legally bill by usage."How much will this cost?""No idea.""What will I get?""No idea.""Can you estimate either?""Absolutely not!"We're one jackpot animation away from calling it a slot machine.
       
 (DIR) Post #B7tJwBWuOqR2ETFUf2 by wolf480pl@mstdn.io
       0 likes, 0 repeats
       
       @fribbledomelectricity company:EDIT: nvm misread as "cloud companies", for AI companies it's indeed next level
       
 (DIR) Post #B7tLiGxTcTyyvHAa7U by fedops@fosstodon.org
       0 likes, 0 repeats
       
       @fribbledom we can call it a slop machine for now.
       
 (DIR) Post #B7tNjvbYBOifBPpa6a by jake4480@c.im
       0 likes, 0 repeats
       
       @fribbledom it IS a slot machine.
       
 (DIR) Post #B7tPbyMQFglHK35awS by happyborg@fosstodon.org
       0 likes, 0 repeats
       
       @fribbledom What's amazing is that companies choose to accept this and spend money based on hype and wishful thinking, because they*hope* to sack workers. As for individuals using #GenAI it's plain stupid (IMO). #LLM
       
 (DIR) Post #B7tR6xuS4yRar2Lfrk by RayeWithAnAye@mas.to
       0 likes, 0 repeats
       
       @fribbledom this is just like every healthcare visit in the US 🫠
       
 (DIR) Post #B7tSTtOd7pxFTHxcMy by BenCotterill@mastodon.social
       1 likes, 0 repeats
       
       @fribbledom Slot machines have regulated odds.
       
 (DIR) Post #B7tSxeO3UglHbmsi7k by inanedirk@metalhead.club
       0 likes, 0 repeats
       
       @fribbledom "They're not loot boxes, they're surprise mechanics" πŸ˜‚
       
 (DIR) Post #B7tUjUyhsYSiHWP8iW by lproven@vivaldi.net
       0 likes, 0 repeats
       
       @fribbledom An attempt at estimating, from @glyph https://blog.glyph.im/2025/08/futzing-fraction.html
       
 (DIR) Post #B7tVBTfCMOlQJRcYgi by jcnotwit@mastodon.social
       0 likes, 0 repeats
       
       @fribbledom Wishing every customer of these AI companies a very unexpected and catastrophic bankruptcy and total ruin.
       
 (DIR) Post #B7tW19ATo5x7nztGdM by bedifferent@mastodon.social
       0 likes, 0 repeats
       
       @fribbledom brilliant analogy but I don't think they care about an estimate.  They don't care who does it so long as they have no idea what they're doing.  They make more money that way by doing it many times over.  (See 2016ish term Puerto Rico and a cabinet member's sons  none of whom knew what they were doing but getting richer and f'ing up the power further (and getting away with all the money and leaving PR in a worse mess than it was.)
       
 (DIR) Post #B7tXbbm71YweuKRVjs by Offbeatmammal@mastodon.social
       0 likes, 0 repeats
       
       @fribbledom and when the damn Socratic Statistic machine comes up with the wrong answer, you have to pay to try again :(
       
 (DIR) Post #B7tXehNycVipB4cwm8 by falken@qoto.org
       0 likes, 0 repeats
       
       @fribbledom slot machines actually have well defined odds and rates !
       
 (DIR) Post #B7tXsrIgYRBJyuu7QO by vSourceCode@mastodon.social
       0 likes, 0 repeats
       
       @fribbledom Every platform eventually faces the same question.Are we building long-term trust or short-term revenue?
       
 (DIR) Post #B7tYafbNPVvsROXOdM by 0x0000000B@mastodon.social
       0 likes, 0 repeats
       
       @fribbledom Don't they bill by the token? (I don't know, I don't use cloud LLMs.)
       
 (DIR) Post #B7tZ8prvkDeW84ZWUK by svavar@masto.svavar.com
       0 likes, 0 repeats
       
       @fribbledom Consulting companies have been doing this for decades.
       
 (DIR) Post #B7tZQ6hKQaXdFKqg9g by Kimota94@mas.to
       0 likes, 0 repeats
       
       @fribbledom Call it a slop machine and you’re there.
       
 (DIR) Post #B7tZeaGRIWbmzzAsuu by CuriousMagpie@beige.party
       0 likes, 0 repeats
       
       @fribbledom you're describing the US health system
       
 (DIR) Post #B7tZgu1ApBxYqL8j7A by fribbledom@mastodon.social
       0 likes, 0 repeats
       
       @0x0000000B They do. But there is no estimate how many tokens any given task will consume. Or a definition of what a token really is.
       
 (DIR) Post #B7tZrjBJSFlaQ62nZo by Epic_Null@infosec.exchange
       0 likes, 0 repeats
       
       @fribbledom @0x0000000B This includes trying to use historical data. The cost of a given task may change unexpectedly
       
 (DIR) Post #B7tcaQ5xSp2QAEyhns by GLaDTheresCake@todon.nl
       0 likes, 0 repeats
       
       @fribbledom @0x0000000B A token does have a specific definition: it is the discrete parts that get spat out by the model, easiest to see them as parts of words like a few letters in it. These get strung together by using the predictions from the model, so it will go by what it expects the next part of the word will look like statistically, and pick one of the most likely options (usually). As an example if you give it the incomplete sentence "I have a bouncy red b" the next token it is most likely to predict is probably "all" to make the sentence "I have a bouncy red ball". If you add a few more tokens to that it might add something like "in" after that then "my" then "gar" then "age" to make the sentence "I have a bouncy red ball in my garage". And so on and so forth for anything it tries to string together until it reaches a termination token and stops answering.
       
 (DIR) Post #B7tcvNvciNCzWe8oBE by Npars01@mstdn.social
       0 likes, 0 repeats
       
       @fribbledom The "Something for Nothing" Economy of fraud & scamshttps://www.theguardian.com/us-news/2026/jun/29/us-supreme-court-ftc-ruling-slaughterhttps://www.motherjones.com/politics/2026/06/supreme-court-amicus-briefs-investigation/https://theintercept.com/2022/06/30/supreme-court-epa-climate-charles-koch/No wonder the rich fund corrupt efforts to kneecap regulatory bodies.https://www.axios.com/2017/12/14/koch-tries-its-hand-at-venture-capital-1513277500https://www.tilburguniversity.edu/current/news/more-news/scam-startups-and-venture-capitalhttps://readsludge.com/2025/07/09/billionaire-pharma-investors-fund-privatized-strike-team-at-the-fda/NFT'shttps://www.newswire.com/news/trustnft-research-pegs-true-annual-cost-of-email-fraud-to-u-s-economy-at-22Cryptocurrency https://www.theartnewspaper.com/2026/04/28/texas-man-23-years-prison-20m-crypto-scamhttps://www.bitdefender.com/en-us/blog/hotforsecurity/two-california-men-charged-in-nft-fraud-case-that-caused-22m-in-losseshttps://www.latimes.com/business/story/2025-12-24/california-crypto-firm-accused-of-inflating-katy-perry-nfts-fraudhttps://www.cpajournal.com/2025/01/08/fraud-risks-in-the-market-for-nonfungible-tokens/https://www.nytimes.com/2023/04/15/business/silicon-valley-fraud.htmlSnake oilhttps://www.bbc.com/news/business-583369981/
       
 (DIR) Post #B7tdGTTrGcBopJD6Uy by fribbledom@mastodon.social
       0 likes, 0 repeats
       
       @GLaDTheresCake @0x0000000B You're right, I oversimplified. Tokens are well-defined technically, but what they actually represent varies a lot between models, which makes them a pretty unintuitive unit for humans to estimate or reason about.Maybe you do develop a gut feeling for it over time, but personally, every time I hit the button, I have absolutely no sense of how long it will take or how many tokens it will burn through.
       
 (DIR) Post #B7tf96hhkenc4EIlUm by Sassinake@mastodon.social
       0 likes, 0 repeats
       
       @fribbledom Pluralistic already called it that. It works the same way.
       
 (DIR) Post #B7tggMT2GSJ6T0rRaa by jerrej@mastodon.social
       0 likes, 0 repeats
       
       @fribbledom @GLaDTheresCake @0x0000000B "You're absolutely right!" /s
       
 (DIR) Post #B7thdIN18mdUqU9EVk by fribbledom@mastodon.social
       1 likes, 0 repeats
       
       @jerrej @GLaDTheresCake @0x0000000B The irony wasn't lost on me when typing this out :D
       
 (DIR) Post #B7tkITbIUW57SOak4m by renardboy@mastodon.social
       0 likes, 0 repeats
       
       @fribbledom Personally I'm more amazed that people buy into this ridiculous proposition 😬
       
 (DIR) Post #B7tojHEiy5hskJ4IMK by larsmb@mastodon.online
       0 likes, 0 repeats
       
       @fribbledom If this was illegal all large consulting companies and law firms would be in trouble
       
 (DIR) Post #B7tqon0AUdLgcW13Vw by gary_alderson@infosec.exchange
       0 likes, 0 repeats
       
       @fribbledom the dgx station pays for itself in 2 months (at current cloud rates)
       
 (DIR) Post #B7tr4fIGv2VPLc3PLU by troed@swecyb.com
       0 likes, 0 repeats
       
       @fribbledom What's different from hiring a contractor?/contractor
       
 (DIR) Post #B7tvBIau74lin3Bhfk by obscurestar@mastodon.social
       0 likes, 0 repeats
       
       @fribbledom Oh, we can estimate the cost. AI companies just don't want to talk about it.  The cost is about O(n^3), where a binary search of the same data is O(log n) Ergo:*An AI search for 'cats' in a list of 100 items is 100*100*100 operations to find the cat.*A binary search for 'cats' in a list of 100 items is at most 8 operations.(*)Neither of these numbers are wholly accurate, just cost estimates for the respective algorithms.
       
 (DIR) Post #B7txDpTj1NO7GZAF8q by claudius@darmstadt.social
       0 likes, 0 repeats
       
       @fribbledom basically gambling.
       
 (DIR) Post #B7u0IVS3CyTSSpIppw by Professor_Stevens@mastodon.gamedev.place
       0 likes, 0 repeats
       
       @fribbledom Last time I talked about a surgical procedure with my doctor, I had much the same conversation.
       
 (DIR) Post #B7u52GazAqT1WiYdrU by zygoon@fosstodon.org
       0 likes, 0 repeats
       
       @fribbledom are you talking about ai or lawyers?
       
 (DIR) Post #B7uBRR1I9Yewp0FLX6 by coderbyheart@chaos.social
       0 likes, 0 repeats
       
       @fribbledom "Will I get a refund if it's wrong?""No way!"
       
 (DIR) Post #B7uG2l5AvIr1Zm7EZc by bitbraindev@mastodon.gamedev.place
       0 likes, 0 repeats
       
       @fribbledom this should be illegal.
       
 (DIR) Post #B8A7LOu3wLXPgz1CV6 by heiglandreas@phpc.social
       0 likes, 0 repeats
       
       @fribbledom Where is that better than developer estimations?
       
 (DIR) Post #B8AcpN6cHijbNrf58K by fribbledom@mastodon.social
       0 likes, 0 repeats
       
       @heiglandreas We at least pretend we can estimate...
       
 (DIR) Post #B8Ada1JN8JJVOts6We by heiglandreas@phpc.social
       0 likes, 0 repeats
       
       @fribbledom 🀣 True that! Often we can and then suddenly the whole spec changes... πŸ™ˆ
       
 (DIR) Post #B8wUImPF8T6ynftlMu by 0x0000000B@mastodon.social
       0 likes, 0 repeats
       
       @fribbledom That's a good point, and another excellent argument for running open-weight local LLMs rather than corpo-hosted ones.It reminds me of whenever I'm on a metered Internet connection and I want to watch a YouTube video; I don't know how many megabytes are going to be consumed (and charged for) until I've finished watching.