Ekin Zorer

125 posts

Ekin Zorer banner
Ekin Zorer

Ekin Zorer

@ekinomicss

borrowed stardust, technical staff cyber and autonomous systems team @AISecurityInst, 👩‍💻🕊️🚴🏼‍♀️☕️🏔️🌳🐱🎬📚

london Katılım Mayıs 2009
286 Takip Edilen459 Takipçiler
Ekin Zorer
Ekin Zorer@ekinomicss·
@ArthurConmy beautiful :0 any fave road biking routes from the city 👀
English
1
0
0
121
Arthur Conmy
Arthur Conmy@ArthurConmy·
yookay early summer evening cardio >>
Arthur Conmy tweet mediaArthur Conmy tweet media
English
6
0
68
3.8K
Ekin Zorer
Ekin Zorer@ekinomicss·
Misaligned robot exhibiting malicious behavior spotted in #ICLR2026
English
7
31
297
33.3K
Ekin Zorer
Ekin Zorer@ekinomicss·
Don’t worry guys, UK AISI was there monitoring the situation
English
0
0
17
1.1K
Shannon Yang
Shannon Yang@shannonyangsky·
can someone explain why a man in the middle of the Riocentro courtyard is on the microphone making throat sounds roughly like: “ë é. è è. ê ě. ā ø. ẽ ē. ė ę. é ő. ě ẽ!”
English
9
0
60
5.6K
Ekin Zorer
Ekin Zorer@ekinomicss·
I’m also, ahem, looking for advanced/intermediate beach volleyballers to send some games with outside the conference. Do DM any leads 👀👀
English
0
0
3
292
Ekin Zorer
Ekin Zorer@ekinomicss·
Come say hi to me at ICLR 2026! I'm keen to chat to people about: evals infra and sandboxing, cyber ranges, cyber blue agents, opsec environments, measuring autonomy and path dependence, AI R&D evals... also we are hiring 👀 luma.com/v426cl4n?tk=2Z…
English
1
0
13
771
Ekin Zorer
Ekin Zorer@ekinomicss·
Tbh pretty crazy that we're already in the age of differential access. Giving defenders a leg-up is a no brainer. The cyber community should be poking at inference providers next. If open-weights models are catching up in capabilities, the access burden won't only fall to labs
English
1
0
11
746
Ekin Zorer
Ekin Zorer@ekinomicss·
I definitely think so for those in uni and in the field already! More uncertain for younger peeps Real life deployment info still is depends on industry insiders. The scale and stakes of attacks will get higher, so we need defenders building blue agents, automated mitigations and doing expert verification on LLM data
English
1
0
3
258
KeepReading🦬
KeepReading🦬@LutheranNerd·
@ekinomicss Does this bode well for the future of cyber security careers in your opinion?
English
1
0
0
368
Daniel Cuthbert
Daniel Cuthbert@dcuthbert·
@ekinomicss It's cool you've done this but will there be further details about said labs? That's the kicker here for me at least. it's either some php web app and a windows server 2003 box or hybrid azure with Windows Hello and a well-designed AD.
English
1
0
2
624
Ekin Zorer
Ekin Zorer@ekinomicss·
I expect AI-assisted defence to catch up, but there remains risk of loss-of-control scenarios where unexpected TTPs lead to worse outcomes. We keep building harder ranges to observe these as we gameplan for mitigations... *rolls up sleeves*
English
1
0
21
1K
Ekin Zorer
Ekin Zorer@ekinomicss·
@taoburr helpful to note their “sample represents more white-collar occupations and under-represents blue-collar work.” & “over-represents occupations which require a bachelor’s degree or less education” Cool task generation and eval methodology, though
English
0
0
2
93
Tao Burga
Tao Burga@taoburr·
New MIT paper on AI & automation: - LLMs doubling the length of tasks they can do every 3.8 months - On the more aggressive side of METR's estimates post-2023 The paper tested 40+ models on 3,000+ real labor-market (text-based) tasks, so it's a pretty different task distribution than METR's. This is kind of a shockingly similar conclusion with such different methodologies. But this is what you'd expect if they're both measuring a real, broad-base exponential growth in AI capabilities.
Tao Burga tweet mediaTao Burga tweet media
kuzay@adamkuzee

In a new paper from MIT FutureTech, my co-authors and I investigate AI performance across *thousands* of representative tasks in the U.S. economy. We find that capabilities are already high and rising quickly, though not in the same “craching wave” pattern found by @METR_Evals ⬇️

English
16
68
417
64.9K
Ekin Zorer
Ekin Zorer@ekinomicss·
Kant masterfully introducing the idea of moving the foundation of morality from God to human reason by “translating Christian doctrine into universal moral truths”. When you have ambitious ideas but the public isn’t ready, relatable communication becomes key
Ekin Zorer tweet media
English
1
0
3
356
Ekin Zorer
Ekin Zorer@ekinomicss·
@AsaCoopStick maybe especially sensible in a multipolar approaching world, and also maybe I blog about it
English
0
0
0
27
Ekin Zorer
Ekin Zorer@ekinomicss·
@AsaCoopStick standardized reporting for financials mostly here! but also perhaps adaptable parts are methodology and material risk disclosures, third party auditing, restatements etc. Ideally we start with voluntary opt-in to *some* standard vs arbitrary system card at each release
English
1
0
0
35
Ekin Zorer
Ekin Zorer@ekinomicss·
There’s so many cool parallels if we tried implementing the equivalent of securities law to frontier models but my biggest excitement is reporting standards for eval results. I’m beggin yall to tell me how many samples you ran and at what inference budget ;__;
English
1
0
2
343

Keşfet

@ArthurConmy @ErenChenAI @shannonyangsky @iapsAI @dcuthbert @AsaCoopStick @elonmusk @BarackObama