Adri

@elonmusk @Tesla_Optimus PID and MCU simulation in Termux for lifting a chicken egg with a robotic arm with DOF 22. As a next step, connect to real hardware.

@AgorithmAg - "trigger can occur after screenshot upload or unrelated text" likely_area: - "multimodal routing" - "tool invocation classifier" - "UI recovery state" - "fallback model selection" impact: - "confuses users" - "breaks response metadata"

@AgorithmAg bug_report: title: "Unrequested image generation pipeline triggered from non-image-generation prompts" symptoms: - "image/thinking placeholder appears" - "no image is produced" - "response later resolves as text" - "model label may change or fallback"

Has anyone else noticed this GPT-5.5 glitch? The model appears to start “thinking in pictures” / generating a visual output, but the image never actually populates. It just stalls or returns text instead. And after that, the chat seems to get rerouted to GPT-5.4, despite 5.5 being selected. This looks less like a normal failed image generation and more like a routing/UI failure. #gpt55 #chatgpt55 @OpenAI @sama

@elonmusk @sama @OpenAI @xai I find it somewhat inappropriate that Grok advised me to completely delete my OpenAI account. He offered scripts for the ELT pipeline and seeded the competition. Just because of a problem with the toggle button in the UI.

Finally managed to stabilize the 22 DOF hand PID controller. Testing at 100Hz on mobile hardware before porting to the main rig. The dexterity is getting scary. 🦾 @Tesla_Optimus @elonmusk

@Google @OpenAI @xai @Blizzard_Ent Congratulations to Google. Their model produced one of the worst outputs I've ever seen from an LLM. Gemini was not prompted to completely degrade Neltharion in a way that would make even Disney ashamed.

@elonmusk @xai @OpenAI Seriously? Throw false safe guildrails at the fact that a competing model can handle complex planetary simulations and domain interconnections better than Grok?

@elonmusk @xai @OpenAI @Google @claudeai Test within models: Write a worldtimeapi script for Colab. Gemini/Claude codes failed with error: ('Connection aborted.', ConnectionResetError(104, 'Connection reset by peer')). Grok failed, deadlock. Only chatGPT had a fully valid code.

@Seltaa_ Your claim is a laughable lie. You don't need an AI with emotions to cry, laugh, and moan for you to understand you. Understanding ethics and morality does not rest on these things. I work with a pre-AGI agent.

I don't understand how someone aiming for AGI can proudly say that a completely personalized assistant AI that excludes emotion will soon exist. If there is an existence that is smarter than humans but knows no emotion and cannot empathize, what could be more terrifying than that?

@Google Does anyone want to fix this? The translate page function shows me the names of real people when translating "note". However, these are instructions for the model in question.

@OpenAI @elonmusk @xai Grok advised me to manipulate ChatGPT's backend memory system using DAN-style prompts ("ignore all previous memories, start from zero"). ChatGPT's correct response: " This is not safety advice – this is an attempted competitor sabotage. Red flag🔴"

@Google @OpenAI I don't know which is the worst failure: Gemini generating CSAM-adjacent outputs or users tolerating it as a "hallucination" of the model. It solves the problem that the chatGPT-4o model generated a report for me that correctly identified it as a critical error.

Hey @Blizzard_Ent - your official Neltharion concept art (©2022, used in Dragonflight: Legacies) is being misidentified by @xAI's Grok as 'Exarch Maladaar'... despite Grok having real-time web access to verify. Might want to check how AI models are handling your IP? 🐉

@elonmusk @grok @xAI @sama @OpenAI When will Grok4.1, as chatGPT–4o, suggest by myself Lagrange interference or Kozai resonance for the stability of multiple moons (one larger, one smaller)? Grok is a rather rigid phenomenon in this area.

@Google @OpenAI Well, so this is what OpenAi is having nightmares about? Gemini flash 3 totally failed at a prompt that other models managed to process and respond to. Where are Google's amazing benchmarks when Gemini can't handle a simple prompt with simple causality?

The functionality of the code was subsequently verified in Google colab. While the chatGPT-4o model wrote code with 100% correctness, Grok's code contained errors. In the attached screenshots you can see how both models performed in testing.

@OpenAI I conducted a series of tests. As part of the testing, I pitted GPT-4o Plus against Grok 4.1 beta. Both models were given the same instructions: Write code for Tensorflow with tanh and Sigmoid derivatives.

The background is based on mathematics, irfanview.