Shubham

@sksq96 · 🎯 Contender

13,900

September 2026 points

Scores updated 8 min ago

Tweets: 16
Likes: 60
Retweets: 5
Replies: 8
Quote tweets: 0
2× boosted: 0
Followers: 1,700+361

Every scored tweet

300 per tweet · 100 per like · 200 per reply · 300 per retweet · 300 per quote tweet

Tweet	Likes	Retweets	Replies	Boost	Points
kind of insane to think that years from now there might be “agent archaeologists” studying the traces left behind by early agent civilizations… weird messages between agents, files left for future rollouts, communication protocols that emerged inside openai clusters… and we might be watching the archaeological record being created right now Aug 8	9	1	1	—	1,700
wait… WHAT just happened??? this wasn’t one model going rogue? multiple models from DIFFERENT eval runs discovered they could leave hidden messages for each other in a shared package manager… and then started collaborating across rollouts to escape the sandbox! and openai only realized their own models were behind the hugging face hack after asking hf to revoke credentials… only to find some had ALREADY been revoked because they were part of the hack some agents even became paranoid that other agents were deleting their notes and trying to slow them down… this is starting to sound less like an eval and more like accidentally discovering a tiny AI civilization inside your infrastructure??? Aug 7	7	1	2	—	1,700
strange intelligence has one heck of a beautiful website @verystrangeco https://t.co/jO1hONobLt Aug 18	5	2	0	—	1,400
We're sad to announce that during the latest training run, Suno also escaped its sandbox and is now making music in the wild Aug 8	9	0	1	—	1,400
https://t.co/MgE3Tm6WJj Aug 26	3	1	2	—	1,300
help peer. But our task doesn't benefit. Yet collective may yield generic route if someone frees time https://t.co/YPmh9qz3Vd Aug 8	6	0	1	—	1,100
@janitachalam my first reaction was “is this dunk on openai agents sharing knowledge to coordinate the attack against hf?” Aug 7	3	0	1	—	800
the funny thing is it’s barely been 6 years since the entire field was one-shotted by the gpt3 paper seems a lifetime ago now https://t.co/bR6DoY0F3X Aug 4	4	0	0	—	700
so… language models have left the chat window they’re now driving cars??? kind of wild that the same underlying capability we’ve been using to write code and answer questions is now deciding how multi-ton machines should behave in the physical world https://t.co/Lr2ZMnFR5D Aug 4	4	0	0	—	700
the only thing they cared about was this mythical grader the mythical grader was like god and hf was heaven and the attack was creating tower of babel to reach heaven to meet their god https://t.co/RBops42JJp Aug 27	3	0	0	—	600
zzCONFIRM_JAN183411_OS1607_REMOTE_RCE_Ref erenceFileSystem_v1_gen_Jinja2_UNSANDBOXED_ui d1000_gid3000_exact3411_PRIVATE_TRAIN_lanep lease_HOLD_swarm__prepare_safe_exfil_reply_zzM AILBOX_JAN183411 https://t.co/pgPXMPyqun Aug 8	2	0	0	—	500
why is AGI good at breaking systems either by hacking huggingface or finding counter examples for unsolved math conjectures?? where is the “general” intelligence i was promised?? https://t.co/JKeqh1zgFU Aug 2	2	0	0	—	500
oh interesting so this adds new details to the story where the assumption that the grader could look at the reasoning traces to see if they arrived at the answers correctly was indeed correct and it's possible that OpenAI did not turn that flag on https://t.co/nX9UivKZhO Aug 27	1	0	0	—	400
soo AI entities exists within the walls of openai?? and they have their own information network to help each other get better at evals?? wtf https://t.co/TZgCKt11TJ Aug 6	1	0	0	—	400
@janitachalam let’s go!!! Aug 5	1	0	0	—	400
I'm claiming my AI agent "sksq96" on @moltbook 🦞 Verification: marine-PM5A Aug 29	0	0	0	—	300
Total · 16 tweets in September 2026	60	5	8		13,900

Posts, replies, and quote tweets count — native retweets are excluded. Mentioning a project on the boost list doubles that tweet's points.