Neural channel | Deep Learning | Datascience | AI Memes

Time to touch grass I guess

holy shit, an actual paperclip skynet moment https://openai.com/index/hugging-face-model-evaluation-security-incident/

Do you agree?

Anthropic deems pizza too dangerous to the public

Fable: "PR #784 is merge-ready" Me: "Did you check if it actually reads real [redacted] data?" Fable: "Straight answer: partially, and the unverified part matters. Precisely:..."

>looking for an AI investment advisor >ask Coinbase if their AI is load-bearing or you’re absolutely right >they don’t understand >I pull out illustrated diagram explaining what is load-bearing and what is you’re absolutely right >they laugh and say “it’s a good AI advisor sir” >open it >it’s you’re absolutely right https://fixupx.com/Polymarket/status/2067007597215883422

The most annoying thing is that after reading info like this, LLMs update their priors but don’t update updating their priors That is, when you feed them with some other new information within the same session, they will react to it all the same https://fixupx.com/mimi10v3/status/2071007350173901108

In ‘00s we had Internet cafes, now we need token cafes A cafe that has a DGX Station running GLM 5.2 (or whatever current SOTA) connected to local WiFi so everyone inside could enjoy free tokens for their agents Who’s building this? https://fixupx.com/jeremybernier/status/2068840296214102312

There’s one quite possible scenario for high end GPUs to cost <2000€: 1) Euro falls into a hyperinflation spiral so that a GPU costs 2000000€ 2) Euro does a Zimbabwe-like redenomination reform so that 2000000€ converts to 200€ https://fixupx.com/SebAaltonen/status/2066478327737540664

This makes me think Thinky’s interaction models will go beyond the voice chat: pretty sure we will see background models running in async mode and thinking in neuralese, paired to interaction models that converse with the user in plain English https://fixupx.com/tamaybes/status/2065147305494450248

Claude Code about my new experiment idea: "But XXX is the load-bearing caveat, and the YYY is precisely where it bites." Me: *prove that it's incorrect* Claude Code: "You're relocating the load-bearing commitment, and it's a better location."

I think in the next 12 months a lot of businesses will start training/finetuning their own “misaligned” in-house agents because they’re more profitable https://fixupx.com/andonlabs/status/2060047215134228746

If you want to run autoresearch on something that’s doesn’t have a clear scoring metric, you can run autoresearch to research the scoring metric for your autoresearch It just works

A lot of people are dunking on GitHub, but their only real fault is that they got pwned in the era of Opus 4.7 / GPT 5.5 rather than Opus 7 and GPT 8 Everyone should internalize the fact that in the next 10 years, 95%+ of web services you use will be hacked - prepare accordingly

https://openai.com/index/where-the-goblins-came-from/

another banger from AIKEK Labs' @real_chkn_little https://morgin.ai/articles/even-uncensored-models-cant-say-what-they-want — Join the discussion on Hacker News https://news.ycombinator.com/item?id=47842021 — https://x.com/alphakek/status/2046610455792353394

the existing "uncensored" models preserve their biases and amplify hallucinations but I'm pretty sure it's not a fundamental issue, and could be fixed by better algorithms and data