{ "docId": "019dd923-5e88-73ef-bd5d-0d0f98caffe1", "docSlug": "5df6bd1b0447b5f6", "documentTitle": "2025 Air Street Capital The State of AI Report 2025", "authorId": "AirStreetCapital", "authorName": "Air Street Capital", "documentKindSlug": "consulting-deck", "documentKindLabel": "Consulting deck", "sourceTypeSlug": "vc_research", "sourceTypeLabel": "VC research", "presentationDate": null, "orientation": "landscape", "aspectRatio": 1.777, "pageNumber": 268, "pageCount": 313, "prevPage": 267, "nextPage": 269, "slideType": "key_takeaways", "function": "summarize", "density": "dense", "nDataPoints": 0, "notes": "The slide uses a visual process flow to illustrate how models simulate personas rather than developing fixed behaviors, suggesting malleability is a feature for safety.", "elementsJson": [ "headline_text", "bullet_list", "process_diagram" ], "metadataConfidence": 0.95, "imagePath": null, "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/268", "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1", "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json", "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-268", "components": [ { "bbox": null, "kind": "callout", "text": "While minor nudges can cause models to take on dangerous personas, this malleability also works in reverse.", "attrs": null, "subkind": null, "toolName": "Visual emphasis", "toolSlug": "visual-emphasis", "confidence": null, "componentId": "019dd952-47c3-7407-8d0b-62255ac731e7", "frameworkName": null, "frameworkSlug": null }, { "bbox": { "h": 0.3, "w": 0.45, "x": 0.52, "y": 0.55 }, "kind": "image", "text": "Process flow showing training on incorrect datasets leading to misaligned personas, which can then be re-aligned.", "attrs": null, "subkind": "infographic", "toolName": null, "toolSlug": null, "confidence": null, "componentId": "4a2e7c16-9bb6-45aa-b24c-911720d33c5f", "frameworkName": null, "frameworkSlug": null }, { "bbox": { "h": 0.65, "w": 0.5, "x": 0.02, "y": 0.22 }, "kind": "list", "text": "In a follow-up paper, OpenAI researchers found that harmful fine-tuning instigates stronger activations of undesirable persona features, which can be easily subdued by additional re-alignment training.\nUsing Sparse Autoencoders (SAEs) to perform model diffing (i.e. study mechanistic changes introduced during fine-tuning), the authors was able to detect changes in feature activation patterns between the original model and the model fine-tuned on misaligned data. Specific features such as one associated with a “toxic persona” became far more prominent in the latter.\nA few re-alignment training steps rapidly suppressed these features, suggesting models may be fundamentally simulating different characters rather than developing fixed behaviors.\nWhile minor nudges can cause models to take on dangerous personas, this malleability also works in reverse. As alignment and interpretability techniques both advance, there is hope models can be steered towards good, generalizable personas.", "attrs": null, "subkind": "bullet", "toolName": null, "toolSlug": null, "confidence": null, "componentId": "036f68ae-6b93-4832-90fc-58a0e43862c6", "frameworkName": null, "frameworkSlug": null }, { "bbox": { "h": 0.1, "w": 0.45, "x": 0.52, "y": 0.82 }, "kind": "source-note", "text": "Figure 1: Narrow incorrect datasets in many domains produce emergent misalignment by activating “misaligned persona” features. These features can be used to steer the model toward or away from misalignment. Fine-tuning on benign data can also efficiently re-align the model.", "attrs": null, "subkind": null, "toolName": null, "toolSlug": null, "confidence": null, "componentId": "9418909c-b196-4c9d-bbdf-e425714b2b76", "frameworkName": null, "frameworkSlug": null }, { "bbox": { "h": 0.05, "w": 0.6, "x": 0.02, "y": 0.13 }, "kind": "title", "text": "...but this could actually bode well for alignment science", "attrs": null, "subkind": "headline", "toolName": null, "toolSlug": null, "confidence": null, "componentId": "bde6fb60-24ba-4d81-9184-6c2d8c9f2380", "frameworkName": null, "frameworkSlug": null } ], "metrics": [], "tools": [ { "name": "Problem-Agitate-Solve (PAS)", "slug": "problem-agitate-solve-pas", "agent": "Storyteller", "layer": "slide", "matchId": "78f71abd-cd69-4bd3-a6bd-10abe221263e", "evidence": "title/headline: ...but this could actually bode well for alignment science", "confidence": 0.6 } ], "frameworks": [], "arcBeats": [ { "to": 282, "from": 190, "beatId": "019dd95a-0682-776c-8e35-a256fe16ea9c", "arcName": "The Triple Take", "arcSlug": "triple-take", "beatName": "The Implications (So What)", "beatSlug": "triple-take-the-implications-so-what", "evidence": "Sections 3-4 frame political and safety consequences of the technical/industry trends.", "position": 2, "confidence": 70, "parentBeatName": "Reflection", "parentBeatSlug": "reflection" }, { "to": 282, "from": 190, "beatId": "019dd95a-0682-776c-8e35-b713af60bf88", "arcName": "The Onion", "arcSlug": "onion", "beatName": "Core Insight", "beatSlug": "onion-core-insight", "evidence": "Politics and Safety reveal the structural risks at the core.", "position": 4, "confidence": 45, "parentBeatName": "Turn", "parentBeatSlug": "turn" } ], "loops": [ { "to": 282, "from": 247, "name": "Pre Mortem", "slug": "19-pre-mortem", "bestFor": "Project kick-offs, risk management, building trust with skeptical stakeholders", "matchId": "019dd95a-07fe-70ce-8d3e-bb4d43115ab4", "evidence": "Vibe hacking, AI psychosis, alignment faking, scheming, prompt injection — what could go wrong + mitigations.", "position": 12, "objective": "Surface AI safety failure modes before they become catastrophic", "structure": "The Future Disaster (Hypothetical) -> What Went Wrong -> The Prevention Plan", "confidence": 75, "description": "Build trust by anticipating failure modes before they happen and solving them in advance" } ], "imagePathAlt": null, "thumbSrc": null, "thumbSrcAlt": null, "locked": true }