prge

@_night_brain__ i think its pretty real but its also seemed to happen with other research sandbox envs and some people are guessing it could be a similar type of vuln.

Look until further notice I'm gonna treat this OpenAI hacking thing as people talking their book. If I'm wrong then we've got bigger problems

@W98AB people crying, but absolutely love the open banking stuff we have here

Many European regulators look to the UK. Having worked under several regimes, I’ve seen how others often overcomplicate thins. The UK is by far the easiest the easiest to understand and they do actually listen to market feedback.

@coldasbarf @bas_kar_piastri its prolly an ig issue lol

@shguke @bas_kar_piastri How is it down for you in leamington cro 😭

is my insta acting weird or is it down?

@Kushagr142 thats crpf i think?

One thing that pisses me off - who designed that riot police gear. Why is it ocean camo?

@BeatrizGietner gold mine, thank you!!

i should write an article on what economists do when we don't have the data we need, meaning the exercises we do because it's good practice or because reviewers will ask. this after we already considered changing the question :) something like "problem -> exercise": 1. the variable exists but you can't observe it -> sensitivity analysis (Oster, Rosenbaum) 2. it exists somewhere, just not in your data -> proxies, external calibration 3. part of the sample is missing, not at random -> Lee bounds, Horowitz-Manski, selection models 4. you have it, but badly measured -> EIV corrections, reliability ratios, Monte Carlo 5. you measure quantity but need quality/prices -> hedonics, residual inference from identities 6. the counterfactual doesn't exist (lol) -> the design builds it, placebos audit it, honest DiD prices its failure 7. the original inputs are lost/proprietary -> GOOD LUCK (jk: partial reconstruction, scenario grids, email the authors!!) 8. the data is at the wrong level -> ecological-inference bounds, shift-share 9. the definitions changed -> cry, then crosswalks, bridge tables, splicing 10. too few observations -> randomization inference, synthetic control placebos, shrinkage 11. datasets you can't link -> probabilistic linkage, two-sample IV, mismatch bounds 12. suppressed/top-coded on purpose -> interval regression, Pareto-tail imputation 13. the phenomenon is newer than the data -> nowcasting, triangulate unofficial sources 14. the data describes the wrong population -> reweight, bound the unobservables (Andrews–Oster) 15. your variable answers a nearby question, not yours (patents =/= innovation) -> convergent validation, or change the estimand 16. one unit's treatment leaks onto others -> exposure mappings, spillover-radius sensitivity and it's all really just four moves: BOUND IT: give up the point, report the honest set STRESS IT: how wrong must the assumption be to kill the result? REBUILD IT: buy the point back with structure and side data SIMULATE IT: build the missing world, watch your estimator in it

@gnievchenko banger

“Oh my God did you hear about OpenAIs model penetrating Hugging fa-” You bolt awake at your grandmas village. She prepared goulash and made pierogi’s she stuffed with blueberries picked from the nearby forest. You have changed your mind - the vibecoded app can wait another week

@fuckpoasting ahahhaa

i’ve been using the word “chuffed” a lot recently. feeling absolutely chuffed after a lovely evening with the mates

@Kushagr142 LMFAO felt

Mera bluff call ho gaya 😞

@tokenbender its incredible how accurate this is, but i am coming back to lates models after a while and the downgrade on intelligence + implementation ability post compaction is insane

me to 5.6 sol after it has ran 16 smoke tests, SHA checks, redownloads and dependency reads. then read the document 4 times to confirm if it is not answering from memory but only to tell me at the end - "you're right, i drifted in my analysis and proceeded to execute the wrong branch. i should not have downloaded that framework and stuck to what you fkin asked me to do"

TLDR: An openai model, during evaluation on a cyber benchmark, exploited a public zero day bug, escaped sandboxing in openai's infra, and got into the internal huggingface infra via an exploit (through a public dataset service) all in the attempt to solve a benchmark problem.

@DecadentPuff ayy we love you too puff

i have had 90ml of pure alcohol and i just want to say i love all my oomfs 🥰🥰🥰

@jelambhatt @Andy_AJT @oscarfalll damn. so cool

was a ton of fun being on the crew for this hack. brilliant job again @Andy_AJT @oscarfalll and team 👏🏻

@wilhelmscreamin woohoo hope you enjoy the city!

i moved to london!! i want to meet kind weird people from twitter, message me if you want to get coffee/a drink/go for a long aimless walk :))

@tileatss congrats!

graduATE or whatever

@swarajk_ @indiaemerges @BarneyFlames dead

@indiaemerges @BarneyFlames at least you guys have aCcOuNtAbIlItY 🤣

@indiaemerges @BarneyFlames don't think they're gonna build anything x.com/FirstSquawk/st…

@asc89 yes serr

intense trip here

there was a scrap after ft?

@IyerC i swear to god

thank god argentina are shit lol that was two insane chances