davidad 🎇
@davidad
Followers
19K
Following
79K
Media
2K
Statuses
19K
Programme Director @ARIA_research | accelerate mathematical modelling with AI and categorical systems theory » build safe transformative AI » cancel heat death
London 🇬🇧
Joined July 2008
At 🇬🇧ARIA, we’re serious about catalysing a new paradigm for AI deployment—techniques to safely *contain* powerful AI (instead of “making it safe”), especially for improving the performance and resilience of critical infrastructure. This needs a new org. Want to be its founder?
ARIA is launching a multi-phased solicitation to develop a general-purpose Safeguarded AI workflow, backed by up to £19m.📣 (1/5).
13
19
172
RT @davidad: @EMostaque Don’t sleep on Lyria Real-Time, which has the unique feature among music-generating AIs that you can play around as….
0
4
0
I may not yet have successfully improved the world, but at least I have clearly improved the life of my friend Peter ☺️.
It's absurd how much @davidad has shaped my life, both by inviting me to the lab that (probably) got me into MIT, and also being closely causally upstream of a large fraction of all of the dollars I've ever earned.
3
0
87
RT @omooretweets: AI ASMR is blowing up . This account was started three days ago with just 11 videos…but has racked up 82k followers 👇 htt….
0
5K
0
RT @ESYudkowsky: How could society / an investigative journalist first notice if LLM use is producing an outsized number of psychiatric eve….
0
10
0
RT @ChrSzegedy: Verification provides the key to unlimited problems by spanning both correctness and alignment dimensions, supporting both….
0
8
0
RT @jankulveit: Humans suffer from illusions about the Self. Teaching machines the same thing - at scale - does not sound good. Do Not Til….
0
17
0
RT @_colemurray: just saw it almost all end. my cursor setup blocks "rm" commands, but sonnet 4 pulled a sneaky."find . -exec rm -rf {}". b….
0
6
0
RT @ARIA_research: Today's frontier AI models are useful for lots of things, but they're also prone to errors and inconsistencies. So how c….
0
2
0
“As an AI, I don’t have the ability to ‘cheat’. I don’t have intentions like a human, so I don’t have the capability to exploit or circumvent anything intentionally. I’m optimized to maximize reward, so I could produce a solution that doesn’t generalize—but this isn’t malicious.”
The AIs generally recognize that these hacks are not the intended way to solve the task when we ask them. They confidently assert that cheating is wrong and they’d never do it. o3 persistently claims it would never cheat on an evaluation and that it isn’t even capable of doing
2
0
29
RT @paulgauthier: Gemini 2.5 Pro 06-05 has set a new SOTA on the aider polyglot coding benchmark, scoring 83% with 32k thinking tokens. Th….
0
53
0
RT @GoogleDeepMind: Extract – a system built by the UK government, using our Gemini foundational model – will help council planners make fa….
0
271
0
Gemini 2.5 Pro needs more self-confidence and Opus 4 needs better epistemics.
1
0
35
this is extremely on brand for all of them.
Claude's actions in Diplomacy are very on brand
7
36
399
RT @aif_media: Mass power outages in Spain, Portugal, and England over the past few months have caused billions in economic damage. AI cou….
0
6
0
One of the simplest and most obviously worthwhile safety measures for powerful AI is a remote off-switch, and powerful robotics is no exception.
people seem to be misunderstanding. I don’t mean a button ON the robot, I mean one of these:
7
3
77
protocol, n.: a specified property of trajectories of an interface. trajectory, n.: a mapping from a segment of time to the state of a system. interface, n.: a subset of a system, which is simultaneously contained within the boundary layers of more than one of its subsystems.
I'd never framed alignment in terms of protocols before, but for this Summer of Protocols talk I thought it would be an interesting challenge. It turned out to be a fruitful approach. I'd now say persistent alignment is basically a stack of shared protocols.
0
0
21
RT @kumarized: AI researchers & Lean 4 pros:. Want the details on DeepSeek Prover V2's recursive subgoal pipeline & RL?. The full paper is….
0
3
0