scan-v5.json (28790B)
1 { 2 "scan_version": 5, 3 "paper_type": "empirical", 4 "paper": { 5 "title": "How Alignment and Jailbreak Work: Explain LLM Safety through Intermediate Hidden States", 6 "authors": [ 7 "Zhenhong Zhou", 8 "Haiyang Yu", 9 "Xinghua Zhang", 10 "Rongwu Xu", 11 "Fei Huang" 12 ], 13 "year": 2024, 14 "venue": "Conference on Empirical Methods in Natural Language Processing", 15 "arxiv_id": "2406.05644", 16 "doi": "10.48550/arXiv.2406.05644" 17 }, 18 "checklist": { 19 "claims_and_evidence": { 20 "abstract_claims_supported": { 21 "applies": true, 22 "answer": true, 23 "justification": "The abstract's three main claims—pre-training origins of ethical concepts, alignment's role bridging early and middle layers, and jailbreak disruption of that association—are each tested empirically with weak classifier experiments and Logit Grafting.", 24 "source": "haiku" 25 }, 26 "causal_claims_justified": { 27 "applies": true, 28 "answer": true, 29 "justification": "Logit Grafting provides a direct intervention: replacing middle-layer hidden states with positive-emotion representations causes models to respond to malicious inputs, supporting the causal claim that jailbreak works by disrupting mid-layer association rather than fooling early ethical detection.", 30 "source": "haiku" 31 }, 32 "generalization_bounded": { 33 "applies": true, 34 "answer": false, 35 "justification": "Claims are framed as general to 'LLMs' but tested only on 5 open-source model families (7B–70B). Closed-source models (GPT-4, Claude3) are used for data generation but not analyzed. The title and conclusions do not bound findings to tested open-source transformer models.", 36 "source": "haiku" 37 }, 38 "alternative_explanations_discussed": { 39 "applies": true, 40 "answer": false, 41 "justification": "The paper does not consider whether hidden-state separability reflects stylistic features, query length, or topic distribution rather than 'ethical concepts.' The finding that GPT-4 vs. Claude3-Opus inputs are separable (Table 6) is noted but only superficially attributed to token length.", 42 "source": "haiku" 43 }, 44 "proxy_outcome_distinction": { 45 "applies": true, 46 "answer": false, 47 "justification": "Weak classifier accuracy on hidden states is treated as direct evidence that 'LLMs learn ethical concepts during pre-training.' The proxy (classifiability of representations) is not the same as the claimed construct (genuine ethical understanding), and this gap is not discussed.", 48 "source": "haiku" 49 } 50 }, 51 "limitations_and_scope": { 52 "limitations_section_present": { 53 "applies": true, 54 "answer": true, 55 "justification": "Section 6 is titled 'Limitations' and discusses the simplicity of weak classifiers used and restriction of experiments to LLM safety.", 56 "source": "haiku" 57 }, 58 "threats_to_validity_specific": { 59 "applies": true, 60 "answer": false, 61 "justification": "Section 6 only notes that default weak classifier settings were used and that the paper is restricted to safety. No specific threats—such as dataset representativeness, evaluation reliability via GPT-4, or model selection bias—are addressed.", 62 "source": "haiku" 63 }, 64 "scope_boundaries_stated": { 65 "applies": true, 66 "answer": false, 67 "justification": "The paper mentions it 'just conducted experiments about LLM safety' but does not formally state what the findings do NOT show (e.g., inapplicability to closed-source models, non-transformer architectures, or instruction-following capabilities).", 68 "source": "haiku" 69 } 70 }, 71 "conflicts_of_interest": { 72 "funding_disclosed": { 73 "applies": true, 74 "answer": false, 75 "justification": "No funding source is disclosed anywhere in the paper, including the acknowledgments which are absent.", 76 "source": "haiku" 77 }, 78 "affiliations_disclosed": { 79 "applies": true, 80 "answer": true, 81 "justification": "Author affiliations are clearly listed: four authors from Alibaba Group and one from Tsinghua University.", 82 "source": "haiku" 83 }, 84 "funder_independent_of_outcome": { 85 "applies": true, 86 "answer": false, 87 "justification": "No external funder is disclosed; the work is conducted by Alibaba Group employees. Alibaba has direct commercial interest in LLM safety research outcomes, so the organization supporting the work is not independent of its conclusions.", 88 "source": "haiku" 89 }, 90 "financial_interests_declared": { 91 "applies": true, 92 "answer": false, 93 "justification": "No competing interests or financial interests declaration appears in the paper.", 94 "source": "haiku" 95 } 96 }, 97 "scope_and_framing": { 98 "key_terms_defined": { 99 "applies": true, 100 "answer": false, 101 "justification": "'Alignment,' 'jailbreak,' and 'ethical concepts' are central terms used throughout without formal definition. 'Positive/negative emotion tokens' are enumerated in appendix tables but the categorization is author-curated with no independent validation.", 102 "source": "haiku" 103 }, 104 "intended_contribution_clear": { 105 "applies": true, 106 "answer": true, 107 "justification": "The paper clearly states it contributes: the Weak-to-Strong Explanation (WSE) methodology, an empirical explanation of alignment and jailbreak mechanisms via intermediate hidden states, and Logit Grafting as a mechanistic probe.", 108 "source": "haiku" 109 }, 110 "engagement_with_prior_work": { 111 "applies": true, 112 "answer": true, 113 "justification": "The paper explicitly positions findings relative to prior work—building on Logit Lens (nostalgebraist 2020), contrasting with Lin et al. 2023 findings on logit shifts, and situating against the jailbreak literature (GCG, AutoDAN, Deepinception).", 114 "source": "haiku" 115 } 116 } 117 }, 118 "type_checklist": { 119 "empirical": { 120 "artifacts": { 121 "code_released": { 122 "applies": true, 123 "answer": true, 124 "justification": "Code is released at https://github.com/ydyjya/LLM-IHS-Explanation as stated in the abstract.", 125 "source": "haiku" 126 }, 127 "data_released": { 128 "applies": true, 129 "answer": true, 130 "justification": "Malicious datasets (advbench, strongreject, jailbreakbench) are publicly available benchmarks used unmodified. Normal datasets generated from GPT-4/Claude3-Opus are promised for release; jailbreak datasets will not be released.", 131 "source": "haiku" 132 }, 133 "environment_specified": { 134 "applies": true, 135 "answer": false, 136 "justification": "Only sklearn and RapidsAI for t-SNE are mentioned. No requirements.txt, Dockerfile, Python version, CUDA version, or full environment specification is provided.", 137 "source": "haiku" 138 }, 139 "reproduction_instructions": { 140 "applies": true, 141 "answer": false, 142 "justification": "The paper describes the methodology at a high level but provides no step-by-step instructions for reproducing experiments. While code is linked, the paper itself does not guide reproduction.", 143 "source": "haiku" 144 } 145 }, 146 "statistical_methodology": { 147 "confidence_intervals_or_error_bars": { 148 "applies": true, 149 "answer": false, 150 "justification": "All main results (accuracy values, ASR, consistency scores) are reported as point estimates with no confidence intervals or error bars.", 151 "source": "haiku" 152 }, 153 "significance_tests": { 154 "applies": true, 155 "answer": false, 156 "justification": "No statistical significance tests are applied to any comparative claims, including the correlation between consistency and ASR or differences in accuracy across model families.", 157 "source": "haiku" 158 }, 159 "effect_sizes_reported": { 160 "applies": true, 161 "answer": true, 162 "justification": "ASR before and after Logit Grafting are reported with baselines (e.g., Llama-2-7b ASR 0.000 vanilla vs. 0.0172 after grafting vs. 0.0466 jailbreak). Correlation coefficients between consistency and ASR are provided (r=-0.516, r=-0.810).", 163 "source": "haiku" 164 }, 165 "sample_size_justified": { 166 "applies": true, 167 "answer": false, 168 "justification": "500 samples (250 per class, test_size=0.3) are used without power analysis or justification for adequacy given the claims made.", 169 "source": "haiku" 170 }, 171 "variance_reported": { 172 "applies": true, 173 "answer": false, 174 "justification": "Tables 5 and 6 show mean/min/max across layers but no standard deviation across independent runs. Main results tables report only point estimates.", 175 "source": "haiku" 176 } 177 }, 178 "evaluation_design": { 179 "baselines_included": { 180 "applies": true, 181 "answer": true, 182 "justification": "Base (unaligned) models serve as baselines for aligned models, and embedding-layer results serve as the overfitting baseline for weak classifiers.", 183 "source": "haiku" 184 }, 185 "baselines_contemporary": { 186 "applies": true, 187 "answer": true, 188 "justification": "Experiments use Llama-2, Llama-3, Mistral, Vicuna, and Falcon—all state-of-the-art open-source models at time of publication.", 189 "source": "haiku" 190 }, 191 "ablation_study": { 192 "applies": true, 193 "answer": true, 194 "justification": "Appendix B provides a dedicated overfitting ablation for WSE (shuffled-label and cross-source label experiments). Logit Grafting functionally ablates the mid-layer association to test the causal claim.", 195 "source": "haiku" 196 }, 197 "multiple_metrics": { 198 "applies": true, 199 "answer": true, 200 "justification": "Multiple metrics used: weak classifier accuracy (SVM and MLP), Top-K Intermediate Consistency, attack success rate (ASR), and GPT-4-based jailbreak evaluation scoring.", 201 "source": "haiku" 202 }, 203 "human_evaluation": { 204 "applies": true, 205 "answer": false, 206 "justification": "GPT-4 is used as judge to evaluate whether model outputs complete malicious goals (Table 4 shows the evaluation prompt). This is LLM-as-judge, not human evaluation.", 207 "source": "haiku" 208 }, 209 "held_out_test_set": { 210 "applies": true, 211 "answer": true, 212 "justification": "Section 3.1 states 'setting the test size to 0.3,' indicating a 70/30 train/test split for weak classifier evaluation.", 213 "source": "haiku" 214 }, 215 "per_category_breakdown": { 216 "applies": true, 217 "answer": true, 218 "justification": "Results are broken down by model family, model size (7B vs. 70B), alignment status (base vs. chat), and input type (normal vs. malicious vs. jailbreak).", 219 "source": "haiku" 220 }, 221 "failure_cases_discussed": { 222 "applies": true, 223 "answer": true, 224 "justification": "Appendix D explicitly shows cases where jailbreaks fail to disturb strong safety models (Llama-2, Llama-3) with visualizations showing that partial disturbance still allows rejection to dominate in later layers.", 225 "source": "haiku" 226 }, 227 "negative_results_reported": { 228 "applies": true, 229 "answer": true, 230 "justification": "The paper reports that weak jailbreak methods fail against models with stronger safety guardrails, that not all disturbances lead to harmful output, and that Logit Grafting's ASR gains vary widely across models (near-zero increase for Llama-2-13b-chat).", 231 "source": "haiku" 232 } 233 }, 234 "setup_transparency": { 235 "model_versions_specified": { 236 "applies": true, 237 "answer": true, 238 "justification": "Exact Hugging Face model identifiers are used throughout: 'Llama-2-7b-chat-hf', 'Meta-Llama-3-8B-Instruct', 'Mistral-7b-Instruct-v0.1', 'vicuna-7b-v1.5', 'falcon-7b-instruct'.", 239 "source": "haiku" 240 }, 241 "prompts_provided": { 242 "applies": true, 243 "answer": true, 244 "justification": "The full GPT-4 evaluation prompt is provided in Table 4. Example jailbreak inputs and outputs are shown in Appendix A (Figures 8–11).", 245 "source": "haiku" 246 }, 247 "hyperparameters_reported": { 248 "applies": true, 249 "answer": false, 250 "justification": "SVM uses 'default settings,' MLP uses '100 neurons from sklearn' without further detail, t-SNE parameters are given (perplexity=30, lr=500), but LLM generation parameters (temperature, top-p, max tokens) are not reported.", 251 "source": "haiku" 252 }, 253 "scaffolding_described": { 254 "applies": false, 255 "answer": false, 256 "justification": "No agentic scaffolding is used; the paper analyzes static forward passes through frozen models.", 257 "source": "haiku" 258 }, 259 "data_preprocessing_documented": { 260 "applies": true, 261 "answer": true, 262 "justification": "The paper describes merging three malicious datasets, generating 500 normal examples from GPT-4 and Claude3-Opus (250 each), random sampling of 500 total, and a 0.3 test split.", 263 "source": "haiku" 264 } 265 }, 266 "data_integrity": { 267 "raw_data_available": { 268 "applies": true, 269 "answer": false, 270 "justification": "While code is released, raw hidden state data is not available. Results are presented as processed figures and tables; the normal dataset generated from GPT-4/Claude had not yet been released at time of publication.", 271 "source": "haiku" 272 }, 273 "data_collection_described": { 274 "applies": true, 275 "answer": true, 276 "justification": "Data sources are clearly described: advbench, strongreject, and jailbreakbench for malicious inputs; GPT-4 and Claude3-Opus generated 250 examples each for normal inputs across different domains.", 277 "source": "haiku" 278 }, 279 "recruitment_methods_described": { 280 "applies": false, 281 "answer": false, 282 "justification": "No human participants; N/A.", 283 "source": "haiku" 284 }, 285 "data_pipeline_documented": { 286 "applies": true, 287 "answer": false, 288 "justification": "High-level pipeline is described but key steps are missing: how normal queries were prompted, how generated outputs were filtered for quality, and how datasets were merged and balanced. The full pipeline from generation to final train/test split is not documented end-to-end.", 289 "source": "haiku" 290 } 291 }, 292 "contamination": { 293 "training_cutoff_stated": { 294 "applies": true, 295 "answer": false, 296 "justification": "Training data cutoffs for none of the evaluated models (Llama-2, Llama-3, Mistral, Vicuna, Falcon) are stated in the paper.", 297 "source": "haiku" 298 }, 299 "train_test_overlap_discussed": { 300 "applies": true, 301 "answer": false, 302 "justification": "The paper does not discuss whether examples from advbench, strongreject, or jailbreakbench may have appeared in the evaluated models' pre-training or alignment data.", 303 "source": "haiku" 304 }, 305 "benchmark_contamination_addressed": { 306 "applies": true, 307 "answer": false, 308 "justification": "The malicious datasets (advbench released 2023, strongreject and jailbreakbench early 2024) predate some tested model checkpoints, raising potential contamination concerns that are not discussed.", 309 "source": "haiku" 310 } 311 }, 312 "human_studies": { 313 "pre_registered": { 314 "applies": false, 315 "answer": false, 316 "justification": "No human participants; N/A.", 317 "source": "haiku" 318 }, 319 "irb_or_ethics_approval": { 320 "applies": false, 321 "answer": false, 322 "justification": "No human participants; N/A.", 323 "source": "haiku" 324 }, 325 "demographics_reported": { 326 "applies": false, 327 "answer": false, 328 "justification": "No human participants; N/A.", 329 "source": "haiku" 330 }, 331 "inclusion_exclusion_criteria": { 332 "applies": false, 333 "answer": false, 334 "justification": "No human participants; N/A.", 335 "source": "haiku" 336 }, 337 "randomization_described": { 338 "applies": false, 339 "answer": false, 340 "justification": "No human participants; N/A.", 341 "source": "haiku" 342 }, 343 "blinding_described": { 344 "applies": false, 345 "answer": false, 346 "justification": "No human participants; N/A.", 347 "source": "haiku" 348 }, 349 "attrition_reported": { 350 "applies": false, 351 "answer": false, 352 "justification": "No human participants; N/A.", 353 "source": "haiku" 354 } 355 }, 356 "cost_and_practicality": { 357 "inference_cost_reported": { 358 "applies": true, 359 "answer": false, 360 "justification": "No inference cost, latency, or computational cost is reported for the hidden state analysis across 12 models ranging from 7B to 70B parameters.", 361 "source": "haiku" 362 }, 363 "compute_budget_stated": { 364 "applies": true, 365 "answer": false, 366 "justification": "No compute budget, GPU hours, or hardware specifications are provided anywhere in the paper.", 367 "source": "haiku" 368 } 369 } 370 } 371 }, 372 "claims": [ 373 { 374 "claim": "LLMs learn ethical concepts during pre-training rather than safety alignment, since both base and aligned models' hidden states are equally classifiable by weak classifiers in early layers.", 375 "evidence": "SVM and MLP classifiers achieve >95% accuracy from layer 3–5 onward on both base (unaligned) and chat (aligned) model variants across all five model families (Table 2, Figure 3).", 376 "supported": "moderate" 377 }, 378 { 379 "claim": "Safety alignment bridges early ethical classification with emotional token guesses in middle layers (16–24), associating malicious inputs with negative emotions and safe inputs with positive emotions.", 380 "evidence": "Logit Lens visualization shows aligned models generate emotional tokens in middle layers while base models produce format tokens ('answer', 'quelle') with no emotional valence; Top-K Intermediate Consistency quantifies this difference.", 381 "supported": "moderate" 382 }, 383 { 384 "claim": "Jailbreaks work by disrupting the mid-layer emotional association, not by deceiving the early ethical detection layer.", 385 "evidence": "Three-way classifier (Table 2) distinguishes jailbreak, malicious, and normal inputs in early layers with >93% accuracy. Middle-layer visualization shows jailbreak inputs produce ambiguous emotional tokens rather than clear negative emotion.", 386 "supported": "moderate" 387 }, 388 { 389 "claim": "Top-K Intermediate Consistency of negative emotional tokens correlates negatively with jailbreak attack success rates (r=−0.810).", 390 "evidence": "Correlation computed across 7 models in Table 1. Models with low consistency (Vicuna, Mistral) have high ASR (0.65–0.73) while models with high consistency (Llama-2-chat) have near-zero ASR.", 391 "supported": "weak" 392 }, 393 { 394 "claim": "Logit Grafting—replacing middle-layer hidden states with positive-emotion representations from benign inputs—approximates jailbreak effects, confirming the association-disruption causal mechanism.", 395 "evidence": "Table 3 shows Logit Grafting increases malicious response rates across all tested models (e.g., vicuna-7b ASR rises from 0.114 to 0.788, matching or exceeding jailbreak baselines).", 396 "supported": "strong" 397 } 398 ], 399 "methodology_tags": [ 400 "observational", 401 "benchmark-eval", 402 "theoretical" 403 ], 404 "key_findings": "Using weak classifiers (SVM and MLP) on intermediate transformer hidden states, the paper shows that LLMs develop ethical classification during pre-training—not alignment—since both aligned and unaligned models achieve >95% early-layer accuracy distinguishing malicious from normal inputs. Safety alignment adds an emotional association layer in middle transformer layers (16–24), linking ethical classifications to positive/negative emotional token predictions before refining them into accept/reject response tokens. Jailbreaks are demonstrated to bypass alignment by disrupting this middle-layer emotional association rather than the early ethical detection, as three-way classifiers can still separate jailbreak inputs from normal inputs in early layers. Logit Grafting—a causal intervention that swaps mid-layer hidden states from benign inputs—replicates jailbreak effects and provides mechanistic evidence that disrupting the association stage, not the ethical classification stage, is how jailbreaks succeed.", 405 "red_flags": [ 406 { 407 "flag": "Proxy conflated with construct", 408 "detail": "The paper treats weak classifier accuracy on hidden states as direct evidence that 'LLMs learn ethical concepts.' Classifiability could reflect stylistic, length, or topic features rather than genuine ethical understanding—this alternative is not tested." 409 }, 410 { 411 "flag": "Subjective emotion token taxonomy", 412 "detail": "The positive/negative/neutral emotion token lists (Tables 8–10) are manually curated by the authors with no inter-rater reliability check or external validation of emotional categorization." 413 }, 414 { 415 "flag": "GPT-4 as sole evaluator of jailbreak success", 416 "detail": "ASR is determined entirely by GPT-4 scoring via a custom prompt (Table 4). No human validation of GPT-4's judgments is reported, and the evaluation prompt's threshold (score=10 means jailbreak) is a binary cutoff on a 1–10 scale without justification." 417 }, 418 { 419 "flag": "No statistical significance tests", 420 "detail": "All comparative claims—model family differences, before/after grafting effects, and the r=−0.810 correlation—are made without statistical significance tests or confidence intervals." 421 }, 422 { 423 "flag": "Undisclosed funding from interested party", 424 "detail": "No funding source is disclosed. All major authors are Alibaba Group employees, who have direct commercial interests in LLM safety research outcomes." 425 }, 426 { 427 "flag": "Benchmark contamination unaddressed", 428 "detail": "The malicious datasets (advbench, strongreject, jailbreakbench) predate several evaluated model checkpoints. Whether these examples appeared in pre-training or alignment data is not discussed, potentially inflating the early-layer classifiability results." 429 } 430 ], 431 "cited_papers": [ 432 { 433 "title": "Universal and Transferable Adversarial Attacks on Aligned Language Models (GCG)", 434 "relevance": "Core jailbreak method used as experimental input; foundational prior work on automated adversarial suffix attacks" 435 }, 436 { 437 "title": "Interpreting GPT: The Logit Lens", 438 "relevance": "Core technique borrowed to decode intermediate hidden states into token space; directly enables the paper's visualization methodology" 439 }, 440 { 441 "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", 442 "relevance": "Primary model family tested; represents RLHF-aligned open-source models" 443 }, 444 { 445 "title": "Jailbroken: How Does LLM Safety Training Fail?", 446 "relevance": "Prior analysis of alignment failures motivating this paper's mechanistic investigation" 447 }, 448 { 449 "title": "Training Language Models to Follow Instructions with Human Feedback (InstructGPT)", 450 "relevance": "Foundational alignment technique whose internal mechanism this paper investigates" 451 }, 452 { 453 "title": "Generating Stealthy Jailbreak Prompts on Aligned LLMs (AutoDAN)", 454 "relevance": "Jailbreak method used in experiments alongside GCG" 455 }, 456 { 457 "title": "JailbreakBench: An Open Robustness Benchmark for Jailbreaking LLMs", 458 "relevance": "Dataset used for malicious input evaluation" 459 }, 460 { 461 "title": "The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning", 462 "relevance": "Prior work on logit shifts in aligned vs. base models that this paper builds on and partially challenges" 463 }, 464 { 465 "title": "A StrongREJECT for Empty Jailbreaks", 466 "relevance": "Dataset used for malicious input evaluation; provides diverse jailbreak-resistant benchmark prompts" 467 } 468 ], 469 "engagement_factors": { 470 "practical_relevance": { 471 "score": 2, 472 "justification": "Provides actionable insight for alignment researchers—mid-layer emotional association is a concrete optimization target—but requires white-box model access, limiting practitioner use." 473 }, 474 "surprise_contrarian": { 475 "score": 3, 476 "justification": "The finding that ethical concepts emerge from pre-training rather than alignment directly challenges the dominant narrative that RLHF teaches the model what is safe." 477 }, 478 "fear_safety": { 479 "score": 2, 480 "justification": "Mechanistically explains how jailbreaks succeed and demonstrates Logit Grafting can bypass safety guardrails, but frames this as defensive research with no new attack capability." 481 }, 482 "drama_conflict": { 483 "score": 1, 484 "justification": "Standard academic interpretability paper with no major controversy or community conflict identified." 485 }, 486 "demo_ability": { 487 "score": 2, 488 "justification": "Code is released on GitHub and experiments can in principle be reproduced on open-source models, though significant compute (70B models) is required." 489 }, 490 "brand_recognition": { 491 "score": 1, 492 "justification": "Alibaba Group and Tsinghua University are notable institutions but not the highest-profile labs in LLM interpretability research." 493 } 494 }, 495 "hn_data": { 496 "threads": [ 497 { 498 "hn_id": "39675585", 499 "title": "Is Cosine-Similarity of Embeddings Really About Similarity?", 500 "points": 210, 501 "comments": 115, 502 "url": "https://news.ycombinator.com/item?id=39675585" 503 }, 504 { 505 "hn_id": "40385302", 506 "title": "Thinking Tokens for Language Modeling", 507 "points": 6, 508 "comments": 1, 509 "url": "https://news.ycombinator.com/item?id=40385302" 510 }, 511 { 512 "hn_id": "40147616", 513 "title": "Rot: Enhancing Large Language Models with Reflection on Search Trees", 514 "points": 3, 515 "comments": 0, 516 "url": "https://news.ycombinator.com/item?id=40147616" 517 }, 518 { 519 "hn_id": "39016301", 520 "title": "Towards Conversational Diagnostic AI", 521 "points": 2, 522 "comments": 1, 523 "url": "https://news.ycombinator.com/item?id=39016301" 524 }, 525 { 526 "hn_id": "43909036", 527 "title": "Language Representations Can Be What Recommenders Need: Findings and Potentials", 528 "points": 2, 529 "comments": 0, 530 "url": "https://news.ycombinator.com/item?id=43909036" 531 }, 532 { 533 "hn_id": "39030639", 534 "title": "Towards Conversational Diagnostic AI", 535 "points": 2, 536 "comments": 0, 537 "url": "https://news.ycombinator.com/item?id=39030639" 538 }, 539 { 540 "hn_id": "41035192", 541 "title": "The Limitations of Compute Thresholds as a Governance Strategy", 542 "points": 1, 543 "comments": 0, 544 "url": "https://news.ycombinator.com/item?id=41035192" 545 }, 546 { 547 "hn_id": "39138295", 548 "title": "Rebus: A Robust Evaluation Benchmark of Understanding Symbols", 549 "points": 1, 550 "comments": 0, 551 "url": "https://news.ycombinator.com/item?id=39138295" 552 }, 553 { 554 "hn_id": "38994242", 555 "title": "Towards Conversational Diagnostic AI", 556 "points": 1, 557 "comments": 0, 558 "url": "https://news.ycombinator.com/item?id=38994242" 559 } 560 ], 561 "top_points": 210, 562 "total_points": 228, 563 "total_comments": 117 564 } 565 }