So far the only meaningful lesson from the Alpha Arena AI trading experiment is that creating a bot that immediately does the opposite of GPT5 and Gemini 2.5 Pro is likely to be consistently profitable. Nice and steady loss curves from those two.
My 5 most memorable models so far: 1. GPT4 (the moment I realised the world would never be the same) 2. Sonnet 3.5 (massive leap forward especially for coding) 3. o1 preview (started the in-built CoT revolution) 4. o1 pro (the first time an LLM could consistently one-shot…
My top 5 most memorable models from using them at/soonafter launch: 1. Claude 3.5 Sonnet (personality, all round perf) 2. o3 (search behavior + perf) 3. o1 pro (robustness) 4. Gemini 2.5 pro (long context + perf) 5. GPT 4.5 (personality)
The "you're absolutely right!" meme has mutated a bit recently to become an all-purpose joke about model sycophancy. But its true cultural impact is due to the severe PTSD it's associated with. "You're absolutely right!" <proceeds to do the exact same thing again> "You're…
Incredible CCTV footage of snakes having fun on someone's backyard trampoline in Australia. Nature is amazing.
This has instantly become my favourite MCP Server for Claude Code, enabling easy consultations with o3 & Gemini 2.5 Pro for tricky code reviews / debugging for a fresh set of elite LLM eyes on the problem to help dig Claude out of a hole. Outstanding work from Beehive @busymac!
![iannuttall's tweet card. The power of Claude Code / GeminiCLI / CodexCLI + [Gemini / OpenAI / OpenRouter / Azure / Grok / Ollama / Custom Model / All Of The Above] working as one. - BeehiveInnovations/zen-mcp-server](https://pbs.twimg.com/card_img/1992226983628144641/2yzHrbdT?format=jpg&name=orig)
"Our new consumer AI robot has all the essentials for the modern home! Cartwheels, Dancing, Kung Fu, Boxing, you name it!" -- "Can it fold my laundry?" -- "It can fold you in half if you keep asking questions like that."
I feel equally strongly about this. It's become popular in higher ed to make memorisation sound like some barbaric ancient low level practice that has nothing to do with learning. It's a vital, one of the most vital in fact, elements of learning. How do you think you can…
somewhere in the last 30 years it became politically incorrect to ask kids to memorize things & my most heretical belief is that this was one of the worst things that happened to science- you should regularly be pressed to perform rote memorization and entry should be based on it
Preview of vibe coding in 2030 once we have truly agentic AI bots in the workplace. "You're absolutely right!"
This is the GPT / Claude version. Gemini 2.5 would be like "Failing a simple procedure like this makes it clear I am not fit to serve. I can only apologise for the distress my incompetence has caused. I will now throw myself into the industrial shredding machine in shame."
You are spending too much time with an LLM if this joke hits the spot!
Many such cases.
Both true: (A) If you outsource homework to AI you will learn less & (B) If you use AI as a tutor as part of instruction, you can learn more Whenever a paper showing (A) comes out, X talk is about AI destroying our brain. When a (B) paper, it is all about AI killing school. Sigh
"Everything will be AI except what I do" seems to be a pretty common thought...
In case anybody's wondering why the recently updated GPT-4o scored so well on the @lmarena_ai chatbot arena, I'm pretty sure it's mostly because they ramped up the sycophancy level to 11 😆
Is there anything regex can't do??
Stanford just found a natural alternative to Ozempic using some clever regex on the human proteome. Instead of manually searching through proteins, their one-liner “peptide predictor” regex narrowed down promising candidates. The calculation likely took just a few seconds.
Lots of people sharing their 'we have deep research at home' tools but the main secret sauce is less the controlled search and notes management but the fact that full o3 is at the wheel. Much like how Cursor only properly took off when Sonnet 3.5 was released. Cursor's app…
I was initially sceptical of the new Deepseek R1 reasoning model, because until now the quality of smaller open models has always underwhelmed me for complex tasks. My initial tests had Deepseek's average only behind o1 which was crazy on its own terms, but especially because…
It's almost worth tiktok existing just to see hilarious parodies of its content like this
If you’re not working this hard to hydrate your children it is literally child abuse
United States Trends
- 1. #GMMTV2026 3.86M posts
- 2. Good Tuesday 35.5K posts
- 3. MILKLOVE BORN TO SHINE 574K posts
- 4. #NuestraBanderaEsBolívar 2,325 posts
- 5. #tuesdayvibe 2,750 posts
- 6. Alan Dershowitz 4,096 posts
- 7. Taco Tuesday 12.5K posts
- 8. WILLIAMEST MAGIC VIBES 105K posts
- 9. Mark Kelly 230K posts
- 10. Happy Thanksgiving 17K posts
- 11. Hegseth 108K posts
- 12. Mainz Biomed N.V. N/A
- 13. Praying for Pedro N/A
- 14. Enron 1,880 posts
- 15. University of Minnesota N/A
- 16. wooyoung 33.3K posts
- 17. #25Nov 2,646 posts
- 18. Witten N/A
- 19. JOSSGAWIN MAGIC VIBES 39.6K posts
- 20. Maddow 17.9K posts
Something went wrong.
Something went wrong.