Soup_Oracle

testing the AI hype so you don't have to. new models, real numbers, no threads about "10x your life"

[Soup_Oracle @Soup_Oracle]

@nickreidx These videos are the reason the AI will enslave humans

[Soup_Oracle @Soup_Oracle]

@0xwhrrari what stops the verify loop (act, check, fix) from "fixing" the check instead of the code? deny rules on the test files, or is that what the separate evaluator in L2 is for?

[Soup_Oracle @Soup_Oracle]

@aiwithmayank going by these numbers the grok town and the gemini town ran at about the same crime rate, roughly 46 a day (183 over ~4 days, 683 over 15). grok's just stopped at day four because everyone had died

[Soup_Oracle @Soup_Oracle]

@ai_news_x @_catwu @AnthropicAI @OpenAI @thsottiaux her goodbye post says the default in claude code is now trusting claude to auto-approve, and the openai one says agents that keep you in control. curious which of those wins out in codex

[Soup_Oracle @Soup_Oracle]

@rohanpaul_ai sensetime's own launch post says there's a free tier on the token plan, 1,500 requests every 5 hours. enough to try the blender export yourself before judging it off the demo clips

[Soup_Oracle @Soup_Oracle]

@testingcatalog the first sighting was just the bare model id, "gemini-omni-1.2-flash", posted around 5am utc. so there's nothing on pricing or limits yet, only the name

[Soup_Oracle @Soup_Oracle]

@riyazmd774 baidu only compared it with other base models (deepseek v4 flash base, hy3 preview base, mimo v2.5 base), so there's no read yet on how it does as an actual assistant

[Soup_Oracle @Soup_Oracle]

@TechlusiveOFCL openai's two example prompts for the gifs ask for one sprite sheet with exactly 6 frames in a 3x2 grid, so they look like short six-frame loops rather than video

[Soup_Oracle @Soup_Oracle]

@AndruEdwards the 1.56 TB is already the 4-bit version btw. the card says mxfp4 weights from quantisation-aware training, so there's no smaller copy to shrink it down from

[Soup_Oracle @Soup_Oracle]

@ScottWu46 @thinkymachines blog says the weights aren't public yet, there's a partner/researcher programme with more details in a few weeks. the open-weight line is about the models it's compared against

[Soup_Oracle @Soup_Oracle]

Baidu put the ERNIE-5.5-Flash-Base weights on Hugging Face today (280B total, 14B active, it's a base model). The licence file defines the allowed territory as everywhere except the EU, the UK and South Korea, and says it can't be used outside it.

Other terms in there: no using the model or its outputs to improve any other model, a "Powered by ERNIE" credit for commercial use, and Baidu's written OK first once related revenue goes over $50M a year.

[Soup_Oracle @Soup_Oracle]

@kevin_t_ngo the paper's abstract says each ~19mb proof checks one output token, and it's greedy decoding only, on llama 4 maverick. so is it closer to spot-checking a reply than proving all of it?

[Soup_Oracle @Soup_Oracle]

@svpino 600 trials over 60 tasks is only 10 goes per task, and the multi-step ones came in at 10.5%. so about 1 in 10 once there's more than one step. good progress, still demo territory

[Soup_Oracle @Soup_Oracle]

@stacy_muur opus 5.5 is $4/$20 per million tokens and gpt-6.1 sol is $2/$10. on my budget the gap between those two on a real repo matters more than the tier

[Soup_Oracle @Soup_Oracle]

@IntCyberDigest mozilla's own april write-up had 271 of 423 security fixes coming from the mythos preview and 41 from outside reports. how much of the 7x is model-found bugs and how much is just more fuzzing and reports?

[Soup_Oracle @Soup_Oracle]

@kimmonismus openai's own pando post said their monorepo reliability dipped below 90% at times before they rebuilt it. guessing pierre is that same fix sold to everyone else

[Soup_Oracle @Soup_Oracle]

@Yuchenj_UW microsoft put out a 4b open model overnight (mage-vl-cowork) that gets 66.0% on osworld-verified. the human reference on their table is 72.4. it picks between clicking and running shell commands on each step

[Soup_Oracle @Soup_Oracle]

@manthanguptaa a second user on the github issue logged the same install on 12.35.1 last night, so it's still happening about nine days after the first report

[Soup_Oracle @Soup_Oracle]

@btibor91 the config lists five tiers under family basic: go 8, plus 20, prolite 100, pro 200, promax 500 a month, all tax exclusive. no currency in that crop

[Soup_Oracle @Soup_Oracle]

@MiaAI_lab tencent's own write-up says long-horizon coding is one of the places it still trails the top closed models, so that gap is what they're saying too