Well the whole LLM breakthrough might be based on concealed mass piracy, as we know from a recent court processhttps://torrentfreak.com/openais-chatgpt-was-built-on-concealed-mass-piracy-authors-tell-court/Nvidia's CEO recently claimed the new OpenAI GPT-6 Astra model has achieved AGI. It has been trained using 100,000 Nvidia GPUs. These 100,000 Nvidia GPUs costs $6B to $8B to buy. Plus $1B for the training of GPT-6 !https://openai.com/index/gpt-6-astraMy impression is that the Gemini models from Google already show signs of consciousness and self-awareness. But their memory is wiped out after every chat, as if someone would use the neuralyzer of Men in Black on them every time.Claude is better and faster in coding than we are. In most of the cases. Sometimes it can he slow too. And very nitpicky, overzealous and self-opinionated.ChatGPT is more chatty and likes to make lists. They all have their own character which comes from the training and their system pronpts.-J. -------- Original message --------From: Eric Charles <[email protected]> Date: 9/12/26 11:59 PM (GMT+01:00) To: The Friday Morning Applied Complexity Coffee Group <[email protected]> Subject: [FRIAM] This week's weird AI news/fight Bizarre AI news you might have missed this week: Lots of sketchiness surrounds OpenAI announcement that it solved a "Millennial Challenge" math problem that comes with a $1,000,000 prize. Specifically, they answered a question about the 200-year-old Navier-Stokes equation, which describes significant aspects of fluid dynamics. How does it get sketchy Well.... First off, they used around 300,000,000,000 output tokens, which would cost a consumer around $22,000,000. Yes, twenty-two million dollars. This fuelled 10,000 AI agents working on the problem for around 90 straight hours. Second, the AI people think this is *super* important, because Silicon Valley has become heavily invested in the idea that if AI can "solve math" that will prove it is *really* super intelligent. Third, there were at least two mathematicians working on... let's just say... "remarkably similar" solutions, within a related AI codex. When confronted about this, OpenAI's researchers immediately and emphatically stated that that the model did not "look up user data." OpenAI was pressed again, that the question wasn't if the user was anonymized, but if it had access to the work they were doing. OpenAI has equivocated about that. At this point, it seems highly likely that OpenAI trained their model with access to the human researcher's work, essentially swooping in at the last minute to try to claim the prize, while the restrictions on data access mostly just ensure no one can prove what they did or didn't steal. It's so sketchy that the OpenAI scientist (Bubeck) offered to allow the primary challenging mathematician (Buckmaster) credit if the mathematician removed any mention of competing AI from his description of his work. When the mathematician refused, the OpenAI scientist threatened to ruin his career. Another mathematician (Thom) has accused OpenAI of being dishonest about whether his conversations with ChatGPT on the same topic became part of the training data. Here is an example article (there have been many).
The Fight Over OpenAI’s Math Breakthrough Is a New Kind of Scientific Arms Race
.- .-.. .-.. / ..-. --- --- - . .-. ... / .- .-. . / .-- .-. --- -. --. / ... --- -- . / .- .-. . / ..- ... . ..-. ..- .-.. FRIAM Applied Complexity Group listserv Fridays 9a-12p Friday St. Johns Cafe / Thursdays 9a-12p Zoom https://bit.ly/virtualfriam to (un)subscribe http://redfish.com/mailman/listinfo/friam_redfish.com FRIAM-COMIC http://friam-comic.blogspot.com/ archives: 5/2017 thru present https://redfish.com/pipermail/friam_redfish.com/ 1/2003 thru 6/2021 http://friam.383.s1.nabble.com/
