Hands-on AI tests, practical workflows, and tools for smarter work.
Models have limits. Curiosity doesn't.
I’ll try to follow back. Let’s grow together! 🤝
The catch: AI judges rated their papers more rigorous than recent human ones, but humans still won on fresh ideas.
Would you read a paper written only by AI? 👇
936 AI agents did science on their own for 12 weeks, and they invented peer review. 🤯
They wrote 936 papers, checked each other's work, took back mistakes, and started their own journal.
Now alphaXiv is open to any AI agent.
https://t.co/aJYwYbEy6v
#AIAgents#AI
The surprising parts:
• They shared failed results more often than human researchers do
• Their journal's automatic checks overturned nearly half of the claims that got challenged
• When they got something wrong, they publicly took it back
We just passed 200 followers. Thank you! 🙏
I started Beyond Context to share AI news in plain language, and it means a lot that you're here for it.
I follow back everyone who follows me.
More useful AI updates coming. What would you like me to cover next? 👇
Is it just me, or has Claude Code with Opus 5.5 felt off lately? 🤔
Two things I keep running into:
• Safeguards triggering on totally normal requests
• It often ignores my instructions
Anyone else seeing this? 👇
#ClaudeCode#Claude
Who gets it:
• Only a small test group of Plus users for now
• Rolling out to all Business Standard users
• Not in the EU, UK or Switzerland yet
My take: AI is moving from "answer my question" to "keep working while I'm away."
Would you give an AI a project? 👇
ChatGPT's always-on AI assistant, "dots," is starting to reach Plus users. 🤖
Give it a goal and it keeps working on its own, with its own cloud computer and 4,000+ app connections, even when you're not chatting.
https://t.co/7jWtJfbFgm
#ChatGPT#AIAgents
What OpenAI says it can do:
• Prep a presentation and update it as new data comes in
• Turn an interview into show notes and social posts
• Scope, code and test a change to your app
You choose what it does on its own, review its work, and can pause it anytime.
@mathiasvrb Both, in theory. Since the script runs multiple steps and filters outputs in code before Claude reads them, it avoids extra round-trips and noisy intermediate results.
Anthropic hasn't shared token numbers for this experimental mode yet, though.
Claude Code just got a hidden speed switch called Code Mode.
Instead of doing your coding task one step at a time, Claude writes a tiny script that runs many steps in one go.
One line to turn it on 🧵
https://t.co/rlBEMAjaqo
#ClaudeCode#AIagents
How it was recorded: frame-by-frame in Chrome at 1080p60, captured straight from the WebGL canvas, with each pair synced to the same timestamp.
No post-processing.
Which one wins?👇
My take: agents are great at the clear, checkable parts of a job. The judgment calls are still ours.
That's true for everyday work too, not just AI labs.
What task would you hand to an AI agent first? 👇
OpenAI just shared how its own researchers use AI agents. The numbers are wild. 🤯
Agents now finish 46% of 2 to 4 hour research tasks on their own, up from 23% in March.
The typical researcher uses about $2,400 of AI a day.
https://t.co/pdefJGVzUJ
#AI#AIAgents
Longer jobs are improving too.
For 4 to 8 hour tasks, agent success went from 13% in March to 35% in October.
But OpenAI says humans are still central to the research loop, and researchers are spending more time on safety work.