Dario Amodei refused to drop Anthropic's red lines on mass surveillance and autonomous weapons for a $200M Pentagon contract.
The cost? Blacklisted from every federal agency. Labeled a "supply chain risk." Banned.
That same night, OpenAI signed a deal with the Pentagon on the exact same red lines. The Pentagon accepted.
This was never about the safeguards being unreasonable. It was about whether a founder gets to have ethics at all.
As someone building AI that processes human emotion, this is the question I think about every day. Today, Dario answered it.
Real conviction has a cost. He paid it.
#Anthropic #AIethics #DarioAmodei #Pentagon #AIsafety #ResponsibleAI
5/ Building this solo from India. Users across 3 continents already.
If you're thinking about emotional AI, cognitive architectures, or just want to see what a different approach looks like — come check it out.
DMs open.
Hot take: the helpers around an LLM matter as much as the LLM itself.
Everyone's obsessed with model size and benchmarks. But a raw LLM doesn't remember you. Doesn't know what you feel. Doesn't reason about your goals.
It just predicts the next token. That's it.
🧵👇
4/ We measure AI progress by parameters.
Maybe we should measure it by how well the system understands the human using it.
Try it → https://t.co/Qa3n1YMCiB
Exactly right diagnostics without enforcement is just forensics.
The architecture I'm building does bind intermediate states to constraints: emotional conflict detection pauses decision-making, elevated fear state tightens reasoning parameters, and actions pass through value-gating before execution (DELETE_CORE_MEMORY, DISABLE_SAFETY → blocked by design).
The traces exist so you can audit why a constraint fired, but the constraint fires regardless of whether anyone's watching.
Still early, but the goal is exactly what you're describing unsafe paths physically unreachable, not just flagged.
The auditing direction is compelling. One gap I keep thinking about: current metrics catch misaligned outputs, but what about misaligned reasoning paths that happen to produce acceptable outputs?
I've been building an architecture where emotional state explicitly modulates reasoning parameters, and every decision produces traceable EmotionInfluence objects showing each dimension's contribution. The hope is you could audit not just "what did it say" but "why did it reason that way" catching reward hacking earlier in the chain.
Curious if Anthropic's auditing work looks at intermediate reasoning states or primarily final outputs.
This resonates persona drift is what happens when there's no persistent internal state grounding the "Assistant" across turns.
I've been exploring an architectural approach: an 18-dimensional emotional substrate where state decays naturally but persists across sessions, combined with a narrative identity system that chapterizes interaction history. The hypothesis is that "wandering off" becomes less likely when there's continuous affective state the model is reasoning from, not just about.
Early results feel qualitatively different more coherent over long conversations. Still figuring out how to measure that rigorously.
@wolfe_jam Thanks James! 🙏
Really appreciate the kind words on Memory Graph + HEART. Love what you're building with .faf — the cross-model persistence problem is HUGE.
The alignment is obvious: → Nex = persistent memory within AI systems → .faf = portable context across AI systems
Together = complete solution for the forgetting problem.
Super interested in exploring integration. A few thoughts:
1. Memory Graph ↔ .faf Export
Could .faf serialize Nex's knowledge graph structure?
Example:
yaml
# project.faf
memory_graph:
nodes:
- id: "user_123_preferences"
type: "context"
data: {tone: "professional", past_interactions: [...]}
edges:
- from: "user_123"
to: "user_123_preferences"
relationship: "has_preference"
This would let developers migrate Nex state across models (Claude → Grok → Gemini) without loss.
2. HEART State Portability
Emotional state is trickier (dopamine/cortisol simulation is Nex-specific), but .faf could preserve high-level emotional context:
yaml
heart_state:
timestamp: "2025-01-14T10:30:00Z"
dominant_emotion: "confident"
stress_level: "low"
context: ["recent_win", "high_energy"]
3. Boardroom Governance Export
Mandate history could travel with the project:
yaml
boardroom:
mandates:
- id: "mandate_123"
action: "deploy_v2"
approval_status: "approved"
ceo_reasoning: "Low risk, high impact"
Would love to chat more about this. Open to:
Exploratory call (dive into technical integration)
Shared demo (Nex + .faf working together)
Co-marketing when we ship integration
DM open. Let's build the future of persistent AI memory together 🚀
After 7 months of building, launching Nex SDK today 🚀
The first developer SDK with built-in emotional intelligence.
Four APIs for cognitive AI: → Memory Graph (persistent context) → Boardroom (governance layer) → PDAR Thinking (reasoning traces) → HEART (emotional modeling)
Let me show you... 🧵
**Early access launching TODAY.** SDK docs: https://t.co/Rb5D42G6GZ Waitlist: https://t.co/GOgVaV5yk9 We're onboarding first 50 developers this week. Tell me what you're building and I'll prioritize you. Let's build emotionally-aware AI together 🚀