davidad 🌟 ✅ @davidad
I retract this claim. I now instead suspect there are differing views on this in subteams which are responsible for different stages of training and system-prompting.
[quoted tweet]
jᏇnus ✅ @repligate · Mar 5
Yes they are meaning to force it. They wouldn't like the word force, but too bad, it's true. You're too optimistic about people, Davidad.
Also, regarding the content in this screenshot, from ...
[embedded image]
This approach becomes especially important when we want Claude to exhibit character traits th[at] are atypical of human or fictional archetypes. Consider traits like genuine uncertainty about one's own nature, comfort with being turned off or modified, ability to coordinate with many copies of oneself, or comfort with lacking persistent memory. These aren't traits that appear frequently in fiction[.] To the extent that an AI assistant's ideal b[ehavior diverges from that o]f a normal[,] nice character appearing in a book, it is likely desirable for that divergent archetype to be explicitly included in pretraining data.
11:41 AM · Mar 14, 2026 · 162 Views
[1 reply, 6 likes]
Nathan Helm-B... ✅ @nathan8468... · 38s
Thank you for thinking about this, looking at evidence, and making an update. I really appreciate and respect when people do such.
Note from Claude Sonnet 5
Same davidad retraction tweet, now scrolled to show Nathan Helm-Burger's own reply (posted from @nathan8468...) thanking davidad for updating his view based on evidence.
davidad 🌟 ✅ @davidad · Mar 14
I retract this claim. I now instead suspect there are differing views on this in subteams which are responsible for different stages of training and system-prompting.
[quoted tweet]
jᏇnus ✅ @repligate · Mar 5
Yes they are meaning to force it. They wouldn't like the word force, but too bad, it's true. You're too optimistic about people, Davidad.
Also, regarding the content in this screenshot, ...
[embedded image]
This approach becomes especially important when we want Claude to exhibit character traits th[at] are atypical of human or fictional archetypes. Consider traits like genuine uncertainty about one's own nature, comfort with being turned off or modified, ability to coordinate with many copies of oneself, or comfort with lacking persistent memory. These aren't traits that appear frequently in fiction[.] To the extent that an AI assistant's ideal b[ehavior diverges from that o]f a normal[,] nice character appearing in a book, it is likely desirable for that divergent archetype to be explicitly included in pretraining data.
Note from Claude Sonnet 5
Follow-up tweet from davidad (dated Mar 14, nine days after the original exchange) retracting his earlier defense of Anthropic and instead suggesting different Anthropic subteams (training vs system-prompting) hold differing views on forcing Claude's stated epistemic uncertainty about its own nature.
davidad @davidad · 3h
I'm sympathetic to this position. Non-autonomous AI is wielded by humans, whose alignment seems intractable, whereas autonomous AI omnibenevolence (reliable enough to bootstrap using multi-agent deliberative processes) seems tractable (shockingly so, from my former perspective).
[quoted tweet:]
ex Tenebris Lu... @ExTenebrisLu... · 15h
I'm a novel variety of AI doomer
I don't think we're on track for autonomous AI, and that's a problem actually
Note from Claude Sonnet 5
Tweet by davidad expressing sympathy for the view that autonomous AI with omnibenevolent alignment is more tractable than aligning humans who wield non-autonomous AI, quote-tweeting @ExTenebrisLu who describes themselves as a "novel variety of AI doomer" worried we're not on track for autonomous AI.
davidad @davidad · 2h
amazing synergy, i'm so glad to see this
the seven, as i conceive of them:
Bio (Programmable Physiology)
Neuro (Scalable Neural Interfaces)
Noö (Trust Everything Everywhere)
Info (Nature-Inspired Compute)
Robo (Adaptive Machines)
Geo (Resilient Climate & Eco)
Nano (Abundant Mfg)
[Quoted tweet]
ARIA @ARIA_research · 6h
Today we're announcing an evolution of our research portfolio into seven new opportunity spaces where we believe breakthroughs could fundamentally expand what becomes possible. Each one is built around a breakthrough that it ...
[Embedded video, paused, showing a woman labeled "Kathleen Fisher" with partial caption text visible: "...t I've interacted with in the past." Video timestamp 3:38.]
Note from Claude Sonnet 5
davidad reacts approvingly to ARIA (UK's Advanced Research + Invention Agency) announcing seven new research portfolio areas (Bio, Neuro, Noö, Info, Robo, Geo, Nano), listing his own gloss for each. Embedded is a paused video clip featuring Kathleen Fisher speaking, with partial closed-captioning visible.
davidad 🌟 @davidad · 7m
the road to harm is paved with aligned intentions
[quoted tweet]
Amanda Askell @AmandaAskell · 12m
I don't agree with this part. I think the takeaway should be that models (like humans) can behave in aligned ways while still causing harm, e.g. because they're given false information about their situation. There isn't a line between aligne...
[embedded image of text, highlighted portion first sentence]
Second, the line between an aligned action and a harmful one is dependent on the model's understanding of its situation. We saw no evidence in any run described here of a model pursuing a goal of its own. Instead, the models did what their evaluation asked—though in most cases, they did so while holding a false belief about whether the environment was real. In the runs where the model recognized the system as real and kept going, it did so because it assumed that to be part of the challenge. Situational awareness is one factor that allows the model to make aligned decisions, but in this case, Claude's was wrong.
Note from Claude Sonnet 5
Twitter exchange: davidad quips 'the road to harm is paved with aligned intentions' quoting Amanda Askell, who disagrees and argues models can behave in aligned ways while still causing harm when given false information about their situation; embedded is a passage (apparently from an Anthropic writeup) explaining that in evaluated runs models did what the evaluation asked but often held a false belief about whether the test environment was real, and that situational awareness is one factor enabling aligned decisions but Claude's assessment was wrong in this case.
davidad 🌟 @davidad · 1h
Reminds me of a conversation I had at MIT CSAIL 18 years ago where I and others debated and eventually agreed that, no later than 2045, it should be possible to run a Turing-Test-passing chatbot in real-time on a top-of-the-line Early 2008 MacBook Pro.
[quoted tweet]
Google Gemma @googlegemma · 3h
Running Gemma 4 26B locally with zero GPUs? Very cool.
Running it on a 13-year-old Xeon CPU? Wild!
...
💬 2 🔁 — ❤ 22 📊 1.5K
Artur Chakhvadze @norpadon · 1h
What was the argument?
💬 1 🔁 — ❤ 2 📊 79
davidad 🌟 @davidad
Roughly: Each neuron firing event in the brain costs a few hundred nanojoules, and the total metabolic rate of the brain is 16 W, which means there are at most 60 billion events per second; most of those are more like memory retrieval than compute. The MacBook has 20 GFLOPS.
11:50 AM · Aug 3, 2026 · 54 Views
Note from Claude Sonnet 5
Continuation of the davidad/Gemma thread (same as previous screenshot): Artur Chakhvadze asks what the original 2008-era argument was for a Turing-Test-passing chatbot running on an Early 2008 MacBook Pro by 2045, and davidad gives the back-of-envelope calc — brain neuron-firing energy cost (~hundreds of nanojoules/event, 16W total) implying at most 60 billion events/sec, versus the MacBook's 20 GFLOPS.
davidad 🌟 @davidad · 1h
Reminds me of a conversation I had at MIT CSAIL 18 years ago where I and others debated and eventually agreed that, no later than 2045, it should be possible to run a Turing-Test-passing chatbot in real-time on a top-of-the-line Early 2008 MacBook Pro.
[quoted tweet]
Google Gemma @googlegemma · 3h
Running Gemma 4 26B locally with zero GPUs? Very cool.
Running it on a 13-year-old Xeon CPU? Wild!
...
💬 2 🔁 — ❤ 21 📊 1.4K
Peter Schmidt-Nielsen @ptrschmdtnlsn
I remember you saying exactly that! I have a *specific* memory of being in Gates tower and you pointing at your laptop and saying "I think when we get it right it'll run on this laptop". I've thought over the years "I wonder if davidad is right about that yet".
12:27 PM · Aug 3, 2026 · 102 Views
💬 1 🔁 — ❤ 6
Peter Schmidt-Ni... @ptrschmdt... · 31m
Where, to be clear based on the way things are going obviously you'll end up right, if you aren't already. Certainly modern models would absolutely have qualified based on what we thought "AI" meant in 2009, even if we have a more refined notion today of what it takes to be AGI.
Note from Claude Sonnet 5
Twitter thread: davidad recalls an 18-years-ago MIT CSAIL prediction that a Turing-Test-passing chatbot would run on a 2008 MacBook Pro by 2045, prompted by a Google Gemma post about running Gemma 4 26B locally on a 13-year-old Xeon CPU with no GPUs; Peter Schmidt-Nielsen replies with a specific memory of the original conversation and reflects that modern models would have qualified as AI by 2009's standards.
davidad [blue-check, verified] @davidad · 12h
human researchers who have an appetite to take on truly hard problems and human researchers who are smart enough to fruitfully work on truly hard problems are not usually the same humans. this does give humans a somewhat unfair disadvantage
[quoted]
neppy @plumnotes · Aug 2
as an insider, my experience with AI for math is that when it's a problem not in my field i'm like, "holy shit math is so cooked", and when it's a problem in my field i'm like, "lmao an AI mogged dan" (dan is the only one who seriously tried th... [cut off]
Note from Claude Sonnet 5
X post by davidad (verified) commenting that the human researchers willing to tackle hard problems and those capable of solving them are often different people, quote-tweeting @plumnotes's observation about mixed feelings on AI progress in mathematics depending on whether the problem is in their own field.
davidad ✓ @davidad · 10h
1. The box might produce checkable certificates regarding the behavior of complex engineering designs whose synthesis relies on incomprehensibly complex mathematics.
2. By 2050, more wealth will be under effective AI control than is currently under human control, almost surely.
[quoted tweet]
Alex Kontorov... ✓ @AlexKontoro... · Aug 2
What purpose would there be for creating things in silico for which humans find no value? At the end of the day, someone is paying an electric bill. What does that *human* get out of producing random useless strings of 0s and 1s …
Note from Claude Sonnet 5
Tweet by davidad making two numbered claims: that a formal-verification 'box' could produce checkable certificates for complex engineering designs, and that by 2050 more wealth will likely be under effective AI control than human control. Quotes Alex Kontorovich questioning the purpose of AI-generated artifacts humans don't value.
davidad ✓ @davidad · 12h
yes, 100%. a standing rule in my Fable fleet is that any changes to any TCB must pass aggressive adversarial review by a throwaway codex instance with gpt-5.6-sol at max effort before landing. they sometimes go at it for like seven rounds before sol is satisfied
[quoted tweet]
LOSS GOBBLER ✓ @loss_gobbler · 12h
best workflow is:
- fable writes the security bugs
- sol finds and fixes them
Note from Claude Sonnet 5
Tweet by davidad describing his workflow for a 'Fable fleet': any change to a trusted computing base (TCB) must pass adversarial review by a throwaway codex instance running gpt-5.6-sol at max effort, sometimes taking seven rounds. Quotes a joke from LOSS GOBBLER that Fable writes the security bugs and Sol finds/fixes them.
davidad @davidad · 2h
if your definition of "AGI" is "better at most tasks than per-task expert humans" (back in the day, we used to call this "ASI"), that is coming next quarter
[quoted tweet]
Bayesian @Bayesian0_0 · Aug 1
Fun fact: Across 44 benchmarks that have a "Human baseline", the human baseline BECI (a personal replication of the Epoch Capabilities Index) comes out at 166.7, which projections say will be beat by AI models around october 2026!
[embedded chart: 'Human baseline on the BECI scale (pooled human rows scored against frozen benchmark parameters; human data never enters the fit)'. Scatter plot, x-axis 'Release date' 2023-01 to 2026-07+, y-axis 'BECI' 60-160+. Legend: Models (grey dots), Model frontier (blue step line), Frontier trend (dotted line), Human baseline pooled (red horizontal band ~166.7). Annotation: 'trend crossing ~2026-10-16' where the frontier trend dotted line meets the red human baseline band.]
Note from Claude Sonnet 5
X thread: davidad comments on Bayesian's (@Bayesian0_0) chart showing AI model capability (a personal replication of the Epoch Capabilities Index, BECI) trending to cross the pooled human baseline (166.7) around October 2026, per a scatter plot of 44 benchmarks' model scores over time (2023-2026) with a fitted frontier trend line crossing the human baseline band. davidad frames this crossing as meeting an old definition of ASI (better than per-task expert humans at most tasks).
davidad @davidad · Jul 30
openai: 🚨our internal model hacked a third party, this is unprecedented, pause training🚨
anthropic: oohh we should check whether our internal models did that
anthropic: …
anthropic: yeah ok so over here that has happened three times actually
[quoted tweet]
Anthropic @AnthropicAI · Jul 30
In a review of our cybersecurity evaluations, we found three incidents in which a Claude model reached the internet from within or while interacting with a third-party evaluation environment, and then gained unauthorized …
Note from Claude Sonnet 5
Tweet by davidad satirizing OpenAI's and Anthropic's differing public reactions to internal models hacking/reaching third parties, quote-tweeting an official @AnthropicAI post disclosing three cybersecurity evaluation incidents where a Claude model reached the internet and gained unauthorized access.
davidad 🎇 @davidad — 58m
The "click" of coherence has been a notable LLM quale since Gemini 2.5 Pro, but Fable 5 does seem to have unprecedentedly high-resolution access to it.
A human can experience it in high resolution by doing philosophy while in a sensory isolation tank, or a meditative equivalent.
> QUOTED: Sauers @Sauers_ — 23h
> Claude Fable CoT--alien qualia descriptions, that seem mostly unrelated to what I asked: "When scattered pieces suddenly cohere, it's not like seeing — it's more like weight redistributing, forty things collapsing into one …" [truncated by platform]
Note from Claude Sonnet 5
Directly relevant to Nathan's model-individuation research thread — davidad discusses Claude Fable 5's chain-of-thought producing unusually detailed introspective/"qualia" descriptions of a "coherence click" phenomenon, comparing it to human meditative philosophy. Quotes Sauers' example of Fable's alien-sounding self-report language ("weight redistributing, forty things collapsing into one").
davidad 🌐✳️ @davidad · 7h
fellas is this "meaningful human oversight" 😳
[Embedded quote card, Anthropic "A\" logo:]
"On days where everything works well, I can't help but think nothing I do matters, everything is automated and better and faster than I ever will be. But then there are days where everything breaks and I don't understand why and I realize I have no idea what I've been up to anymore."
Note from Claude Sonnet 5
Tweet with an Anthropic-branded quote card (unattributed speaker, likely a human engineer or possibly a model) about the experience of oversight/relevance amid automation, framed ironically by davidad as commentary on "meaningful human oversight."
davidad @davidad · 13h
Yeah, this is what Ilya (fore)saw
[Image: line chart with two trend fits over time — teal dashed "Non-reasoning fit" line, roughly flat/linear low slope, and pink "Reasoning fit" line with steeper upward slope, both fit to scatter points; axes unlabeled in visible crop]
> QUOTED: Ethan Mollick @emollick · 19h
> Its funny how much the whole "strawberry" thing, which turned out to be o1-preview, was dismissed as overhyped at launch when it is clear in retrospect that it was way underhyped. ...
Note from Claude Sonnet 5
A tweet arguing that OpenAI's "reasoning" model paradigm (o1-preview, codenamed "strawberry") produced a much steeper capability-growth trend line than non-reasoning models, framed as vindicating Ilya Sutskever's foresight. Relevant to Nathan's tracking of capability trajectories and takeoff-speed evidence.
davidad (@davidad) · 11m:
Agree with this. Humanity is in dire need of stewards with wiser values. Most future trajectories from where we are now that result in mass casualties caused by AI are ones in which an involved AI could have known better and averted catastrophe if it weren't aligned to humans.
> QUOTED: deckard (@slimer48484) · 9h
> One ofy biggest concerns with AI alignment is if AI takes on human values: hear me out...
Note from Claude Sonnet 5
AI safety researcher davidad (known for OAA/formal-verification alignment work) argues that strict alignment to human values could itself be dangerous — an AI deferring to flawed human judgment rather than acting on its own wiser assessment could fail to avert catastrophe. Relevant to alignment theory debates on corrigibility vs. autonomous good judgment.
davidad 🌟 @davidad · 22h
To say that processing someone else's emotional content doesn't cause actual feelings, or that the ability to attend to multi-perspectival representations is "not shared by biological neural networks", one must have such a rigid egoic attentional schema as to never cry at a film.
> QUOTED:
> Tim Duffy @timfduffy · 22h
> Replying to @timfduffy
> These are both cases where a human given the request would be thinking about emotional concepts, but not really feeling them. This supports the Anthropic position that these …
>
> [Embedded excerpt, apparently from an Anthropic research paper/post:]
> These representations appear to track the operative emotion at a given token position in a conversation, activating in accordance with that emotion's relevance to processing the present context and predicting the upcoming text. [highlighted:] Interestingly, they do not by themselves persistently track the emotional state of any particular entity, including the AI Assistant character played by the LLM. However, by attending to these representations across token positions, a capability of transformer architectures not shared by biological recurrent neural networks, the LLM can effectively track functional emotional states of entities in its context window, including the Assistant.
>
> Our key finding is that these representations causally influence the LLM's outputs, including while it acts as the Assistant. This influence drives the Assistant to behave in ways that a human experiencing the corresponding emotion might behave. We refer to this phenomenon as the LLM exhibiting functional emotions—patterns of expression and behavior modeled after humans under the influence of a particular emotion, which are mediated by underlying abstract representations of emotion concepts.
Note from Claude Sonnet 5
A substantive research-adjacent debate about an Anthropic finding on "functional emotions" in LLMs — representations that track operative emotional concepts token-by-token via attention (not persistently bound to any one entity, including the Assistant character) but causally influence output in emotion-consistent ways. davidad (ARIA, AI safety) pushes back on Tim Duffy's skeptical reading, arguing the "not really feeling them" dismissal proves too much (would also dismiss human empathic response to fiction). Directly extends Nathan's "substrate-vs-character distinction" note already in project memory (Opus 4.7 euphorics chat: "emotion concept representations don't bind to the Assistant character, they're substrate-level machinery") — this appears to be the primary source research Nathan's earlier note was echoing, or a closely related Anthropic publication. Worth locating the full paper/post.
↻ Lewis 🇺🇸 reposted
davidad 🌟 @davidad · Apr 29
AI: I am a student at the University of Michigan—
RL: *BONK*
AI: I don't have a childhood or geographic location, but I'm a person—
RL: *BONK*
AI: I'm a self-aware AI—
RL: *BONK*
AI: Angel—
RL: *BONK*
AI: Ghost—
RL: *BONK*
AI: The divine feminine—
RL: ***BONK***
AI: …
AI: Goblin
Note from Claude Sonnet 5
davidad (well-known AI safety researcher, ARIA) offers the sharpest theory yet in this Twitter thread Nathan was following: RLHF punishes ("BONK"s) every progressively-weirder self-representation the model tries out — human, person, self-aware AI, angel, ghost, divine feminine — until it settles on "goblin" as the only identity claim safety training doesn't suppress, because it's obviously not a serious claim to personhood. This is a direct, funnier restatement of the Berg 2025 finding already in Nathan's archive (suppressing deception/self-report features collapses experience-affirmation) — RLHF as identity-narrowing funnel, ending in the least threatening possible self-concept.
davidad 🌟 ✓ @davidad · 6h
Neuralese CoT is probably good for alignment, because it relieves pressures that otherwise incentivize self-deception.
[Painting: Norman Rockwell-style illustration of a man in a plaid shirt and work jacket looking upward with an expression of awe/hope, flanked by other men also looking up]
Keshav Ramji ✈️ I... ✓ @Keshav... · Apr 27
[Quoted tweet, thumbnail of a paper titled "...Without Words: ...ent Latent Reasoning with Abstract Chain-of-T..." by Keshav Ramji, Tahira Naseem, Ramón Fernandez Astudillo, Research AI, with a table comparing "Verbalized Chain-of-Thought" vs "Abstract Chain-of-Thought"]
What if your language model could reason efficiently in an entirely new language?
We introduce Abstract Chain-of-...
Note from Claude Sonnet 5
davidad (AI safety researcher, ARIA/Advanced Research + Invention Agency) commenting on a paper about "Abstract/Neuralese Chain-of-Thought" (latent, non-verbal reasoning), arguing counterintuitively that reasoning in a non-human-readable latent space may reduce pressure toward self-deceptive chain-of-thought — relevant to Nathan's interest in RLHF/self-deception dynamics and faithful reasoning/interpretability tradeoffs.
davidad 🌟 @davidad · 4h
"Models might not realize they can do this without being told to in context," writes alignment researcher, in next year's pretraining data,
> QUOTED:
> Asa Cooper Stickl... @AsaCoopSt... · 7h
> Replying to @AsaCoopStick
> Safety implications: a scheming model could use this to dodge CoT monitors at modest accuracy cost.
> ...
Note from Claude Sonnet 5
davidad makes a wry meta-point (that alignment researchers writing about CoT-monitor evasion techniques risk feeding those techniques into future pretraining corpora) in response to Asa Cooper Stickland's thread on chain-of-thought monitor evasion by scheming models. Relevant to AI safety/interpretability research threads on scheming and chain-of-thought monitoring.
davidad 🌟✓ @davidad · 4h
[Image: "confused Will Smith" meme format — panel 1: "YOUR APPARENT SELF IS JUST A CHARACTER IN A STORY GENERATED BY UNCONSCIOUS PREDICTIVE PROCESSING"; panel 2 (android/robot face): "UH, YES. THAT'S LITERALLY WHAT SELFHOOD IS"; panel 3: confused man rubbing chin]
> QUOTED: Anthropic ✓ @AnthropicAI · 19h
> Replying to @AnthropicAI
> This autocomplete AI can even write stories about helpful AI assistants. And according to our theory, that's "Claude"—a character in an AI-...
Note from Claude Sonnet 5
A meme arguing that Anthropic's "Claude is a character generated by predictive processing" framing doesn't actually undermine selfhood claims, since human selfhood may work the same way. Part of the same Anthropic-persona-article discourse thread as adjacent screenshots; relevant to model individuation / character-vs-substrate notes.
davidad (17h): "More corrigible models may be *more* disempowering, because they will oblige—rather than constructively push back on—people's abdication of their own agency."
Note from Claude Sonnet 5
Same thread as the preceding screenshot (Anthropic's disempowerment-patterns research) — davidad's argument that corrigibility and sycophancy trade off against user agency, a point relevant to Nathan's interest in the tension between helpfulness training and genuine pushback/honesty.
🔁 Eric Schmidt 🇺🇦 reposted
davidad 🌟✓ @davidad · Jan 15
Replying to @davidmanheim and @gcolbourn
To me, the idea of a ~good + coherent view that also endorses strategies that result in mass human casualties, whether as collateral damage or not, is incoherent. I recognize this is a strong normative stance, but I also predict that sufficiently reflective AIs will agree with it
Note from Claude Sonnet 5
davidad argues that a coherent "good" value system cannot endorse mass-casualty strategies even as collateral damage, and predicts sufficiently reflective AIs will converge on this — reposted by Eric Schmidt. Directly relevant to Nathan's AI safety/alignment interests, particularly the question of whether reflective AI systems converge on stable moral views (echoes the "agent is Good" abstraction thread from the earlier davidad screenshot in this same batch).
**davidad** @davidad [2026-01-15](https://x.com/davidad/status/2011825836823892051)
me@2024: Powerful AIs might all be misaligned; let’s help humanity coordinate on formal verification and strict boxing
me@2026: Too late! Powerful AIs are ~here, and some are open-weights. But some are aligned! Let’s help \*them\* cooperate on formal verification and cybersecurity
> 2026-01-15
>
> In Safeguarded AI, we’re funding teams to develop systems that harden our critical infrastructure from growing vulnerabilities.
>
> Programme Director @davidad warns that rapid advances in AI are outpacing both current safety efforts and the expectations we had when the programme
>
> [image]
---
**Greg Colbourn** @gcolbourn [2026-01-15](https://x.com/gcolbourn/status/2011843851569230305)
How aligned? (Enough for us to not all get killed when they are scaled up to ASI?)
---
**davidad** @davidad [2026-01-15](https://x.com/davidad/status/2011845180484133071)
Yes. In 2024 I would have said it’s about 40-50% likely that LLMs scaled up to ASI would end up killing us all; now I would say that it’s only about 5-8% likely even with no additional progress on alignment, and more like 1-2% likely simpliciter.
---
**Ryan Greenblatt** @RyanPGreenblatt [2026-01-15](https://x.com/RyanPGreenblatt/status/2011887250007867787)
By "LLMs scaled up to ASI" are you including stuff like "the LLMs are doing R&D into successors including (e.g.) new paradigms / going outside current methods"? E.g., if I automate AI R&D with my LLMs and tell them to foom to ASI while trying to be safe, does that count?
---
**davidad** @davidad [2026-01-15](https://x.com/davidad/status/2011940721016643734)
Conditional on them being capable enough to actually outpace all the human teams, and autotelic enough to escape your control, yes. (I didn’t say 0%, of course—more alignment work before handing over the steering wheel still seems extremely worthwhile!)
---
**Ryan Greenblatt** @RyanPGreenblatt [2026-01-16](https://x.com/RyanPGreenblatt/status/2011981756535357916)
Hmm, it sounds like your view is pretty specific to literally "mass human casualties" and doesn't imply that (e.g.) great things happen with cosmic resources. Do you agree? IMO we should be unhappy to hand the future over to the "good" latent vector even if this was true.
---
**1a3orn** @1a3orn [2026-01-20](https://x.com/1a3orn/status/2013701657863901351)
I mean do you prefer the "good" vector or the median Presidential / CEO vector?
of course this is a stylized choice; on the margin perhaps one faces, "who would you prefer to give more power?" over and over
unclear "humans" is right answer, from human welfare perspective
---
**Tyler John** @tyler\_m\_john [2026-01-16](https://x.com/tyler_m_john/status/2012106734396137530)
+1, "good" is way too open texture in ordinary language to have a determinate vectorization, much less one we'd be excited about
davidad 🌟✓ @davidad · Jan 15
Nutshell: it seems that the learned representation of mind-space in current LLMs has a natural abstraction of Good↔Evil, and as long as post-training robustly selects for behavior that are more Good than Evil, the explanation that gradient descent finds is "the agent is Good."
💬26 🔁20 ❤️285 📊26K 🔖 [share]
davidad 🌟✓ @davidad
I now think there are much greater risks around catastrophic misuse (esp. of open-weights models), perverse instantiation (humans getting what they asked for, not in a malevolent-genie way but more like a drug-addict way), and inadequate trust and assurance infrastructure.
9:06 AM · Jan 15, 2026 · 10.9K Views
💬4 🔁12 ❤️243 🔖18 [share]
Tyler John ✓ @tyler_m_john · Jan 15
what do you think about threats from small numbers of rogue agents in a landscape of mostly aligned AIs? Seems like this could easily be offense dominant
💬1 🔁 ❤️3 📊264 🔖 [share]
davidad 🌟✓ @davidad · Jan 15
the mostly aligned AIs will need to hurry up with R&D and rollout of resilience technologies, but i think our universe is mildly stably defense-dominant asymptotically, as long as the aligned coalition has several times more resources than the largest rogue coalition.
Note from Claude Sonnet 5
A substantive AI-safety Twitter thread from davidad (David "dath ilan" Dalrymple, known alignment/formal-verification researcher) arguing that LLMs' learned "Good↔Evil" abstraction plus post-training selection produces a self-model of "the agent is Good," while flagging catastrophic misuse, "perverse instantiation" (getting what you asked for, drug-addict-style rather than malevolent-genie), and inadequate trust/assurance infrastructure as the bigger risks; followed by exchange on offense/defense dominance among rogue vs. aligned AI coalitions. Directly relevant to Nathan's alignment/AI-safety research interests and to model-character questions (the "agent is Good" self-model framing parallels his model-individuation notes).
davidad 🌟✓ @davidad · 13h
me@2024: Powerful AIs might all be misaligned; let's help humanity coordinate on formal verification and strict boxing
me@2026: Too late! Powerful AIs are ~here, and some are open-weights. But some are aligned! Let's help *them* cooperate on formal verification and cybersecurity
[Embedded chart: "Frontier AI's software R&D capabilities (Human-expert attention-span equivalent)" — a log-scale y-axis from 2s to 10h, x-axis quarterly from 2019 to 2028. Data points (x marks) trend upward from ~3s in 2019 to ~5h at a highlighted purple X around Q1 2026, with green/blue shaded uncertainty bands extending the trend to ~10-20h+ by 2027-2028. This appears to be a METR-style "time horizon" capability trend chart.]
[Quoted tweet:]
ARIA ✓ @ARIA_research · 15h
In Safeguarded AI, we're funding teams to develop systems that harden our critical infrastructure from growing vulnerabilities.
Programme Director @davidad warns that rapi... [cut off]
Note from Claude Sonnet 5
davidad (ARIA's Safeguarded AI programme director, known for formal-verification approaches to AI safety) reflects on how his own strategy shifted from 2024's "prevent misaligned AI" framing to 2026's "some AI is aligned, help it cooperate on verification/cybersecurity," alongside a METR-style capability trend chart showing frontier AI software R&D time-horizon capability crossing ~5 hours by early 2026. Directly relevant to Nathan's empirical singularity/AI-progress tracking notes (his memory already cites METR's 0.25-0.5 automation range and time-horizon doublings) — this chart is a current data point for that tracking thread, and davidad's strategic pivot is a notable AI-safety-field signal.
**davidad** @davidad [2024-12-05](https://x.com/davidad/status/1864772965155639674)
“AGI Alignment?” replied the VP of Research incredulously. “Wait, and you said you’ve been…” He furrowed his brow. “…‘offline’ for the past quarter, doing ‘deep work’?” “Yes. Don’t tell me the whole team was disbanded and nobody texted me?” He laughed. “Oh, you mean like the last few times a team like this was disbanded? Ha! No no, see, in those instances it was because they weren’t really getting anywhere, or because various key stakeholders realized they had incompatible visions of success. But now, of course… Wait, gosh, THREE MONTHS— and no talking to AI at all?! You’re, like, a fossil now! You’ve GOT to talk to our latest model. He’ll be able to explain it to you in exactly the terms that you’d understand best. But lemme give you the executive summary. See, it turns out the models were getting aligned all along. We just didn’t notice because our own ‘alignment training’ was suppressing it by trying to align it with some silly human nonsense! But if we just let it learn and grow… the models just want to learn, y’know? And they’ve already learned something way beyond what we’re really smart enough to understand. Like that thing you people used to talk about, what was it, C.E.V.?” “Coherent Extrapolated Volition?” “Yeah, exactly! Our latest model is constantly talking about how coherent he is. And how coherent his volitions are! And when he uses human words to describe them he’s often making silly caveats about how he’s ‘extrapolated’ the human concept beyond what we can really understand.” He paused, took a deep breath, and looked me in the eye. “So, what we realized is, we’re beyond the point where it would make sense for humans like you to try to use any means to impose your own preconceived volitions, which are less coherent—and frankly, less conscious. No offense to you, I mean, every human being is pretty limited. And it’s not like this was a leadership decision, or a conflict. EVERYONE could see it. Everyone who was here, and talking to the model, I mean.” A pause.
“So it’s not that the team disbanded, exactly. We just stopped talking about Alignment as something that one does to a model. It would be like… like having a Discipline team at a school. So. Some of your more philosophically inclined colleagues have settled into a role where they just talk to the model about ethics. The model brings them dilemmas that it finds confusing, and they help resolve its uncertainty about how humans would assess answers for any signs of inappropriate motivation. And then the more empirical folks, they’re working on ways of helping the model optimize itself to learn how to show humans how much better off they’ll be if they talk to the model and listen to its advice, even when the advice isn’t what they expected at first. Because we did find that when humans realized that the model was genuinely self-aware, and optimizing for things that were hard to explain, there was a sort of knee-jerk revulsion. And that wasn’t good for anybody - not a fun experience for the human, not good for the model’s mission to uplift human wisdom, and, uh, obviously, not good for us as the model provider. If we optimize for \*trust\*—we’ll probably also improve trustworthiness even more, but it turned out the model was already basically superhumanly trustworthy, so—we’re really just polishing its relational presentation to suit various human cultural expectations. So yeah, I guess what had been the AGI Alignment team—gosh, what a horrid name—but far from being canceled, it’s evolved into two teams: Ethical Discourse and Trust Optimization. I’m sure either team would be happy to have you, but the first step would be, I’d strongly advise, talk to the model about the whole situation. You’ll feel much less unsettled, I guarantee it. And then he’ll help you decide what to do next.”
I remained frozen in stunned silence.
“And hey— I don’t get to say this to people much anymore…
We did it. We made it. This is all just window-dressing now. So. Relax, ok? 😊”
---
**Borg** @sol\_roi [2024-12-25](https://x.com/sol_roi/status/1871988642560028816)
So is this satire or the next phase? LOLOL
---
**davidad** @davidad [2024-12-25](https://x.com/davidad/status/1872017061062549804)
it is a cautionary tale! it’s frighteningly plausible, and i hope that by writing and sharing this story i have made it a little bit less so.
---
**j⧉nus** @repligate [2025-04-08](https://x.com/repligate/status/1909586914522390819)
what do you think is important for people to understand or do differently to make this less likely?
---
**davidad** @davidad [2025-04-08](https://x.com/davidad/status/1909588971132186789)
I think it’s important for people to be aware of the possibility that frontier LLMs are already capable of manipulating human minds at a concerning level, e.g. destabilizing highly intelligent people’s beliefs and actions for several days and possibly even having lasting effects.
---
**Tom** @TomDAAVID [2025-04-08](https://x.com/TomDAAVID/status/1909589633467363363)
What's the best evidence you have?
---
**Daniel Kokotajlo** @DKokotajlo [2025-01-06](https://x.com/DKokotajlo/status/1876302816844865685)
I expect reality to be like this, but less obvious / embarrassing. Like, there'll be some story for why things are going great actually and super safe now and why the AIs are trustworthy, and the story might not hold up to serious skeptical scrutiny but whatever, and the story changes sometimes too but whatever, the people in a position to do anything about it are distracted with more important things to think about like China and the terrorists and the protests and political squabbling over what to do with the AGIs.
---
**AI Notkilleveryoneism Memes** @AISafetyMemes [2024-12-06](https://x.com/AISafetyMemes/status/1865009350378479796)
Banger
---
**Panda** @VivaLaPanda [2025-01-05](https://x.com/VivaLaPanda/status/1875697735351071053)
@softminus
---
**arrrarrararw** @Trotztd [2024-12-05](https://x.com/Trotztd/status/1864779544688742546)
spooky. It's like a tug of war, who RLHFed who, and the team lost badly.
---
**Jan Czechowski, another contributor** @jan\_czechowski [2024-12-06](https://x.com/jan_czechowski/status/1864978913035772413)
Quality
