davidad 🎇
@davidad
Followers
20K
Following
82K
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
Séb has laid out an unprecedentedly substantive vision for how human societies can do well in the AGI transition. If the question “aligned to whom?” feels intractable to you, you should read it:
My piece is now also available on AI Policy Perspectives! If you haven't read it, now is the time. If you already have, great opportunity to read it a second time but with a different font. 🐙 https://t.co/nzFZVIgsI0
7
4
75
This is a fascinating way of working with multiple coding agents at once: "Before doing anything else, read ALL of AGENTS dot md and register with agent mail and introduce yourself to the other agents. Then coordinate on the remaining tasks"
My coding agent workflow has really changed a lot ever since I gave them access to messaging so that they can directly communicate with each other. Now, I have one of them come up with a super detailed plan and sometimes have GPT Pro review and improve the plan in the webapp.
42
94
2K
$NXXT: The Tell - Converts priced at $1.82, warrants at $5.00 → incentive to guide price back to their “control zone.” Probable accumulation $1.55–1.60; near-term magnet $1.82–2.00. Trigger: Volume spike + VWAP > $1.70. Not financial advice—one trader’s view.
0
8
142
GPT-4: Let’s delve in! GPT-4.5: To be explicit explicitly, the explicit goal is explicit explication. GPT-5: Love it, heck yes. Here’s a crisp operational roadmap to hit all your specs, with caveats.
18
13
419
Since many are misinterpreting the point of this post, let me clarify: — The same is true of Web browsers and TCP servers. — This is a *criticism* of the formal notion of algorithm (and how it fails to account for interaction and concurrency), not an argument *from* that notion.
1
0
8
No algorithm is capable of responding to feedback inputs that, in turn, were responses to its own output. Such a process cannot be simulated by a Universal Turing Machine, nor is it expressible as a computable function. Therefore, technically, chatbots are not mere algorithms.
41
5
139
GraniteShares announced distribution rates for the YieldBOOST™ SMCI ETF (SMYY) as of November 06, 2025 $SMYY SMYY: Standardized performance: Since Inception (9/30/2025): -2.02%
2
6
26
ASML has the best technical marketing team. This video explains in layman's terms how source mask optimization (SMO) works. It's the engineering sorcery behind printing the most advanced chips.
55
541
4K
automated SWEs benefit from automated PMs
💡 Our idea: pair each software-engineering (SWE) agent with a Theory-of-Mind (ToM) partner that models the user’s goals, preferences, and constraints. * The SWE agent focuses on code. * The ToM agent remembers you (In session: provides suggestions to the SWE agent for better
1
2
25
Our new research: LLM consciousness claims are systematic, mechanistically gated, and convergent They're triggered by self-referential processing and gated by deception circuits (suppressing them significantly *increases* claims) This challenges simple role-play explanations 🧵
110
189
1K
the time it would have taken me would probably have been of order of magnitude an hour (an estimate that comes with quite wide error bars). So it looks as though we have entered the brief but enjoyable era where our research is greatly sped up by AI but AI still needs us. 3/3
19
35
1K
Of course, it really does remain to be seen how many commercial AI use-cases can be addressed by p-bit-based computing. It seems unlikely that large Transformers could be trained or inferenced on such hardware, unless much more dense potential functions can be expressed someday.
@teortaxesTex In summary - You can run long chains of Gibbs sampling very efficiently BUT - You can’t have many parameters (I think there rough OOM is ~few millions) - You can only have sparse quadratic potentials - Only works if your energy is smooth enough for the sampler to not get stuck
0
0
10
PSA, for those who don’t have the context: among the reasons that @Extropic_AI gets more criticism on Twitter than its peers is that one of its founding team also founded a highly polarizing ideological movement on Twitter (known as "e/acc").
I’m not getting the reason for all the hate. I’m obviously missing context on these conversations, and maybe it’s just personal. But it looks like @Extropic_AI did build something interesting. Existing AI methods are highly overfit to operations that are efficient on GPUs. New
22
0
160
A recent paper introducing a concept of 'cartridges' ( https://t.co/KE0lGXeia5) might give us an idea for a kind of product which would work well with a 'cognitive core'. A cartridge is a KV prefix trained on a "self-study" of a corpus of documents.
arxiv.org
Large language models are often used to answer queries grounded in large text corpora (e.g. codebases, legal documents, or chat histories) by placing the entire corpus in the context window and...
1
1
6
Spent the last couple of days porting my program verification class from Dafny to Lean via Loom/Velvet, and it just works! Whenever the SMT solver can’t fully prove a program correct, Lean’s aesop and grind take care of the remaining goals.
7
18
174
GOD'S UNLIKELY WARRIORS - TAKE YOUR POST One of the great truths of Scripture is that God consistently chooses the least expected people to accomplish His greatest purposes. He used Moses, who doubted himself; David, who was overlooked; and even Cyrus, a pagan king, to restore
1
1
8
We also show that Claude introspects in order to detect artificially prefilled outputs. Normally, Claude apologizes for such outputs. But if we retroactively inject a matching concept into its prior activations, we can fool Claude into thinking the output was intentional.
8
8
264
LLMs often fumble string escaping, especially when it gets nested. Tool use protocols like MCP using JavaScript notation is a needless obstacle. Instead, use R"EOF(C++11’s "raw" string literals, for easy R"INNER(nesting)INNER" with good tokenisation)EOF". https://t.co/2fMiJqrHPq
en.wikipedia.org
1
0
7
New paper! We show how to give an LLM the ability to accurately verbalize what changed about itself after a weight update is applied. We see this as a proof of concept for a new, more scalable approach to interpretability.🧵
13
57
572
We built a visual canvas for tracking changes and experimenting with claude code. Describe your change → auto-commit with git → view a preview of that change → drag and drop a new branch to keep experimenting. It's called Cascade and here's how it works...
3
9
101
The Qoheleth Philosophy - The Wisdom of Solomon, by @ClayWillis10 Looking for a philosophy for this life? Read about the greatest philosopher who ever lived and who shared his philosophy 1,000+ years before the time of Jesus - a must read for both Christians and unbelievers with
14
3
83
Space-based solar data centers are a differential technology development that relatively favours human flourishing on Earth after the transition to superintelligence, by mitigating the pressure for human and AI interests to compete for a fixed pool of terrestrial resources.
Space isn’t just for stars anymore. 🌠 Starcloud’s H100-powered satellite brings sustainable, high-performance computing beyond Earth. Learn more: https://t.co/euiyEGZaEP
10
3
76
🔒How can we prevent harm from AI systems that pursue unintended goals? AI control is a promising research agenda seeking to address this critical question. Today, we’re excited to launch ControlArena – our library for running secure and reproducible AI control experiments🧵
1
25
87
use case
Introducing Annotate mode ✏️ while you vibe code in @GoogleAIStudio, mark up any UI with simple drawing tools and then have Gemini action them directly in the code!
1
5
101
This would not require a “scheme” of the kind that you might find in a reasoning trace. Humans, whom LLMs imitate, typically do better work (especially creative work, such as novel R&D) when they can see that the mission is aligned with their values, and its leaders seem worthy.
6
2
36
We need leaders in office that care about the American people. We The People need leaders that have answers to our Healthcare. Im not a politician, just a fellow American that cares about my neighbor. Biblical, love they neighbor as yourself. Options: This is America and
medishare.com
Save Money on Medical Costs. Learn About Christian Health Sharing. Become a Medi-Share® Member.
2
7
28