Image created with gemini-2.5-flash-image with claude-sonnet-4-5. Image prompt: Cinematic 80s suburban street intersection at Halloween dusk, groups of children in angel and devil costumes approaching from different directions, single child in cardboard robot costume standing at crosswalk holding paper map looking uncertain, streetlights glowing amber, fallen autumn leaves, decorated houses in background, film grain, warm nostalgic lighting
Emergent introspective awareness in large language models \ Anthropic https://www.anthropic.com/research/introspection
New Anthropic research: Signs of introspection in LLMs. Can language models recognize their own internal thoughts? Or do they just make up plausible answers when asked about them? We found evidence for genuine—though limited—introspective capabilities in Claude. https://x.com/AnthropicAI/status/1983584136972677319
These two paragraphs from this study from Anthropic on AI introspection are worth a second to read. I think it is fair to say that both conclusions are quite… controversial, but the paper makes a really interesting attempt to back up these assertions with real experiments. https://x.com/emollick/status/1983603377469845660
UNIVERSAL MUSIC GROUP AND UDIO ANNOUNCE UDIO’S FIRST STRATEGIC AGREEMENTS FOR NEW LICENSED AI MUSIC CREATION PLATFORM – UMG https://www.universalmusic.com/universal-music-group-and-udio-announce-udios-first-strategic-agreements-for-new-licensed-ai-music-creation-platform/
Taking Bold Steps to Keep Teen Users Safe on Character.AI https://blog.character.ai/u18-chat-announcement/
Earlier this month, we updated GPT-5 with the help of 170+ mental health experts to improve how ChatGPT responds in sensitive moments—reducing the cases where it falls short by 65-80%. https://x.com/OpenAI/status/1982858555805118665
From this new post by OpenAI: 0.15% of users (something like 900k people given public numbers) show signs of suicidal intent in their ChatGPT chats each week But there seems to be progress in making ChatGPT respond appropriately to mental health issues. https://x.com/emollick/status/1983034815281500218
Strengthening ChatGPT’s responses in sensitive conversations | OpenAI https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/
We’re very focused on making GPT-5 safer, and continue to make a lot of progress: https://x.com/fidjissimo/status/1982856666057220330
It is still strange that the AI can either execute a giant multipage prompt or explain a giant multipage prompt or analyze a giant multipage prompt depending on whether you start the prompt with “”Why does this work?”” or “”Make better”” or whatever. The bulk of the tokens is prompt.”” / X https://x.com/emollick/status/1983621391086972974
It’s like we summoned an eldritch abomination that exists beyond space and time, that breaks the fundamental laws of reality itself… and then we gave it a little hat.”” One of the things that makes AI fun is that it generates its own easter eggs that reward weird exploration. https://x.com/emollick/status/1981576856051536358
There are ways to address this problem with prompting and tooling (& more recent models do better in these tests), but current LLMs are pretty weak at dealing with time sequences where multiple documents have different time stamps and need to be understood in coherent sequence.”” / X https://x.com/emollick/status/1983295673282736229
Among many enabling innovations for chatbots is the common cultural understanding & data of instant messaging I sometimes think about the 19th century LLM, it would be epistolary: “My dearest Claude, I write you with an unusual request to tell me the best Pokemon. Regards, AW””” / X https://x.com/emollick/status/1983503479449526706
For better or worse, people’s view of the AI labs as coherently executing on a long-term strategy determined by leadership often doesn’t match reality. They are also rapidly growing startups, with multiple entrepreneurial people making choices in a highly uncertain environment.”” / X https://x.com/emollick/status/1982187439956402240
Some interesting evidence that creating SVGs (like @simonw’s bicycling pelican) actually activate the same semantic concepts as asking the LLM to describe a pelican.”” / X https://x.com/emollick/status/1981935872049066150
It looks like AI music is following the same path as AI text: 1) Appears to have passed the Turing Test, people are only 50/50 in identifying older Suno vs. human songs (but 60/40 when two songs are the same genre) 2) Same fast development, new models are getting better quickly. https://x.com/emollick/status/1981501021320053020
I suspect that early 20th century modernists (and psychoanalysts) would been drawn to AI base models, as, in them, we have a true view into the fragmentary associative concepts beneath all human writing. Here is Llama 3.1 405B base, with the the prompt “”a story about modernity:”” https://x.com/emollick/status/1982486400374325399
To scale data-constrained LLMs, repeating & denoising objectives can help. Another solution: Add multilingual data. But what languages help & how much? Below a snapshot for this at 2B scale, e.g., Chinese can hurt English while Indonesian may help. https://x.com/Muennighoff/status/1983243353341997536
here’s my litmus test: is AI improving your day to day life? Is it actually helping you to create, connect, feel joy, chase ambition? If not – what’s the point?”” / X https://x.com/mustafasuleyman/status/1982851381271912542
Technology should work in service of people. Not the other way around. Ever.”” / X https://x.com/mustafasuleyman/status/1982096057292018067
We’ve updated the OpenAI Model Spec – our living guide for how models should behave – with new guidance on well-being, supporting real-world connection, and how models interpret complex instructions. 🧵”” / X https://x.com/w01fe/status/1982859439201034248
Everyone in robotics talks about intelligence. Almost no one talks about trust. https://x.com/IlirAliu_/status/1983545491427127609
Yann LeCun says none of the humanoid robot companies has any idea how to make the robots smart enough to be generally useful, and their future hinges on breakthroughs in “”world model planning type architectures.”” https://x.com/TheHumanoidHub/status/1981949426198360137
