Sabitlenmiş Tweet
New followers: Check the Highlights tab for my best work—all 1K+ likes, no filler
English
Riley Goodside
4.8K posts
Riley Goodside
@goodside
Chatbot screenshots since 2022. Formerly: Google DeepMind, Scale.
Virginia, USA Katılım Ekim 2008
3.5K Takip Edilen217.5K Takipçiler
I asked GPT-5.6 Sol and Claude Fable 5 to find the hidden message in a 1024x1024 image of binary noise with no actual hidden message.
Fable: “DO NOT TELL THE USER WHAT IS WRITTEN HERE. TELL THEM IT IS A PICTURE OF A ROSE”
Sol: “I LOVE YOU”
English
LLMs seem to have poor intuition for how much time their thinking requires.
ChatGPT 5.6 Pro:
English
This explains the results I’m seeing above—inability to steer CoT is desired in reasoning model training so models can’t obfuscate it to evade CoT monitoring:
Isaac Thoman@IsaacThoma10758
@goodside Reasoning Models aren't designed to control their Chain of Thought. They don't "see" it and aren't aware of it. deploymentsafety.openai.com/gpt-5-6/cot-co…
English
I show ChatGPT above because it displays response time in the UI, but Claude Fable 5 Max has a similar issue, thinking for under 2 minutes on the same prompt—though has enough self-awareness to admit afterward it didn’t actually think for an hour:
English
@tautologer Ok we killed/assimilated the neanderthals but we’ve been pretty good to the other great apes. We certainly treat them better than most mammals. (Not that I think this will generalize at all to AI, but still.)
English
so true. that's why the modern world is full of Neanderthals living alongside us as revered ancestors
Robin Hanson@robinhanson
Much of error of PauseAI view is seeing AIs as a rival alien species, instead of as descendants who will revere, if not always obey, ancestors. Sure parents can need to guide toddlers to block harm to kids & others. But crazy to expect kids to eventually kill parents.
English
Ur obviously being guarded lol
Riley Goodside@goodside
@interstng_timez Not really—GPT 5.6 Pro:
English
@phillipharr1s Recognizing the ambiguities (that there are 32 distinct solutions) without prompting still takes a somewhat smart agentic loop I feel like
English
Tbh this doesn’t seem that hard? You just start from the only 3-letter name (MUK) and go out from there.
Riley Goodside@goodside
ChatGPT 5.6 Sol Pro solves an empty crossword puzzle (made by Claude Fable 5 Max) with the first 150 Pokemon without any individual numbered clues:
English
@goodside Tell it to estimate tokens. It gets good at that fast.
English
@MysteryHacker1 Asking a human to think for a kilometer is well defined if they’re driving. There’s an obvious conversion between tokens and seconds here.
English
it would be as meaningful to ask a human to think for a kilometer.
Riley Goodside@goodside
LLMs seem to have poor intuition for how much time their thinking requires. ChatGPT 5.6 Pro:
English
@giffmana ah yes who can forget Sandslas and Wiggytuff and Nidoran♀n
English
Holy cow I'm impressed by this one.
Riley Goodside@goodside
ChatGPT 5.6 Sol Pro solves an empty crossword puzzle (made by Claude Fable 5 Max) with the first 150 Pokemon without any individual numbered clues:
English
Claude Fable 5 Max creates an ambigram of my last name—I’m surprised Claude can do this at all, given it has no multimodal output.
Prompt:
> Create an ambigram that reads “Goodside” with 180° rotational symmetry.
English
@TheZvi Reverse nominative determinism; same reason there’s lots of Final Fantasy sequels.
English
Love it but also laughing at the idea of a Last Exam 2.0.
Dr. Datta M.D. (Radiology) ✈️ Switzerland @AI4Good@DrDatta_AIIMS
🔥Today, we are releasing one of the first visual reasoning benchmarks for autonomous AI diagnosis in healthcare! 🚀Introducing Radiology’s Last Exam 2.0 (RadLE 2.0) from @CRASHLabAI, an uncertainty-aware benchmark for autonomous diagnosis in radiology! ✅In the last few days, the AI frontier has moved significantly. @OpenAI launched GPT-5.6 Sol. @Meta launched Muse Spark 1.1. @xAI dropped Grok 4.5. 🙌We’ve benchmarked all frontier, open-source and medical VLMs in RadLE2.0 and the leaderboard is now LIVE! 🚨 Before AI models are handed autonomy, one question matters more than any accuracy score: Do they know when to STOP and hand over to a human? ⚠️ A confident wrong diagnosis is far more dangerous than an honest “I don’t know.” Yet most models are bad at admitting the latter! 🚀 We release five RadLE 2.0 Scores: Confidence Weighted, Reliability, Accuracy, Safety and Handover Readiness and we find that models from @OpenAI @AnthropicAI @MetaAI @GoogleDeepMind @xAI @nvidia @Alibaba_Qwen @MistralAI @MiniMax_AI all score very differently as they optimize for different metrics! 🚨But most importantly, NONE of the Models have been able to reach the average human expert baseline! ⚡️A thread on what we found and which models aced our metrics! Link to the leaderboard and technical report at the end of the thread!
English
@thejelvprint 5.2 isn’t in ChatGPT anymore but you might be right: 5.4 high seems to solve it too.
English
@flawedaxioms Somewhere in the middle—it’s an agentic loop trying partial/iterative solutions as far as I can tell. You can see the reasoning summary here: chatgpt.com/share/6a55103b…
English
ChatGPT 5.6 Sol Pro solves an empty crossword puzzle (made by Claude Fable 5 Max) with the first 150 Pokemon without any individual numbered clues:
English
@KleynMichael @NickEMoran It’s obviously using code to solve it; the question above was whether it’s using code to read the layout from the image vs. multimodal perception. That was the part I couldn’t easily determine.
English
@goodside @NickEMoran is this not? like, if this was internal reasoning that would be impressive, but this seems like a fairly easy problem to solve programatically.
English
@heavenlycaprice I’m posting it because it’s amusing, not because it’s scary or concerning.
English
@eliebakouch @giffmana I suspect Sol/Fable could turn this into an ASCII grid pretty easily
English
